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 /// UnfoldedOffset - An additional constant offset which added near the
213 /// use. This requires a temporary register, but the offset itself can
214 /// live in an add immediate field rather than a register.
215 int64_t UnfoldedOffset;
217 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
219 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
221 unsigned getNumRegs() const;
222 const Type *getType() const;
224 void DeleteBaseReg(const SCEV *&S);
226 bool referencesReg(const SCEV *S) const;
227 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
228 const RegUseTracker &RegUses) const;
230 void print(raw_ostream &OS) const;
236 /// DoInitialMatch - Recursion helper for InitialMatch.
237 static void DoInitialMatch(const SCEV *S, Loop *L,
238 SmallVectorImpl<const SCEV *> &Good,
239 SmallVectorImpl<const SCEV *> &Bad,
240 ScalarEvolution &SE) {
241 // Collect expressions which properly dominate the loop header.
242 if (SE.properlyDominates(S, L->getHeader())) {
247 // Look at add operands.
248 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
249 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
251 DoInitialMatch(*I, L, Good, Bad, SE);
255 // Look at addrec operands.
256 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
257 if (!AR->getStart()->isZero()) {
258 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
259 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
260 AR->getStepRecurrence(SE),
261 // FIXME: AR->getNoWrapFlags()
262 AR->getLoop(), SCEV::FlagAnyWrap),
267 // Handle a multiplication by -1 (negation) if it didn't fold.
268 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
269 if (Mul->getOperand(0)->isAllOnesValue()) {
270 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
271 const SCEV *NewMul = SE.getMulExpr(Ops);
273 SmallVector<const SCEV *, 4> MyGood;
274 SmallVector<const SCEV *, 4> MyBad;
275 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
276 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
277 SE.getEffectiveSCEVType(NewMul->getType())));
278 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
279 E = MyGood.end(); I != E; ++I)
280 Good.push_back(SE.getMulExpr(NegOne, *I));
281 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
282 E = MyBad.end(); I != E; ++I)
283 Bad.push_back(SE.getMulExpr(NegOne, *I));
287 // Ok, we can't do anything interesting. Just stuff the whole thing into a
288 // register and hope for the best.
292 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
293 /// attempting to keep all loop-invariant and loop-computable values in a
294 /// single base register.
295 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
296 SmallVector<const SCEV *, 4> Good;
297 SmallVector<const SCEV *, 4> Bad;
298 DoInitialMatch(S, L, Good, Bad, SE);
300 const SCEV *Sum = SE.getAddExpr(Good);
302 BaseRegs.push_back(Sum);
303 AM.HasBaseReg = true;
306 const SCEV *Sum = SE.getAddExpr(Bad);
308 BaseRegs.push_back(Sum);
309 AM.HasBaseReg = true;
313 /// getNumRegs - Return the total number of register operands used by this
314 /// formula. This does not include register uses implied by non-constant
316 unsigned Formula::getNumRegs() const {
317 return !!ScaledReg + BaseRegs.size();
320 /// getType - Return the type of this formula, if it has one, or null
321 /// otherwise. This type is meaningless except for the bit size.
322 const Type *Formula::getType() const {
323 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
324 ScaledReg ? ScaledReg->getType() :
325 AM.BaseGV ? AM.BaseGV->getType() :
329 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
330 void Formula::DeleteBaseReg(const SCEV *&S) {
331 if (&S != &BaseRegs.back())
332 std::swap(S, BaseRegs.back());
336 /// referencesReg - Test if this formula references the given register.
337 bool Formula::referencesReg(const SCEV *S) const {
338 return S == ScaledReg ||
339 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
342 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
343 /// which are used by uses other than the use with the given index.
344 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
345 const RegUseTracker &RegUses) const {
347 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
349 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
350 E = BaseRegs.end(); I != E; ++I)
351 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
356 void Formula::print(raw_ostream &OS) const {
359 if (!First) OS << " + "; else First = false;
360 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
362 if (AM.BaseOffs != 0) {
363 if (!First) OS << " + "; else First = false;
366 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
367 E = BaseRegs.end(); I != E; ++I) {
368 if (!First) OS << " + "; else First = false;
369 OS << "reg(" << **I << ')';
371 if (AM.HasBaseReg && BaseRegs.empty()) {
372 if (!First) OS << " + "; else First = false;
373 OS << "**error: HasBaseReg**";
374 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
375 if (!First) OS << " + "; else First = false;
376 OS << "**error: !HasBaseReg**";
379 if (!First) OS << " + "; else First = false;
380 OS << AM.Scale << "*reg(";
387 if (UnfoldedOffset != 0) {
388 if (!First) OS << " + "; else First = false;
389 OS << "imm(" << UnfoldedOffset << ')';
393 void Formula::dump() const {
394 print(errs()); errs() << '\n';
397 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
398 /// without changing its value.
399 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
401 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
402 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
405 /// isAddSExtable - Return true if the given add can be sign-extended
406 /// without changing its value.
407 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
409 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
410 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
413 /// isMulSExtable - Return true if the given mul can be sign-extended
414 /// without changing its value.
415 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
417 IntegerType::get(SE.getContext(),
418 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
419 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
422 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
423 /// and if the remainder is known to be zero, or null otherwise. If
424 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
425 /// to Y, ignoring that the multiplication may overflow, which is useful when
426 /// the result will be used in a context where the most significant bits are
428 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
430 bool IgnoreSignificantBits = false) {
431 // Handle the trivial case, which works for any SCEV type.
433 return SE.getConstant(LHS->getType(), 1);
435 // Handle a few RHS special cases.
436 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
438 const APInt &RA = RC->getValue()->getValue();
439 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
441 if (RA.isAllOnesValue())
442 return SE.getMulExpr(LHS, RC);
443 // Handle x /s 1 as x.
448 // Check for a division of a constant by a constant.
449 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
452 const APInt &LA = C->getValue()->getValue();
453 const APInt &RA = RC->getValue()->getValue();
454 if (LA.srem(RA) != 0)
456 return SE.getConstant(LA.sdiv(RA));
459 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
460 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
461 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
462 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
463 IgnoreSignificantBits);
465 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
466 IgnoreSignificantBits);
467 if (!Start) return 0;
468 // FlagNW is independent of the start value, step direction, and is
469 // preserved with smaller magnitude steps.
470 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
471 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
476 // Distribute the sdiv over add operands, if the add doesn't overflow.
477 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
478 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
479 SmallVector<const SCEV *, 8> Ops;
480 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
482 const SCEV *Op = getExactSDiv(*I, RHS, SE,
483 IgnoreSignificantBits);
487 return SE.getAddExpr(Ops);
492 // Check for a multiply operand that we can pull RHS out of.
493 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
494 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
495 SmallVector<const SCEV *, 4> Ops;
497 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
501 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
502 IgnoreSignificantBits)) {
508 return Found ? SE.getMulExpr(Ops) : 0;
513 // Otherwise we don't know.
517 /// ExtractImmediate - If S involves the addition of a constant integer value,
518 /// return that integer value, and mutate S to point to a new SCEV with that
520 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
521 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
522 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
523 S = SE.getConstant(C->getType(), 0);
524 return C->getValue()->getSExtValue();
526 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
527 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
528 int64_t Result = ExtractImmediate(NewOps.front(), SE);
530 S = SE.getAddExpr(NewOps);
532 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
533 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
534 int64_t Result = ExtractImmediate(NewOps.front(), SE);
536 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
537 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
544 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
545 /// return that symbol, and mutate S to point to a new SCEV with that
547 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
548 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
549 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
550 S = SE.getConstant(GV->getType(), 0);
553 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
554 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
555 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
557 S = SE.getAddExpr(NewOps);
559 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
560 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
561 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
563 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
564 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
571 /// isAddressUse - Returns true if the specified instruction is using the
572 /// specified value as an address.
573 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
574 bool isAddress = isa<LoadInst>(Inst);
575 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
576 if (SI->getOperand(1) == OperandVal)
578 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
579 // Addressing modes can also be folded into prefetches and a variety
581 switch (II->getIntrinsicID()) {
583 case Intrinsic::prefetch:
584 case Intrinsic::x86_sse_storeu_ps:
585 case Intrinsic::x86_sse2_storeu_pd:
586 case Intrinsic::x86_sse2_storeu_dq:
587 case Intrinsic::x86_sse2_storel_dq:
588 if (II->getArgOperand(0) == OperandVal)
596 /// getAccessType - Return the type of the memory being accessed.
597 static const Type *getAccessType(const Instruction *Inst) {
598 const Type *AccessTy = Inst->getType();
599 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
600 AccessTy = SI->getOperand(0)->getType();
601 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
602 // Addressing modes can also be folded into prefetches and a variety
604 switch (II->getIntrinsicID()) {
606 case Intrinsic::x86_sse_storeu_ps:
607 case Intrinsic::x86_sse2_storeu_pd:
608 case Intrinsic::x86_sse2_storeu_dq:
609 case Intrinsic::x86_sse2_storel_dq:
610 AccessTy = II->getArgOperand(0)->getType();
615 // All pointers have the same requirements, so canonicalize them to an
616 // arbitrary pointer type to minimize variation.
617 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
618 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
619 PTy->getAddressSpace());
624 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
625 /// specified set are trivially dead, delete them and see if this makes any of
626 /// their operands subsequently dead.
628 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
629 bool Changed = false;
631 while (!DeadInsts.empty()) {
632 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
634 if (I == 0 || !isInstructionTriviallyDead(I))
637 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
638 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
641 DeadInsts.push_back(U);
644 I->eraseFromParent();
653 /// Cost - This class is used to measure and compare candidate formulae.
655 /// TODO: Some of these could be merged. Also, a lexical ordering
656 /// isn't always optimal.
660 unsigned NumBaseAdds;
666 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
669 bool operator<(const Cost &Other) const;
673 void RateFormula(const Formula &F,
674 SmallPtrSet<const SCEV *, 16> &Regs,
675 const DenseSet<const SCEV *> &VisitedRegs,
677 const SmallVectorImpl<int64_t> &Offsets,
678 ScalarEvolution &SE, DominatorTree &DT);
680 void print(raw_ostream &OS) const;
684 void RateRegister(const SCEV *Reg,
685 SmallPtrSet<const SCEV *, 16> &Regs,
687 ScalarEvolution &SE, DominatorTree &DT);
688 void RatePrimaryRegister(const SCEV *Reg,
689 SmallPtrSet<const SCEV *, 16> &Regs,
691 ScalarEvolution &SE, DominatorTree &DT);
696 /// RateRegister - Tally up interesting quantities from the given register.
697 void Cost::RateRegister(const SCEV *Reg,
698 SmallPtrSet<const SCEV *, 16> &Regs,
700 ScalarEvolution &SE, DominatorTree &DT) {
701 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
702 if (AR->getLoop() == L)
703 AddRecCost += 1; /// TODO: This should be a function of the stride.
705 // If this is an addrec for a loop that's already been visited by LSR,
706 // don't second-guess its addrec phi nodes. LSR isn't currently smart
707 // enough to reason about more than one loop at a time. Consider these
708 // registers free and leave them alone.
709 else if (L->contains(AR->getLoop()) ||
710 (!AR->getLoop()->contains(L) &&
711 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
712 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
713 PHINode *PN = dyn_cast<PHINode>(I); ++I)
714 if (SE.isSCEVable(PN->getType()) &&
715 (SE.getEffectiveSCEVType(PN->getType()) ==
716 SE.getEffectiveSCEVType(AR->getType())) &&
717 SE.getSCEV(PN) == AR)
720 // If this isn't one of the addrecs that the loop already has, it
721 // would require a costly new phi and add. TODO: This isn't
722 // precisely modeled right now.
724 if (!Regs.count(AR->getStart()))
725 RateRegister(AR->getStart(), Regs, L, SE, DT);
728 // Add the step value register, if it needs one.
729 // TODO: The non-affine case isn't precisely modeled here.
730 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
731 if (!Regs.count(AR->getStart()))
732 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
736 // Rough heuristic; favor registers which don't require extra setup
737 // instructions in the preheader.
738 if (!isa<SCEVUnknown>(Reg) &&
739 !isa<SCEVConstant>(Reg) &&
740 !(isa<SCEVAddRecExpr>(Reg) &&
741 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
742 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
745 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
746 SE.hasComputableLoopEvolution(Reg, L);
749 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
751 void Cost::RatePrimaryRegister(const SCEV *Reg,
752 SmallPtrSet<const SCEV *, 16> &Regs,
754 ScalarEvolution &SE, DominatorTree &DT) {
755 if (Regs.insert(Reg))
756 RateRegister(Reg, Regs, L, SE, DT);
759 void Cost::RateFormula(const Formula &F,
760 SmallPtrSet<const SCEV *, 16> &Regs,
761 const DenseSet<const SCEV *> &VisitedRegs,
763 const SmallVectorImpl<int64_t> &Offsets,
764 ScalarEvolution &SE, DominatorTree &DT) {
765 // Tally up the registers.
766 if (const SCEV *ScaledReg = F.ScaledReg) {
767 if (VisitedRegs.count(ScaledReg)) {
771 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
773 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
774 E = F.BaseRegs.end(); I != E; ++I) {
775 const SCEV *BaseReg = *I;
776 if (VisitedRegs.count(BaseReg)) {
780 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
783 // Determine how many (unfolded) adds we'll need inside the loop.
784 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
785 if (NumBaseParts > 1)
786 NumBaseAdds += NumBaseParts - 1;
788 // Tally up the non-zero immediates.
789 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
790 E = Offsets.end(); I != E; ++I) {
791 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
793 ImmCost += 64; // Handle symbolic values conservatively.
794 // TODO: This should probably be the pointer size.
795 else if (Offset != 0)
796 ImmCost += APInt(64, Offset, true).getMinSignedBits();
800 /// Loose - Set this cost to a losing value.
810 /// operator< - Choose the lower cost.
811 bool Cost::operator<(const Cost &Other) const {
812 if (NumRegs != Other.NumRegs)
813 return NumRegs < Other.NumRegs;
814 if (AddRecCost != Other.AddRecCost)
815 return AddRecCost < Other.AddRecCost;
816 if (NumIVMuls != Other.NumIVMuls)
817 return NumIVMuls < Other.NumIVMuls;
818 if (NumBaseAdds != Other.NumBaseAdds)
819 return NumBaseAdds < Other.NumBaseAdds;
820 if (ImmCost != Other.ImmCost)
821 return ImmCost < Other.ImmCost;
822 if (SetupCost != Other.SetupCost)
823 return SetupCost < Other.SetupCost;
827 void Cost::print(raw_ostream &OS) const {
828 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
830 OS << ", with addrec cost " << AddRecCost;
832 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
833 if (NumBaseAdds != 0)
834 OS << ", plus " << NumBaseAdds << " base add"
835 << (NumBaseAdds == 1 ? "" : "s");
837 OS << ", plus " << ImmCost << " imm cost";
839 OS << ", plus " << SetupCost << " setup cost";
842 void Cost::dump() const {
843 print(errs()); errs() << '\n';
848 /// LSRFixup - An operand value in an instruction which is to be replaced
849 /// with some equivalent, possibly strength-reduced, replacement.
851 /// UserInst - The instruction which will be updated.
852 Instruction *UserInst;
854 /// OperandValToReplace - The operand of the instruction which will
855 /// be replaced. The operand may be used more than once; every instance
856 /// will be replaced.
857 Value *OperandValToReplace;
859 /// PostIncLoops - If this user is to use the post-incremented value of an
860 /// induction variable, this variable is non-null and holds the loop
861 /// associated with the induction variable.
862 PostIncLoopSet PostIncLoops;
864 /// LUIdx - The index of the LSRUse describing the expression which
865 /// this fixup needs, minus an offset (below).
868 /// Offset - A constant offset to be added to the LSRUse expression.
869 /// This allows multiple fixups to share the same LSRUse with different
870 /// offsets, for example in an unrolled loop.
873 bool isUseFullyOutsideLoop(const Loop *L) const;
877 void print(raw_ostream &OS) const;
884 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
886 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
887 /// value outside of the given loop.
888 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
889 // PHI nodes use their value in their incoming blocks.
890 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
891 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
892 if (PN->getIncomingValue(i) == OperandValToReplace &&
893 L->contains(PN->getIncomingBlock(i)))
898 return !L->contains(UserInst);
901 void LSRFixup::print(raw_ostream &OS) const {
903 // Store is common and interesting enough to be worth special-casing.
904 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
906 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
907 } else if (UserInst->getType()->isVoidTy())
908 OS << UserInst->getOpcodeName();
910 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
912 OS << ", OperandValToReplace=";
913 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
915 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
916 E = PostIncLoops.end(); I != E; ++I) {
917 OS << ", PostIncLoop=";
918 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
921 if (LUIdx != ~size_t(0))
922 OS << ", LUIdx=" << LUIdx;
925 OS << ", Offset=" << Offset;
928 void LSRFixup::dump() const {
929 print(errs()); errs() << '\n';
934 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
935 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
936 struct UniquifierDenseMapInfo {
937 static SmallVector<const SCEV *, 2> getEmptyKey() {
938 SmallVector<const SCEV *, 2> V;
939 V.push_back(reinterpret_cast<const SCEV *>(-1));
943 static SmallVector<const SCEV *, 2> getTombstoneKey() {
944 SmallVector<const SCEV *, 2> V;
945 V.push_back(reinterpret_cast<const SCEV *>(-2));
949 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
951 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
952 E = V.end(); I != E; ++I)
953 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
957 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
958 const SmallVector<const SCEV *, 2> &RHS) {
963 /// LSRUse - This class holds the state that LSR keeps for each use in
964 /// IVUsers, as well as uses invented by LSR itself. It includes information
965 /// about what kinds of things can be folded into the user, information about
966 /// the user itself, and information about how the use may be satisfied.
967 /// TODO: Represent multiple users of the same expression in common?
969 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
972 /// KindType - An enum for a kind of use, indicating what types of
973 /// scaled and immediate operands it might support.
975 Basic, ///< A normal use, with no folding.
976 Special, ///< A special case of basic, allowing -1 scales.
977 Address, ///< An address use; folding according to TargetLowering
978 ICmpZero ///< An equality icmp with both operands folded into one.
979 // TODO: Add a generic icmp too?
983 const Type *AccessTy;
985 SmallVector<int64_t, 8> Offsets;
989 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
990 /// LSRUse are outside of the loop, in which case some special-case heuristics
992 bool AllFixupsOutsideLoop;
994 /// WidestFixupType - This records the widest use type for any fixup using
995 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
996 /// max fixup widths to be equivalent, because the narrower one may be relying
997 /// on the implicit truncation to truncate away bogus bits.
998 const Type *WidestFixupType;
1000 /// Formulae - A list of ways to build a value that can satisfy this user.
1001 /// After the list is populated, one of these is selected heuristically and
1002 /// used to formulate a replacement for OperandValToReplace in UserInst.
1003 SmallVector<Formula, 12> Formulae;
1005 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1006 SmallPtrSet<const SCEV *, 4> Regs;
1008 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
1009 MinOffset(INT64_MAX),
1010 MaxOffset(INT64_MIN),
1011 AllFixupsOutsideLoop(true),
1012 WidestFixupType(0) {}
1014 bool HasFormulaWithSameRegs(const Formula &F) const;
1015 bool InsertFormula(const Formula &F);
1016 void DeleteFormula(Formula &F);
1017 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1019 void print(raw_ostream &OS) const;
1025 /// HasFormula - Test whether this use as a formula which has the same
1026 /// registers as the given formula.
1027 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1028 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1029 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1030 // Unstable sort by host order ok, because this is only used for uniquifying.
1031 std::sort(Key.begin(), Key.end());
1032 return Uniquifier.count(Key);
1035 /// InsertFormula - If the given formula has not yet been inserted, add it to
1036 /// the list, and return true. Return false otherwise.
1037 bool LSRUse::InsertFormula(const Formula &F) {
1038 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1039 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1040 // Unstable sort by host order ok, because this is only used for uniquifying.
1041 std::sort(Key.begin(), Key.end());
1043 if (!Uniquifier.insert(Key).second)
1046 // Using a register to hold the value of 0 is not profitable.
1047 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1048 "Zero allocated in a scaled register!");
1050 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1051 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1052 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1055 // Add the formula to the list.
1056 Formulae.push_back(F);
1058 // Record registers now being used by this use.
1059 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1060 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1065 /// DeleteFormula - Remove the given formula from this use's list.
1066 void LSRUse::DeleteFormula(Formula &F) {
1067 if (&F != &Formulae.back())
1068 std::swap(F, Formulae.back());
1069 Formulae.pop_back();
1070 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1073 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1074 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1075 // Now that we've filtered out some formulae, recompute the Regs set.
1076 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1078 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1079 E = Formulae.end(); I != E; ++I) {
1080 const Formula &F = *I;
1081 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1082 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1085 // Update the RegTracker.
1086 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1087 E = OldRegs.end(); I != E; ++I)
1088 if (!Regs.count(*I))
1089 RegUses.DropRegister(*I, LUIdx);
1092 void LSRUse::print(raw_ostream &OS) const {
1093 OS << "LSR Use: Kind=";
1095 case Basic: OS << "Basic"; break;
1096 case Special: OS << "Special"; break;
1097 case ICmpZero: OS << "ICmpZero"; break;
1099 OS << "Address of ";
1100 if (AccessTy->isPointerTy())
1101 OS << "pointer"; // the full pointer type could be really verbose
1106 OS << ", Offsets={";
1107 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1108 E = Offsets.end(); I != E; ++I) {
1110 if (llvm::next(I) != E)
1115 if (AllFixupsOutsideLoop)
1116 OS << ", all-fixups-outside-loop";
1118 if (WidestFixupType)
1119 OS << ", widest fixup type: " << *WidestFixupType;
1122 void LSRUse::dump() const {
1123 print(errs()); errs() << '\n';
1126 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1127 /// be completely folded into the user instruction at isel time. This includes
1128 /// address-mode folding and special icmp tricks.
1129 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1130 LSRUse::KindType Kind, const Type *AccessTy,
1131 const TargetLowering *TLI) {
1133 case LSRUse::Address:
1134 // If we have low-level target information, ask the target if it can
1135 // completely fold this address.
1136 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1138 // Otherwise, just guess that reg+reg addressing is legal.
1139 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1141 case LSRUse::ICmpZero:
1142 // There's not even a target hook for querying whether it would be legal to
1143 // fold a GV into an ICmp.
1147 // ICmp only has two operands; don't allow more than two non-trivial parts.
1148 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1151 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1152 // putting the scaled register in the other operand of the icmp.
1153 if (AM.Scale != 0 && AM.Scale != -1)
1156 // If we have low-level target information, ask the target if it can fold an
1157 // integer immediate on an icmp.
1158 if (AM.BaseOffs != 0) {
1159 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1166 // Only handle single-register values.
1167 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1169 case LSRUse::Special:
1170 // Only handle -1 scales, or no scale.
1171 return AM.Scale == 0 || AM.Scale == -1;
1177 static bool isLegalUse(TargetLowering::AddrMode AM,
1178 int64_t MinOffset, int64_t MaxOffset,
1179 LSRUse::KindType Kind, const Type *AccessTy,
1180 const TargetLowering *TLI) {
1181 // Check for overflow.
1182 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1185 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1186 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1187 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1188 // Check for overflow.
1189 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1192 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1193 return isLegalUse(AM, Kind, AccessTy, TLI);
1198 static bool isAlwaysFoldable(int64_t BaseOffs,
1199 GlobalValue *BaseGV,
1201 LSRUse::KindType Kind, const Type *AccessTy,
1202 const TargetLowering *TLI) {
1203 // Fast-path: zero is always foldable.
1204 if (BaseOffs == 0 && !BaseGV) return true;
1206 // Conservatively, create an address with an immediate and a
1207 // base and a scale.
1208 TargetLowering::AddrMode AM;
1209 AM.BaseOffs = BaseOffs;
1211 AM.HasBaseReg = HasBaseReg;
1212 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1214 // Canonicalize a scale of 1 to a base register if the formula doesn't
1215 // already have a base register.
1216 if (!AM.HasBaseReg && AM.Scale == 1) {
1218 AM.HasBaseReg = true;
1221 return isLegalUse(AM, Kind, AccessTy, TLI);
1224 static bool isAlwaysFoldable(const SCEV *S,
1225 int64_t MinOffset, int64_t MaxOffset,
1227 LSRUse::KindType Kind, const Type *AccessTy,
1228 const TargetLowering *TLI,
1229 ScalarEvolution &SE) {
1230 // Fast-path: zero is always foldable.
1231 if (S->isZero()) return true;
1233 // Conservatively, create an address with an immediate and a
1234 // base and a scale.
1235 int64_t BaseOffs = ExtractImmediate(S, SE);
1236 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1238 // If there's anything else involved, it's not foldable.
1239 if (!S->isZero()) return false;
1241 // Fast-path: zero is always foldable.
1242 if (BaseOffs == 0 && !BaseGV) return true;
1244 // Conservatively, create an address with an immediate and a
1245 // base and a scale.
1246 TargetLowering::AddrMode AM;
1247 AM.BaseOffs = BaseOffs;
1249 AM.HasBaseReg = HasBaseReg;
1250 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1252 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1257 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1258 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1259 struct UseMapDenseMapInfo {
1260 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1261 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1264 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1265 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1269 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1270 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1271 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1275 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1276 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1281 /// LSRInstance - This class holds state for the main loop strength reduction
1285 ScalarEvolution &SE;
1288 const TargetLowering *const TLI;
1292 /// IVIncInsertPos - This is the insert position that the current loop's
1293 /// induction variable increment should be placed. In simple loops, this is
1294 /// the latch block's terminator. But in more complicated cases, this is a
1295 /// position which will dominate all the in-loop post-increment users.
1296 Instruction *IVIncInsertPos;
1298 /// Factors - Interesting factors between use strides.
1299 SmallSetVector<int64_t, 8> Factors;
1301 /// Types - Interesting use types, to facilitate truncation reuse.
1302 SmallSetVector<const Type *, 4> Types;
1304 /// Fixups - The list of operands which are to be replaced.
1305 SmallVector<LSRFixup, 16> Fixups;
1307 /// Uses - The list of interesting uses.
1308 SmallVector<LSRUse, 16> Uses;
1310 /// RegUses - Track which uses use which register candidates.
1311 RegUseTracker RegUses;
1313 void OptimizeShadowIV();
1314 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1315 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1316 void OptimizeLoopTermCond();
1318 void CollectInterestingTypesAndFactors();
1319 void CollectFixupsAndInitialFormulae();
1321 LSRFixup &getNewFixup() {
1322 Fixups.push_back(LSRFixup());
1323 return Fixups.back();
1326 // Support for sharing of LSRUses between LSRFixups.
1327 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1329 UseMapDenseMapInfo> UseMapTy;
1332 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1333 LSRUse::KindType Kind, const Type *AccessTy);
1335 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1336 LSRUse::KindType Kind,
1337 const Type *AccessTy);
1339 void DeleteUse(LSRUse &LU, size_t LUIdx);
1341 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1344 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1345 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1346 void CountRegisters(const Formula &F, size_t LUIdx);
1347 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1349 void CollectLoopInvariantFixupsAndFormulae();
1351 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1352 unsigned Depth = 0);
1353 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1354 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1355 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1356 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1357 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1358 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1359 void GenerateCrossUseConstantOffsets();
1360 void GenerateAllReuseFormulae();
1362 void FilterOutUndesirableDedicatedRegisters();
1364 size_t EstimateSearchSpaceComplexity() const;
1365 void NarrowSearchSpaceByDetectingSupersets();
1366 void NarrowSearchSpaceByCollapsingUnrolledCode();
1367 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1368 void NarrowSearchSpaceByPickingWinnerRegs();
1369 void NarrowSearchSpaceUsingHeuristics();
1371 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1373 SmallVectorImpl<const Formula *> &Workspace,
1374 const Cost &CurCost,
1375 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1376 DenseSet<const SCEV *> &VisitedRegs) const;
1377 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1379 BasicBlock::iterator
1380 HoistInsertPosition(BasicBlock::iterator IP,
1381 const SmallVectorImpl<Instruction *> &Inputs) const;
1382 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1384 const LSRUse &LU) const;
1386 Value *Expand(const LSRFixup &LF,
1388 BasicBlock::iterator IP,
1389 SCEVExpander &Rewriter,
1390 SmallVectorImpl<WeakVH> &DeadInsts) const;
1391 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1393 SCEVExpander &Rewriter,
1394 SmallVectorImpl<WeakVH> &DeadInsts,
1396 void Rewrite(const LSRFixup &LF,
1398 SCEVExpander &Rewriter,
1399 SmallVectorImpl<WeakVH> &DeadInsts,
1401 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1404 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1406 bool getChanged() const { return Changed; }
1408 void print_factors_and_types(raw_ostream &OS) const;
1409 void print_fixups(raw_ostream &OS) const;
1410 void print_uses(raw_ostream &OS) const;
1411 void print(raw_ostream &OS) const;
1417 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1418 /// inside the loop then try to eliminate the cast operation.
1419 void LSRInstance::OptimizeShadowIV() {
1420 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1421 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1424 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1425 UI != E; /* empty */) {
1426 IVUsers::const_iterator CandidateUI = UI;
1428 Instruction *ShadowUse = CandidateUI->getUser();
1429 const Type *DestTy = NULL;
1431 /* If shadow use is a int->float cast then insert a second IV
1432 to eliminate this cast.
1434 for (unsigned i = 0; i < n; ++i)
1440 for (unsigned i = 0; i < n; ++i, ++d)
1443 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1444 DestTy = UCast->getDestTy();
1445 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1446 DestTy = SCast->getDestTy();
1447 if (!DestTy) continue;
1450 // If target does not support DestTy natively then do not apply
1451 // this transformation.
1452 EVT DVT = TLI->getValueType(DestTy);
1453 if (!TLI->isTypeLegal(DVT)) continue;
1456 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1458 if (PH->getNumIncomingValues() != 2) continue;
1460 const Type *SrcTy = PH->getType();
1461 int Mantissa = DestTy->getFPMantissaWidth();
1462 if (Mantissa == -1) continue;
1463 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1466 unsigned Entry, Latch;
1467 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1475 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1476 if (!Init) continue;
1477 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1479 BinaryOperator *Incr =
1480 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1481 if (!Incr) continue;
1482 if (Incr->getOpcode() != Instruction::Add
1483 && Incr->getOpcode() != Instruction::Sub)
1486 /* Initialize new IV, double d = 0.0 in above example. */
1487 ConstantInt *C = NULL;
1488 if (Incr->getOperand(0) == PH)
1489 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1490 else if (Incr->getOperand(1) == PH)
1491 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1497 // Ignore negative constants, as the code below doesn't handle them
1498 // correctly. TODO: Remove this restriction.
1499 if (!C->getValue().isStrictlyPositive()) continue;
1501 /* Add new PHINode. */
1502 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1504 /* create new increment. '++d' in above example. */
1505 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1506 BinaryOperator *NewIncr =
1507 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1508 Instruction::FAdd : Instruction::FSub,
1509 NewPH, CFP, "IV.S.next.", Incr);
1511 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1512 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1514 /* Remove cast operation */
1515 ShadowUse->replaceAllUsesWith(NewPH);
1516 ShadowUse->eraseFromParent();
1522 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1523 /// set the IV user and stride information and return true, otherwise return
1525 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1526 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1527 if (UI->getUser() == Cond) {
1528 // NOTE: we could handle setcc instructions with multiple uses here, but
1529 // InstCombine does it as well for simple uses, it's not clear that it
1530 // occurs enough in real life to handle.
1537 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1538 /// a max computation.
1540 /// This is a narrow solution to a specific, but acute, problem. For loops
1546 /// } while (++i < n);
1548 /// the trip count isn't just 'n', because 'n' might not be positive. And
1549 /// unfortunately this can come up even for loops where the user didn't use
1550 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1551 /// will commonly be lowered like this:
1557 /// } while (++i < n);
1560 /// and then it's possible for subsequent optimization to obscure the if
1561 /// test in such a way that indvars can't find it.
1563 /// When indvars can't find the if test in loops like this, it creates a
1564 /// max expression, which allows it to give the loop a canonical
1565 /// induction variable:
1568 /// max = n < 1 ? 1 : n;
1571 /// } while (++i != max);
1573 /// Canonical induction variables are necessary because the loop passes
1574 /// are designed around them. The most obvious example of this is the
1575 /// LoopInfo analysis, which doesn't remember trip count values. It
1576 /// expects to be able to rediscover the trip count each time it is
1577 /// needed, and it does this using a simple analysis that only succeeds if
1578 /// the loop has a canonical induction variable.
1580 /// However, when it comes time to generate code, the maximum operation
1581 /// can be quite costly, especially if it's inside of an outer loop.
1583 /// This function solves this problem by detecting this type of loop and
1584 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1585 /// the instructions for the maximum computation.
1587 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1588 // Check that the loop matches the pattern we're looking for.
1589 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1590 Cond->getPredicate() != CmpInst::ICMP_NE)
1593 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1594 if (!Sel || !Sel->hasOneUse()) return Cond;
1596 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1597 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1599 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1601 // Add one to the backedge-taken count to get the trip count.
1602 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1603 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1605 // Check for a max calculation that matches the pattern. There's no check
1606 // for ICMP_ULE here because the comparison would be with zero, which
1607 // isn't interesting.
1608 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1609 const SCEVNAryExpr *Max = 0;
1610 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1611 Pred = ICmpInst::ICMP_SLE;
1613 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1614 Pred = ICmpInst::ICMP_SLT;
1616 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1617 Pred = ICmpInst::ICMP_ULT;
1624 // To handle a max with more than two operands, this optimization would
1625 // require additional checking and setup.
1626 if (Max->getNumOperands() != 2)
1629 const SCEV *MaxLHS = Max->getOperand(0);
1630 const SCEV *MaxRHS = Max->getOperand(1);
1632 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1633 // for a comparison with 1. For <= and >=, a comparison with zero.
1635 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1638 // Check the relevant induction variable for conformance to
1640 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1641 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1642 if (!AR || !AR->isAffine() ||
1643 AR->getStart() != One ||
1644 AR->getStepRecurrence(SE) != One)
1647 assert(AR->getLoop() == L &&
1648 "Loop condition operand is an addrec in a different loop!");
1650 // Check the right operand of the select, and remember it, as it will
1651 // be used in the new comparison instruction.
1653 if (ICmpInst::isTrueWhenEqual(Pred)) {
1654 // Look for n+1, and grab n.
1655 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1656 if (isa<ConstantInt>(BO->getOperand(1)) &&
1657 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1658 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1659 NewRHS = BO->getOperand(0);
1660 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1661 if (isa<ConstantInt>(BO->getOperand(1)) &&
1662 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1663 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1664 NewRHS = BO->getOperand(0);
1667 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1668 NewRHS = Sel->getOperand(1);
1669 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1670 NewRHS = Sel->getOperand(2);
1671 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1672 NewRHS = SU->getValue();
1674 // Max doesn't match expected pattern.
1677 // Determine the new comparison opcode. It may be signed or unsigned,
1678 // and the original comparison may be either equality or inequality.
1679 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1680 Pred = CmpInst::getInversePredicate(Pred);
1682 // Ok, everything looks ok to change the condition into an SLT or SGE and
1683 // delete the max calculation.
1685 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1687 // Delete the max calculation instructions.
1688 Cond->replaceAllUsesWith(NewCond);
1689 CondUse->setUser(NewCond);
1690 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1691 Cond->eraseFromParent();
1692 Sel->eraseFromParent();
1693 if (Cmp->use_empty())
1694 Cmp->eraseFromParent();
1698 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1699 /// postinc iv when possible.
1701 LSRInstance::OptimizeLoopTermCond() {
1702 SmallPtrSet<Instruction *, 4> PostIncs;
1704 BasicBlock *LatchBlock = L->getLoopLatch();
1705 SmallVector<BasicBlock*, 8> ExitingBlocks;
1706 L->getExitingBlocks(ExitingBlocks);
1708 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1709 BasicBlock *ExitingBlock = ExitingBlocks[i];
1711 // Get the terminating condition for the loop if possible. If we
1712 // can, we want to change it to use a post-incremented version of its
1713 // induction variable, to allow coalescing the live ranges for the IV into
1714 // one register value.
1716 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1719 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1720 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1723 // Search IVUsesByStride to find Cond's IVUse if there is one.
1724 IVStrideUse *CondUse = 0;
1725 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1726 if (!FindIVUserForCond(Cond, CondUse))
1729 // If the trip count is computed in terms of a max (due to ScalarEvolution
1730 // being unable to find a sufficient guard, for example), change the loop
1731 // comparison to use SLT or ULT instead of NE.
1732 // One consequence of doing this now is that it disrupts the count-down
1733 // optimization. That's not always a bad thing though, because in such
1734 // cases it may still be worthwhile to avoid a max.
1735 Cond = OptimizeMax(Cond, CondUse);
1737 // If this exiting block dominates the latch block, it may also use
1738 // the post-inc value if it won't be shared with other uses.
1739 // Check for dominance.
1740 if (!DT.dominates(ExitingBlock, LatchBlock))
1743 // Conservatively avoid trying to use the post-inc value in non-latch
1744 // exits if there may be pre-inc users in intervening blocks.
1745 if (LatchBlock != ExitingBlock)
1746 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1747 // Test if the use is reachable from the exiting block. This dominator
1748 // query is a conservative approximation of reachability.
1749 if (&*UI != CondUse &&
1750 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1751 // Conservatively assume there may be reuse if the quotient of their
1752 // strides could be a legal scale.
1753 const SCEV *A = IU.getStride(*CondUse, L);
1754 const SCEV *B = IU.getStride(*UI, L);
1755 if (!A || !B) continue;
1756 if (SE.getTypeSizeInBits(A->getType()) !=
1757 SE.getTypeSizeInBits(B->getType())) {
1758 if (SE.getTypeSizeInBits(A->getType()) >
1759 SE.getTypeSizeInBits(B->getType()))
1760 B = SE.getSignExtendExpr(B, A->getType());
1762 A = SE.getSignExtendExpr(A, B->getType());
1764 if (const SCEVConstant *D =
1765 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1766 const ConstantInt *C = D->getValue();
1767 // Stride of one or negative one can have reuse with non-addresses.
1768 if (C->isOne() || C->isAllOnesValue())
1769 goto decline_post_inc;
1770 // Avoid weird situations.
1771 if (C->getValue().getMinSignedBits() >= 64 ||
1772 C->getValue().isMinSignedValue())
1773 goto decline_post_inc;
1774 // Without TLI, assume that any stride might be valid, and so any
1775 // use might be shared.
1777 goto decline_post_inc;
1778 // Check for possible scaled-address reuse.
1779 const Type *AccessTy = getAccessType(UI->getUser());
1780 TargetLowering::AddrMode AM;
1781 AM.Scale = C->getSExtValue();
1782 if (TLI->isLegalAddressingMode(AM, AccessTy))
1783 goto decline_post_inc;
1784 AM.Scale = -AM.Scale;
1785 if (TLI->isLegalAddressingMode(AM, AccessTy))
1786 goto decline_post_inc;
1790 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1793 // It's possible for the setcc instruction to be anywhere in the loop, and
1794 // possible for it to have multiple users. If it is not immediately before
1795 // the exiting block branch, move it.
1796 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1797 if (Cond->hasOneUse()) {
1798 Cond->moveBefore(TermBr);
1800 // Clone the terminating condition and insert into the loopend.
1801 ICmpInst *OldCond = Cond;
1802 Cond = cast<ICmpInst>(Cond->clone());
1803 Cond->setName(L->getHeader()->getName() + ".termcond");
1804 ExitingBlock->getInstList().insert(TermBr, Cond);
1806 // Clone the IVUse, as the old use still exists!
1807 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1808 TermBr->replaceUsesOfWith(OldCond, Cond);
1812 // If we get to here, we know that we can transform the setcc instruction to
1813 // use the post-incremented version of the IV, allowing us to coalesce the
1814 // live ranges for the IV correctly.
1815 CondUse->transformToPostInc(L);
1818 PostIncs.insert(Cond);
1822 // Determine an insertion point for the loop induction variable increment. It
1823 // must dominate all the post-inc comparisons we just set up, and it must
1824 // dominate the loop latch edge.
1825 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1826 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1827 E = PostIncs.end(); I != E; ++I) {
1829 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1831 if (BB == (*I)->getParent())
1832 IVIncInsertPos = *I;
1833 else if (BB != IVIncInsertPos->getParent())
1834 IVIncInsertPos = BB->getTerminator();
1838 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
1839 /// at the given offset and other details. If so, update the use and
1842 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1843 LSRUse::KindType Kind, const Type *AccessTy) {
1844 int64_t NewMinOffset = LU.MinOffset;
1845 int64_t NewMaxOffset = LU.MaxOffset;
1846 const Type *NewAccessTy = AccessTy;
1848 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1849 // something conservative, however this can pessimize in the case that one of
1850 // the uses will have all its uses outside the loop, for example.
1851 if (LU.Kind != Kind)
1853 // Conservatively assume HasBaseReg is true for now.
1854 if (NewOffset < LU.MinOffset) {
1855 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1856 Kind, AccessTy, TLI))
1858 NewMinOffset = NewOffset;
1859 } else if (NewOffset > LU.MaxOffset) {
1860 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1861 Kind, AccessTy, TLI))
1863 NewMaxOffset = NewOffset;
1865 // Check for a mismatched access type, and fall back conservatively as needed.
1866 // TODO: Be less conservative when the type is similar and can use the same
1867 // addressing modes.
1868 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1869 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1872 LU.MinOffset = NewMinOffset;
1873 LU.MaxOffset = NewMaxOffset;
1874 LU.AccessTy = NewAccessTy;
1875 if (NewOffset != LU.Offsets.back())
1876 LU.Offsets.push_back(NewOffset);
1880 /// getUse - Return an LSRUse index and an offset value for a fixup which
1881 /// needs the given expression, with the given kind and optional access type.
1882 /// Either reuse an existing use or create a new one, as needed.
1883 std::pair<size_t, int64_t>
1884 LSRInstance::getUse(const SCEV *&Expr,
1885 LSRUse::KindType Kind, const Type *AccessTy) {
1886 const SCEV *Copy = Expr;
1887 int64_t Offset = ExtractImmediate(Expr, SE);
1889 // Basic uses can't accept any offset, for example.
1890 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1895 std::pair<UseMapTy::iterator, bool> P =
1896 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1898 // A use already existed with this base.
1899 size_t LUIdx = P.first->second;
1900 LSRUse &LU = Uses[LUIdx];
1901 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1903 return std::make_pair(LUIdx, Offset);
1906 // Create a new use.
1907 size_t LUIdx = Uses.size();
1908 P.first->second = LUIdx;
1909 Uses.push_back(LSRUse(Kind, AccessTy));
1910 LSRUse &LU = Uses[LUIdx];
1912 // We don't need to track redundant offsets, but we don't need to go out
1913 // of our way here to avoid them.
1914 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1915 LU.Offsets.push_back(Offset);
1917 LU.MinOffset = Offset;
1918 LU.MaxOffset = Offset;
1919 return std::make_pair(LUIdx, Offset);
1922 /// DeleteUse - Delete the given use from the Uses list.
1923 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
1924 if (&LU != &Uses.back())
1925 std::swap(LU, Uses.back());
1929 RegUses.SwapAndDropUse(LUIdx, Uses.size());
1932 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1933 /// a formula that has the same registers as the given formula.
1935 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1936 const LSRUse &OrigLU) {
1937 // Search all uses for the formula. This could be more clever.
1938 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1939 LSRUse &LU = Uses[LUIdx];
1940 // Check whether this use is close enough to OrigLU, to see whether it's
1941 // worthwhile looking through its formulae.
1942 // Ignore ICmpZero uses because they may contain formulae generated by
1943 // GenerateICmpZeroScales, in which case adding fixup offsets may
1945 if (&LU != &OrigLU &&
1946 LU.Kind != LSRUse::ICmpZero &&
1947 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1948 LU.WidestFixupType == OrigLU.WidestFixupType &&
1949 LU.HasFormulaWithSameRegs(OrigF)) {
1950 // Scan through this use's formulae.
1951 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1952 E = LU.Formulae.end(); I != E; ++I) {
1953 const Formula &F = *I;
1954 // Check to see if this formula has the same registers and symbols
1956 if (F.BaseRegs == OrigF.BaseRegs &&
1957 F.ScaledReg == OrigF.ScaledReg &&
1958 F.AM.BaseGV == OrigF.AM.BaseGV &&
1959 F.AM.Scale == OrigF.AM.Scale &&
1960 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
1961 if (F.AM.BaseOffs == 0)
1963 // This is the formula where all the registers and symbols matched;
1964 // there aren't going to be any others. Since we declined it, we
1965 // can skip the rest of the formulae and procede to the next LSRUse.
1972 // Nothing looked good.
1976 void LSRInstance::CollectInterestingTypesAndFactors() {
1977 SmallSetVector<const SCEV *, 4> Strides;
1979 // Collect interesting types and strides.
1980 SmallVector<const SCEV *, 4> Worklist;
1981 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1982 const SCEV *Expr = IU.getExpr(*UI);
1984 // Collect interesting types.
1985 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1987 // Add strides for mentioned loops.
1988 Worklist.push_back(Expr);
1990 const SCEV *S = Worklist.pop_back_val();
1991 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1992 Strides.insert(AR->getStepRecurrence(SE));
1993 Worklist.push_back(AR->getStart());
1994 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1995 Worklist.append(Add->op_begin(), Add->op_end());
1997 } while (!Worklist.empty());
2000 // Compute interesting factors from the set of interesting strides.
2001 for (SmallSetVector<const SCEV *, 4>::const_iterator
2002 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2003 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2004 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2005 const SCEV *OldStride = *I;
2006 const SCEV *NewStride = *NewStrideIter;
2008 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2009 SE.getTypeSizeInBits(NewStride->getType())) {
2010 if (SE.getTypeSizeInBits(OldStride->getType()) >
2011 SE.getTypeSizeInBits(NewStride->getType()))
2012 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2014 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2016 if (const SCEVConstant *Factor =
2017 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2019 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2020 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2021 } else if (const SCEVConstant *Factor =
2022 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2025 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2026 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2030 // If all uses use the same type, don't bother looking for truncation-based
2032 if (Types.size() == 1)
2035 DEBUG(print_factors_and_types(dbgs()));
2038 void LSRInstance::CollectFixupsAndInitialFormulae() {
2039 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2041 LSRFixup &LF = getNewFixup();
2042 LF.UserInst = UI->getUser();
2043 LF.OperandValToReplace = UI->getOperandValToReplace();
2044 LF.PostIncLoops = UI->getPostIncLoops();
2046 LSRUse::KindType Kind = LSRUse::Basic;
2047 const Type *AccessTy = 0;
2048 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2049 Kind = LSRUse::Address;
2050 AccessTy = getAccessType(LF.UserInst);
2053 const SCEV *S = IU.getExpr(*UI);
2055 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2056 // (N - i == 0), and this allows (N - i) to be the expression that we work
2057 // with rather than just N or i, so we can consider the register
2058 // requirements for both N and i at the same time. Limiting this code to
2059 // equality icmps is not a problem because all interesting loops use
2060 // equality icmps, thanks to IndVarSimplify.
2061 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2062 if (CI->isEquality()) {
2063 // Swap the operands if needed to put the OperandValToReplace on the
2064 // left, for consistency.
2065 Value *NV = CI->getOperand(1);
2066 if (NV == LF.OperandValToReplace) {
2067 CI->setOperand(1, CI->getOperand(0));
2068 CI->setOperand(0, NV);
2069 NV = CI->getOperand(1);
2073 // x == y --> x - y == 0
2074 const SCEV *N = SE.getSCEV(NV);
2075 if (SE.isLoopInvariant(N, L)) {
2076 Kind = LSRUse::ICmpZero;
2077 S = SE.getMinusSCEV(N, S);
2080 // -1 and the negations of all interesting strides (except the negation
2081 // of -1) are now also interesting.
2082 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2083 if (Factors[i] != -1)
2084 Factors.insert(-(uint64_t)Factors[i]);
2088 // Set up the initial formula for this use.
2089 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2091 LF.Offset = P.second;
2092 LSRUse &LU = Uses[LF.LUIdx];
2093 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2094 if (!LU.WidestFixupType ||
2095 SE.getTypeSizeInBits(LU.WidestFixupType) <
2096 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2097 LU.WidestFixupType = LF.OperandValToReplace->getType();
2099 // If this is the first use of this LSRUse, give it a formula.
2100 if (LU.Formulae.empty()) {
2101 InsertInitialFormula(S, LU, LF.LUIdx);
2102 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2106 DEBUG(print_fixups(dbgs()));
2109 /// InsertInitialFormula - Insert a formula for the given expression into
2110 /// the given use, separating out loop-variant portions from loop-invariant
2111 /// and loop-computable portions.
2113 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2115 F.InitialMatch(S, L, SE);
2116 bool Inserted = InsertFormula(LU, LUIdx, F);
2117 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2120 /// InsertSupplementalFormula - Insert a simple single-register formula for
2121 /// the given expression into the given use.
2123 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2124 LSRUse &LU, size_t LUIdx) {
2126 F.BaseRegs.push_back(S);
2127 F.AM.HasBaseReg = true;
2128 bool Inserted = InsertFormula(LU, LUIdx, F);
2129 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2132 /// CountRegisters - Note which registers are used by the given formula,
2133 /// updating RegUses.
2134 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2136 RegUses.CountRegister(F.ScaledReg, LUIdx);
2137 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2138 E = F.BaseRegs.end(); I != E; ++I)
2139 RegUses.CountRegister(*I, LUIdx);
2142 /// InsertFormula - If the given formula has not yet been inserted, add it to
2143 /// the list, and return true. Return false otherwise.
2144 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2145 if (!LU.InsertFormula(F))
2148 CountRegisters(F, LUIdx);
2152 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2153 /// loop-invariant values which we're tracking. These other uses will pin these
2154 /// values in registers, making them less profitable for elimination.
2155 /// TODO: This currently misses non-constant addrec step registers.
2156 /// TODO: Should this give more weight to users inside the loop?
2158 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2159 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2160 SmallPtrSet<const SCEV *, 8> Inserted;
2162 while (!Worklist.empty()) {
2163 const SCEV *S = Worklist.pop_back_val();
2165 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2166 Worklist.append(N->op_begin(), N->op_end());
2167 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2168 Worklist.push_back(C->getOperand());
2169 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2170 Worklist.push_back(D->getLHS());
2171 Worklist.push_back(D->getRHS());
2172 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2173 if (!Inserted.insert(U)) continue;
2174 const Value *V = U->getValue();
2175 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2176 // Look for instructions defined outside the loop.
2177 if (L->contains(Inst)) continue;
2178 } else if (isa<UndefValue>(V))
2179 // Undef doesn't have a live range, so it doesn't matter.
2181 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2183 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2184 // Ignore non-instructions.
2187 // Ignore instructions in other functions (as can happen with
2189 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2191 // Ignore instructions not dominated by the loop.
2192 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2193 UserInst->getParent() :
2194 cast<PHINode>(UserInst)->getIncomingBlock(
2195 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2196 if (!DT.dominates(L->getHeader(), UseBB))
2198 // Ignore uses which are part of other SCEV expressions, to avoid
2199 // analyzing them multiple times.
2200 if (SE.isSCEVable(UserInst->getType())) {
2201 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2202 // If the user is a no-op, look through to its uses.
2203 if (!isa<SCEVUnknown>(UserS))
2207 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2211 // Ignore icmp instructions which are already being analyzed.
2212 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2213 unsigned OtherIdx = !UI.getOperandNo();
2214 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2215 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2219 LSRFixup &LF = getNewFixup();
2220 LF.UserInst = const_cast<Instruction *>(UserInst);
2221 LF.OperandValToReplace = UI.getUse();
2222 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2224 LF.Offset = P.second;
2225 LSRUse &LU = Uses[LF.LUIdx];
2226 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2227 if (!LU.WidestFixupType ||
2228 SE.getTypeSizeInBits(LU.WidestFixupType) <
2229 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2230 LU.WidestFixupType = LF.OperandValToReplace->getType();
2231 InsertSupplementalFormula(U, LU, LF.LUIdx);
2232 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2239 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2240 /// separate registers. If C is non-null, multiply each subexpression by C.
2241 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2242 SmallVectorImpl<const SCEV *> &Ops,
2244 ScalarEvolution &SE) {
2245 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2246 // Break out add operands.
2247 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2249 CollectSubexprs(*I, C, Ops, L, SE);
2251 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2252 // Split a non-zero base out of an addrec.
2253 if (!AR->getStart()->isZero()) {
2254 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2255 AR->getStepRecurrence(SE),
2257 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2260 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2263 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2264 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2265 if (Mul->getNumOperands() == 2)
2266 if (const SCEVConstant *Op0 =
2267 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2268 CollectSubexprs(Mul->getOperand(1),
2269 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2275 // Otherwise use the value itself, optionally with a scale applied.
2276 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2279 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2281 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2284 // Arbitrarily cap recursion to protect compile time.
2285 if (Depth >= 3) return;
2287 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2288 const SCEV *BaseReg = Base.BaseRegs[i];
2290 SmallVector<const SCEV *, 8> AddOps;
2291 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2293 if (AddOps.size() == 1) continue;
2295 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2296 JE = AddOps.end(); J != JE; ++J) {
2298 // Loop-variant "unknown" values are uninteresting; we won't be able to
2299 // do anything meaningful with them.
2300 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2303 // Don't pull a constant into a register if the constant could be folded
2304 // into an immediate field.
2305 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2306 Base.getNumRegs() > 1,
2307 LU.Kind, LU.AccessTy, TLI, SE))
2310 // Collect all operands except *J.
2311 SmallVector<const SCEV *, 8> InnerAddOps
2312 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2314 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2316 // Don't leave just a constant behind in a register if the constant could
2317 // be folded into an immediate field.
2318 if (InnerAddOps.size() == 1 &&
2319 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2320 Base.getNumRegs() > 1,
2321 LU.Kind, LU.AccessTy, TLI, SE))
2324 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2325 if (InnerSum->isZero())
2329 // Add the remaining pieces of the add back into the new formula.
2330 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
2331 if (TLI && InnerSumSC &&
2332 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
2333 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2334 InnerSumSC->getValue()->getZExtValue())) {
2335 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2336 InnerSumSC->getValue()->getZExtValue();
2337 F.BaseRegs.erase(F.BaseRegs.begin() + i);
2339 F.BaseRegs[i] = InnerSum;
2341 // Add J as its own register, or an unfolded immediate.
2342 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
2343 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
2344 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2345 SC->getValue()->getZExtValue()))
2346 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2347 SC->getValue()->getZExtValue();
2349 F.BaseRegs.push_back(*J);
2351 if (InsertFormula(LU, LUIdx, F))
2352 // If that formula hadn't been seen before, recurse to find more like
2354 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2359 /// GenerateCombinations - Generate a formula consisting of all of the
2360 /// loop-dominating registers added into a single register.
2361 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2363 // This method is only interesting on a plurality of registers.
2364 if (Base.BaseRegs.size() <= 1) return;
2368 SmallVector<const SCEV *, 4> Ops;
2369 for (SmallVectorImpl<const SCEV *>::const_iterator
2370 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2371 const SCEV *BaseReg = *I;
2372 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2373 !SE.hasComputableLoopEvolution(BaseReg, L))
2374 Ops.push_back(BaseReg);
2376 F.BaseRegs.push_back(BaseReg);
2378 if (Ops.size() > 1) {
2379 const SCEV *Sum = SE.getAddExpr(Ops);
2380 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2381 // opportunity to fold something. For now, just ignore such cases
2382 // rather than proceed with zero in a register.
2383 if (!Sum->isZero()) {
2384 F.BaseRegs.push_back(Sum);
2385 (void)InsertFormula(LU, LUIdx, F);
2390 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2391 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2393 // We can't add a symbolic offset if the address already contains one.
2394 if (Base.AM.BaseGV) return;
2396 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2397 const SCEV *G = Base.BaseRegs[i];
2398 GlobalValue *GV = ExtractSymbol(G, SE);
2399 if (G->isZero() || !GV)
2403 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2404 LU.Kind, LU.AccessTy, TLI))
2407 (void)InsertFormula(LU, LUIdx, F);
2411 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2412 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2414 // TODO: For now, just add the min and max offset, because it usually isn't
2415 // worthwhile looking at everything inbetween.
2416 SmallVector<int64_t, 2> Worklist;
2417 Worklist.push_back(LU.MinOffset);
2418 if (LU.MaxOffset != LU.MinOffset)
2419 Worklist.push_back(LU.MaxOffset);
2421 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2422 const SCEV *G = Base.BaseRegs[i];
2424 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2425 E = Worklist.end(); I != E; ++I) {
2427 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2428 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2429 LU.Kind, LU.AccessTy, TLI)) {
2430 // Add the offset to the base register.
2431 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2432 // If it cancelled out, drop the base register, otherwise update it.
2433 if (NewG->isZero()) {
2434 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2435 F.BaseRegs.pop_back();
2437 F.BaseRegs[i] = NewG;
2439 (void)InsertFormula(LU, LUIdx, F);
2443 int64_t Imm = ExtractImmediate(G, SE);
2444 if (G->isZero() || Imm == 0)
2447 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2448 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2449 LU.Kind, LU.AccessTy, TLI))
2452 (void)InsertFormula(LU, LUIdx, F);
2456 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2457 /// the comparison. For example, x == y -> x*c == y*c.
2458 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2460 if (LU.Kind != LSRUse::ICmpZero) return;
2462 // Determine the integer type for the base formula.
2463 const Type *IntTy = Base.getType();
2465 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2467 // Don't do this if there is more than one offset.
2468 if (LU.MinOffset != LU.MaxOffset) return;
2470 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2472 // Check each interesting stride.
2473 for (SmallSetVector<int64_t, 8>::const_iterator
2474 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2475 int64_t Factor = *I;
2477 // Check that the multiplication doesn't overflow.
2478 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2480 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2481 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2484 // Check that multiplying with the use offset doesn't overflow.
2485 int64_t Offset = LU.MinOffset;
2486 if (Offset == INT64_MIN && Factor == -1)
2488 Offset = (uint64_t)Offset * Factor;
2489 if (Offset / Factor != LU.MinOffset)
2493 F.AM.BaseOffs = NewBaseOffs;
2495 // Check that this scale is legal.
2496 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2499 // Compensate for the use having MinOffset built into it.
2500 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2502 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2504 // Check that multiplying with each base register doesn't overflow.
2505 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2506 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2507 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2511 // Check that multiplying with the scaled register doesn't overflow.
2513 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2514 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2518 // Check that multiplying with the unfolded offset doesn't overflow.
2519 if (F.UnfoldedOffset != 0) {
2520 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
2521 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
2525 // If we make it here and it's legal, add it.
2526 (void)InsertFormula(LU, LUIdx, F);
2531 /// GenerateScales - Generate stride factor reuse formulae by making use of
2532 /// scaled-offset address modes, for example.
2533 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2534 // Determine the integer type for the base formula.
2535 const Type *IntTy = Base.getType();
2538 // If this Formula already has a scaled register, we can't add another one.
2539 if (Base.AM.Scale != 0) return;
2541 // Check each interesting stride.
2542 for (SmallSetVector<int64_t, 8>::const_iterator
2543 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2544 int64_t Factor = *I;
2546 Base.AM.Scale = Factor;
2547 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2548 // Check whether this scale is going to be legal.
2549 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2550 LU.Kind, LU.AccessTy, TLI)) {
2551 // As a special-case, handle special out-of-loop Basic users specially.
2552 // TODO: Reconsider this special case.
2553 if (LU.Kind == LSRUse::Basic &&
2554 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2555 LSRUse::Special, LU.AccessTy, TLI) &&
2556 LU.AllFixupsOutsideLoop)
2557 LU.Kind = LSRUse::Special;
2561 // For an ICmpZero, negating a solitary base register won't lead to
2563 if (LU.Kind == LSRUse::ICmpZero &&
2564 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2566 // For each addrec base reg, apply the scale, if possible.
2567 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2568 if (const SCEVAddRecExpr *AR =
2569 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2570 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2571 if (FactorS->isZero())
2573 // Divide out the factor, ignoring high bits, since we'll be
2574 // scaling the value back up in the end.
2575 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2576 // TODO: This could be optimized to avoid all the copying.
2578 F.ScaledReg = Quotient;
2579 F.DeleteBaseReg(F.BaseRegs[i]);
2580 (void)InsertFormula(LU, LUIdx, F);
2586 /// GenerateTruncates - Generate reuse formulae from different IV types.
2587 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2588 // This requires TargetLowering to tell us which truncates are free.
2591 // Don't bother truncating symbolic values.
2592 if (Base.AM.BaseGV) return;
2594 // Determine the integer type for the base formula.
2595 const Type *DstTy = Base.getType();
2597 DstTy = SE.getEffectiveSCEVType(DstTy);
2599 for (SmallSetVector<const Type *, 4>::const_iterator
2600 I = Types.begin(), E = Types.end(); I != E; ++I) {
2601 const Type *SrcTy = *I;
2602 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2605 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2606 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2607 JE = F.BaseRegs.end(); J != JE; ++J)
2608 *J = SE.getAnyExtendExpr(*J, SrcTy);
2610 // TODO: This assumes we've done basic processing on all uses and
2611 // have an idea what the register usage is.
2612 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2615 (void)InsertFormula(LU, LUIdx, F);
2622 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2623 /// defer modifications so that the search phase doesn't have to worry about
2624 /// the data structures moving underneath it.
2628 const SCEV *OrigReg;
2630 WorkItem(size_t LI, int64_t I, const SCEV *R)
2631 : LUIdx(LI), Imm(I), OrigReg(R) {}
2633 void print(raw_ostream &OS) const;
2639 void WorkItem::print(raw_ostream &OS) const {
2640 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2641 << " , add offset " << Imm;
2644 void WorkItem::dump() const {
2645 print(errs()); errs() << '\n';
2648 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2649 /// distance apart and try to form reuse opportunities between them.
2650 void LSRInstance::GenerateCrossUseConstantOffsets() {
2651 // Group the registers by their value without any added constant offset.
2652 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2653 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2655 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2656 SmallVector<const SCEV *, 8> Sequence;
2657 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2659 const SCEV *Reg = *I;
2660 int64_t Imm = ExtractImmediate(Reg, SE);
2661 std::pair<RegMapTy::iterator, bool> Pair =
2662 Map.insert(std::make_pair(Reg, ImmMapTy()));
2664 Sequence.push_back(Reg);
2665 Pair.first->second.insert(std::make_pair(Imm, *I));
2666 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2669 // Now examine each set of registers with the same base value. Build up
2670 // a list of work to do and do the work in a separate step so that we're
2671 // not adding formulae and register counts while we're searching.
2672 SmallVector<WorkItem, 32> WorkItems;
2673 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2674 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2675 E = Sequence.end(); I != E; ++I) {
2676 const SCEV *Reg = *I;
2677 const ImmMapTy &Imms = Map.find(Reg)->second;
2679 // It's not worthwhile looking for reuse if there's only one offset.
2680 if (Imms.size() == 1)
2683 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2684 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2686 dbgs() << ' ' << J->first;
2689 // Examine each offset.
2690 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2692 const SCEV *OrigReg = J->second;
2694 int64_t JImm = J->first;
2695 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2697 if (!isa<SCEVConstant>(OrigReg) &&
2698 UsedByIndicesMap[Reg].count() == 1) {
2699 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2703 // Conservatively examine offsets between this orig reg a few selected
2705 ImmMapTy::const_iterator OtherImms[] = {
2706 Imms.begin(), prior(Imms.end()),
2707 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2709 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2710 ImmMapTy::const_iterator M = OtherImms[i];
2711 if (M == J || M == JE) continue;
2713 // Compute the difference between the two.
2714 int64_t Imm = (uint64_t)JImm - M->first;
2715 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2716 LUIdx = UsedByIndices.find_next(LUIdx))
2717 // Make a memo of this use, offset, and register tuple.
2718 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2719 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2726 UsedByIndicesMap.clear();
2727 UniqueItems.clear();
2729 // Now iterate through the worklist and add new formulae.
2730 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2731 E = WorkItems.end(); I != E; ++I) {
2732 const WorkItem &WI = *I;
2733 size_t LUIdx = WI.LUIdx;
2734 LSRUse &LU = Uses[LUIdx];
2735 int64_t Imm = WI.Imm;
2736 const SCEV *OrigReg = WI.OrigReg;
2738 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2739 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2740 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2742 // TODO: Use a more targeted data structure.
2743 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2744 const Formula &F = LU.Formulae[L];
2745 // Use the immediate in the scaled register.
2746 if (F.ScaledReg == OrigReg) {
2747 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2748 Imm * (uint64_t)F.AM.Scale;
2749 // Don't create 50 + reg(-50).
2750 if (F.referencesReg(SE.getSCEV(
2751 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2754 NewF.AM.BaseOffs = Offs;
2755 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2756 LU.Kind, LU.AccessTy, TLI))
2758 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2760 // If the new scale is a constant in a register, and adding the constant
2761 // value to the immediate would produce a value closer to zero than the
2762 // immediate itself, then the formula isn't worthwhile.
2763 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2764 if (C->getValue()->getValue().isNegative() !=
2765 (NewF.AM.BaseOffs < 0) &&
2766 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2767 .ule(abs64(NewF.AM.BaseOffs)))
2771 (void)InsertFormula(LU, LUIdx, NewF);
2773 // Use the immediate in a base register.
2774 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2775 const SCEV *BaseReg = F.BaseRegs[N];
2776 if (BaseReg != OrigReg)
2779 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2780 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2781 LU.Kind, LU.AccessTy, TLI)) {
2783 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
2786 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
2788 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2790 // If the new formula has a constant in a register, and adding the
2791 // constant value to the immediate would produce a value closer to
2792 // zero than the immediate itself, then the formula isn't worthwhile.
2793 for (SmallVectorImpl<const SCEV *>::const_iterator
2794 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2796 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2797 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2798 abs64(NewF.AM.BaseOffs)) &&
2799 (C->getValue()->getValue() +
2800 NewF.AM.BaseOffs).countTrailingZeros() >=
2801 CountTrailingZeros_64(NewF.AM.BaseOffs))
2805 (void)InsertFormula(LU, LUIdx, NewF);
2814 /// GenerateAllReuseFormulae - Generate formulae for each use.
2816 LSRInstance::GenerateAllReuseFormulae() {
2817 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2818 // queries are more precise.
2819 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2820 LSRUse &LU = Uses[LUIdx];
2821 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2822 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2823 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2824 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2826 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2827 LSRUse &LU = Uses[LUIdx];
2828 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2829 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2830 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2831 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2832 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2833 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2834 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2835 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2837 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2838 LSRUse &LU = Uses[LUIdx];
2839 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2840 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2843 GenerateCrossUseConstantOffsets();
2845 DEBUG(dbgs() << "\n"
2846 "After generating reuse formulae:\n";
2847 print_uses(dbgs()));
2850 /// If there are multiple formulae with the same set of registers used
2851 /// by other uses, pick the best one and delete the others.
2852 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2853 DenseSet<const SCEV *> VisitedRegs;
2854 SmallPtrSet<const SCEV *, 16> Regs;
2856 bool ChangedFormulae = false;
2859 // Collect the best formula for each unique set of shared registers. This
2860 // is reset for each use.
2861 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2863 BestFormulaeTy BestFormulae;
2865 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2866 LSRUse &LU = Uses[LUIdx];
2867 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2870 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2871 FIdx != NumForms; ++FIdx) {
2872 Formula &F = LU.Formulae[FIdx];
2874 SmallVector<const SCEV *, 2> Key;
2875 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2876 JE = F.BaseRegs.end(); J != JE; ++J) {
2877 const SCEV *Reg = *J;
2878 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2882 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2883 Key.push_back(F.ScaledReg);
2884 // Unstable sort by host order ok, because this is only used for
2886 std::sort(Key.begin(), Key.end());
2888 std::pair<BestFormulaeTy::const_iterator, bool> P =
2889 BestFormulae.insert(std::make_pair(Key, FIdx));
2891 Formula &Best = LU.Formulae[P.first->second];
2894 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2897 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2899 if (CostF < CostBest)
2901 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2903 " in favor of formula "; Best.print(dbgs());
2906 ChangedFormulae = true;
2908 LU.DeleteFormula(F);
2916 // Now that we've filtered out some formulae, recompute the Regs set.
2918 LU.RecomputeRegs(LUIdx, RegUses);
2920 // Reset this to prepare for the next use.
2921 BestFormulae.clear();
2924 DEBUG(if (ChangedFormulae) {
2926 "After filtering out undesirable candidates:\n";
2931 // This is a rough guess that seems to work fairly well.
2932 static const size_t ComplexityLimit = UINT16_MAX;
2934 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2935 /// solutions the solver might have to consider. It almost never considers
2936 /// this many solutions because it prune the search space, but the pruning
2937 /// isn't always sufficient.
2938 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2940 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2941 E = Uses.end(); I != E; ++I) {
2942 size_t FSize = I->Formulae.size();
2943 if (FSize >= ComplexityLimit) {
2944 Power = ComplexityLimit;
2948 if (Power >= ComplexityLimit)
2954 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
2955 /// of the registers of another formula, it won't help reduce register
2956 /// pressure (though it may not necessarily hurt register pressure); remove
2957 /// it to simplify the system.
2958 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
2959 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2960 DEBUG(dbgs() << "The search space is too complex.\n");
2962 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2963 "which use a superset of registers used by other "
2966 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2967 LSRUse &LU = Uses[LUIdx];
2969 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2970 Formula &F = LU.Formulae[i];
2971 // Look for a formula with a constant or GV in a register. If the use
2972 // also has a formula with that same value in an immediate field,
2973 // delete the one that uses a register.
2974 for (SmallVectorImpl<const SCEV *>::const_iterator
2975 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
2976 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
2978 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
2979 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2980 (I - F.BaseRegs.begin()));
2981 if (LU.HasFormulaWithSameRegs(NewF)) {
2982 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2983 LU.DeleteFormula(F);
2989 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
2990 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
2993 NewF.AM.BaseGV = GV;
2994 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
2995 (I - F.BaseRegs.begin()));
2996 if (LU.HasFormulaWithSameRegs(NewF)) {
2997 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
2999 LU.DeleteFormula(F);
3010 LU.RecomputeRegs(LUIdx, RegUses);
3013 DEBUG(dbgs() << "After pre-selection:\n";
3014 print_uses(dbgs()));
3018 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3019 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3021 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3022 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3023 DEBUG(dbgs() << "The search space is too complex.\n");
3025 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3026 "separated by a constant offset will use the same "
3029 // This is especially useful for unrolled loops.
3031 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3032 LSRUse &LU = Uses[LUIdx];
3033 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3034 E = LU.Formulae.end(); I != E; ++I) {
3035 const Formula &F = *I;
3036 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3037 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3038 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3039 /*HasBaseReg=*/false,
3040 LU.Kind, LU.AccessTy)) {
3041 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3044 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3046 // Update the relocs to reference the new use.
3047 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3048 E = Fixups.end(); I != E; ++I) {
3049 LSRFixup &Fixup = *I;
3050 if (Fixup.LUIdx == LUIdx) {
3051 Fixup.LUIdx = LUThatHas - &Uses.front();
3052 Fixup.Offset += F.AM.BaseOffs;
3053 // Add the new offset to LUThatHas' offset list.
3054 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3055 LUThatHas->Offsets.push_back(Fixup.Offset);
3056 if (Fixup.Offset > LUThatHas->MaxOffset)
3057 LUThatHas->MaxOffset = Fixup.Offset;
3058 if (Fixup.Offset < LUThatHas->MinOffset)
3059 LUThatHas->MinOffset = Fixup.Offset;
3061 DEBUG(dbgs() << "New fixup has offset "
3062 << Fixup.Offset << '\n');
3064 if (Fixup.LUIdx == NumUses-1)
3065 Fixup.LUIdx = LUIdx;
3068 // Delete formulae from the new use which are no longer legal.
3070 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3071 Formula &F = LUThatHas->Formulae[i];
3072 if (!isLegalUse(F.AM,
3073 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3074 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3075 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3077 LUThatHas->DeleteFormula(F);
3084 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3086 // Delete the old use.
3087 DeleteUse(LU, LUIdx);
3097 DEBUG(dbgs() << "After pre-selection:\n";
3098 print_uses(dbgs()));
3102 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3103 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3104 /// we've done more filtering, as it may be able to find more formulae to
3106 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3107 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3108 DEBUG(dbgs() << "The search space is too complex.\n");
3110 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3111 "undesirable dedicated registers.\n");
3113 FilterOutUndesirableDedicatedRegisters();
3115 DEBUG(dbgs() << "After pre-selection:\n";
3116 print_uses(dbgs()));
3120 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3121 /// to be profitable, and then in any use which has any reference to that
3122 /// register, delete all formulae which do not reference that register.
3123 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3124 // With all other options exhausted, loop until the system is simple
3125 // enough to handle.
3126 SmallPtrSet<const SCEV *, 4> Taken;
3127 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3128 // Ok, we have too many of formulae on our hands to conveniently handle.
3129 // Use a rough heuristic to thin out the list.
3130 DEBUG(dbgs() << "The search space is too complex.\n");
3132 // Pick the register which is used by the most LSRUses, which is likely
3133 // to be a good reuse register candidate.
3134 const SCEV *Best = 0;
3135 unsigned BestNum = 0;
3136 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3138 const SCEV *Reg = *I;
3139 if (Taken.count(Reg))
3144 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3145 if (Count > BestNum) {
3152 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3153 << " will yield profitable reuse.\n");
3156 // In any use with formulae which references this register, delete formulae
3157 // which don't reference it.
3158 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3159 LSRUse &LU = Uses[LUIdx];
3160 if (!LU.Regs.count(Best)) continue;
3163 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3164 Formula &F = LU.Formulae[i];
3165 if (!F.referencesReg(Best)) {
3166 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3167 LU.DeleteFormula(F);
3171 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3177 LU.RecomputeRegs(LUIdx, RegUses);
3180 DEBUG(dbgs() << "After pre-selection:\n";
3181 print_uses(dbgs()));
3185 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3186 /// formulae to choose from, use some rough heuristics to prune down the number
3187 /// of formulae. This keeps the main solver from taking an extraordinary amount
3188 /// of time in some worst-case scenarios.
3189 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3190 NarrowSearchSpaceByDetectingSupersets();
3191 NarrowSearchSpaceByCollapsingUnrolledCode();
3192 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3193 NarrowSearchSpaceByPickingWinnerRegs();
3196 /// SolveRecurse - This is the recursive solver.
3197 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3199 SmallVectorImpl<const Formula *> &Workspace,
3200 const Cost &CurCost,
3201 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3202 DenseSet<const SCEV *> &VisitedRegs) const {
3205 // - use more aggressive filtering
3206 // - sort the formula so that the most profitable solutions are found first
3207 // - sort the uses too
3209 // - don't compute a cost, and then compare. compare while computing a cost
3211 // - track register sets with SmallBitVector
3213 const LSRUse &LU = Uses[Workspace.size()];
3215 // If this use references any register that's already a part of the
3216 // in-progress solution, consider it a requirement that a formula must
3217 // reference that register in order to be considered. This prunes out
3218 // unprofitable searching.
3219 SmallSetVector<const SCEV *, 4> ReqRegs;
3220 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3221 E = CurRegs.end(); I != E; ++I)
3222 if (LU.Regs.count(*I))
3225 bool AnySatisfiedReqRegs = false;
3226 SmallPtrSet<const SCEV *, 16> NewRegs;
3229 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3230 E = LU.Formulae.end(); I != E; ++I) {
3231 const Formula &F = *I;
3233 // Ignore formulae which do not use any of the required registers.
3234 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3235 JE = ReqRegs.end(); J != JE; ++J) {
3236 const SCEV *Reg = *J;
3237 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3238 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3242 AnySatisfiedReqRegs = true;
3244 // Evaluate the cost of the current formula. If it's already worse than
3245 // the current best, prune the search at that point.
3248 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3249 if (NewCost < SolutionCost) {
3250 Workspace.push_back(&F);
3251 if (Workspace.size() != Uses.size()) {
3252 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3253 NewRegs, VisitedRegs);
3254 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3255 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3257 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3258 dbgs() << ". Regs:";
3259 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3260 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3261 dbgs() << ' ' << **I;
3264 SolutionCost = NewCost;
3265 Solution = Workspace;
3267 Workspace.pop_back();
3272 // If none of the formulae had all of the required registers, relax the
3273 // constraint so that we don't exclude all formulae.
3274 if (!AnySatisfiedReqRegs) {
3275 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3281 /// Solve - Choose one formula from each use. Return the results in the given
3282 /// Solution vector.
3283 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3284 SmallVector<const Formula *, 8> Workspace;
3286 SolutionCost.Loose();
3288 SmallPtrSet<const SCEV *, 16> CurRegs;
3289 DenseSet<const SCEV *> VisitedRegs;
3290 Workspace.reserve(Uses.size());
3292 // SolveRecurse does all the work.
3293 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3294 CurRegs, VisitedRegs);
3296 // Ok, we've now made all our decisions.
3297 DEBUG(dbgs() << "\n"
3298 "The chosen solution requires "; SolutionCost.print(dbgs());
3300 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3302 Uses[i].print(dbgs());
3305 Solution[i]->print(dbgs());
3309 assert(Solution.size() == Uses.size() && "Malformed solution!");
3312 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3313 /// the dominator tree far as we can go while still being dominated by the
3314 /// input positions. This helps canonicalize the insert position, which
3315 /// encourages sharing.
3316 BasicBlock::iterator
3317 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3318 const SmallVectorImpl<Instruction *> &Inputs)
3321 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3322 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3325 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3326 if (!Rung) return IP;
3327 Rung = Rung->getIDom();
3328 if (!Rung) return IP;
3329 IDom = Rung->getBlock();
3331 // Don't climb into a loop though.
3332 const Loop *IDomLoop = LI.getLoopFor(IDom);
3333 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3334 if (IDomDepth <= IPLoopDepth &&
3335 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3339 bool AllDominate = true;
3340 Instruction *BetterPos = 0;
3341 Instruction *Tentative = IDom->getTerminator();
3342 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3343 E = Inputs.end(); I != E; ++I) {
3344 Instruction *Inst = *I;
3345 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3346 AllDominate = false;
3349 // Attempt to find an insert position in the middle of the block,
3350 // instead of at the end, so that it can be used for other expansions.
3351 if (IDom == Inst->getParent() &&
3352 (!BetterPos || DT.dominates(BetterPos, Inst)))
3353 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3366 /// AdjustInsertPositionForExpand - Determine an input position which will be
3367 /// dominated by the operands and which will dominate the result.
3368 BasicBlock::iterator
3369 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3371 const LSRUse &LU) const {
3372 // Collect some instructions which must be dominated by the
3373 // expanding replacement. These must be dominated by any operands that
3374 // will be required in the expansion.
3375 SmallVector<Instruction *, 4> Inputs;
3376 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3377 Inputs.push_back(I);
3378 if (LU.Kind == LSRUse::ICmpZero)
3379 if (Instruction *I =
3380 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3381 Inputs.push_back(I);
3382 if (LF.PostIncLoops.count(L)) {
3383 if (LF.isUseFullyOutsideLoop(L))
3384 Inputs.push_back(L->getLoopLatch()->getTerminator());
3386 Inputs.push_back(IVIncInsertPos);
3388 // The expansion must also be dominated by the increment positions of any
3389 // loops it for which it is using post-inc mode.
3390 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3391 E = LF.PostIncLoops.end(); I != E; ++I) {
3392 const Loop *PIL = *I;
3393 if (PIL == L) continue;
3395 // Be dominated by the loop exit.
3396 SmallVector<BasicBlock *, 4> ExitingBlocks;
3397 PIL->getExitingBlocks(ExitingBlocks);
3398 if (!ExitingBlocks.empty()) {
3399 BasicBlock *BB = ExitingBlocks[0];
3400 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3401 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3402 Inputs.push_back(BB->getTerminator());
3406 // Then, climb up the immediate dominator tree as far as we can go while
3407 // still being dominated by the input positions.
3408 IP = HoistInsertPosition(IP, Inputs);
3410 // Don't insert instructions before PHI nodes.
3411 while (isa<PHINode>(IP)) ++IP;
3413 // Ignore debug intrinsics.
3414 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3419 /// Expand - Emit instructions for the leading candidate expression for this
3420 /// LSRUse (this is called "expanding").
3421 Value *LSRInstance::Expand(const LSRFixup &LF,
3423 BasicBlock::iterator IP,
3424 SCEVExpander &Rewriter,
3425 SmallVectorImpl<WeakVH> &DeadInsts) const {
3426 const LSRUse &LU = Uses[LF.LUIdx];
3428 // Determine an input position which will be dominated by the operands and
3429 // which will dominate the result.
3430 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3432 // Inform the Rewriter if we have a post-increment use, so that it can
3433 // perform an advantageous expansion.
3434 Rewriter.setPostInc(LF.PostIncLoops);
3436 // This is the type that the user actually needs.
3437 const Type *OpTy = LF.OperandValToReplace->getType();
3438 // This will be the type that we'll initially expand to.
3439 const Type *Ty = F.getType();
3441 // No type known; just expand directly to the ultimate type.
3443 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3444 // Expand directly to the ultimate type if it's the right size.
3446 // This is the type to do integer arithmetic in.
3447 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
3449 // Build up a list of operands to add together to form the full base.
3450 SmallVector<const SCEV *, 8> Ops;
3452 // Expand the BaseRegs portion.
3453 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3454 E = F.BaseRegs.end(); I != E; ++I) {
3455 const SCEV *Reg = *I;
3456 assert(!Reg->isZero() && "Zero allocated in a base register!");
3458 // If we're expanding for a post-inc user, make the post-inc adjustment.
3459 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3460 Reg = TransformForPostIncUse(Denormalize, Reg,
3461 LF.UserInst, LF.OperandValToReplace,
3464 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3467 // Flush the operand list to suppress SCEVExpander hoisting.
3469 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3471 Ops.push_back(SE.getUnknown(FullV));
3474 // Expand the ScaledReg portion.
3475 Value *ICmpScaledV = 0;
3476 if (F.AM.Scale != 0) {
3477 const SCEV *ScaledS = F.ScaledReg;
3479 // If we're expanding for a post-inc user, make the post-inc adjustment.
3480 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3481 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3482 LF.UserInst, LF.OperandValToReplace,
3485 if (LU.Kind == LSRUse::ICmpZero) {
3486 // An interesting way of "folding" with an icmp is to use a negated
3487 // scale, which we'll implement by inserting it into the other operand
3489 assert(F.AM.Scale == -1 &&
3490 "The only scale supported by ICmpZero uses is -1!");
3491 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3493 // Otherwise just expand the scaled register and an explicit scale,
3494 // which is expected to be matched as part of the address.
3495 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3496 ScaledS = SE.getMulExpr(ScaledS,
3497 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3498 Ops.push_back(ScaledS);
3500 // Flush the operand list to suppress SCEVExpander hoisting.
3501 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3503 Ops.push_back(SE.getUnknown(FullV));
3507 // Expand the GV portion.
3509 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3511 // Flush the operand list to suppress SCEVExpander hoisting.
3512 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3514 Ops.push_back(SE.getUnknown(FullV));
3517 // Expand the immediate portion.
3518 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3520 if (LU.Kind == LSRUse::ICmpZero) {
3521 // The other interesting way of "folding" with an ICmpZero is to use a
3522 // negated immediate.
3524 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3526 Ops.push_back(SE.getUnknown(ICmpScaledV));
3527 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3530 // Just add the immediate values. These again are expected to be matched
3531 // as part of the address.
3532 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3536 // Expand the unfolded offset portion.
3537 int64_t UnfoldedOffset = F.UnfoldedOffset;
3538 if (UnfoldedOffset != 0) {
3539 // Just add the immediate values.
3540 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
3544 // Emit instructions summing all the operands.
3545 const SCEV *FullS = Ops.empty() ?
3546 SE.getConstant(IntTy, 0) :
3548 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3550 // We're done expanding now, so reset the rewriter.
3551 Rewriter.clearPostInc();
3553 // An ICmpZero Formula represents an ICmp which we're handling as a
3554 // comparison against zero. Now that we've expanded an expression for that
3555 // form, update the ICmp's other operand.
3556 if (LU.Kind == LSRUse::ICmpZero) {
3557 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3558 DeadInsts.push_back(CI->getOperand(1));
3559 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3560 "a scale at the same time!");
3561 if (F.AM.Scale == -1) {
3562 if (ICmpScaledV->getType() != OpTy) {
3564 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3566 ICmpScaledV, OpTy, "tmp", CI);
3569 CI->setOperand(1, ICmpScaledV);
3571 assert(F.AM.Scale == 0 &&
3572 "ICmp does not support folding a global value and "
3573 "a scale at the same time!");
3574 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3576 if (C->getType() != OpTy)
3577 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3581 CI->setOperand(1, C);
3588 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3589 /// of their operands effectively happens in their predecessor blocks, so the
3590 /// expression may need to be expanded in multiple places.
3591 void LSRInstance::RewriteForPHI(PHINode *PN,
3594 SCEVExpander &Rewriter,
3595 SmallVectorImpl<WeakVH> &DeadInsts,
3597 DenseMap<BasicBlock *, Value *> Inserted;
3598 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3599 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3600 BasicBlock *BB = PN->getIncomingBlock(i);
3602 // If this is a critical edge, split the edge so that we do not insert
3603 // the code on all predecessor/successor paths. We do this unless this
3604 // is the canonical backedge for this loop, which complicates post-inc
3606 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3607 !isa<IndirectBrInst>(BB->getTerminator())) {
3608 Loop *PNLoop = LI.getLoopFor(PN->getParent());
3609 if (!PNLoop || PN->getParent() != PNLoop->getHeader()) {
3610 // Split the critical edge.
3611 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3613 // If PN is outside of the loop and BB is in the loop, we want to
3614 // move the block to be immediately before the PHI block, not
3615 // immediately after BB.
3616 if (L->contains(BB) && !L->contains(PN))
3617 NewBB->moveBefore(PN->getParent());
3619 // Splitting the edge can reduce the number of PHI entries we have.
3620 e = PN->getNumIncomingValues();
3622 i = PN->getBasicBlockIndex(BB);
3626 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3627 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3629 PN->setIncomingValue(i, Pair.first->second);
3631 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3633 // If this is reuse-by-noop-cast, insert the noop cast.
3634 const Type *OpTy = LF.OperandValToReplace->getType();
3635 if (FullV->getType() != OpTy)
3637 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3639 FullV, LF.OperandValToReplace->getType(),
3640 "tmp", BB->getTerminator());
3642 PN->setIncomingValue(i, FullV);
3643 Pair.first->second = FullV;
3648 /// Rewrite - Emit instructions for the leading candidate expression for this
3649 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3650 /// the newly expanded value.
3651 void LSRInstance::Rewrite(const LSRFixup &LF,
3653 SCEVExpander &Rewriter,
3654 SmallVectorImpl<WeakVH> &DeadInsts,
3656 // First, find an insertion point that dominates UserInst. For PHI nodes,
3657 // find the nearest block which dominates all the relevant uses.
3658 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3659 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3661 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3663 // If this is reuse-by-noop-cast, insert the noop cast.
3664 const Type *OpTy = LF.OperandValToReplace->getType();
3665 if (FullV->getType() != OpTy) {
3667 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3668 FullV, OpTy, "tmp", LF.UserInst);
3672 // Update the user. ICmpZero is handled specially here (for now) because
3673 // Expand may have updated one of the operands of the icmp already, and
3674 // its new value may happen to be equal to LF.OperandValToReplace, in
3675 // which case doing replaceUsesOfWith leads to replacing both operands
3676 // with the same value. TODO: Reorganize this.
3677 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3678 LF.UserInst->setOperand(0, FullV);
3680 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3683 DeadInsts.push_back(LF.OperandValToReplace);
3686 /// ImplementSolution - Rewrite all the fixup locations with new values,
3687 /// following the chosen solution.
3689 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3691 // Keep track of instructions we may have made dead, so that
3692 // we can remove them after we are done working.
3693 SmallVector<WeakVH, 16> DeadInsts;
3695 SCEVExpander Rewriter(SE);
3696 Rewriter.disableCanonicalMode();
3697 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3699 // Expand the new value definitions and update the users.
3700 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3701 E = Fixups.end(); I != E; ++I) {
3702 const LSRFixup &Fixup = *I;
3704 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3709 // Clean up after ourselves. This must be done before deleting any
3713 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3716 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3717 : IU(P->getAnalysis<IVUsers>()),
3718 SE(P->getAnalysis<ScalarEvolution>()),
3719 DT(P->getAnalysis<DominatorTree>()),
3720 LI(P->getAnalysis<LoopInfo>()),
3721 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3723 // If LoopSimplify form is not available, stay out of trouble.
3724 if (!L->isLoopSimplifyForm()) return;
3726 // If there's no interesting work to be done, bail early.
3727 if (IU.empty()) return;
3729 DEBUG(dbgs() << "\nLSR on loop ";
3730 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3733 // First, perform some low-level loop optimizations.
3735 OptimizeLoopTermCond();
3737 // Start collecting data and preparing for the solver.
3738 CollectInterestingTypesAndFactors();
3739 CollectFixupsAndInitialFormulae();
3740 CollectLoopInvariantFixupsAndFormulae();
3742 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3743 print_uses(dbgs()));
3745 // Now use the reuse data to generate a bunch of interesting ways
3746 // to formulate the values needed for the uses.
3747 GenerateAllReuseFormulae();
3749 FilterOutUndesirableDedicatedRegisters();
3750 NarrowSearchSpaceUsingHeuristics();
3752 SmallVector<const Formula *, 8> Solution;
3755 // Release memory that is no longer needed.
3761 // Formulae should be legal.
3762 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3763 E = Uses.end(); I != E; ++I) {
3764 const LSRUse &LU = *I;
3765 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3766 JE = LU.Formulae.end(); J != JE; ++J)
3767 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3768 LU.Kind, LU.AccessTy, TLI) &&
3769 "Illegal formula generated!");
3773 // Now that we've decided what we want, make it so.
3774 ImplementSolution(Solution, P);
3777 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3778 if (Factors.empty() && Types.empty()) return;
3780 OS << "LSR has identified the following interesting factors and types: ";
3783 for (SmallSetVector<int64_t, 8>::const_iterator
3784 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3785 if (!First) OS << ", ";
3790 for (SmallSetVector<const Type *, 4>::const_iterator
3791 I = Types.begin(), E = Types.end(); I != E; ++I) {
3792 if (!First) OS << ", ";
3794 OS << '(' << **I << ')';
3799 void LSRInstance::print_fixups(raw_ostream &OS) const {
3800 OS << "LSR is examining the following fixup sites:\n";
3801 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3802 E = Fixups.end(); I != E; ++I) {
3809 void LSRInstance::print_uses(raw_ostream &OS) const {
3810 OS << "LSR is examining the following uses:\n";
3811 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3812 E = Uses.end(); I != E; ++I) {
3813 const LSRUse &LU = *I;
3817 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3818 JE = LU.Formulae.end(); J != JE; ++J) {
3826 void LSRInstance::print(raw_ostream &OS) const {
3827 print_factors_and_types(OS);
3832 void LSRInstance::dump() const {
3833 print(errs()); errs() << '\n';
3838 class LoopStrengthReduce : public LoopPass {
3839 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3840 /// transformation profitability.
3841 const TargetLowering *const TLI;
3844 static char ID; // Pass ID, replacement for typeid
3845 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3848 bool runOnLoop(Loop *L, LPPassManager &LPM);
3849 void getAnalysisUsage(AnalysisUsage &AU) const;
3854 char LoopStrengthReduce::ID = 0;
3855 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
3856 "Loop Strength Reduction", false, false)
3857 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3858 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3859 INITIALIZE_PASS_DEPENDENCY(IVUsers)
3860 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
3861 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3862 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
3863 "Loop Strength Reduction", false, false)
3866 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3867 return new LoopStrengthReduce(TLI);
3870 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3871 : LoopPass(ID), TLI(tli) {
3872 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
3875 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3876 // We split critical edges, so we change the CFG. However, we do update
3877 // many analyses if they are around.
3878 AU.addPreservedID(LoopSimplifyID);
3880 AU.addRequired<LoopInfo>();
3881 AU.addPreserved<LoopInfo>();
3882 AU.addRequiredID(LoopSimplifyID);
3883 AU.addRequired<DominatorTree>();
3884 AU.addPreserved<DominatorTree>();
3885 AU.addRequired<ScalarEvolution>();
3886 AU.addPreserved<ScalarEvolution>();
3887 // Requiring LoopSimplify a second time here prevents IVUsers from running
3888 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
3889 AU.addRequiredID(LoopSimplifyID);
3890 AU.addRequired<IVUsers>();
3891 AU.addPreserved<IVUsers>();
3894 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3895 bool Changed = false;
3897 // Run the main LSR transformation.
3898 Changed |= LSRInstance(TLI, L, this).getChanged();
3900 // At this point, it is worth checking to see if any recurrence PHIs are also
3901 // dead, so that we can remove them as well.
3902 Changed |= DeleteDeadPHIs(L->getHeader());