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 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 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 Type *getAccessType(const Instruction *Inst) {
598 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 (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;
674 // Once any of the metrics loses, they must all remain losers.
676 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
677 | ImmCost | SetupCost) != ~0u)
678 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
679 & ImmCost & SetupCost) == ~0u);
684 assert(isValid() && "invalid cost");
685 return NumRegs == ~0u;
688 void RateFormula(const Formula &F,
689 SmallPtrSet<const SCEV *, 16> &Regs,
690 const DenseSet<const SCEV *> &VisitedRegs,
692 const SmallVectorImpl<int64_t> &Offsets,
693 ScalarEvolution &SE, DominatorTree &DT);
695 void print(raw_ostream &OS) const;
699 void RateRegister(const SCEV *Reg,
700 SmallPtrSet<const SCEV *, 16> &Regs,
702 ScalarEvolution &SE, DominatorTree &DT);
703 void RatePrimaryRegister(const SCEV *Reg,
704 SmallPtrSet<const SCEV *, 16> &Regs,
706 ScalarEvolution &SE, DominatorTree &DT);
711 /// RateRegister - Tally up interesting quantities from the given register.
712 void Cost::RateRegister(const SCEV *Reg,
713 SmallPtrSet<const SCEV *, 16> &Regs,
715 ScalarEvolution &SE, DominatorTree &DT) {
716 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
717 if (AR->getLoop() == L)
718 AddRecCost += 1; /// TODO: This should be a function of the stride.
720 // If this is an addrec for a loop that's already been visited by LSR,
721 // don't second-guess its addrec phi nodes. LSR isn't currently smart
722 // enough to reason about more than one loop at a time. Consider these
723 // registers free and leave them alone.
724 else if (L->contains(AR->getLoop()) ||
725 (!AR->getLoop()->contains(L) &&
726 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
727 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
728 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
729 if (SE.isSCEVable(PN->getType()) &&
730 (SE.getEffectiveSCEVType(PN->getType()) ==
731 SE.getEffectiveSCEVType(AR->getType())) &&
732 SE.getSCEV(PN) == AR)
735 // If this isn't one of the addrecs that the loop already has, it
736 // would require a costly new phi and add. TODO: This isn't
737 // precisely modeled right now.
739 if (!Regs.count(AR->getStart())) {
740 RateRegister(AR->getStart(), Regs, L, SE, DT);
746 // Add the step value register, if it needs one.
747 // TODO: The non-affine case isn't precisely modeled here.
748 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
749 if (!Regs.count(AR->getOperand(1))) {
750 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
758 // Rough heuristic; favor registers which don't require extra setup
759 // instructions in the preheader.
760 if (!isa<SCEVUnknown>(Reg) &&
761 !isa<SCEVConstant>(Reg) &&
762 !(isa<SCEVAddRecExpr>(Reg) &&
763 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
764 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
767 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
768 SE.hasComputableLoopEvolution(Reg, L);
771 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
773 void Cost::RatePrimaryRegister(const SCEV *Reg,
774 SmallPtrSet<const SCEV *, 16> &Regs,
776 ScalarEvolution &SE, DominatorTree &DT) {
777 if (Regs.insert(Reg))
778 RateRegister(Reg, Regs, L, SE, DT);
781 void Cost::RateFormula(const Formula &F,
782 SmallPtrSet<const SCEV *, 16> &Regs,
783 const DenseSet<const SCEV *> &VisitedRegs,
785 const SmallVectorImpl<int64_t> &Offsets,
786 ScalarEvolution &SE, DominatorTree &DT) {
787 // Tally up the registers.
788 if (const SCEV *ScaledReg = F.ScaledReg) {
789 if (VisitedRegs.count(ScaledReg)) {
793 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
797 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
798 E = F.BaseRegs.end(); I != E; ++I) {
799 const SCEV *BaseReg = *I;
800 if (VisitedRegs.count(BaseReg)) {
804 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
809 // Determine how many (unfolded) adds we'll need inside the loop.
810 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
811 if (NumBaseParts > 1)
812 NumBaseAdds += NumBaseParts - 1;
814 // Tally up the non-zero immediates.
815 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
816 E = Offsets.end(); I != E; ++I) {
817 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
819 ImmCost += 64; // Handle symbolic values conservatively.
820 // TODO: This should probably be the pointer size.
821 else if (Offset != 0)
822 ImmCost += APInt(64, Offset, true).getMinSignedBits();
824 assert(isValid() && "invalid cost");
827 /// Loose - Set this cost to a losing value.
837 /// operator< - Choose the lower cost.
838 bool Cost::operator<(const Cost &Other) const {
839 if (NumRegs != Other.NumRegs)
840 return NumRegs < Other.NumRegs;
841 if (AddRecCost != Other.AddRecCost)
842 return AddRecCost < Other.AddRecCost;
843 if (NumIVMuls != Other.NumIVMuls)
844 return NumIVMuls < Other.NumIVMuls;
845 if (NumBaseAdds != Other.NumBaseAdds)
846 return NumBaseAdds < Other.NumBaseAdds;
847 if (ImmCost != Other.ImmCost)
848 return ImmCost < Other.ImmCost;
849 if (SetupCost != Other.SetupCost)
850 return SetupCost < Other.SetupCost;
854 void Cost::print(raw_ostream &OS) const {
855 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
857 OS << ", with addrec cost " << AddRecCost;
859 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
860 if (NumBaseAdds != 0)
861 OS << ", plus " << NumBaseAdds << " base add"
862 << (NumBaseAdds == 1 ? "" : "s");
864 OS << ", plus " << ImmCost << " imm cost";
866 OS << ", plus " << SetupCost << " setup cost";
869 void Cost::dump() const {
870 print(errs()); errs() << '\n';
875 /// LSRFixup - An operand value in an instruction which is to be replaced
876 /// with some equivalent, possibly strength-reduced, replacement.
878 /// UserInst - The instruction which will be updated.
879 Instruction *UserInst;
881 /// OperandValToReplace - The operand of the instruction which will
882 /// be replaced. The operand may be used more than once; every instance
883 /// will be replaced.
884 Value *OperandValToReplace;
886 /// PostIncLoops - If this user is to use the post-incremented value of an
887 /// induction variable, this variable is non-null and holds the loop
888 /// associated with the induction variable.
889 PostIncLoopSet PostIncLoops;
891 /// LUIdx - The index of the LSRUse describing the expression which
892 /// this fixup needs, minus an offset (below).
895 /// Offset - A constant offset to be added to the LSRUse expression.
896 /// This allows multiple fixups to share the same LSRUse with different
897 /// offsets, for example in an unrolled loop.
900 bool isUseFullyOutsideLoop(const Loop *L) const;
904 void print(raw_ostream &OS) const;
911 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
913 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
914 /// value outside of the given loop.
915 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
916 // PHI nodes use their value in their incoming blocks.
917 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
918 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
919 if (PN->getIncomingValue(i) == OperandValToReplace &&
920 L->contains(PN->getIncomingBlock(i)))
925 return !L->contains(UserInst);
928 void LSRFixup::print(raw_ostream &OS) const {
930 // Store is common and interesting enough to be worth special-casing.
931 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
933 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
934 } else if (UserInst->getType()->isVoidTy())
935 OS << UserInst->getOpcodeName();
937 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
939 OS << ", OperandValToReplace=";
940 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
942 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
943 E = PostIncLoops.end(); I != E; ++I) {
944 OS << ", PostIncLoop=";
945 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
948 if (LUIdx != ~size_t(0))
949 OS << ", LUIdx=" << LUIdx;
952 OS << ", Offset=" << Offset;
955 void LSRFixup::dump() const {
956 print(errs()); errs() << '\n';
961 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
962 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
963 struct UniquifierDenseMapInfo {
964 static SmallVector<const SCEV *, 2> getEmptyKey() {
965 SmallVector<const SCEV *, 2> V;
966 V.push_back(reinterpret_cast<const SCEV *>(-1));
970 static SmallVector<const SCEV *, 2> getTombstoneKey() {
971 SmallVector<const SCEV *, 2> V;
972 V.push_back(reinterpret_cast<const SCEV *>(-2));
976 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
978 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
979 E = V.end(); I != E; ++I)
980 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
984 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
985 const SmallVector<const SCEV *, 2> &RHS) {
990 /// LSRUse - This class holds the state that LSR keeps for each use in
991 /// IVUsers, as well as uses invented by LSR itself. It includes information
992 /// about what kinds of things can be folded into the user, information about
993 /// the user itself, and information about how the use may be satisfied.
994 /// TODO: Represent multiple users of the same expression in common?
996 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
999 /// KindType - An enum for a kind of use, indicating what types of
1000 /// scaled and immediate operands it might support.
1002 Basic, ///< A normal use, with no folding.
1003 Special, ///< A special case of basic, allowing -1 scales.
1004 Address, ///< An address use; folding according to TargetLowering
1005 ICmpZero ///< An equality icmp with both operands folded into one.
1006 // TODO: Add a generic icmp too?
1012 SmallVector<int64_t, 8> Offsets;
1016 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1017 /// LSRUse are outside of the loop, in which case some special-case heuristics
1019 bool AllFixupsOutsideLoop;
1021 /// WidestFixupType - This records the widest use type for any fixup using
1022 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1023 /// max fixup widths to be equivalent, because the narrower one may be relying
1024 /// on the implicit truncation to truncate away bogus bits.
1025 Type *WidestFixupType;
1027 /// Formulae - A list of ways to build a value that can satisfy this user.
1028 /// After the list is populated, one of these is selected heuristically and
1029 /// used to formulate a replacement for OperandValToReplace in UserInst.
1030 SmallVector<Formula, 12> Formulae;
1032 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1033 SmallPtrSet<const SCEV *, 4> Regs;
1035 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1036 MinOffset(INT64_MAX),
1037 MaxOffset(INT64_MIN),
1038 AllFixupsOutsideLoop(true),
1039 WidestFixupType(0) {}
1041 bool HasFormulaWithSameRegs(const Formula &F) const;
1042 bool InsertFormula(const Formula &F);
1043 void DeleteFormula(Formula &F);
1044 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1046 void print(raw_ostream &OS) const;
1052 /// HasFormula - Test whether this use as a formula which has the same
1053 /// registers as the given formula.
1054 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1055 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1056 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1057 // Unstable sort by host order ok, because this is only used for uniquifying.
1058 std::sort(Key.begin(), Key.end());
1059 return Uniquifier.count(Key);
1062 /// InsertFormula - If the given formula has not yet been inserted, add it to
1063 /// the list, and return true. Return false otherwise.
1064 bool LSRUse::InsertFormula(const Formula &F) {
1065 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1066 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1067 // Unstable sort by host order ok, because this is only used for uniquifying.
1068 std::sort(Key.begin(), Key.end());
1070 if (!Uniquifier.insert(Key).second)
1073 // Using a register to hold the value of 0 is not profitable.
1074 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1075 "Zero allocated in a scaled register!");
1077 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1078 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1079 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1082 // Add the formula to the list.
1083 Formulae.push_back(F);
1085 // Record registers now being used by this use.
1086 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1087 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1092 /// DeleteFormula - Remove the given formula from this use's list.
1093 void LSRUse::DeleteFormula(Formula &F) {
1094 if (&F != &Formulae.back())
1095 std::swap(F, Formulae.back());
1096 Formulae.pop_back();
1097 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1100 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1101 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1102 // Now that we've filtered out some formulae, recompute the Regs set.
1103 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1105 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1106 E = Formulae.end(); I != E; ++I) {
1107 const Formula &F = *I;
1108 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1109 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1112 // Update the RegTracker.
1113 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1114 E = OldRegs.end(); I != E; ++I)
1115 if (!Regs.count(*I))
1116 RegUses.DropRegister(*I, LUIdx);
1119 void LSRUse::print(raw_ostream &OS) const {
1120 OS << "LSR Use: Kind=";
1122 case Basic: OS << "Basic"; break;
1123 case Special: OS << "Special"; break;
1124 case ICmpZero: OS << "ICmpZero"; break;
1126 OS << "Address of ";
1127 if (AccessTy->isPointerTy())
1128 OS << "pointer"; // the full pointer type could be really verbose
1133 OS << ", Offsets={";
1134 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1135 E = Offsets.end(); I != E; ++I) {
1137 if (llvm::next(I) != E)
1142 if (AllFixupsOutsideLoop)
1143 OS << ", all-fixups-outside-loop";
1145 if (WidestFixupType)
1146 OS << ", widest fixup type: " << *WidestFixupType;
1149 void LSRUse::dump() const {
1150 print(errs()); errs() << '\n';
1153 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1154 /// be completely folded into the user instruction at isel time. This includes
1155 /// address-mode folding and special icmp tricks.
1156 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1157 LSRUse::KindType Kind, Type *AccessTy,
1158 const TargetLowering *TLI) {
1160 case LSRUse::Address:
1161 // If we have low-level target information, ask the target if it can
1162 // completely fold this address.
1163 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1165 // Otherwise, just guess that reg+reg addressing is legal.
1166 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1168 case LSRUse::ICmpZero:
1169 // There's not even a target hook for querying whether it would be legal to
1170 // fold a GV into an ICmp.
1174 // ICmp only has two operands; don't allow more than two non-trivial parts.
1175 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1178 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1179 // putting the scaled register in the other operand of the icmp.
1180 if (AM.Scale != 0 && AM.Scale != -1)
1183 // If we have low-level target information, ask the target if it can fold an
1184 // integer immediate on an icmp.
1185 if (AM.BaseOffs != 0) {
1186 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1193 // Only handle single-register values.
1194 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1196 case LSRUse::Special:
1197 // Only handle -1 scales, or no scale.
1198 return AM.Scale == 0 || AM.Scale == -1;
1204 static bool isLegalUse(TargetLowering::AddrMode AM,
1205 int64_t MinOffset, int64_t MaxOffset,
1206 LSRUse::KindType Kind, Type *AccessTy,
1207 const TargetLowering *TLI) {
1208 // Check for overflow.
1209 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1212 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1213 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1214 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1215 // Check for overflow.
1216 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1219 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1220 return isLegalUse(AM, Kind, AccessTy, TLI);
1225 static bool isAlwaysFoldable(int64_t BaseOffs,
1226 GlobalValue *BaseGV,
1228 LSRUse::KindType Kind, Type *AccessTy,
1229 const TargetLowering *TLI) {
1230 // Fast-path: zero is always foldable.
1231 if (BaseOffs == 0 && !BaseGV) return true;
1233 // Conservatively, create an address with an immediate and a
1234 // base and a scale.
1235 TargetLowering::AddrMode AM;
1236 AM.BaseOffs = BaseOffs;
1238 AM.HasBaseReg = HasBaseReg;
1239 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1241 // Canonicalize a scale of 1 to a base register if the formula doesn't
1242 // already have a base register.
1243 if (!AM.HasBaseReg && AM.Scale == 1) {
1245 AM.HasBaseReg = true;
1248 return isLegalUse(AM, Kind, AccessTy, TLI);
1251 static bool isAlwaysFoldable(const SCEV *S,
1252 int64_t MinOffset, int64_t MaxOffset,
1254 LSRUse::KindType Kind, Type *AccessTy,
1255 const TargetLowering *TLI,
1256 ScalarEvolution &SE) {
1257 // Fast-path: zero is always foldable.
1258 if (S->isZero()) return true;
1260 // Conservatively, create an address with an immediate and a
1261 // base and a scale.
1262 int64_t BaseOffs = ExtractImmediate(S, SE);
1263 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1265 // If there's anything else involved, it's not foldable.
1266 if (!S->isZero()) return false;
1268 // Fast-path: zero is always foldable.
1269 if (BaseOffs == 0 && !BaseGV) return true;
1271 // Conservatively, create an address with an immediate and a
1272 // base and a scale.
1273 TargetLowering::AddrMode AM;
1274 AM.BaseOffs = BaseOffs;
1276 AM.HasBaseReg = HasBaseReg;
1277 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1279 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1284 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1285 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1286 struct UseMapDenseMapInfo {
1287 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1288 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1291 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1292 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1296 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1297 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1298 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1302 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1303 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1308 /// LSRInstance - This class holds state for the main loop strength reduction
1312 ScalarEvolution &SE;
1315 const TargetLowering *const TLI;
1319 /// IVIncInsertPos - This is the insert position that the current loop's
1320 /// induction variable increment should be placed. In simple loops, this is
1321 /// the latch block's terminator. But in more complicated cases, this is a
1322 /// position which will dominate all the in-loop post-increment users.
1323 Instruction *IVIncInsertPos;
1325 /// Factors - Interesting factors between use strides.
1326 SmallSetVector<int64_t, 8> Factors;
1328 /// Types - Interesting use types, to facilitate truncation reuse.
1329 SmallSetVector<Type *, 4> Types;
1331 /// Fixups - The list of operands which are to be replaced.
1332 SmallVector<LSRFixup, 16> Fixups;
1334 /// Uses - The list of interesting uses.
1335 SmallVector<LSRUse, 16> Uses;
1337 /// RegUses - Track which uses use which register candidates.
1338 RegUseTracker RegUses;
1340 void OptimizeShadowIV();
1341 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1342 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1343 void OptimizeLoopTermCond();
1345 void CollectInterestingTypesAndFactors();
1346 void CollectFixupsAndInitialFormulae();
1348 LSRFixup &getNewFixup() {
1349 Fixups.push_back(LSRFixup());
1350 return Fixups.back();
1353 // Support for sharing of LSRUses between LSRFixups.
1354 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1356 UseMapDenseMapInfo> UseMapTy;
1359 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1360 LSRUse::KindType Kind, Type *AccessTy);
1362 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1363 LSRUse::KindType Kind,
1366 void DeleteUse(LSRUse &LU, size_t LUIdx);
1368 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1371 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1372 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1373 void CountRegisters(const Formula &F, size_t LUIdx);
1374 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1376 void CollectLoopInvariantFixupsAndFormulae();
1378 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1379 unsigned Depth = 0);
1380 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1381 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1382 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1383 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1384 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1385 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1386 void GenerateCrossUseConstantOffsets();
1387 void GenerateAllReuseFormulae();
1389 void FilterOutUndesirableDedicatedRegisters();
1391 size_t EstimateSearchSpaceComplexity() const;
1392 void NarrowSearchSpaceByDetectingSupersets();
1393 void NarrowSearchSpaceByCollapsingUnrolledCode();
1394 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1395 void NarrowSearchSpaceByPickingWinnerRegs();
1396 void NarrowSearchSpaceUsingHeuristics();
1398 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1400 SmallVectorImpl<const Formula *> &Workspace,
1401 const Cost &CurCost,
1402 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1403 DenseSet<const SCEV *> &VisitedRegs) const;
1404 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1406 BasicBlock::iterator
1407 HoistInsertPosition(BasicBlock::iterator IP,
1408 const SmallVectorImpl<Instruction *> &Inputs) const;
1409 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1411 const LSRUse &LU) const;
1413 Value *Expand(const LSRFixup &LF,
1415 BasicBlock::iterator IP,
1416 SCEVExpander &Rewriter,
1417 SmallVectorImpl<WeakVH> &DeadInsts) const;
1418 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1420 SCEVExpander &Rewriter,
1421 SmallVectorImpl<WeakVH> &DeadInsts,
1423 void Rewrite(const LSRFixup &LF,
1425 SCEVExpander &Rewriter,
1426 SmallVectorImpl<WeakVH> &DeadInsts,
1428 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1431 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1433 bool getChanged() const { return Changed; }
1435 void print_factors_and_types(raw_ostream &OS) const;
1436 void print_fixups(raw_ostream &OS) const;
1437 void print_uses(raw_ostream &OS) const;
1438 void print(raw_ostream &OS) const;
1444 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1445 /// inside the loop then try to eliminate the cast operation.
1446 void LSRInstance::OptimizeShadowIV() {
1447 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1448 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1451 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1452 UI != E; /* empty */) {
1453 IVUsers::const_iterator CandidateUI = UI;
1455 Instruction *ShadowUse = CandidateUI->getUser();
1456 Type *DestTy = NULL;
1457 bool IsSigned = false;
1459 /* If shadow use is a int->float cast then insert a second IV
1460 to eliminate this cast.
1462 for (unsigned i = 0; i < n; ++i)
1468 for (unsigned i = 0; i < n; ++i, ++d)
1471 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1473 DestTy = UCast->getDestTy();
1475 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1477 DestTy = SCast->getDestTy();
1479 if (!DestTy) continue;
1482 // If target does not support DestTy natively then do not apply
1483 // this transformation.
1484 EVT DVT = TLI->getValueType(DestTy);
1485 if (!TLI->isTypeLegal(DVT)) continue;
1488 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1490 if (PH->getNumIncomingValues() != 2) continue;
1492 Type *SrcTy = PH->getType();
1493 int Mantissa = DestTy->getFPMantissaWidth();
1494 if (Mantissa == -1) continue;
1495 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1498 unsigned Entry, Latch;
1499 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1507 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1508 if (!Init) continue;
1509 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1510 (double)Init->getSExtValue() :
1511 (double)Init->getZExtValue());
1513 BinaryOperator *Incr =
1514 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1515 if (!Incr) continue;
1516 if (Incr->getOpcode() != Instruction::Add
1517 && Incr->getOpcode() != Instruction::Sub)
1520 /* Initialize new IV, double d = 0.0 in above example. */
1521 ConstantInt *C = NULL;
1522 if (Incr->getOperand(0) == PH)
1523 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1524 else if (Incr->getOperand(1) == PH)
1525 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1531 // Ignore negative constants, as the code below doesn't handle them
1532 // correctly. TODO: Remove this restriction.
1533 if (!C->getValue().isStrictlyPositive()) continue;
1535 /* Add new PHINode. */
1536 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1538 /* create new increment. '++d' in above example. */
1539 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1540 BinaryOperator *NewIncr =
1541 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1542 Instruction::FAdd : Instruction::FSub,
1543 NewPH, CFP, "IV.S.next.", Incr);
1545 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1546 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1548 /* Remove cast operation */
1549 ShadowUse->replaceAllUsesWith(NewPH);
1550 ShadowUse->eraseFromParent();
1556 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1557 /// set the IV user and stride information and return true, otherwise return
1559 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1560 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1561 if (UI->getUser() == Cond) {
1562 // NOTE: we could handle setcc instructions with multiple uses here, but
1563 // InstCombine does it as well for simple uses, it's not clear that it
1564 // occurs enough in real life to handle.
1571 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1572 /// a max computation.
1574 /// This is a narrow solution to a specific, but acute, problem. For loops
1580 /// } while (++i < n);
1582 /// the trip count isn't just 'n', because 'n' might not be positive. And
1583 /// unfortunately this can come up even for loops where the user didn't use
1584 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1585 /// will commonly be lowered like this:
1591 /// } while (++i < n);
1594 /// and then it's possible for subsequent optimization to obscure the if
1595 /// test in such a way that indvars can't find it.
1597 /// When indvars can't find the if test in loops like this, it creates a
1598 /// max expression, which allows it to give the loop a canonical
1599 /// induction variable:
1602 /// max = n < 1 ? 1 : n;
1605 /// } while (++i != max);
1607 /// Canonical induction variables are necessary because the loop passes
1608 /// are designed around them. The most obvious example of this is the
1609 /// LoopInfo analysis, which doesn't remember trip count values. It
1610 /// expects to be able to rediscover the trip count each time it is
1611 /// needed, and it does this using a simple analysis that only succeeds if
1612 /// the loop has a canonical induction variable.
1614 /// However, when it comes time to generate code, the maximum operation
1615 /// can be quite costly, especially if it's inside of an outer loop.
1617 /// This function solves this problem by detecting this type of loop and
1618 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1619 /// the instructions for the maximum computation.
1621 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1622 // Check that the loop matches the pattern we're looking for.
1623 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1624 Cond->getPredicate() != CmpInst::ICMP_NE)
1627 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1628 if (!Sel || !Sel->hasOneUse()) return Cond;
1630 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1631 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1633 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1635 // Add one to the backedge-taken count to get the trip count.
1636 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1637 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1639 // Check for a max calculation that matches the pattern. There's no check
1640 // for ICMP_ULE here because the comparison would be with zero, which
1641 // isn't interesting.
1642 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1643 const SCEVNAryExpr *Max = 0;
1644 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1645 Pred = ICmpInst::ICMP_SLE;
1647 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1648 Pred = ICmpInst::ICMP_SLT;
1650 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1651 Pred = ICmpInst::ICMP_ULT;
1658 // To handle a max with more than two operands, this optimization would
1659 // require additional checking and setup.
1660 if (Max->getNumOperands() != 2)
1663 const SCEV *MaxLHS = Max->getOperand(0);
1664 const SCEV *MaxRHS = Max->getOperand(1);
1666 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1667 // for a comparison with 1. For <= and >=, a comparison with zero.
1669 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1672 // Check the relevant induction variable for conformance to
1674 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1675 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1676 if (!AR || !AR->isAffine() ||
1677 AR->getStart() != One ||
1678 AR->getStepRecurrence(SE) != One)
1681 assert(AR->getLoop() == L &&
1682 "Loop condition operand is an addrec in a different loop!");
1684 // Check the right operand of the select, and remember it, as it will
1685 // be used in the new comparison instruction.
1687 if (ICmpInst::isTrueWhenEqual(Pred)) {
1688 // Look for n+1, and grab n.
1689 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1690 if (isa<ConstantInt>(BO->getOperand(1)) &&
1691 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1692 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1693 NewRHS = BO->getOperand(0);
1694 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1695 if (isa<ConstantInt>(BO->getOperand(1)) &&
1696 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1697 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1698 NewRHS = BO->getOperand(0);
1701 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1702 NewRHS = Sel->getOperand(1);
1703 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1704 NewRHS = Sel->getOperand(2);
1705 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1706 NewRHS = SU->getValue();
1708 // Max doesn't match expected pattern.
1711 // Determine the new comparison opcode. It may be signed or unsigned,
1712 // and the original comparison may be either equality or inequality.
1713 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1714 Pred = CmpInst::getInversePredicate(Pred);
1716 // Ok, everything looks ok to change the condition into an SLT or SGE and
1717 // delete the max calculation.
1719 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1721 // Delete the max calculation instructions.
1722 Cond->replaceAllUsesWith(NewCond);
1723 CondUse->setUser(NewCond);
1724 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1725 Cond->eraseFromParent();
1726 Sel->eraseFromParent();
1727 if (Cmp->use_empty())
1728 Cmp->eraseFromParent();
1732 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1733 /// postinc iv when possible.
1735 LSRInstance::OptimizeLoopTermCond() {
1736 SmallPtrSet<Instruction *, 4> PostIncs;
1738 BasicBlock *LatchBlock = L->getLoopLatch();
1739 SmallVector<BasicBlock*, 8> ExitingBlocks;
1740 L->getExitingBlocks(ExitingBlocks);
1742 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1743 BasicBlock *ExitingBlock = ExitingBlocks[i];
1745 // Get the terminating condition for the loop if possible. If we
1746 // can, we want to change it to use a post-incremented version of its
1747 // induction variable, to allow coalescing the live ranges for the IV into
1748 // one register value.
1750 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1753 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1754 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1757 // Search IVUsesByStride to find Cond's IVUse if there is one.
1758 IVStrideUse *CondUse = 0;
1759 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1760 if (!FindIVUserForCond(Cond, CondUse))
1763 // If the trip count is computed in terms of a max (due to ScalarEvolution
1764 // being unable to find a sufficient guard, for example), change the loop
1765 // comparison to use SLT or ULT instead of NE.
1766 // One consequence of doing this now is that it disrupts the count-down
1767 // optimization. That's not always a bad thing though, because in such
1768 // cases it may still be worthwhile to avoid a max.
1769 Cond = OptimizeMax(Cond, CondUse);
1771 // If this exiting block dominates the latch block, it may also use
1772 // the post-inc value if it won't be shared with other uses.
1773 // Check for dominance.
1774 if (!DT.dominates(ExitingBlock, LatchBlock))
1777 // Conservatively avoid trying to use the post-inc value in non-latch
1778 // exits if there may be pre-inc users in intervening blocks.
1779 if (LatchBlock != ExitingBlock)
1780 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1781 // Test if the use is reachable from the exiting block. This dominator
1782 // query is a conservative approximation of reachability.
1783 if (&*UI != CondUse &&
1784 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1785 // Conservatively assume there may be reuse if the quotient of their
1786 // strides could be a legal scale.
1787 const SCEV *A = IU.getStride(*CondUse, L);
1788 const SCEV *B = IU.getStride(*UI, L);
1789 if (!A || !B) continue;
1790 if (SE.getTypeSizeInBits(A->getType()) !=
1791 SE.getTypeSizeInBits(B->getType())) {
1792 if (SE.getTypeSizeInBits(A->getType()) >
1793 SE.getTypeSizeInBits(B->getType()))
1794 B = SE.getSignExtendExpr(B, A->getType());
1796 A = SE.getSignExtendExpr(A, B->getType());
1798 if (const SCEVConstant *D =
1799 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1800 const ConstantInt *C = D->getValue();
1801 // Stride of one or negative one can have reuse with non-addresses.
1802 if (C->isOne() || C->isAllOnesValue())
1803 goto decline_post_inc;
1804 // Avoid weird situations.
1805 if (C->getValue().getMinSignedBits() >= 64 ||
1806 C->getValue().isMinSignedValue())
1807 goto decline_post_inc;
1808 // Without TLI, assume that any stride might be valid, and so any
1809 // use might be shared.
1811 goto decline_post_inc;
1812 // Check for possible scaled-address reuse.
1813 Type *AccessTy = getAccessType(UI->getUser());
1814 TargetLowering::AddrMode AM;
1815 AM.Scale = C->getSExtValue();
1816 if (TLI->isLegalAddressingMode(AM, AccessTy))
1817 goto decline_post_inc;
1818 AM.Scale = -AM.Scale;
1819 if (TLI->isLegalAddressingMode(AM, AccessTy))
1820 goto decline_post_inc;
1824 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1827 // It's possible for the setcc instruction to be anywhere in the loop, and
1828 // possible for it to have multiple users. If it is not immediately before
1829 // the exiting block branch, move it.
1830 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1831 if (Cond->hasOneUse()) {
1832 Cond->moveBefore(TermBr);
1834 // Clone the terminating condition and insert into the loopend.
1835 ICmpInst *OldCond = Cond;
1836 Cond = cast<ICmpInst>(Cond->clone());
1837 Cond->setName(L->getHeader()->getName() + ".termcond");
1838 ExitingBlock->getInstList().insert(TermBr, Cond);
1840 // Clone the IVUse, as the old use still exists!
1841 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1842 TermBr->replaceUsesOfWith(OldCond, Cond);
1846 // If we get to here, we know that we can transform the setcc instruction to
1847 // use the post-incremented version of the IV, allowing us to coalesce the
1848 // live ranges for the IV correctly.
1849 CondUse->transformToPostInc(L);
1852 PostIncs.insert(Cond);
1856 // Determine an insertion point for the loop induction variable increment. It
1857 // must dominate all the post-inc comparisons we just set up, and it must
1858 // dominate the loop latch edge.
1859 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1860 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1861 E = PostIncs.end(); I != E; ++I) {
1863 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1865 if (BB == (*I)->getParent())
1866 IVIncInsertPos = *I;
1867 else if (BB != IVIncInsertPos->getParent())
1868 IVIncInsertPos = BB->getTerminator();
1872 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
1873 /// at the given offset and other details. If so, update the use and
1876 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1877 LSRUse::KindType Kind, Type *AccessTy) {
1878 int64_t NewMinOffset = LU.MinOffset;
1879 int64_t NewMaxOffset = LU.MaxOffset;
1880 Type *NewAccessTy = AccessTy;
1882 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1883 // something conservative, however this can pessimize in the case that one of
1884 // the uses will have all its uses outside the loop, for example.
1885 if (LU.Kind != Kind)
1887 // Conservatively assume HasBaseReg is true for now.
1888 if (NewOffset < LU.MinOffset) {
1889 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1890 Kind, AccessTy, TLI))
1892 NewMinOffset = NewOffset;
1893 } else if (NewOffset > LU.MaxOffset) {
1894 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1895 Kind, AccessTy, TLI))
1897 NewMaxOffset = NewOffset;
1899 // Check for a mismatched access type, and fall back conservatively as needed.
1900 // TODO: Be less conservative when the type is similar and can use the same
1901 // addressing modes.
1902 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1903 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1906 LU.MinOffset = NewMinOffset;
1907 LU.MaxOffset = NewMaxOffset;
1908 LU.AccessTy = NewAccessTy;
1909 if (NewOffset != LU.Offsets.back())
1910 LU.Offsets.push_back(NewOffset);
1914 /// getUse - Return an LSRUse index and an offset value for a fixup which
1915 /// needs the given expression, with the given kind and optional access type.
1916 /// Either reuse an existing use or create a new one, as needed.
1917 std::pair<size_t, int64_t>
1918 LSRInstance::getUse(const SCEV *&Expr,
1919 LSRUse::KindType Kind, Type *AccessTy) {
1920 const SCEV *Copy = Expr;
1921 int64_t Offset = ExtractImmediate(Expr, SE);
1923 // Basic uses can't accept any offset, for example.
1924 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1929 std::pair<UseMapTy::iterator, bool> P =
1930 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1932 // A use already existed with this base.
1933 size_t LUIdx = P.first->second;
1934 LSRUse &LU = Uses[LUIdx];
1935 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1937 return std::make_pair(LUIdx, Offset);
1940 // Create a new use.
1941 size_t LUIdx = Uses.size();
1942 P.first->second = LUIdx;
1943 Uses.push_back(LSRUse(Kind, AccessTy));
1944 LSRUse &LU = Uses[LUIdx];
1946 // We don't need to track redundant offsets, but we don't need to go out
1947 // of our way here to avoid them.
1948 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1949 LU.Offsets.push_back(Offset);
1951 LU.MinOffset = Offset;
1952 LU.MaxOffset = Offset;
1953 return std::make_pair(LUIdx, Offset);
1956 /// DeleteUse - Delete the given use from the Uses list.
1957 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
1958 if (&LU != &Uses.back())
1959 std::swap(LU, Uses.back());
1963 RegUses.SwapAndDropUse(LUIdx, Uses.size());
1966 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1967 /// a formula that has the same registers as the given formula.
1969 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1970 const LSRUse &OrigLU) {
1971 // Search all uses for the formula. This could be more clever.
1972 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1973 LSRUse &LU = Uses[LUIdx];
1974 // Check whether this use is close enough to OrigLU, to see whether it's
1975 // worthwhile looking through its formulae.
1976 // Ignore ICmpZero uses because they may contain formulae generated by
1977 // GenerateICmpZeroScales, in which case adding fixup offsets may
1979 if (&LU != &OrigLU &&
1980 LU.Kind != LSRUse::ICmpZero &&
1981 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1982 LU.WidestFixupType == OrigLU.WidestFixupType &&
1983 LU.HasFormulaWithSameRegs(OrigF)) {
1984 // Scan through this use's formulae.
1985 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1986 E = LU.Formulae.end(); I != E; ++I) {
1987 const Formula &F = *I;
1988 // Check to see if this formula has the same registers and symbols
1990 if (F.BaseRegs == OrigF.BaseRegs &&
1991 F.ScaledReg == OrigF.ScaledReg &&
1992 F.AM.BaseGV == OrigF.AM.BaseGV &&
1993 F.AM.Scale == OrigF.AM.Scale &&
1994 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
1995 if (F.AM.BaseOffs == 0)
1997 // This is the formula where all the registers and symbols matched;
1998 // there aren't going to be any others. Since we declined it, we
1999 // can skip the rest of the formulae and procede to the next LSRUse.
2006 // Nothing looked good.
2010 void LSRInstance::CollectInterestingTypesAndFactors() {
2011 SmallSetVector<const SCEV *, 4> Strides;
2013 // Collect interesting types and strides.
2014 SmallVector<const SCEV *, 4> Worklist;
2015 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2016 const SCEV *Expr = IU.getExpr(*UI);
2018 // Collect interesting types.
2019 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2021 // Add strides for mentioned loops.
2022 Worklist.push_back(Expr);
2024 const SCEV *S = Worklist.pop_back_val();
2025 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2026 Strides.insert(AR->getStepRecurrence(SE));
2027 Worklist.push_back(AR->getStart());
2028 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2029 Worklist.append(Add->op_begin(), Add->op_end());
2031 } while (!Worklist.empty());
2034 // Compute interesting factors from the set of interesting strides.
2035 for (SmallSetVector<const SCEV *, 4>::const_iterator
2036 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2037 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2038 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2039 const SCEV *OldStride = *I;
2040 const SCEV *NewStride = *NewStrideIter;
2042 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2043 SE.getTypeSizeInBits(NewStride->getType())) {
2044 if (SE.getTypeSizeInBits(OldStride->getType()) >
2045 SE.getTypeSizeInBits(NewStride->getType()))
2046 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2048 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2050 if (const SCEVConstant *Factor =
2051 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2053 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2054 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2055 } else if (const SCEVConstant *Factor =
2056 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2059 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2060 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2064 // If all uses use the same type, don't bother looking for truncation-based
2066 if (Types.size() == 1)
2069 DEBUG(print_factors_and_types(dbgs()));
2072 void LSRInstance::CollectFixupsAndInitialFormulae() {
2073 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2075 LSRFixup &LF = getNewFixup();
2076 LF.UserInst = UI->getUser();
2077 LF.OperandValToReplace = UI->getOperandValToReplace();
2078 LF.PostIncLoops = UI->getPostIncLoops();
2080 LSRUse::KindType Kind = LSRUse::Basic;
2082 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2083 Kind = LSRUse::Address;
2084 AccessTy = getAccessType(LF.UserInst);
2087 const SCEV *S = IU.getExpr(*UI);
2089 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2090 // (N - i == 0), and this allows (N - i) to be the expression that we work
2091 // with rather than just N or i, so we can consider the register
2092 // requirements for both N and i at the same time. Limiting this code to
2093 // equality icmps is not a problem because all interesting loops use
2094 // equality icmps, thanks to IndVarSimplify.
2095 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2096 if (CI->isEquality()) {
2097 // Swap the operands if needed to put the OperandValToReplace on the
2098 // left, for consistency.
2099 Value *NV = CI->getOperand(1);
2100 if (NV == LF.OperandValToReplace) {
2101 CI->setOperand(1, CI->getOperand(0));
2102 CI->setOperand(0, NV);
2103 NV = CI->getOperand(1);
2107 // x == y --> x - y == 0
2108 const SCEV *N = SE.getSCEV(NV);
2109 if (SE.isLoopInvariant(N, L)) {
2110 // S is normalized, so normalize N before folding it into S
2111 // to keep the result normalized.
2112 N = TransformForPostIncUse(Normalize, N, CI, 0,
2113 LF.PostIncLoops, SE, DT);
2114 Kind = LSRUse::ICmpZero;
2115 S = SE.getMinusSCEV(N, S);
2118 // -1 and the negations of all interesting strides (except the negation
2119 // of -1) are now also interesting.
2120 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2121 if (Factors[i] != -1)
2122 Factors.insert(-(uint64_t)Factors[i]);
2126 // Set up the initial formula for this use.
2127 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2129 LF.Offset = P.second;
2130 LSRUse &LU = Uses[LF.LUIdx];
2131 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2132 if (!LU.WidestFixupType ||
2133 SE.getTypeSizeInBits(LU.WidestFixupType) <
2134 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2135 LU.WidestFixupType = LF.OperandValToReplace->getType();
2137 // If this is the first use of this LSRUse, give it a formula.
2138 if (LU.Formulae.empty()) {
2139 InsertInitialFormula(S, LU, LF.LUIdx);
2140 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2144 DEBUG(print_fixups(dbgs()));
2147 /// InsertInitialFormula - Insert a formula for the given expression into
2148 /// the given use, separating out loop-variant portions from loop-invariant
2149 /// and loop-computable portions.
2151 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2153 F.InitialMatch(S, L, SE);
2154 bool Inserted = InsertFormula(LU, LUIdx, F);
2155 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2158 /// InsertSupplementalFormula - Insert a simple single-register formula for
2159 /// the given expression into the given use.
2161 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2162 LSRUse &LU, size_t LUIdx) {
2164 F.BaseRegs.push_back(S);
2165 F.AM.HasBaseReg = true;
2166 bool Inserted = InsertFormula(LU, LUIdx, F);
2167 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2170 /// CountRegisters - Note which registers are used by the given formula,
2171 /// updating RegUses.
2172 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2174 RegUses.CountRegister(F.ScaledReg, LUIdx);
2175 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2176 E = F.BaseRegs.end(); I != E; ++I)
2177 RegUses.CountRegister(*I, LUIdx);
2180 /// InsertFormula - If the given formula has not yet been inserted, add it to
2181 /// the list, and return true. Return false otherwise.
2182 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2183 if (!LU.InsertFormula(F))
2186 CountRegisters(F, LUIdx);
2190 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2191 /// loop-invariant values which we're tracking. These other uses will pin these
2192 /// values in registers, making them less profitable for elimination.
2193 /// TODO: This currently misses non-constant addrec step registers.
2194 /// TODO: Should this give more weight to users inside the loop?
2196 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2197 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2198 SmallPtrSet<const SCEV *, 8> Inserted;
2200 while (!Worklist.empty()) {
2201 const SCEV *S = Worklist.pop_back_val();
2203 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2204 Worklist.append(N->op_begin(), N->op_end());
2205 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2206 Worklist.push_back(C->getOperand());
2207 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2208 Worklist.push_back(D->getLHS());
2209 Worklist.push_back(D->getRHS());
2210 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2211 if (!Inserted.insert(U)) continue;
2212 const Value *V = U->getValue();
2213 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2214 // Look for instructions defined outside the loop.
2215 if (L->contains(Inst)) continue;
2216 } else if (isa<UndefValue>(V))
2217 // Undef doesn't have a live range, so it doesn't matter.
2219 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2221 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2222 // Ignore non-instructions.
2225 // Ignore instructions in other functions (as can happen with
2227 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2229 // Ignore instructions not dominated by the loop.
2230 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2231 UserInst->getParent() :
2232 cast<PHINode>(UserInst)->getIncomingBlock(
2233 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2234 if (!DT.dominates(L->getHeader(), UseBB))
2236 // Ignore uses which are part of other SCEV expressions, to avoid
2237 // analyzing them multiple times.
2238 if (SE.isSCEVable(UserInst->getType())) {
2239 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2240 // If the user is a no-op, look through to its uses.
2241 if (!isa<SCEVUnknown>(UserS))
2245 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2249 // Ignore icmp instructions which are already being analyzed.
2250 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2251 unsigned OtherIdx = !UI.getOperandNo();
2252 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2253 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2257 LSRFixup &LF = getNewFixup();
2258 LF.UserInst = const_cast<Instruction *>(UserInst);
2259 LF.OperandValToReplace = UI.getUse();
2260 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2262 LF.Offset = P.second;
2263 LSRUse &LU = Uses[LF.LUIdx];
2264 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2265 if (!LU.WidestFixupType ||
2266 SE.getTypeSizeInBits(LU.WidestFixupType) <
2267 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2268 LU.WidestFixupType = LF.OperandValToReplace->getType();
2269 InsertSupplementalFormula(U, LU, LF.LUIdx);
2270 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2277 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2278 /// separate registers. If C is non-null, multiply each subexpression by C.
2279 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2280 SmallVectorImpl<const SCEV *> &Ops,
2282 ScalarEvolution &SE) {
2283 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2284 // Break out add operands.
2285 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2287 CollectSubexprs(*I, C, Ops, L, SE);
2289 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2290 // Split a non-zero base out of an addrec.
2291 if (!AR->getStart()->isZero()) {
2292 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2293 AR->getStepRecurrence(SE),
2295 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2298 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2301 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2302 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2303 if (Mul->getNumOperands() == 2)
2304 if (const SCEVConstant *Op0 =
2305 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2306 CollectSubexprs(Mul->getOperand(1),
2307 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2313 // Otherwise use the value itself, optionally with a scale applied.
2314 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2317 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2319 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2322 // Arbitrarily cap recursion to protect compile time.
2323 if (Depth >= 3) return;
2325 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2326 const SCEV *BaseReg = Base.BaseRegs[i];
2328 SmallVector<const SCEV *, 8> AddOps;
2329 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2331 if (AddOps.size() == 1) continue;
2333 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2334 JE = AddOps.end(); J != JE; ++J) {
2336 // Loop-variant "unknown" values are uninteresting; we won't be able to
2337 // do anything meaningful with them.
2338 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2341 // Don't pull a constant into a register if the constant could be folded
2342 // into an immediate field.
2343 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2344 Base.getNumRegs() > 1,
2345 LU.Kind, LU.AccessTy, TLI, SE))
2348 // Collect all operands except *J.
2349 SmallVector<const SCEV *, 8> InnerAddOps
2350 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2352 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2354 // Don't leave just a constant behind in a register if the constant could
2355 // be folded into an immediate field.
2356 if (InnerAddOps.size() == 1 &&
2357 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2358 Base.getNumRegs() > 1,
2359 LU.Kind, LU.AccessTy, TLI, SE))
2362 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2363 if (InnerSum->isZero())
2367 // Add the remaining pieces of the add back into the new formula.
2368 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
2369 if (TLI && InnerSumSC &&
2370 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
2371 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2372 InnerSumSC->getValue()->getZExtValue())) {
2373 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2374 InnerSumSC->getValue()->getZExtValue();
2375 F.BaseRegs.erase(F.BaseRegs.begin() + i);
2377 F.BaseRegs[i] = InnerSum;
2379 // Add J as its own register, or an unfolded immediate.
2380 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
2381 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
2382 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2383 SC->getValue()->getZExtValue()))
2384 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2385 SC->getValue()->getZExtValue();
2387 F.BaseRegs.push_back(*J);
2389 if (InsertFormula(LU, LUIdx, F))
2390 // If that formula hadn't been seen before, recurse to find more like
2392 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2397 /// GenerateCombinations - Generate a formula consisting of all of the
2398 /// loop-dominating registers added into a single register.
2399 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2401 // This method is only interesting on a plurality of registers.
2402 if (Base.BaseRegs.size() <= 1) return;
2406 SmallVector<const SCEV *, 4> Ops;
2407 for (SmallVectorImpl<const SCEV *>::const_iterator
2408 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2409 const SCEV *BaseReg = *I;
2410 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2411 !SE.hasComputableLoopEvolution(BaseReg, L))
2412 Ops.push_back(BaseReg);
2414 F.BaseRegs.push_back(BaseReg);
2416 if (Ops.size() > 1) {
2417 const SCEV *Sum = SE.getAddExpr(Ops);
2418 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2419 // opportunity to fold something. For now, just ignore such cases
2420 // rather than proceed with zero in a register.
2421 if (!Sum->isZero()) {
2422 F.BaseRegs.push_back(Sum);
2423 (void)InsertFormula(LU, LUIdx, F);
2428 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2429 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2431 // We can't add a symbolic offset if the address already contains one.
2432 if (Base.AM.BaseGV) return;
2434 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2435 const SCEV *G = Base.BaseRegs[i];
2436 GlobalValue *GV = ExtractSymbol(G, SE);
2437 if (G->isZero() || !GV)
2441 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2442 LU.Kind, LU.AccessTy, TLI))
2445 (void)InsertFormula(LU, LUIdx, F);
2449 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2450 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2452 // TODO: For now, just add the min and max offset, because it usually isn't
2453 // worthwhile looking at everything inbetween.
2454 SmallVector<int64_t, 2> Worklist;
2455 Worklist.push_back(LU.MinOffset);
2456 if (LU.MaxOffset != LU.MinOffset)
2457 Worklist.push_back(LU.MaxOffset);
2459 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2460 const SCEV *G = Base.BaseRegs[i];
2462 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2463 E = Worklist.end(); I != E; ++I) {
2465 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2466 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2467 LU.Kind, LU.AccessTy, TLI)) {
2468 // Add the offset to the base register.
2469 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2470 // If it cancelled out, drop the base register, otherwise update it.
2471 if (NewG->isZero()) {
2472 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2473 F.BaseRegs.pop_back();
2475 F.BaseRegs[i] = NewG;
2477 (void)InsertFormula(LU, LUIdx, F);
2481 int64_t Imm = ExtractImmediate(G, SE);
2482 if (G->isZero() || Imm == 0)
2485 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2486 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2487 LU.Kind, LU.AccessTy, TLI))
2490 (void)InsertFormula(LU, LUIdx, F);
2494 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2495 /// the comparison. For example, x == y -> x*c == y*c.
2496 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2498 if (LU.Kind != LSRUse::ICmpZero) return;
2500 // Determine the integer type for the base formula.
2501 Type *IntTy = Base.getType();
2503 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2505 // Don't do this if there is more than one offset.
2506 if (LU.MinOffset != LU.MaxOffset) return;
2508 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2510 // Check each interesting stride.
2511 for (SmallSetVector<int64_t, 8>::const_iterator
2512 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2513 int64_t Factor = *I;
2515 // Check that the multiplication doesn't overflow.
2516 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2518 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2519 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2522 // Check that multiplying with the use offset doesn't overflow.
2523 int64_t Offset = LU.MinOffset;
2524 if (Offset == INT64_MIN && Factor == -1)
2526 Offset = (uint64_t)Offset * Factor;
2527 if (Offset / Factor != LU.MinOffset)
2531 F.AM.BaseOffs = NewBaseOffs;
2533 // Check that this scale is legal.
2534 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2537 // Compensate for the use having MinOffset built into it.
2538 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2540 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2542 // Check that multiplying with each base register doesn't overflow.
2543 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2544 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2545 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2549 // Check that multiplying with the scaled register doesn't overflow.
2551 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2552 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2556 // Check that multiplying with the unfolded offset doesn't overflow.
2557 if (F.UnfoldedOffset != 0) {
2558 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
2560 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
2561 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
2565 // If we make it here and it's legal, add it.
2566 (void)InsertFormula(LU, LUIdx, F);
2571 /// GenerateScales - Generate stride factor reuse formulae by making use of
2572 /// scaled-offset address modes, for example.
2573 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2574 // Determine the integer type for the base formula.
2575 Type *IntTy = Base.getType();
2578 // If this Formula already has a scaled register, we can't add another one.
2579 if (Base.AM.Scale != 0) return;
2581 // Check each interesting stride.
2582 for (SmallSetVector<int64_t, 8>::const_iterator
2583 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2584 int64_t Factor = *I;
2586 Base.AM.Scale = Factor;
2587 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2588 // Check whether this scale is going to be legal.
2589 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2590 LU.Kind, LU.AccessTy, TLI)) {
2591 // As a special-case, handle special out-of-loop Basic users specially.
2592 // TODO: Reconsider this special case.
2593 if (LU.Kind == LSRUse::Basic &&
2594 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2595 LSRUse::Special, LU.AccessTy, TLI) &&
2596 LU.AllFixupsOutsideLoop)
2597 LU.Kind = LSRUse::Special;
2601 // For an ICmpZero, negating a solitary base register won't lead to
2603 if (LU.Kind == LSRUse::ICmpZero &&
2604 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2606 // For each addrec base reg, apply the scale, if possible.
2607 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2608 if (const SCEVAddRecExpr *AR =
2609 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2610 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2611 if (FactorS->isZero())
2613 // Divide out the factor, ignoring high bits, since we'll be
2614 // scaling the value back up in the end.
2615 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2616 // TODO: This could be optimized to avoid all the copying.
2618 F.ScaledReg = Quotient;
2619 F.DeleteBaseReg(F.BaseRegs[i]);
2620 (void)InsertFormula(LU, LUIdx, F);
2626 /// GenerateTruncates - Generate reuse formulae from different IV types.
2627 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2628 // This requires TargetLowering to tell us which truncates are free.
2631 // Don't bother truncating symbolic values.
2632 if (Base.AM.BaseGV) return;
2634 // Determine the integer type for the base formula.
2635 Type *DstTy = Base.getType();
2637 DstTy = SE.getEffectiveSCEVType(DstTy);
2639 for (SmallSetVector<Type *, 4>::const_iterator
2640 I = Types.begin(), E = Types.end(); I != E; ++I) {
2642 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2645 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2646 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2647 JE = F.BaseRegs.end(); J != JE; ++J)
2648 *J = SE.getAnyExtendExpr(*J, SrcTy);
2650 // TODO: This assumes we've done basic processing on all uses and
2651 // have an idea what the register usage is.
2652 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2655 (void)InsertFormula(LU, LUIdx, F);
2662 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2663 /// defer modifications so that the search phase doesn't have to worry about
2664 /// the data structures moving underneath it.
2668 const SCEV *OrigReg;
2670 WorkItem(size_t LI, int64_t I, const SCEV *R)
2671 : LUIdx(LI), Imm(I), OrigReg(R) {}
2673 void print(raw_ostream &OS) const;
2679 void WorkItem::print(raw_ostream &OS) const {
2680 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2681 << " , add offset " << Imm;
2684 void WorkItem::dump() const {
2685 print(errs()); errs() << '\n';
2688 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2689 /// distance apart and try to form reuse opportunities between them.
2690 void LSRInstance::GenerateCrossUseConstantOffsets() {
2691 // Group the registers by their value without any added constant offset.
2692 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2693 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2695 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2696 SmallVector<const SCEV *, 8> Sequence;
2697 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2699 const SCEV *Reg = *I;
2700 int64_t Imm = ExtractImmediate(Reg, SE);
2701 std::pair<RegMapTy::iterator, bool> Pair =
2702 Map.insert(std::make_pair(Reg, ImmMapTy()));
2704 Sequence.push_back(Reg);
2705 Pair.first->second.insert(std::make_pair(Imm, *I));
2706 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2709 // Now examine each set of registers with the same base value. Build up
2710 // a list of work to do and do the work in a separate step so that we're
2711 // not adding formulae and register counts while we're searching.
2712 SmallVector<WorkItem, 32> WorkItems;
2713 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2714 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2715 E = Sequence.end(); I != E; ++I) {
2716 const SCEV *Reg = *I;
2717 const ImmMapTy &Imms = Map.find(Reg)->second;
2719 // It's not worthwhile looking for reuse if there's only one offset.
2720 if (Imms.size() == 1)
2723 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2724 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2726 dbgs() << ' ' << J->first;
2729 // Examine each offset.
2730 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2732 const SCEV *OrigReg = J->second;
2734 int64_t JImm = J->first;
2735 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2737 if (!isa<SCEVConstant>(OrigReg) &&
2738 UsedByIndicesMap[Reg].count() == 1) {
2739 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2743 // Conservatively examine offsets between this orig reg a few selected
2745 ImmMapTy::const_iterator OtherImms[] = {
2746 Imms.begin(), prior(Imms.end()),
2747 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2749 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2750 ImmMapTy::const_iterator M = OtherImms[i];
2751 if (M == J || M == JE) continue;
2753 // Compute the difference between the two.
2754 int64_t Imm = (uint64_t)JImm - M->first;
2755 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2756 LUIdx = UsedByIndices.find_next(LUIdx))
2757 // Make a memo of this use, offset, and register tuple.
2758 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2759 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2766 UsedByIndicesMap.clear();
2767 UniqueItems.clear();
2769 // Now iterate through the worklist and add new formulae.
2770 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2771 E = WorkItems.end(); I != E; ++I) {
2772 const WorkItem &WI = *I;
2773 size_t LUIdx = WI.LUIdx;
2774 LSRUse &LU = Uses[LUIdx];
2775 int64_t Imm = WI.Imm;
2776 const SCEV *OrigReg = WI.OrigReg;
2778 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2779 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2780 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2782 // TODO: Use a more targeted data structure.
2783 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2784 const Formula &F = LU.Formulae[L];
2785 // Use the immediate in the scaled register.
2786 if (F.ScaledReg == OrigReg) {
2787 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2788 Imm * (uint64_t)F.AM.Scale;
2789 // Don't create 50 + reg(-50).
2790 if (F.referencesReg(SE.getSCEV(
2791 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2794 NewF.AM.BaseOffs = Offs;
2795 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2796 LU.Kind, LU.AccessTy, TLI))
2798 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2800 // If the new scale is a constant in a register, and adding the constant
2801 // value to the immediate would produce a value closer to zero than the
2802 // immediate itself, then the formula isn't worthwhile.
2803 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2804 if (C->getValue()->isNegative() !=
2805 (NewF.AM.BaseOffs < 0) &&
2806 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2807 .ule(abs64(NewF.AM.BaseOffs)))
2811 (void)InsertFormula(LU, LUIdx, NewF);
2813 // Use the immediate in a base register.
2814 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2815 const SCEV *BaseReg = F.BaseRegs[N];
2816 if (BaseReg != OrigReg)
2819 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2820 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2821 LU.Kind, LU.AccessTy, TLI)) {
2823 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
2826 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
2828 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2830 // If the new formula has a constant in a register, and adding the
2831 // constant value to the immediate would produce a value closer to
2832 // zero than the immediate itself, then the formula isn't worthwhile.
2833 for (SmallVectorImpl<const SCEV *>::const_iterator
2834 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2836 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2837 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2838 abs64(NewF.AM.BaseOffs)) &&
2839 (C->getValue()->getValue() +
2840 NewF.AM.BaseOffs).countTrailingZeros() >=
2841 CountTrailingZeros_64(NewF.AM.BaseOffs))
2845 (void)InsertFormula(LU, LUIdx, NewF);
2854 /// GenerateAllReuseFormulae - Generate formulae for each use.
2856 LSRInstance::GenerateAllReuseFormulae() {
2857 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2858 // queries are more precise.
2859 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2860 LSRUse &LU = Uses[LUIdx];
2861 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2862 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2863 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2864 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2866 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2867 LSRUse &LU = Uses[LUIdx];
2868 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2869 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2870 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2871 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2872 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2873 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2874 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2875 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2877 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2878 LSRUse &LU = Uses[LUIdx];
2879 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2880 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2883 GenerateCrossUseConstantOffsets();
2885 DEBUG(dbgs() << "\n"
2886 "After generating reuse formulae:\n";
2887 print_uses(dbgs()));
2890 /// If there are multiple formulae with the same set of registers used
2891 /// by other uses, pick the best one and delete the others.
2892 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2893 DenseSet<const SCEV *> VisitedRegs;
2894 SmallPtrSet<const SCEV *, 16> Regs;
2896 bool ChangedFormulae = false;
2899 // Collect the best formula for each unique set of shared registers. This
2900 // is reset for each use.
2901 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2903 BestFormulaeTy BestFormulae;
2905 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2906 LSRUse &LU = Uses[LUIdx];
2907 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2910 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2911 FIdx != NumForms; ++FIdx) {
2912 Formula &F = LU.Formulae[FIdx];
2914 SmallVector<const SCEV *, 2> Key;
2915 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2916 JE = F.BaseRegs.end(); J != JE; ++J) {
2917 const SCEV *Reg = *J;
2918 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2922 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2923 Key.push_back(F.ScaledReg);
2924 // Unstable sort by host order ok, because this is only used for
2926 std::sort(Key.begin(), Key.end());
2928 std::pair<BestFormulaeTy::const_iterator, bool> P =
2929 BestFormulae.insert(std::make_pair(Key, FIdx));
2931 Formula &Best = LU.Formulae[P.first->second];
2934 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2937 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2939 if (CostF < CostBest)
2941 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2943 " in favor of formula "; Best.print(dbgs());
2946 ChangedFormulae = true;
2948 LU.DeleteFormula(F);
2956 // Now that we've filtered out some formulae, recompute the Regs set.
2958 LU.RecomputeRegs(LUIdx, RegUses);
2960 // Reset this to prepare for the next use.
2961 BestFormulae.clear();
2964 DEBUG(if (ChangedFormulae) {
2966 "After filtering out undesirable candidates:\n";
2971 // This is a rough guess that seems to work fairly well.
2972 static const size_t ComplexityLimit = UINT16_MAX;
2974 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2975 /// solutions the solver might have to consider. It almost never considers
2976 /// this many solutions because it prune the search space, but the pruning
2977 /// isn't always sufficient.
2978 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2980 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2981 E = Uses.end(); I != E; ++I) {
2982 size_t FSize = I->Formulae.size();
2983 if (FSize >= ComplexityLimit) {
2984 Power = ComplexityLimit;
2988 if (Power >= ComplexityLimit)
2994 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
2995 /// of the registers of another formula, it won't help reduce register
2996 /// pressure (though it may not necessarily hurt register pressure); remove
2997 /// it to simplify the system.
2998 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
2999 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3000 DEBUG(dbgs() << "The search space is too complex.\n");
3002 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3003 "which use a superset of registers used by other "
3006 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3007 LSRUse &LU = Uses[LUIdx];
3009 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3010 Formula &F = LU.Formulae[i];
3011 // Look for a formula with a constant or GV in a register. If the use
3012 // also has a formula with that same value in an immediate field,
3013 // delete the one that uses a register.
3014 for (SmallVectorImpl<const SCEV *>::const_iterator
3015 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3016 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3018 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3019 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3020 (I - F.BaseRegs.begin()));
3021 if (LU.HasFormulaWithSameRegs(NewF)) {
3022 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3023 LU.DeleteFormula(F);
3029 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3030 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3033 NewF.AM.BaseGV = GV;
3034 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3035 (I - F.BaseRegs.begin()));
3036 if (LU.HasFormulaWithSameRegs(NewF)) {
3037 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3039 LU.DeleteFormula(F);
3050 LU.RecomputeRegs(LUIdx, RegUses);
3053 DEBUG(dbgs() << "After pre-selection:\n";
3054 print_uses(dbgs()));
3058 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3059 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3061 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3062 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3063 DEBUG(dbgs() << "The search space is too complex.\n");
3065 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3066 "separated by a constant offset will use the same "
3069 // This is especially useful for unrolled loops.
3071 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3072 LSRUse &LU = Uses[LUIdx];
3073 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3074 E = LU.Formulae.end(); I != E; ++I) {
3075 const Formula &F = *I;
3076 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3077 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3078 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3079 /*HasBaseReg=*/false,
3080 LU.Kind, LU.AccessTy)) {
3081 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3084 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3086 // Update the relocs to reference the new use.
3087 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3088 E = Fixups.end(); I != E; ++I) {
3089 LSRFixup &Fixup = *I;
3090 if (Fixup.LUIdx == LUIdx) {
3091 Fixup.LUIdx = LUThatHas - &Uses.front();
3092 Fixup.Offset += F.AM.BaseOffs;
3093 // Add the new offset to LUThatHas' offset list.
3094 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3095 LUThatHas->Offsets.push_back(Fixup.Offset);
3096 if (Fixup.Offset > LUThatHas->MaxOffset)
3097 LUThatHas->MaxOffset = Fixup.Offset;
3098 if (Fixup.Offset < LUThatHas->MinOffset)
3099 LUThatHas->MinOffset = Fixup.Offset;
3101 DEBUG(dbgs() << "New fixup has offset "
3102 << Fixup.Offset << '\n');
3104 if (Fixup.LUIdx == NumUses-1)
3105 Fixup.LUIdx = LUIdx;
3108 // Delete formulae from the new use which are no longer legal.
3110 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3111 Formula &F = LUThatHas->Formulae[i];
3112 if (!isLegalUse(F.AM,
3113 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3114 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3115 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3117 LUThatHas->DeleteFormula(F);
3124 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3126 // Delete the old use.
3127 DeleteUse(LU, LUIdx);
3137 DEBUG(dbgs() << "After pre-selection:\n";
3138 print_uses(dbgs()));
3142 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3143 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3144 /// we've done more filtering, as it may be able to find more formulae to
3146 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3147 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3148 DEBUG(dbgs() << "The search space is too complex.\n");
3150 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3151 "undesirable dedicated registers.\n");
3153 FilterOutUndesirableDedicatedRegisters();
3155 DEBUG(dbgs() << "After pre-selection:\n";
3156 print_uses(dbgs()));
3160 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3161 /// to be profitable, and then in any use which has any reference to that
3162 /// register, delete all formulae which do not reference that register.
3163 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3164 // With all other options exhausted, loop until the system is simple
3165 // enough to handle.
3166 SmallPtrSet<const SCEV *, 4> Taken;
3167 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3168 // Ok, we have too many of formulae on our hands to conveniently handle.
3169 // Use a rough heuristic to thin out the list.
3170 DEBUG(dbgs() << "The search space is too complex.\n");
3172 // Pick the register which is used by the most LSRUses, which is likely
3173 // to be a good reuse register candidate.
3174 const SCEV *Best = 0;
3175 unsigned BestNum = 0;
3176 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3178 const SCEV *Reg = *I;
3179 if (Taken.count(Reg))
3184 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3185 if (Count > BestNum) {
3192 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3193 << " will yield profitable reuse.\n");
3196 // In any use with formulae which references this register, delete formulae
3197 // which don't reference it.
3198 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3199 LSRUse &LU = Uses[LUIdx];
3200 if (!LU.Regs.count(Best)) continue;
3203 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3204 Formula &F = LU.Formulae[i];
3205 if (!F.referencesReg(Best)) {
3206 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3207 LU.DeleteFormula(F);
3211 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3217 LU.RecomputeRegs(LUIdx, RegUses);
3220 DEBUG(dbgs() << "After pre-selection:\n";
3221 print_uses(dbgs()));
3225 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3226 /// formulae to choose from, use some rough heuristics to prune down the number
3227 /// of formulae. This keeps the main solver from taking an extraordinary amount
3228 /// of time in some worst-case scenarios.
3229 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3230 NarrowSearchSpaceByDetectingSupersets();
3231 NarrowSearchSpaceByCollapsingUnrolledCode();
3232 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3233 NarrowSearchSpaceByPickingWinnerRegs();
3236 /// SolveRecurse - This is the recursive solver.
3237 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3239 SmallVectorImpl<const Formula *> &Workspace,
3240 const Cost &CurCost,
3241 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3242 DenseSet<const SCEV *> &VisitedRegs) const {
3245 // - use more aggressive filtering
3246 // - sort the formula so that the most profitable solutions are found first
3247 // - sort the uses too
3249 // - don't compute a cost, and then compare. compare while computing a cost
3251 // - track register sets with SmallBitVector
3253 const LSRUse &LU = Uses[Workspace.size()];
3255 // If this use references any register that's already a part of the
3256 // in-progress solution, consider it a requirement that a formula must
3257 // reference that register in order to be considered. This prunes out
3258 // unprofitable searching.
3259 SmallSetVector<const SCEV *, 4> ReqRegs;
3260 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3261 E = CurRegs.end(); I != E; ++I)
3262 if (LU.Regs.count(*I))
3265 bool AnySatisfiedReqRegs = false;
3266 SmallPtrSet<const SCEV *, 16> NewRegs;
3269 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3270 E = LU.Formulae.end(); I != E; ++I) {
3271 const Formula &F = *I;
3273 // Ignore formulae which do not use any of the required registers.
3274 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3275 JE = ReqRegs.end(); J != JE; ++J) {
3276 const SCEV *Reg = *J;
3277 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3278 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3282 AnySatisfiedReqRegs = true;
3284 // Evaluate the cost of the current formula. If it's already worse than
3285 // the current best, prune the search at that point.
3288 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3289 if (NewCost < SolutionCost) {
3290 Workspace.push_back(&F);
3291 if (Workspace.size() != Uses.size()) {
3292 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3293 NewRegs, VisitedRegs);
3294 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3295 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3297 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3298 dbgs() << ". Regs:";
3299 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3300 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3301 dbgs() << ' ' << **I;
3304 SolutionCost = NewCost;
3305 Solution = Workspace;
3307 Workspace.pop_back();
3312 // If none of the formulae had all of the required registers, relax the
3313 // constraint so that we don't exclude all formulae.
3314 if (!AnySatisfiedReqRegs) {
3315 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3321 /// Solve - Choose one formula from each use. Return the results in the given
3322 /// Solution vector.
3323 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3324 SmallVector<const Formula *, 8> Workspace;
3326 SolutionCost.Loose();
3328 SmallPtrSet<const SCEV *, 16> CurRegs;
3329 DenseSet<const SCEV *> VisitedRegs;
3330 Workspace.reserve(Uses.size());
3332 // SolveRecurse does all the work.
3333 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3334 CurRegs, VisitedRegs);
3336 // Ok, we've now made all our decisions.
3337 DEBUG(dbgs() << "\n"
3338 "The chosen solution requires "; SolutionCost.print(dbgs());
3340 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3342 Uses[i].print(dbgs());
3345 Solution[i]->print(dbgs());
3349 assert(Solution.size() == Uses.size() && "Malformed solution!");
3352 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3353 /// the dominator tree far as we can go while still being dominated by the
3354 /// input positions. This helps canonicalize the insert position, which
3355 /// encourages sharing.
3356 BasicBlock::iterator
3357 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3358 const SmallVectorImpl<Instruction *> &Inputs)
3361 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3362 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3365 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3366 if (!Rung) return IP;
3367 Rung = Rung->getIDom();
3368 if (!Rung) return IP;
3369 IDom = Rung->getBlock();
3371 // Don't climb into a loop though.
3372 const Loop *IDomLoop = LI.getLoopFor(IDom);
3373 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3374 if (IDomDepth <= IPLoopDepth &&
3375 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3379 bool AllDominate = true;
3380 Instruction *BetterPos = 0;
3381 Instruction *Tentative = IDom->getTerminator();
3382 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3383 E = Inputs.end(); I != E; ++I) {
3384 Instruction *Inst = *I;
3385 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3386 AllDominate = false;
3389 // Attempt to find an insert position in the middle of the block,
3390 // instead of at the end, so that it can be used for other expansions.
3391 if (IDom == Inst->getParent() &&
3392 (!BetterPos || DT.dominates(BetterPos, Inst)))
3393 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3406 /// AdjustInsertPositionForExpand - Determine an input position which will be
3407 /// dominated by the operands and which will dominate the result.
3408 BasicBlock::iterator
3409 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3411 const LSRUse &LU) const {
3412 // Collect some instructions which must be dominated by the
3413 // expanding replacement. These must be dominated by any operands that
3414 // will be required in the expansion.
3415 SmallVector<Instruction *, 4> Inputs;
3416 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3417 Inputs.push_back(I);
3418 if (LU.Kind == LSRUse::ICmpZero)
3419 if (Instruction *I =
3420 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3421 Inputs.push_back(I);
3422 if (LF.PostIncLoops.count(L)) {
3423 if (LF.isUseFullyOutsideLoop(L))
3424 Inputs.push_back(L->getLoopLatch()->getTerminator());
3426 Inputs.push_back(IVIncInsertPos);
3428 // The expansion must also be dominated by the increment positions of any
3429 // loops it for which it is using post-inc mode.
3430 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3431 E = LF.PostIncLoops.end(); I != E; ++I) {
3432 const Loop *PIL = *I;
3433 if (PIL == L) continue;
3435 // Be dominated by the loop exit.
3436 SmallVector<BasicBlock *, 4> ExitingBlocks;
3437 PIL->getExitingBlocks(ExitingBlocks);
3438 if (!ExitingBlocks.empty()) {
3439 BasicBlock *BB = ExitingBlocks[0];
3440 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3441 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3442 Inputs.push_back(BB->getTerminator());
3446 // Then, climb up the immediate dominator tree as far as we can go while
3447 // still being dominated by the input positions.
3448 IP = HoistInsertPosition(IP, Inputs);
3450 // Don't insert instructions before PHI nodes.
3451 while (isa<PHINode>(IP)) ++IP;
3453 // Ignore landingpad instructions.
3454 while (isa<LandingPadInst>(IP)) ++IP;
3456 // Ignore debug intrinsics.
3457 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3462 /// Expand - Emit instructions for the leading candidate expression for this
3463 /// LSRUse (this is called "expanding").
3464 Value *LSRInstance::Expand(const LSRFixup &LF,
3466 BasicBlock::iterator IP,
3467 SCEVExpander &Rewriter,
3468 SmallVectorImpl<WeakVH> &DeadInsts) const {
3469 const LSRUse &LU = Uses[LF.LUIdx];
3471 // Determine an input position which will be dominated by the operands and
3472 // which will dominate the result.
3473 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3475 // Inform the Rewriter if we have a post-increment use, so that it can
3476 // perform an advantageous expansion.
3477 Rewriter.setPostInc(LF.PostIncLoops);
3479 // This is the type that the user actually needs.
3480 Type *OpTy = LF.OperandValToReplace->getType();
3481 // This will be the type that we'll initially expand to.
3482 Type *Ty = F.getType();
3484 // No type known; just expand directly to the ultimate type.
3486 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3487 // Expand directly to the ultimate type if it's the right size.
3489 // This is the type to do integer arithmetic in.
3490 Type *IntTy = SE.getEffectiveSCEVType(Ty);
3492 // Build up a list of operands to add together to form the full base.
3493 SmallVector<const SCEV *, 8> Ops;
3495 // Expand the BaseRegs portion.
3496 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3497 E = F.BaseRegs.end(); I != E; ++I) {
3498 const SCEV *Reg = *I;
3499 assert(!Reg->isZero() && "Zero allocated in a base register!");
3501 // If we're expanding for a post-inc user, make the post-inc adjustment.
3502 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3503 Reg = TransformForPostIncUse(Denormalize, Reg,
3504 LF.UserInst, LF.OperandValToReplace,
3507 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3510 // 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 ScaledReg portion.
3518 Value *ICmpScaledV = 0;
3519 if (F.AM.Scale != 0) {
3520 const SCEV *ScaledS = F.ScaledReg;
3522 // If we're expanding for a post-inc user, make the post-inc adjustment.
3523 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3524 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3525 LF.UserInst, LF.OperandValToReplace,
3528 if (LU.Kind == LSRUse::ICmpZero) {
3529 // An interesting way of "folding" with an icmp is to use a negated
3530 // scale, which we'll implement by inserting it into the other operand
3532 assert(F.AM.Scale == -1 &&
3533 "The only scale supported by ICmpZero uses is -1!");
3534 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3536 // Otherwise just expand the scaled register and an explicit scale,
3537 // which is expected to be matched as part of the address.
3538 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3539 ScaledS = SE.getMulExpr(ScaledS,
3540 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3541 Ops.push_back(ScaledS);
3543 // Flush the operand list to suppress SCEVExpander hoisting.
3544 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3546 Ops.push_back(SE.getUnknown(FullV));
3550 // Expand the GV portion.
3552 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3554 // Flush the operand list to suppress SCEVExpander hoisting.
3555 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3557 Ops.push_back(SE.getUnknown(FullV));
3560 // Expand the immediate portion.
3561 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3563 if (LU.Kind == LSRUse::ICmpZero) {
3564 // The other interesting way of "folding" with an ICmpZero is to use a
3565 // negated immediate.
3567 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3569 Ops.push_back(SE.getUnknown(ICmpScaledV));
3570 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3573 // Just add the immediate values. These again are expected to be matched
3574 // as part of the address.
3575 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3579 // Expand the unfolded offset portion.
3580 int64_t UnfoldedOffset = F.UnfoldedOffset;
3581 if (UnfoldedOffset != 0) {
3582 // Just add the immediate values.
3583 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
3587 // Emit instructions summing all the operands.
3588 const SCEV *FullS = Ops.empty() ?
3589 SE.getConstant(IntTy, 0) :
3591 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3593 // We're done expanding now, so reset the rewriter.
3594 Rewriter.clearPostInc();
3596 // An ICmpZero Formula represents an ICmp which we're handling as a
3597 // comparison against zero. Now that we've expanded an expression for that
3598 // form, update the ICmp's other operand.
3599 if (LU.Kind == LSRUse::ICmpZero) {
3600 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3601 DeadInsts.push_back(CI->getOperand(1));
3602 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3603 "a scale at the same time!");
3604 if (F.AM.Scale == -1) {
3605 if (ICmpScaledV->getType() != OpTy) {
3607 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3609 ICmpScaledV, OpTy, "tmp", CI);
3612 CI->setOperand(1, ICmpScaledV);
3614 assert(F.AM.Scale == 0 &&
3615 "ICmp does not support folding a global value and "
3616 "a scale at the same time!");
3617 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3619 if (C->getType() != OpTy)
3620 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3624 CI->setOperand(1, C);
3631 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3632 /// of their operands effectively happens in their predecessor blocks, so the
3633 /// expression may need to be expanded in multiple places.
3634 void LSRInstance::RewriteForPHI(PHINode *PN,
3637 SCEVExpander &Rewriter,
3638 SmallVectorImpl<WeakVH> &DeadInsts,
3640 DenseMap<BasicBlock *, Value *> Inserted;
3641 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3642 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3643 BasicBlock *BB = PN->getIncomingBlock(i);
3645 // If this is a critical edge, split the edge so that we do not insert
3646 // the code on all predecessor/successor paths. We do this unless this
3647 // is the canonical backedge for this loop, which complicates post-inc
3649 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3650 !isa<IndirectBrInst>(BB->getTerminator())) {
3651 BasicBlock *Parent = PN->getParent();
3652 Loop *PNLoop = LI.getLoopFor(Parent);
3653 if (!PNLoop || Parent != PNLoop->getHeader()) {
3654 // Split the critical edge.
3655 BasicBlock *NewBB = 0;
3656 if (!Parent->isLandingPad()) {
3657 NewBB = SplitCriticalEdge(BB, Parent, P);
3659 SmallVector<BasicBlock*, 2> NewBBs;
3660 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
3664 // If PN is outside of the loop and BB is in the loop, we want to
3665 // move the block to be immediately before the PHI block, not
3666 // immediately after BB.
3667 if (L->contains(BB) && !L->contains(PN))
3668 NewBB->moveBefore(PN->getParent());
3670 // Splitting the edge can reduce the number of PHI entries we have.
3671 e = PN->getNumIncomingValues();
3673 i = PN->getBasicBlockIndex(BB);
3677 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3678 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3680 PN->setIncomingValue(i, Pair.first->second);
3682 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3684 // If this is reuse-by-noop-cast, insert the noop cast.
3685 Type *OpTy = LF.OperandValToReplace->getType();
3686 if (FullV->getType() != OpTy)
3688 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3690 FullV, LF.OperandValToReplace->getType(),
3691 "tmp", BB->getTerminator());
3693 PN->setIncomingValue(i, FullV);
3694 Pair.first->second = FullV;
3699 /// Rewrite - Emit instructions for the leading candidate expression for this
3700 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3701 /// the newly expanded value.
3702 void LSRInstance::Rewrite(const LSRFixup &LF,
3704 SCEVExpander &Rewriter,
3705 SmallVectorImpl<WeakVH> &DeadInsts,
3707 // First, find an insertion point that dominates UserInst. For PHI nodes,
3708 // find the nearest block which dominates all the relevant uses.
3709 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3710 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3712 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3714 // If this is reuse-by-noop-cast, insert the noop cast.
3715 Type *OpTy = LF.OperandValToReplace->getType();
3716 if (FullV->getType() != OpTy) {
3718 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3719 FullV, OpTy, "tmp", LF.UserInst);
3723 // Update the user. ICmpZero is handled specially here (for now) because
3724 // Expand may have updated one of the operands of the icmp already, and
3725 // its new value may happen to be equal to LF.OperandValToReplace, in
3726 // which case doing replaceUsesOfWith leads to replacing both operands
3727 // with the same value. TODO: Reorganize this.
3728 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3729 LF.UserInst->setOperand(0, FullV);
3731 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3734 DeadInsts.push_back(LF.OperandValToReplace);
3737 /// ImplementSolution - Rewrite all the fixup locations with new values,
3738 /// following the chosen solution.
3740 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3742 // Keep track of instructions we may have made dead, so that
3743 // we can remove them after we are done working.
3744 SmallVector<WeakVH, 16> DeadInsts;
3746 SCEVExpander Rewriter(SE, "lsr");
3747 Rewriter.disableCanonicalMode();
3748 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3750 // Expand the new value definitions and update the users.
3751 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3752 E = Fixups.end(); I != E; ++I) {
3753 const LSRFixup &Fixup = *I;
3755 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3760 // Clean up after ourselves. This must be done before deleting any
3764 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3767 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3768 : IU(P->getAnalysis<IVUsers>()),
3769 SE(P->getAnalysis<ScalarEvolution>()),
3770 DT(P->getAnalysis<DominatorTree>()),
3771 LI(P->getAnalysis<LoopInfo>()),
3772 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3774 // If LoopSimplify form is not available, stay out of trouble.
3775 if (!L->isLoopSimplifyForm()) return;
3777 // If there's no interesting work to be done, bail early.
3778 if (IU.empty()) return;
3780 DEBUG(dbgs() << "\nLSR on loop ";
3781 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3784 // First, perform some low-level loop optimizations.
3786 OptimizeLoopTermCond();
3788 // If loop preparation eliminates all interesting IV users, bail.
3789 if (IU.empty()) return;
3791 // Start collecting data and preparing for the solver.
3792 CollectInterestingTypesAndFactors();
3793 CollectFixupsAndInitialFormulae();
3794 CollectLoopInvariantFixupsAndFormulae();
3796 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3797 print_uses(dbgs()));
3799 // Now use the reuse data to generate a bunch of interesting ways
3800 // to formulate the values needed for the uses.
3801 GenerateAllReuseFormulae();
3803 FilterOutUndesirableDedicatedRegisters();
3804 NarrowSearchSpaceUsingHeuristics();
3806 SmallVector<const Formula *, 8> Solution;
3809 // Release memory that is no longer needed.
3815 // Formulae should be legal.
3816 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3817 E = Uses.end(); I != E; ++I) {
3818 const LSRUse &LU = *I;
3819 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3820 JE = LU.Formulae.end(); J != JE; ++J)
3821 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3822 LU.Kind, LU.AccessTy, TLI) &&
3823 "Illegal formula generated!");
3827 // Now that we've decided what we want, make it so.
3828 ImplementSolution(Solution, P);
3831 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3832 if (Factors.empty() && Types.empty()) return;
3834 OS << "LSR has identified the following interesting factors and types: ";
3837 for (SmallSetVector<int64_t, 8>::const_iterator
3838 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3839 if (!First) OS << ", ";
3844 for (SmallSetVector<Type *, 4>::const_iterator
3845 I = Types.begin(), E = Types.end(); I != E; ++I) {
3846 if (!First) OS << ", ";
3848 OS << '(' << **I << ')';
3853 void LSRInstance::print_fixups(raw_ostream &OS) const {
3854 OS << "LSR is examining the following fixup sites:\n";
3855 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3856 E = Fixups.end(); I != E; ++I) {
3863 void LSRInstance::print_uses(raw_ostream &OS) const {
3864 OS << "LSR is examining the following uses:\n";
3865 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3866 E = Uses.end(); I != E; ++I) {
3867 const LSRUse &LU = *I;
3871 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3872 JE = LU.Formulae.end(); J != JE; ++J) {
3880 void LSRInstance::print(raw_ostream &OS) const {
3881 print_factors_and_types(OS);
3886 void LSRInstance::dump() const {
3887 print(errs()); errs() << '\n';
3892 class LoopStrengthReduce : public LoopPass {
3893 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3894 /// transformation profitability.
3895 const TargetLowering *const TLI;
3898 static char ID; // Pass ID, replacement for typeid
3899 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3902 bool runOnLoop(Loop *L, LPPassManager &LPM);
3903 void getAnalysisUsage(AnalysisUsage &AU) const;
3908 char LoopStrengthReduce::ID = 0;
3909 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
3910 "Loop Strength Reduction", false, false)
3911 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3912 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3913 INITIALIZE_PASS_DEPENDENCY(IVUsers)
3914 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
3915 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3916 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
3917 "Loop Strength Reduction", false, false)
3920 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3921 return new LoopStrengthReduce(TLI);
3924 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3925 : LoopPass(ID), TLI(tli) {
3926 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
3929 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3930 // We split critical edges, so we change the CFG. However, we do update
3931 // many analyses if they are around.
3932 AU.addPreservedID(LoopSimplifyID);
3934 AU.addRequired<LoopInfo>();
3935 AU.addPreserved<LoopInfo>();
3936 AU.addRequiredID(LoopSimplifyID);
3937 AU.addRequired<DominatorTree>();
3938 AU.addPreserved<DominatorTree>();
3939 AU.addRequired<ScalarEvolution>();
3940 AU.addPreserved<ScalarEvolution>();
3941 // Requiring LoopSimplify a second time here prevents IVUsers from running
3942 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
3943 AU.addRequiredID(LoopSimplifyID);
3944 AU.addRequired<IVUsers>();
3945 AU.addPreserved<IVUsers>();
3948 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3949 bool Changed = false;
3951 // Run the main LSR transformation.
3952 Changed |= LSRInstance(TLI, L, this).getChanged();
3954 // At this point, it is worth checking to see if any recurrence PHIs are also
3955 // dead, so that we can remove them as well.
3956 Changed |= DeleteDeadPHIs(L->getHeader());