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/CommandLine.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Target/TargetLowering.h"
81 cl::opt<bool> EnableNested(
82 "enable-lsr-nested", cl::Hidden, cl::desc("Enable LSR on nested loops"));
84 cl::opt<bool> EnableRetry(
85 "enable-lsr-retry", cl::Hidden, cl::desc("Enable LSR retry"));
90 /// RegSortData - This class holds data which is used to order reuse candidates.
93 /// UsedByIndices - This represents the set of LSRUse indices which reference
94 /// a particular register.
95 SmallBitVector UsedByIndices;
99 void print(raw_ostream &OS) const;
105 void RegSortData::print(raw_ostream &OS) const {
106 OS << "[NumUses=" << UsedByIndices.count() << ']';
109 void RegSortData::dump() const {
110 print(errs()); errs() << '\n';
115 /// RegUseTracker - Map register candidates to information about how they are
117 class RegUseTracker {
118 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
120 RegUsesTy RegUsesMap;
121 SmallVector<const SCEV *, 16> RegSequence;
124 void CountRegister(const SCEV *Reg, size_t LUIdx);
125 void DropRegister(const SCEV *Reg, size_t LUIdx);
126 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
128 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
130 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
134 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
135 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
136 iterator begin() { return RegSequence.begin(); }
137 iterator end() { return RegSequence.end(); }
138 const_iterator begin() const { return RegSequence.begin(); }
139 const_iterator end() const { return RegSequence.end(); }
145 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
146 std::pair<RegUsesTy::iterator, bool> Pair =
147 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
148 RegSortData &RSD = Pair.first->second;
150 RegSequence.push_back(Reg);
151 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
152 RSD.UsedByIndices.set(LUIdx);
156 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
157 RegUsesTy::iterator It = RegUsesMap.find(Reg);
158 assert(It != RegUsesMap.end());
159 RegSortData &RSD = It->second;
160 assert(RSD.UsedByIndices.size() > LUIdx);
161 RSD.UsedByIndices.reset(LUIdx);
165 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
166 assert(LUIdx <= LastLUIdx);
168 // Update RegUses. The data structure is not optimized for this purpose;
169 // we must iterate through it and update each of the bit vectors.
170 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
172 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
173 if (LUIdx < UsedByIndices.size())
174 UsedByIndices[LUIdx] =
175 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
176 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
181 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
182 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
183 if (I == RegUsesMap.end())
185 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
186 int i = UsedByIndices.find_first();
187 if (i == -1) return false;
188 if ((size_t)i != LUIdx) return true;
189 return UsedByIndices.find_next(i) != -1;
192 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
193 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
194 assert(I != RegUsesMap.end() && "Unknown register!");
195 return I->second.UsedByIndices;
198 void RegUseTracker::clear() {
205 /// Formula - This class holds information that describes a formula for
206 /// computing satisfying a use. It may include broken-out immediates and scaled
209 /// AM - This is used to represent complex addressing, as well as other kinds
210 /// of interesting uses.
211 TargetLowering::AddrMode AM;
213 /// BaseRegs - The list of "base" registers for this use. When this is
214 /// non-empty, AM.HasBaseReg should be set to true.
215 SmallVector<const SCEV *, 2> BaseRegs;
217 /// ScaledReg - The 'scaled' register for this use. This should be non-null
218 /// when AM.Scale is not zero.
219 const SCEV *ScaledReg;
221 /// UnfoldedOffset - An additional constant offset which added near the
222 /// use. This requires a temporary register, but the offset itself can
223 /// live in an add immediate field rather than a register.
224 int64_t UnfoldedOffset;
226 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
228 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
230 unsigned getNumRegs() const;
231 Type *getType() const;
233 void DeleteBaseReg(const SCEV *&S);
235 bool referencesReg(const SCEV *S) const;
236 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
237 const RegUseTracker &RegUses) const;
239 void print(raw_ostream &OS) const;
245 /// DoInitialMatch - Recursion helper for InitialMatch.
246 static void DoInitialMatch(const SCEV *S, Loop *L,
247 SmallVectorImpl<const SCEV *> &Good,
248 SmallVectorImpl<const SCEV *> &Bad,
249 ScalarEvolution &SE) {
250 // Collect expressions which properly dominate the loop header.
251 if (SE.properlyDominates(S, L->getHeader())) {
256 // Look at add operands.
257 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
258 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
260 DoInitialMatch(*I, L, Good, Bad, SE);
264 // Look at addrec operands.
265 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
266 if (!AR->getStart()->isZero()) {
267 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
268 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
269 AR->getStepRecurrence(SE),
270 // FIXME: AR->getNoWrapFlags()
271 AR->getLoop(), SCEV::FlagAnyWrap),
276 // Handle a multiplication by -1 (negation) if it didn't fold.
277 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
278 if (Mul->getOperand(0)->isAllOnesValue()) {
279 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
280 const SCEV *NewMul = SE.getMulExpr(Ops);
282 SmallVector<const SCEV *, 4> MyGood;
283 SmallVector<const SCEV *, 4> MyBad;
284 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
285 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
286 SE.getEffectiveSCEVType(NewMul->getType())));
287 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
288 E = MyGood.end(); I != E; ++I)
289 Good.push_back(SE.getMulExpr(NegOne, *I));
290 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
291 E = MyBad.end(); I != E; ++I)
292 Bad.push_back(SE.getMulExpr(NegOne, *I));
296 // Ok, we can't do anything interesting. Just stuff the whole thing into a
297 // register and hope for the best.
301 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
302 /// attempting to keep all loop-invariant and loop-computable values in a
303 /// single base register.
304 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
305 SmallVector<const SCEV *, 4> Good;
306 SmallVector<const SCEV *, 4> Bad;
307 DoInitialMatch(S, L, Good, Bad, SE);
309 const SCEV *Sum = SE.getAddExpr(Good);
311 BaseRegs.push_back(Sum);
312 AM.HasBaseReg = true;
315 const SCEV *Sum = SE.getAddExpr(Bad);
317 BaseRegs.push_back(Sum);
318 AM.HasBaseReg = true;
322 /// getNumRegs - Return the total number of register operands used by this
323 /// formula. This does not include register uses implied by non-constant
325 unsigned Formula::getNumRegs() const {
326 return !!ScaledReg + BaseRegs.size();
329 /// getType - Return the type of this formula, if it has one, or null
330 /// otherwise. This type is meaningless except for the bit size.
331 Type *Formula::getType() const {
332 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
333 ScaledReg ? ScaledReg->getType() :
334 AM.BaseGV ? AM.BaseGV->getType() :
338 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
339 void Formula::DeleteBaseReg(const SCEV *&S) {
340 if (&S != &BaseRegs.back())
341 std::swap(S, BaseRegs.back());
345 /// referencesReg - Test if this formula references the given register.
346 bool Formula::referencesReg(const SCEV *S) const {
347 return S == ScaledReg ||
348 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
351 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
352 /// which are used by uses other than the use with the given index.
353 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
354 const RegUseTracker &RegUses) const {
356 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
358 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
359 E = BaseRegs.end(); I != E; ++I)
360 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
365 void Formula::print(raw_ostream &OS) const {
368 if (!First) OS << " + "; else First = false;
369 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
371 if (AM.BaseOffs != 0) {
372 if (!First) OS << " + "; else First = false;
375 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
376 E = BaseRegs.end(); I != E; ++I) {
377 if (!First) OS << " + "; else First = false;
378 OS << "reg(" << **I << ')';
380 if (AM.HasBaseReg && BaseRegs.empty()) {
381 if (!First) OS << " + "; else First = false;
382 OS << "**error: HasBaseReg**";
383 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
384 if (!First) OS << " + "; else First = false;
385 OS << "**error: !HasBaseReg**";
388 if (!First) OS << " + "; else First = false;
389 OS << AM.Scale << "*reg(";
396 if (UnfoldedOffset != 0) {
397 if (!First) OS << " + "; else First = false;
398 OS << "imm(" << UnfoldedOffset << ')';
402 void Formula::dump() const {
403 print(errs()); errs() << '\n';
406 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
407 /// without changing its value.
408 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
410 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
411 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
414 /// isAddSExtable - Return true if the given add can be sign-extended
415 /// without changing its value.
416 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
418 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
419 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
422 /// isMulSExtable - Return true if the given mul can be sign-extended
423 /// without changing its value.
424 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
426 IntegerType::get(SE.getContext(),
427 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
428 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
431 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
432 /// and if the remainder is known to be zero, or null otherwise. If
433 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
434 /// to Y, ignoring that the multiplication may overflow, which is useful when
435 /// the result will be used in a context where the most significant bits are
437 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
439 bool IgnoreSignificantBits = false) {
440 // Handle the trivial case, which works for any SCEV type.
442 return SE.getConstant(LHS->getType(), 1);
444 // Handle a few RHS special cases.
445 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
447 const APInt &RA = RC->getValue()->getValue();
448 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
450 if (RA.isAllOnesValue())
451 return SE.getMulExpr(LHS, RC);
452 // Handle x /s 1 as x.
457 // Check for a division of a constant by a constant.
458 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
461 const APInt &LA = C->getValue()->getValue();
462 const APInt &RA = RC->getValue()->getValue();
463 if (LA.srem(RA) != 0)
465 return SE.getConstant(LA.sdiv(RA));
468 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
469 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
470 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
471 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
472 IgnoreSignificantBits);
474 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
475 IgnoreSignificantBits);
476 if (!Start) return 0;
477 // FlagNW is independent of the start value, step direction, and is
478 // preserved with smaller magnitude steps.
479 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
480 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
485 // Distribute the sdiv over add operands, if the add doesn't overflow.
486 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
487 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
488 SmallVector<const SCEV *, 8> Ops;
489 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
491 const SCEV *Op = getExactSDiv(*I, RHS, SE,
492 IgnoreSignificantBits);
496 return SE.getAddExpr(Ops);
501 // Check for a multiply operand that we can pull RHS out of.
502 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
503 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
504 SmallVector<const SCEV *, 4> Ops;
506 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
510 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
511 IgnoreSignificantBits)) {
517 return Found ? SE.getMulExpr(Ops) : 0;
522 // Otherwise we don't know.
526 /// ExtractImmediate - If S involves the addition of a constant integer value,
527 /// return that integer value, and mutate S to point to a new SCEV with that
529 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
530 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
531 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
532 S = SE.getConstant(C->getType(), 0);
533 return C->getValue()->getSExtValue();
535 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
536 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
537 int64_t Result = ExtractImmediate(NewOps.front(), SE);
539 S = SE.getAddExpr(NewOps);
541 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
542 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
543 int64_t Result = ExtractImmediate(NewOps.front(), SE);
545 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
546 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
553 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
554 /// return that symbol, and mutate S to point to a new SCEV with that
556 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
557 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
558 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
559 S = SE.getConstant(GV->getType(), 0);
562 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
563 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
564 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
566 S = SE.getAddExpr(NewOps);
568 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
569 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
570 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
572 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
573 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
580 /// isAddressUse - Returns true if the specified instruction is using the
581 /// specified value as an address.
582 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
583 bool isAddress = isa<LoadInst>(Inst);
584 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
585 if (SI->getOperand(1) == OperandVal)
587 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
588 // Addressing modes can also be folded into prefetches and a variety
590 switch (II->getIntrinsicID()) {
592 case Intrinsic::prefetch:
593 case Intrinsic::x86_sse_storeu_ps:
594 case Intrinsic::x86_sse2_storeu_pd:
595 case Intrinsic::x86_sse2_storeu_dq:
596 case Intrinsic::x86_sse2_storel_dq:
597 if (II->getArgOperand(0) == OperandVal)
605 /// getAccessType - Return the type of the memory being accessed.
606 static Type *getAccessType(const Instruction *Inst) {
607 Type *AccessTy = Inst->getType();
608 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
609 AccessTy = SI->getOperand(0)->getType();
610 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
611 // Addressing modes can also be folded into prefetches and a variety
613 switch (II->getIntrinsicID()) {
615 case Intrinsic::x86_sse_storeu_ps:
616 case Intrinsic::x86_sse2_storeu_pd:
617 case Intrinsic::x86_sse2_storeu_dq:
618 case Intrinsic::x86_sse2_storel_dq:
619 AccessTy = II->getArgOperand(0)->getType();
624 // All pointers have the same requirements, so canonicalize them to an
625 // arbitrary pointer type to minimize variation.
626 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
627 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
628 PTy->getAddressSpace());
633 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
634 /// specified set are trivially dead, delete them and see if this makes any of
635 /// their operands subsequently dead.
637 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
638 bool Changed = false;
640 while (!DeadInsts.empty()) {
641 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
643 if (I == 0 || !isInstructionTriviallyDead(I))
646 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
647 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
650 DeadInsts.push_back(U);
653 I->eraseFromParent();
662 /// Cost - This class is used to measure and compare candidate formulae.
664 /// TODO: Some of these could be merged. Also, a lexical ordering
665 /// isn't always optimal.
669 unsigned NumBaseAdds;
675 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
678 bool operator<(const Cost &Other) const;
683 // Once any of the metrics loses, they must all remain losers.
685 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
686 | ImmCost | SetupCost) != ~0u)
687 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
688 & ImmCost & SetupCost) == ~0u);
693 assert(isValid() && "invalid cost");
694 return NumRegs == ~0u;
697 void RateFormula(const Formula &F,
698 SmallPtrSet<const SCEV *, 16> &Regs,
699 const DenseSet<const SCEV *> &VisitedRegs,
701 const SmallVectorImpl<int64_t> &Offsets,
702 ScalarEvolution &SE, DominatorTree &DT);
704 void print(raw_ostream &OS) const;
708 void RateRegister(const SCEV *Reg,
709 SmallPtrSet<const SCEV *, 16> &Regs,
711 ScalarEvolution &SE, DominatorTree &DT);
712 void RatePrimaryRegister(const SCEV *Reg,
713 SmallPtrSet<const SCEV *, 16> &Regs,
715 ScalarEvolution &SE, DominatorTree &DT);
720 /// RateRegister - Tally up interesting quantities from the given register.
721 void Cost::RateRegister(const SCEV *Reg,
722 SmallPtrSet<const SCEV *, 16> &Regs,
724 ScalarEvolution &SE, DominatorTree &DT) {
725 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
726 if (AR->getLoop() == L)
727 AddRecCost += 1; /// TODO: This should be a function of the stride.
729 // If this is an addrec for another loop, don't second-guess its addrec phi
730 // nodes. LSR isn't currently smart enough to reason about more than one
731 // loop at a time. LSR has either already run on inner loops, will not run
732 // on other loops, and cannot be expected to change sibling loops. If the
733 // AddRec exists, consider it's register free and leave it alone. Otherwise,
734 // do not consider this formula at all.
735 // FIXME: why do we need to generate such fomulae?
736 else if (!EnableNested || L->contains(AR->getLoop()) ||
737 (!AR->getLoop()->contains(L) &&
738 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
739 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
740 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
741 if (SE.isSCEVable(PN->getType()) &&
742 (SE.getEffectiveSCEVType(PN->getType()) ==
743 SE.getEffectiveSCEVType(AR->getType())) &&
744 SE.getSCEV(PN) == AR)
751 // If this isn't one of the addrecs that the loop already has, it
752 // would require a costly new phi and add. TODO: This isn't
753 // precisely modeled right now.
755 if (!Regs.count(AR->getStart())) {
756 RateRegister(AR->getStart(), Regs, L, SE, DT);
762 // Add the step value register, if it needs one.
763 // TODO: The non-affine case isn't precisely modeled here.
764 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
765 if (!Regs.count(AR->getOperand(1))) {
766 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
774 // Rough heuristic; favor registers which don't require extra setup
775 // instructions in the preheader.
776 if (!isa<SCEVUnknown>(Reg) &&
777 !isa<SCEVConstant>(Reg) &&
778 !(isa<SCEVAddRecExpr>(Reg) &&
779 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
780 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
783 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
784 SE.hasComputableLoopEvolution(Reg, L);
787 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
789 void Cost::RatePrimaryRegister(const SCEV *Reg,
790 SmallPtrSet<const SCEV *, 16> &Regs,
792 ScalarEvolution &SE, DominatorTree &DT) {
793 if (Regs.insert(Reg))
794 RateRegister(Reg, Regs, L, SE, DT);
797 void Cost::RateFormula(const Formula &F,
798 SmallPtrSet<const SCEV *, 16> &Regs,
799 const DenseSet<const SCEV *> &VisitedRegs,
801 const SmallVectorImpl<int64_t> &Offsets,
802 ScalarEvolution &SE, DominatorTree &DT) {
803 // Tally up the registers.
804 if (const SCEV *ScaledReg = F.ScaledReg) {
805 if (VisitedRegs.count(ScaledReg)) {
809 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
813 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
814 E = F.BaseRegs.end(); I != E; ++I) {
815 const SCEV *BaseReg = *I;
816 if (VisitedRegs.count(BaseReg)) {
820 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
825 // Determine how many (unfolded) adds we'll need inside the loop.
826 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
827 if (NumBaseParts > 1)
828 NumBaseAdds += NumBaseParts - 1;
830 // Tally up the non-zero immediates.
831 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
832 E = Offsets.end(); I != E; ++I) {
833 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
835 ImmCost += 64; // Handle symbolic values conservatively.
836 // TODO: This should probably be the pointer size.
837 else if (Offset != 0)
838 ImmCost += APInt(64, Offset, true).getMinSignedBits();
840 assert(isValid() && "invalid cost");
843 /// Loose - Set this cost to a losing value.
853 /// operator< - Choose the lower cost.
854 bool Cost::operator<(const Cost &Other) const {
855 if (NumRegs != Other.NumRegs)
856 return NumRegs < Other.NumRegs;
857 if (AddRecCost != Other.AddRecCost)
858 return AddRecCost < Other.AddRecCost;
859 if (NumIVMuls != Other.NumIVMuls)
860 return NumIVMuls < Other.NumIVMuls;
861 if (NumBaseAdds != Other.NumBaseAdds)
862 return NumBaseAdds < Other.NumBaseAdds;
863 if (ImmCost != Other.ImmCost)
864 return ImmCost < Other.ImmCost;
865 if (SetupCost != Other.SetupCost)
866 return SetupCost < Other.SetupCost;
870 void Cost::print(raw_ostream &OS) const {
871 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
873 OS << ", with addrec cost " << AddRecCost;
875 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
876 if (NumBaseAdds != 0)
877 OS << ", plus " << NumBaseAdds << " base add"
878 << (NumBaseAdds == 1 ? "" : "s");
880 OS << ", plus " << ImmCost << " imm cost";
882 OS << ", plus " << SetupCost << " setup cost";
885 void Cost::dump() const {
886 print(errs()); errs() << '\n';
891 /// LSRFixup - An operand value in an instruction which is to be replaced
892 /// with some equivalent, possibly strength-reduced, replacement.
894 /// UserInst - The instruction which will be updated.
895 Instruction *UserInst;
897 /// OperandValToReplace - The operand of the instruction which will
898 /// be replaced. The operand may be used more than once; every instance
899 /// will be replaced.
900 Value *OperandValToReplace;
902 /// PostIncLoops - If this user is to use the post-incremented value of an
903 /// induction variable, this variable is non-null and holds the loop
904 /// associated with the induction variable.
905 PostIncLoopSet PostIncLoops;
907 /// LUIdx - The index of the LSRUse describing the expression which
908 /// this fixup needs, minus an offset (below).
911 /// Offset - A constant offset to be added to the LSRUse expression.
912 /// This allows multiple fixups to share the same LSRUse with different
913 /// offsets, for example in an unrolled loop.
916 bool isUseFullyOutsideLoop(const Loop *L) const;
920 void print(raw_ostream &OS) const;
927 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
929 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
930 /// value outside of the given loop.
931 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
932 // PHI nodes use their value in their incoming blocks.
933 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
934 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
935 if (PN->getIncomingValue(i) == OperandValToReplace &&
936 L->contains(PN->getIncomingBlock(i)))
941 return !L->contains(UserInst);
944 void LSRFixup::print(raw_ostream &OS) const {
946 // Store is common and interesting enough to be worth special-casing.
947 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
949 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
950 } else if (UserInst->getType()->isVoidTy())
951 OS << UserInst->getOpcodeName();
953 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
955 OS << ", OperandValToReplace=";
956 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
958 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
959 E = PostIncLoops.end(); I != E; ++I) {
960 OS << ", PostIncLoop=";
961 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
964 if (LUIdx != ~size_t(0))
965 OS << ", LUIdx=" << LUIdx;
968 OS << ", Offset=" << Offset;
971 void LSRFixup::dump() const {
972 print(errs()); errs() << '\n';
977 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
978 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
979 struct UniquifierDenseMapInfo {
980 static SmallVector<const SCEV *, 2> getEmptyKey() {
981 SmallVector<const SCEV *, 2> V;
982 V.push_back(reinterpret_cast<const SCEV *>(-1));
986 static SmallVector<const SCEV *, 2> getTombstoneKey() {
987 SmallVector<const SCEV *, 2> V;
988 V.push_back(reinterpret_cast<const SCEV *>(-2));
992 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
994 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
995 E = V.end(); I != E; ++I)
996 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1000 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1001 const SmallVector<const SCEV *, 2> &RHS) {
1006 /// LSRUse - This class holds the state that LSR keeps for each use in
1007 /// IVUsers, as well as uses invented by LSR itself. It includes information
1008 /// about what kinds of things can be folded into the user, information about
1009 /// the user itself, and information about how the use may be satisfied.
1010 /// TODO: Represent multiple users of the same expression in common?
1012 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1015 /// KindType - An enum for a kind of use, indicating what types of
1016 /// scaled and immediate operands it might support.
1018 Basic, ///< A normal use, with no folding.
1019 Special, ///< A special case of basic, allowing -1 scales.
1020 Address, ///< An address use; folding according to TargetLowering
1021 ICmpZero ///< An equality icmp with both operands folded into one.
1022 // TODO: Add a generic icmp too?
1028 SmallVector<int64_t, 8> Offsets;
1032 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1033 /// LSRUse are outside of the loop, in which case some special-case heuristics
1035 bool AllFixupsOutsideLoop;
1037 /// WidestFixupType - This records the widest use type for any fixup using
1038 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1039 /// max fixup widths to be equivalent, because the narrower one may be relying
1040 /// on the implicit truncation to truncate away bogus bits.
1041 Type *WidestFixupType;
1043 /// Formulae - A list of ways to build a value that can satisfy this user.
1044 /// After the list is populated, one of these is selected heuristically and
1045 /// used to formulate a replacement for OperandValToReplace in UserInst.
1046 SmallVector<Formula, 12> Formulae;
1048 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1049 SmallPtrSet<const SCEV *, 4> Regs;
1051 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1052 MinOffset(INT64_MAX),
1053 MaxOffset(INT64_MIN),
1054 AllFixupsOutsideLoop(true),
1055 WidestFixupType(0) {}
1057 bool HasFormulaWithSameRegs(const Formula &F) const;
1058 bool InsertFormula(const Formula &F);
1059 void DeleteFormula(Formula &F);
1060 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1062 void print(raw_ostream &OS) const;
1068 /// HasFormula - Test whether this use as a formula which has the same
1069 /// registers as the given formula.
1070 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1071 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1072 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1073 // Unstable sort by host order ok, because this is only used for uniquifying.
1074 std::sort(Key.begin(), Key.end());
1075 return Uniquifier.count(Key);
1078 /// InsertFormula - If the given formula has not yet been inserted, add it to
1079 /// the list, and return true. Return false otherwise.
1080 bool LSRUse::InsertFormula(const Formula &F) {
1081 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1082 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1083 // Unstable sort by host order ok, because this is only used for uniquifying.
1084 std::sort(Key.begin(), Key.end());
1086 if (!Uniquifier.insert(Key).second)
1089 // Using a register to hold the value of 0 is not profitable.
1090 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1091 "Zero allocated in a scaled register!");
1093 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1094 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1095 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1098 // Add the formula to the list.
1099 Formulae.push_back(F);
1101 // Record registers now being used by this use.
1102 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1103 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1108 /// DeleteFormula - Remove the given formula from this use's list.
1109 void LSRUse::DeleteFormula(Formula &F) {
1110 if (&F != &Formulae.back())
1111 std::swap(F, Formulae.back());
1112 Formulae.pop_back();
1113 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1116 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1117 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1118 // Now that we've filtered out some formulae, recompute the Regs set.
1119 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1121 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1122 E = Formulae.end(); I != E; ++I) {
1123 const Formula &F = *I;
1124 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1125 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1128 // Update the RegTracker.
1129 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1130 E = OldRegs.end(); I != E; ++I)
1131 if (!Regs.count(*I))
1132 RegUses.DropRegister(*I, LUIdx);
1135 void LSRUse::print(raw_ostream &OS) const {
1136 OS << "LSR Use: Kind=";
1138 case Basic: OS << "Basic"; break;
1139 case Special: OS << "Special"; break;
1140 case ICmpZero: OS << "ICmpZero"; break;
1142 OS << "Address of ";
1143 if (AccessTy->isPointerTy())
1144 OS << "pointer"; // the full pointer type could be really verbose
1149 OS << ", Offsets={";
1150 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1151 E = Offsets.end(); I != E; ++I) {
1153 if (llvm::next(I) != E)
1158 if (AllFixupsOutsideLoop)
1159 OS << ", all-fixups-outside-loop";
1161 if (WidestFixupType)
1162 OS << ", widest fixup type: " << *WidestFixupType;
1165 void LSRUse::dump() const {
1166 print(errs()); errs() << '\n';
1169 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1170 /// be completely folded into the user instruction at isel time. This includes
1171 /// address-mode folding and special icmp tricks.
1172 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1173 LSRUse::KindType Kind, Type *AccessTy,
1174 const TargetLowering *TLI) {
1176 case LSRUse::Address:
1177 // If we have low-level target information, ask the target if it can
1178 // completely fold this address.
1179 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1181 // Otherwise, just guess that reg+reg addressing is legal.
1182 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1184 case LSRUse::ICmpZero:
1185 // There's not even a target hook for querying whether it would be legal to
1186 // fold a GV into an ICmp.
1190 // ICmp only has two operands; don't allow more than two non-trivial parts.
1191 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1194 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1195 // putting the scaled register in the other operand of the icmp.
1196 if (AM.Scale != 0 && AM.Scale != -1)
1199 // If we have low-level target information, ask the target if it can fold an
1200 // integer immediate on an icmp.
1201 if (AM.BaseOffs != 0) {
1202 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1209 // Only handle single-register values.
1210 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1212 case LSRUse::Special:
1213 // Only handle -1 scales, or no scale.
1214 return AM.Scale == 0 || AM.Scale == -1;
1220 static bool isLegalUse(TargetLowering::AddrMode AM,
1221 int64_t MinOffset, int64_t MaxOffset,
1222 LSRUse::KindType Kind, Type *AccessTy,
1223 const TargetLowering *TLI) {
1224 // Check for overflow.
1225 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1228 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1229 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1230 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1231 // Check for overflow.
1232 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1235 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1236 return isLegalUse(AM, Kind, AccessTy, TLI);
1241 static bool isAlwaysFoldable(int64_t BaseOffs,
1242 GlobalValue *BaseGV,
1244 LSRUse::KindType Kind, Type *AccessTy,
1245 const TargetLowering *TLI) {
1246 // Fast-path: zero is always foldable.
1247 if (BaseOffs == 0 && !BaseGV) return true;
1249 // Conservatively, create an address with an immediate and a
1250 // base and a scale.
1251 TargetLowering::AddrMode AM;
1252 AM.BaseOffs = BaseOffs;
1254 AM.HasBaseReg = HasBaseReg;
1255 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1257 // Canonicalize a scale of 1 to a base register if the formula doesn't
1258 // already have a base register.
1259 if (!AM.HasBaseReg && AM.Scale == 1) {
1261 AM.HasBaseReg = true;
1264 return isLegalUse(AM, Kind, AccessTy, TLI);
1267 static bool isAlwaysFoldable(const SCEV *S,
1268 int64_t MinOffset, int64_t MaxOffset,
1270 LSRUse::KindType Kind, Type *AccessTy,
1271 const TargetLowering *TLI,
1272 ScalarEvolution &SE) {
1273 // Fast-path: zero is always foldable.
1274 if (S->isZero()) return true;
1276 // Conservatively, create an address with an immediate and a
1277 // base and a scale.
1278 int64_t BaseOffs = ExtractImmediate(S, SE);
1279 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1281 // If there's anything else involved, it's not foldable.
1282 if (!S->isZero()) return false;
1284 // Fast-path: zero is always foldable.
1285 if (BaseOffs == 0 && !BaseGV) return true;
1287 // Conservatively, create an address with an immediate and a
1288 // base and a scale.
1289 TargetLowering::AddrMode AM;
1290 AM.BaseOffs = BaseOffs;
1292 AM.HasBaseReg = HasBaseReg;
1293 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1295 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1300 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1301 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1302 struct UseMapDenseMapInfo {
1303 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1304 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1307 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1308 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1312 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1313 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1314 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1318 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1319 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1324 /// LSRInstance - This class holds state for the main loop strength reduction
1328 ScalarEvolution &SE;
1331 const TargetLowering *const TLI;
1335 /// IVIncInsertPos - This is the insert position that the current loop's
1336 /// induction variable increment should be placed. In simple loops, this is
1337 /// the latch block's terminator. But in more complicated cases, this is a
1338 /// position which will dominate all the in-loop post-increment users.
1339 Instruction *IVIncInsertPos;
1341 /// Factors - Interesting factors between use strides.
1342 SmallSetVector<int64_t, 8> Factors;
1344 /// Types - Interesting use types, to facilitate truncation reuse.
1345 SmallSetVector<Type *, 4> Types;
1347 /// Fixups - The list of operands which are to be replaced.
1348 SmallVector<LSRFixup, 16> Fixups;
1350 /// Uses - The list of interesting uses.
1351 SmallVector<LSRUse, 16> Uses;
1353 /// RegUses - Track which uses use which register candidates.
1354 RegUseTracker RegUses;
1356 void OptimizeShadowIV();
1357 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1358 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1359 void OptimizeLoopTermCond();
1361 void CollectInterestingTypesAndFactors();
1362 void CollectFixupsAndInitialFormulae();
1364 LSRFixup &getNewFixup() {
1365 Fixups.push_back(LSRFixup());
1366 return Fixups.back();
1369 // Support for sharing of LSRUses between LSRFixups.
1370 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1372 UseMapDenseMapInfo> UseMapTy;
1375 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1376 LSRUse::KindType Kind, Type *AccessTy);
1378 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1379 LSRUse::KindType Kind,
1382 void DeleteUse(LSRUse &LU, size_t LUIdx);
1384 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1387 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1388 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1389 void CountRegisters(const Formula &F, size_t LUIdx);
1390 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1392 void CollectLoopInvariantFixupsAndFormulae();
1394 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1395 unsigned Depth = 0);
1396 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1397 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1398 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1399 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1400 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1401 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1402 void GenerateCrossUseConstantOffsets();
1403 void GenerateAllReuseFormulae();
1405 void FilterOutUndesirableDedicatedRegisters();
1407 size_t EstimateSearchSpaceComplexity() const;
1408 void NarrowSearchSpaceByDetectingSupersets();
1409 void NarrowSearchSpaceByCollapsingUnrolledCode();
1410 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1411 void NarrowSearchSpaceByPickingWinnerRegs();
1412 void NarrowSearchSpaceUsingHeuristics();
1414 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1416 SmallVectorImpl<const Formula *> &Workspace,
1417 const Cost &CurCost,
1418 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1419 DenseSet<const SCEV *> &VisitedRegs) const;
1420 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1422 BasicBlock::iterator
1423 HoistInsertPosition(BasicBlock::iterator IP,
1424 const SmallVectorImpl<Instruction *> &Inputs) const;
1425 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1427 const LSRUse &LU) const;
1429 Value *Expand(const LSRFixup &LF,
1431 BasicBlock::iterator IP,
1432 SCEVExpander &Rewriter,
1433 SmallVectorImpl<WeakVH> &DeadInsts) const;
1434 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1436 SCEVExpander &Rewriter,
1437 SmallVectorImpl<WeakVH> &DeadInsts,
1439 void Rewrite(const LSRFixup &LF,
1441 SCEVExpander &Rewriter,
1442 SmallVectorImpl<WeakVH> &DeadInsts,
1444 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1447 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1449 bool getChanged() const { return Changed; }
1451 void print_factors_and_types(raw_ostream &OS) const;
1452 void print_fixups(raw_ostream &OS) const;
1453 void print_uses(raw_ostream &OS) const;
1454 void print(raw_ostream &OS) const;
1460 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1461 /// inside the loop then try to eliminate the cast operation.
1462 void LSRInstance::OptimizeShadowIV() {
1463 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1464 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1467 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1468 UI != E; /* empty */) {
1469 IVUsers::const_iterator CandidateUI = UI;
1471 Instruction *ShadowUse = CandidateUI->getUser();
1472 Type *DestTy = NULL;
1473 bool IsSigned = false;
1475 /* If shadow use is a int->float cast then insert a second IV
1476 to eliminate this cast.
1478 for (unsigned i = 0; i < n; ++i)
1484 for (unsigned i = 0; i < n; ++i, ++d)
1487 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1489 DestTy = UCast->getDestTy();
1491 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1493 DestTy = SCast->getDestTy();
1495 if (!DestTy) continue;
1498 // If target does not support DestTy natively then do not apply
1499 // this transformation.
1500 EVT DVT = TLI->getValueType(DestTy);
1501 if (!TLI->isTypeLegal(DVT)) continue;
1504 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1506 if (PH->getNumIncomingValues() != 2) continue;
1508 Type *SrcTy = PH->getType();
1509 int Mantissa = DestTy->getFPMantissaWidth();
1510 if (Mantissa == -1) continue;
1511 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1514 unsigned Entry, Latch;
1515 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1523 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1524 if (!Init) continue;
1525 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1526 (double)Init->getSExtValue() :
1527 (double)Init->getZExtValue());
1529 BinaryOperator *Incr =
1530 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1531 if (!Incr) continue;
1532 if (Incr->getOpcode() != Instruction::Add
1533 && Incr->getOpcode() != Instruction::Sub)
1536 /* Initialize new IV, double d = 0.0 in above example. */
1537 ConstantInt *C = NULL;
1538 if (Incr->getOperand(0) == PH)
1539 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1540 else if (Incr->getOperand(1) == PH)
1541 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1547 // Ignore negative constants, as the code below doesn't handle them
1548 // correctly. TODO: Remove this restriction.
1549 if (!C->getValue().isStrictlyPositive()) continue;
1551 /* Add new PHINode. */
1552 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1554 /* create new increment. '++d' in above example. */
1555 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1556 BinaryOperator *NewIncr =
1557 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1558 Instruction::FAdd : Instruction::FSub,
1559 NewPH, CFP, "IV.S.next.", Incr);
1561 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1562 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1564 /* Remove cast operation */
1565 ShadowUse->replaceAllUsesWith(NewPH);
1566 ShadowUse->eraseFromParent();
1572 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1573 /// set the IV user and stride information and return true, otherwise return
1575 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1576 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1577 if (UI->getUser() == Cond) {
1578 // NOTE: we could handle setcc instructions with multiple uses here, but
1579 // InstCombine does it as well for simple uses, it's not clear that it
1580 // occurs enough in real life to handle.
1587 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1588 /// a max computation.
1590 /// This is a narrow solution to a specific, but acute, problem. For loops
1596 /// } while (++i < n);
1598 /// the trip count isn't just 'n', because 'n' might not be positive. And
1599 /// unfortunately this can come up even for loops where the user didn't use
1600 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1601 /// will commonly be lowered like this:
1607 /// } while (++i < n);
1610 /// and then it's possible for subsequent optimization to obscure the if
1611 /// test in such a way that indvars can't find it.
1613 /// When indvars can't find the if test in loops like this, it creates a
1614 /// max expression, which allows it to give the loop a canonical
1615 /// induction variable:
1618 /// max = n < 1 ? 1 : n;
1621 /// } while (++i != max);
1623 /// Canonical induction variables are necessary because the loop passes
1624 /// are designed around them. The most obvious example of this is the
1625 /// LoopInfo analysis, which doesn't remember trip count values. It
1626 /// expects to be able to rediscover the trip count each time it is
1627 /// needed, and it does this using a simple analysis that only succeeds if
1628 /// the loop has a canonical induction variable.
1630 /// However, when it comes time to generate code, the maximum operation
1631 /// can be quite costly, especially if it's inside of an outer loop.
1633 /// This function solves this problem by detecting this type of loop and
1634 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1635 /// the instructions for the maximum computation.
1637 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1638 // Check that the loop matches the pattern we're looking for.
1639 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1640 Cond->getPredicate() != CmpInst::ICMP_NE)
1643 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1644 if (!Sel || !Sel->hasOneUse()) return Cond;
1646 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1647 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1649 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1651 // Add one to the backedge-taken count to get the trip count.
1652 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1653 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1655 // Check for a max calculation that matches the pattern. There's no check
1656 // for ICMP_ULE here because the comparison would be with zero, which
1657 // isn't interesting.
1658 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1659 const SCEVNAryExpr *Max = 0;
1660 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1661 Pred = ICmpInst::ICMP_SLE;
1663 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1664 Pred = ICmpInst::ICMP_SLT;
1666 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1667 Pred = ICmpInst::ICMP_ULT;
1674 // To handle a max with more than two operands, this optimization would
1675 // require additional checking and setup.
1676 if (Max->getNumOperands() != 2)
1679 const SCEV *MaxLHS = Max->getOperand(0);
1680 const SCEV *MaxRHS = Max->getOperand(1);
1682 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1683 // for a comparison with 1. For <= and >=, a comparison with zero.
1685 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1688 // Check the relevant induction variable for conformance to
1690 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1691 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1692 if (!AR || !AR->isAffine() ||
1693 AR->getStart() != One ||
1694 AR->getStepRecurrence(SE) != One)
1697 assert(AR->getLoop() == L &&
1698 "Loop condition operand is an addrec in a different loop!");
1700 // Check the right operand of the select, and remember it, as it will
1701 // be used in the new comparison instruction.
1703 if (ICmpInst::isTrueWhenEqual(Pred)) {
1704 // Look for n+1, and grab n.
1705 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1706 if (isa<ConstantInt>(BO->getOperand(1)) &&
1707 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1708 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1709 NewRHS = BO->getOperand(0);
1710 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1711 if (isa<ConstantInt>(BO->getOperand(1)) &&
1712 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1713 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1714 NewRHS = BO->getOperand(0);
1717 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1718 NewRHS = Sel->getOperand(1);
1719 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1720 NewRHS = Sel->getOperand(2);
1721 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1722 NewRHS = SU->getValue();
1724 // Max doesn't match expected pattern.
1727 // Determine the new comparison opcode. It may be signed or unsigned,
1728 // and the original comparison may be either equality or inequality.
1729 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1730 Pred = CmpInst::getInversePredicate(Pred);
1732 // Ok, everything looks ok to change the condition into an SLT or SGE and
1733 // delete the max calculation.
1735 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1737 // Delete the max calculation instructions.
1738 Cond->replaceAllUsesWith(NewCond);
1739 CondUse->setUser(NewCond);
1740 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1741 Cond->eraseFromParent();
1742 Sel->eraseFromParent();
1743 if (Cmp->use_empty())
1744 Cmp->eraseFromParent();
1748 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1749 /// postinc iv when possible.
1751 LSRInstance::OptimizeLoopTermCond() {
1752 SmallPtrSet<Instruction *, 4> PostIncs;
1754 BasicBlock *LatchBlock = L->getLoopLatch();
1755 SmallVector<BasicBlock*, 8> ExitingBlocks;
1756 L->getExitingBlocks(ExitingBlocks);
1758 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1759 BasicBlock *ExitingBlock = ExitingBlocks[i];
1761 // Get the terminating condition for the loop if possible. If we
1762 // can, we want to change it to use a post-incremented version of its
1763 // induction variable, to allow coalescing the live ranges for the IV into
1764 // one register value.
1766 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1769 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1770 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1773 // Search IVUsesByStride to find Cond's IVUse if there is one.
1774 IVStrideUse *CondUse = 0;
1775 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1776 if (!FindIVUserForCond(Cond, CondUse))
1779 // If the trip count is computed in terms of a max (due to ScalarEvolution
1780 // being unable to find a sufficient guard, for example), change the loop
1781 // comparison to use SLT or ULT instead of NE.
1782 // One consequence of doing this now is that it disrupts the count-down
1783 // optimization. That's not always a bad thing though, because in such
1784 // cases it may still be worthwhile to avoid a max.
1785 Cond = OptimizeMax(Cond, CondUse);
1787 // If this exiting block dominates the latch block, it may also use
1788 // the post-inc value if it won't be shared with other uses.
1789 // Check for dominance.
1790 if (!DT.dominates(ExitingBlock, LatchBlock))
1793 // Conservatively avoid trying to use the post-inc value in non-latch
1794 // exits if there may be pre-inc users in intervening blocks.
1795 if (LatchBlock != ExitingBlock)
1796 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1797 // Test if the use is reachable from the exiting block. This dominator
1798 // query is a conservative approximation of reachability.
1799 if (&*UI != CondUse &&
1800 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1801 // Conservatively assume there may be reuse if the quotient of their
1802 // strides could be a legal scale.
1803 const SCEV *A = IU.getStride(*CondUse, L);
1804 const SCEV *B = IU.getStride(*UI, L);
1805 if (!A || !B) continue;
1806 if (SE.getTypeSizeInBits(A->getType()) !=
1807 SE.getTypeSizeInBits(B->getType())) {
1808 if (SE.getTypeSizeInBits(A->getType()) >
1809 SE.getTypeSizeInBits(B->getType()))
1810 B = SE.getSignExtendExpr(B, A->getType());
1812 A = SE.getSignExtendExpr(A, B->getType());
1814 if (const SCEVConstant *D =
1815 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1816 const ConstantInt *C = D->getValue();
1817 // Stride of one or negative one can have reuse with non-addresses.
1818 if (C->isOne() || C->isAllOnesValue())
1819 goto decline_post_inc;
1820 // Avoid weird situations.
1821 if (C->getValue().getMinSignedBits() >= 64 ||
1822 C->getValue().isMinSignedValue())
1823 goto decline_post_inc;
1824 // Without TLI, assume that any stride might be valid, and so any
1825 // use might be shared.
1827 goto decline_post_inc;
1828 // Check for possible scaled-address reuse.
1829 Type *AccessTy = getAccessType(UI->getUser());
1830 TargetLowering::AddrMode AM;
1831 AM.Scale = C->getSExtValue();
1832 if (TLI->isLegalAddressingMode(AM, AccessTy))
1833 goto decline_post_inc;
1834 AM.Scale = -AM.Scale;
1835 if (TLI->isLegalAddressingMode(AM, AccessTy))
1836 goto decline_post_inc;
1840 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1843 // It's possible for the setcc instruction to be anywhere in the loop, and
1844 // possible for it to have multiple users. If it is not immediately before
1845 // the exiting block branch, move it.
1846 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1847 if (Cond->hasOneUse()) {
1848 Cond->moveBefore(TermBr);
1850 // Clone the terminating condition and insert into the loopend.
1851 ICmpInst *OldCond = Cond;
1852 Cond = cast<ICmpInst>(Cond->clone());
1853 Cond->setName(L->getHeader()->getName() + ".termcond");
1854 ExitingBlock->getInstList().insert(TermBr, Cond);
1856 // Clone the IVUse, as the old use still exists!
1857 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1858 TermBr->replaceUsesOfWith(OldCond, Cond);
1862 // If we get to here, we know that we can transform the setcc instruction to
1863 // use the post-incremented version of the IV, allowing us to coalesce the
1864 // live ranges for the IV correctly.
1865 CondUse->transformToPostInc(L);
1868 PostIncs.insert(Cond);
1872 // Determine an insertion point for the loop induction variable increment. It
1873 // must dominate all the post-inc comparisons we just set up, and it must
1874 // dominate the loop latch edge.
1875 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1876 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1877 E = PostIncs.end(); I != E; ++I) {
1879 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1881 if (BB == (*I)->getParent())
1882 IVIncInsertPos = *I;
1883 else if (BB != IVIncInsertPos->getParent())
1884 IVIncInsertPos = BB->getTerminator();
1888 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
1889 /// at the given offset and other details. If so, update the use and
1892 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1893 LSRUse::KindType Kind, Type *AccessTy) {
1894 int64_t NewMinOffset = LU.MinOffset;
1895 int64_t NewMaxOffset = LU.MaxOffset;
1896 Type *NewAccessTy = AccessTy;
1898 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1899 // something conservative, however this can pessimize in the case that one of
1900 // the uses will have all its uses outside the loop, for example.
1901 if (LU.Kind != Kind)
1903 // Conservatively assume HasBaseReg is true for now.
1904 if (NewOffset < LU.MinOffset) {
1905 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1906 Kind, AccessTy, TLI))
1908 NewMinOffset = NewOffset;
1909 } else if (NewOffset > LU.MaxOffset) {
1910 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1911 Kind, AccessTy, TLI))
1913 NewMaxOffset = NewOffset;
1915 // Check for a mismatched access type, and fall back conservatively as needed.
1916 // TODO: Be less conservative when the type is similar and can use the same
1917 // addressing modes.
1918 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1919 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1922 LU.MinOffset = NewMinOffset;
1923 LU.MaxOffset = NewMaxOffset;
1924 LU.AccessTy = NewAccessTy;
1925 if (NewOffset != LU.Offsets.back())
1926 LU.Offsets.push_back(NewOffset);
1930 /// getUse - Return an LSRUse index and an offset value for a fixup which
1931 /// needs the given expression, with the given kind and optional access type.
1932 /// Either reuse an existing use or create a new one, as needed.
1933 std::pair<size_t, int64_t>
1934 LSRInstance::getUse(const SCEV *&Expr,
1935 LSRUse::KindType Kind, Type *AccessTy) {
1936 const SCEV *Copy = Expr;
1937 int64_t Offset = ExtractImmediate(Expr, SE);
1939 // Basic uses can't accept any offset, for example.
1940 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1945 std::pair<UseMapTy::iterator, bool> P =
1946 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1948 // A use already existed with this base.
1949 size_t LUIdx = P.first->second;
1950 LSRUse &LU = Uses[LUIdx];
1951 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1953 return std::make_pair(LUIdx, Offset);
1956 // Create a new use.
1957 size_t LUIdx = Uses.size();
1958 P.first->second = LUIdx;
1959 Uses.push_back(LSRUse(Kind, AccessTy));
1960 LSRUse &LU = Uses[LUIdx];
1962 // We don't need to track redundant offsets, but we don't need to go out
1963 // of our way here to avoid them.
1964 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1965 LU.Offsets.push_back(Offset);
1967 LU.MinOffset = Offset;
1968 LU.MaxOffset = Offset;
1969 return std::make_pair(LUIdx, Offset);
1972 /// DeleteUse - Delete the given use from the Uses list.
1973 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
1974 if (&LU != &Uses.back())
1975 std::swap(LU, Uses.back());
1979 RegUses.SwapAndDropUse(LUIdx, Uses.size());
1982 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1983 /// a formula that has the same registers as the given formula.
1985 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1986 const LSRUse &OrigLU) {
1987 // Search all uses for the formula. This could be more clever.
1988 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1989 LSRUse &LU = Uses[LUIdx];
1990 // Check whether this use is close enough to OrigLU, to see whether it's
1991 // worthwhile looking through its formulae.
1992 // Ignore ICmpZero uses because they may contain formulae generated by
1993 // GenerateICmpZeroScales, in which case adding fixup offsets may
1995 if (&LU != &OrigLU &&
1996 LU.Kind != LSRUse::ICmpZero &&
1997 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1998 LU.WidestFixupType == OrigLU.WidestFixupType &&
1999 LU.HasFormulaWithSameRegs(OrigF)) {
2000 // Scan through this use's formulae.
2001 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2002 E = LU.Formulae.end(); I != E; ++I) {
2003 const Formula &F = *I;
2004 // Check to see if this formula has the same registers and symbols
2006 if (F.BaseRegs == OrigF.BaseRegs &&
2007 F.ScaledReg == OrigF.ScaledReg &&
2008 F.AM.BaseGV == OrigF.AM.BaseGV &&
2009 F.AM.Scale == OrigF.AM.Scale &&
2010 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2011 if (F.AM.BaseOffs == 0)
2013 // This is the formula where all the registers and symbols matched;
2014 // there aren't going to be any others. Since we declined it, we
2015 // can skip the rest of the formulae and procede to the next LSRUse.
2022 // Nothing looked good.
2026 void LSRInstance::CollectInterestingTypesAndFactors() {
2027 SmallSetVector<const SCEV *, 4> Strides;
2029 // Collect interesting types and strides.
2030 SmallVector<const SCEV *, 4> Worklist;
2031 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2032 const SCEV *Expr = IU.getExpr(*UI);
2034 // Collect interesting types.
2035 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2037 // Add strides for mentioned loops.
2038 Worklist.push_back(Expr);
2040 const SCEV *S = Worklist.pop_back_val();
2041 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2042 Strides.insert(AR->getStepRecurrence(SE));
2043 Worklist.push_back(AR->getStart());
2044 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2045 Worklist.append(Add->op_begin(), Add->op_end());
2047 } while (!Worklist.empty());
2050 // Compute interesting factors from the set of interesting strides.
2051 for (SmallSetVector<const SCEV *, 4>::const_iterator
2052 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2053 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2054 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2055 const SCEV *OldStride = *I;
2056 const SCEV *NewStride = *NewStrideIter;
2058 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2059 SE.getTypeSizeInBits(NewStride->getType())) {
2060 if (SE.getTypeSizeInBits(OldStride->getType()) >
2061 SE.getTypeSizeInBits(NewStride->getType()))
2062 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2064 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2066 if (const SCEVConstant *Factor =
2067 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2069 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2070 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2071 } else if (const SCEVConstant *Factor =
2072 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2075 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2076 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2080 // If all uses use the same type, don't bother looking for truncation-based
2082 if (Types.size() == 1)
2085 DEBUG(print_factors_and_types(dbgs()));
2088 void LSRInstance::CollectFixupsAndInitialFormulae() {
2089 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2091 LSRFixup &LF = getNewFixup();
2092 LF.UserInst = UI->getUser();
2093 LF.OperandValToReplace = UI->getOperandValToReplace();
2094 LF.PostIncLoops = UI->getPostIncLoops();
2096 LSRUse::KindType Kind = LSRUse::Basic;
2098 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2099 Kind = LSRUse::Address;
2100 AccessTy = getAccessType(LF.UserInst);
2103 const SCEV *S = IU.getExpr(*UI);
2105 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2106 // (N - i == 0), and this allows (N - i) to be the expression that we work
2107 // with rather than just N or i, so we can consider the register
2108 // requirements for both N and i at the same time. Limiting this code to
2109 // equality icmps is not a problem because all interesting loops use
2110 // equality icmps, thanks to IndVarSimplify.
2111 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2112 if (CI->isEquality()) {
2113 // Swap the operands if needed to put the OperandValToReplace on the
2114 // left, for consistency.
2115 Value *NV = CI->getOperand(1);
2116 if (NV == LF.OperandValToReplace) {
2117 CI->setOperand(1, CI->getOperand(0));
2118 CI->setOperand(0, NV);
2119 NV = CI->getOperand(1);
2123 // x == y --> x - y == 0
2124 const SCEV *N = SE.getSCEV(NV);
2125 if (SE.isLoopInvariant(N, L)) {
2126 // S is normalized, so normalize N before folding it into S
2127 // to keep the result normalized.
2128 N = TransformForPostIncUse(Normalize, N, CI, 0,
2129 LF.PostIncLoops, SE, DT);
2130 Kind = LSRUse::ICmpZero;
2131 S = SE.getMinusSCEV(N, S);
2134 // -1 and the negations of all interesting strides (except the negation
2135 // of -1) are now also interesting.
2136 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2137 if (Factors[i] != -1)
2138 Factors.insert(-(uint64_t)Factors[i]);
2142 // Set up the initial formula for this use.
2143 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2145 LF.Offset = P.second;
2146 LSRUse &LU = Uses[LF.LUIdx];
2147 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2148 if (!LU.WidestFixupType ||
2149 SE.getTypeSizeInBits(LU.WidestFixupType) <
2150 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2151 LU.WidestFixupType = LF.OperandValToReplace->getType();
2153 // If this is the first use of this LSRUse, give it a formula.
2154 if (LU.Formulae.empty()) {
2155 InsertInitialFormula(S, LU, LF.LUIdx);
2156 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2160 DEBUG(print_fixups(dbgs()));
2163 /// InsertInitialFormula - Insert a formula for the given expression into
2164 /// the given use, separating out loop-variant portions from loop-invariant
2165 /// and loop-computable portions.
2167 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2169 F.InitialMatch(S, L, SE);
2170 bool Inserted = InsertFormula(LU, LUIdx, F);
2171 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2174 /// InsertSupplementalFormula - Insert a simple single-register formula for
2175 /// the given expression into the given use.
2177 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2178 LSRUse &LU, size_t LUIdx) {
2180 F.BaseRegs.push_back(S);
2181 F.AM.HasBaseReg = true;
2182 bool Inserted = InsertFormula(LU, LUIdx, F);
2183 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2186 /// CountRegisters - Note which registers are used by the given formula,
2187 /// updating RegUses.
2188 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2190 RegUses.CountRegister(F.ScaledReg, LUIdx);
2191 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2192 E = F.BaseRegs.end(); I != E; ++I)
2193 RegUses.CountRegister(*I, LUIdx);
2196 /// InsertFormula - If the given formula has not yet been inserted, add it to
2197 /// the list, and return true. Return false otherwise.
2198 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2199 if (!LU.InsertFormula(F))
2202 CountRegisters(F, LUIdx);
2206 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2207 /// loop-invariant values which we're tracking. These other uses will pin these
2208 /// values in registers, making them less profitable for elimination.
2209 /// TODO: This currently misses non-constant addrec step registers.
2210 /// TODO: Should this give more weight to users inside the loop?
2212 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2213 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2214 SmallPtrSet<const SCEV *, 8> Inserted;
2216 while (!Worklist.empty()) {
2217 const SCEV *S = Worklist.pop_back_val();
2219 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2220 Worklist.append(N->op_begin(), N->op_end());
2221 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2222 Worklist.push_back(C->getOperand());
2223 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2224 Worklist.push_back(D->getLHS());
2225 Worklist.push_back(D->getRHS());
2226 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2227 if (!Inserted.insert(U)) continue;
2228 const Value *V = U->getValue();
2229 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2230 // Look for instructions defined outside the loop.
2231 if (L->contains(Inst)) continue;
2232 } else if (isa<UndefValue>(V))
2233 // Undef doesn't have a live range, so it doesn't matter.
2235 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2237 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2238 // Ignore non-instructions.
2241 // Ignore instructions in other functions (as can happen with
2243 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2245 // Ignore instructions not dominated by the loop.
2246 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2247 UserInst->getParent() :
2248 cast<PHINode>(UserInst)->getIncomingBlock(
2249 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2250 if (!DT.dominates(L->getHeader(), UseBB))
2252 // Ignore uses which are part of other SCEV expressions, to avoid
2253 // analyzing them multiple times.
2254 if (SE.isSCEVable(UserInst->getType())) {
2255 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2256 // If the user is a no-op, look through to its uses.
2257 if (!isa<SCEVUnknown>(UserS))
2261 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2265 // Ignore icmp instructions which are already being analyzed.
2266 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2267 unsigned OtherIdx = !UI.getOperandNo();
2268 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2269 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2273 LSRFixup &LF = getNewFixup();
2274 LF.UserInst = const_cast<Instruction *>(UserInst);
2275 LF.OperandValToReplace = UI.getUse();
2276 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2278 LF.Offset = P.second;
2279 LSRUse &LU = Uses[LF.LUIdx];
2280 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2281 if (!LU.WidestFixupType ||
2282 SE.getTypeSizeInBits(LU.WidestFixupType) <
2283 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2284 LU.WidestFixupType = LF.OperandValToReplace->getType();
2285 InsertSupplementalFormula(U, LU, LF.LUIdx);
2286 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2293 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2294 /// separate registers. If C is non-null, multiply each subexpression by C.
2295 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2296 SmallVectorImpl<const SCEV *> &Ops,
2298 ScalarEvolution &SE) {
2299 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2300 // Break out add operands.
2301 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2303 CollectSubexprs(*I, C, Ops, L, SE);
2305 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2306 // Split a non-zero base out of an addrec.
2307 if (!AR->getStart()->isZero()) {
2308 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2309 AR->getStepRecurrence(SE),
2311 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2314 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2317 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2318 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2319 if (Mul->getNumOperands() == 2)
2320 if (const SCEVConstant *Op0 =
2321 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2322 CollectSubexprs(Mul->getOperand(1),
2323 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2329 // Otherwise use the value itself, optionally with a scale applied.
2330 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2333 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2335 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2338 // Arbitrarily cap recursion to protect compile time.
2339 if (Depth >= 3) return;
2341 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2342 const SCEV *BaseReg = Base.BaseRegs[i];
2344 SmallVector<const SCEV *, 8> AddOps;
2345 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2347 if (AddOps.size() == 1) continue;
2349 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2350 JE = AddOps.end(); J != JE; ++J) {
2352 // Loop-variant "unknown" values are uninteresting; we won't be able to
2353 // do anything meaningful with them.
2354 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2357 // Don't pull a constant into a register if the constant could be folded
2358 // into an immediate field.
2359 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2360 Base.getNumRegs() > 1,
2361 LU.Kind, LU.AccessTy, TLI, SE))
2364 // Collect all operands except *J.
2365 SmallVector<const SCEV *, 8> InnerAddOps
2366 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2368 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2370 // Don't leave just a constant behind in a register if the constant could
2371 // be folded into an immediate field.
2372 if (InnerAddOps.size() == 1 &&
2373 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2374 Base.getNumRegs() > 1,
2375 LU.Kind, LU.AccessTy, TLI, SE))
2378 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2379 if (InnerSum->isZero())
2383 // Add the remaining pieces of the add back into the new formula.
2384 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
2385 if (TLI && InnerSumSC &&
2386 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
2387 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2388 InnerSumSC->getValue()->getZExtValue())) {
2389 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2390 InnerSumSC->getValue()->getZExtValue();
2391 F.BaseRegs.erase(F.BaseRegs.begin() + i);
2393 F.BaseRegs[i] = InnerSum;
2395 // Add J as its own register, or an unfolded immediate.
2396 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
2397 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
2398 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2399 SC->getValue()->getZExtValue()))
2400 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2401 SC->getValue()->getZExtValue();
2403 F.BaseRegs.push_back(*J);
2405 if (InsertFormula(LU, LUIdx, F))
2406 // If that formula hadn't been seen before, recurse to find more like
2408 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2413 /// GenerateCombinations - Generate a formula consisting of all of the
2414 /// loop-dominating registers added into a single register.
2415 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2417 // This method is only interesting on a plurality of registers.
2418 if (Base.BaseRegs.size() <= 1) return;
2422 SmallVector<const SCEV *, 4> Ops;
2423 for (SmallVectorImpl<const SCEV *>::const_iterator
2424 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2425 const SCEV *BaseReg = *I;
2426 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2427 !SE.hasComputableLoopEvolution(BaseReg, L))
2428 Ops.push_back(BaseReg);
2430 F.BaseRegs.push_back(BaseReg);
2432 if (Ops.size() > 1) {
2433 const SCEV *Sum = SE.getAddExpr(Ops);
2434 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2435 // opportunity to fold something. For now, just ignore such cases
2436 // rather than proceed with zero in a register.
2437 if (!Sum->isZero()) {
2438 F.BaseRegs.push_back(Sum);
2439 (void)InsertFormula(LU, LUIdx, F);
2444 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2445 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2447 // We can't add a symbolic offset if the address already contains one.
2448 if (Base.AM.BaseGV) return;
2450 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2451 const SCEV *G = Base.BaseRegs[i];
2452 GlobalValue *GV = ExtractSymbol(G, SE);
2453 if (G->isZero() || !GV)
2457 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2458 LU.Kind, LU.AccessTy, TLI))
2461 (void)InsertFormula(LU, LUIdx, F);
2465 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2466 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2468 // TODO: For now, just add the min and max offset, because it usually isn't
2469 // worthwhile looking at everything inbetween.
2470 SmallVector<int64_t, 2> Worklist;
2471 Worklist.push_back(LU.MinOffset);
2472 if (LU.MaxOffset != LU.MinOffset)
2473 Worklist.push_back(LU.MaxOffset);
2475 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2476 const SCEV *G = Base.BaseRegs[i];
2478 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2479 E = Worklist.end(); I != E; ++I) {
2481 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2482 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2483 LU.Kind, LU.AccessTy, TLI)) {
2484 // Add the offset to the base register.
2485 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2486 // If it cancelled out, drop the base register, otherwise update it.
2487 if (NewG->isZero()) {
2488 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2489 F.BaseRegs.pop_back();
2491 F.BaseRegs[i] = NewG;
2493 (void)InsertFormula(LU, LUIdx, F);
2497 int64_t Imm = ExtractImmediate(G, SE);
2498 if (G->isZero() || Imm == 0)
2501 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2502 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2503 LU.Kind, LU.AccessTy, TLI))
2506 (void)InsertFormula(LU, LUIdx, F);
2510 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2511 /// the comparison. For example, x == y -> x*c == y*c.
2512 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2514 if (LU.Kind != LSRUse::ICmpZero) return;
2516 // Determine the integer type for the base formula.
2517 Type *IntTy = Base.getType();
2519 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2521 // Don't do this if there is more than one offset.
2522 if (LU.MinOffset != LU.MaxOffset) return;
2524 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2526 // Check each interesting stride.
2527 for (SmallSetVector<int64_t, 8>::const_iterator
2528 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2529 int64_t Factor = *I;
2531 // Check that the multiplication doesn't overflow.
2532 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2534 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2535 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2538 // Check that multiplying with the use offset doesn't overflow.
2539 int64_t Offset = LU.MinOffset;
2540 if (Offset == INT64_MIN && Factor == -1)
2542 Offset = (uint64_t)Offset * Factor;
2543 if (Offset / Factor != LU.MinOffset)
2547 F.AM.BaseOffs = NewBaseOffs;
2549 // Check that this scale is legal.
2550 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2553 // Compensate for the use having MinOffset built into it.
2554 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2556 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2558 // Check that multiplying with each base register doesn't overflow.
2559 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2560 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2561 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2565 // Check that multiplying with the scaled register doesn't overflow.
2567 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2568 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2572 // Check that multiplying with the unfolded offset doesn't overflow.
2573 if (F.UnfoldedOffset != 0) {
2574 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
2576 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
2577 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
2581 // If we make it here and it's legal, add it.
2582 (void)InsertFormula(LU, LUIdx, F);
2587 /// GenerateScales - Generate stride factor reuse formulae by making use of
2588 /// scaled-offset address modes, for example.
2589 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2590 // Determine the integer type for the base formula.
2591 Type *IntTy = Base.getType();
2594 // If this Formula already has a scaled register, we can't add another one.
2595 if (Base.AM.Scale != 0) return;
2597 // Check each interesting stride.
2598 for (SmallSetVector<int64_t, 8>::const_iterator
2599 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2600 int64_t Factor = *I;
2602 Base.AM.Scale = Factor;
2603 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2604 // Check whether this scale is going to be legal.
2605 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2606 LU.Kind, LU.AccessTy, TLI)) {
2607 // As a special-case, handle special out-of-loop Basic users specially.
2608 // TODO: Reconsider this special case.
2609 if (LU.Kind == LSRUse::Basic &&
2610 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2611 LSRUse::Special, LU.AccessTy, TLI) &&
2612 LU.AllFixupsOutsideLoop)
2613 LU.Kind = LSRUse::Special;
2617 // For an ICmpZero, negating a solitary base register won't lead to
2619 if (LU.Kind == LSRUse::ICmpZero &&
2620 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2622 // For each addrec base reg, apply the scale, if possible.
2623 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2624 if (const SCEVAddRecExpr *AR =
2625 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2626 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2627 if (FactorS->isZero())
2629 // Divide out the factor, ignoring high bits, since we'll be
2630 // scaling the value back up in the end.
2631 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2632 // TODO: This could be optimized to avoid all the copying.
2634 F.ScaledReg = Quotient;
2635 F.DeleteBaseReg(F.BaseRegs[i]);
2636 (void)InsertFormula(LU, LUIdx, F);
2642 /// GenerateTruncates - Generate reuse formulae from different IV types.
2643 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2644 // This requires TargetLowering to tell us which truncates are free.
2647 // Don't bother truncating symbolic values.
2648 if (Base.AM.BaseGV) return;
2650 // Determine the integer type for the base formula.
2651 Type *DstTy = Base.getType();
2653 DstTy = SE.getEffectiveSCEVType(DstTy);
2655 for (SmallSetVector<Type *, 4>::const_iterator
2656 I = Types.begin(), E = Types.end(); I != E; ++I) {
2658 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2661 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2662 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2663 JE = F.BaseRegs.end(); J != JE; ++J)
2664 *J = SE.getAnyExtendExpr(*J, SrcTy);
2666 // TODO: This assumes we've done basic processing on all uses and
2667 // have an idea what the register usage is.
2668 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2671 (void)InsertFormula(LU, LUIdx, F);
2678 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2679 /// defer modifications so that the search phase doesn't have to worry about
2680 /// the data structures moving underneath it.
2684 const SCEV *OrigReg;
2686 WorkItem(size_t LI, int64_t I, const SCEV *R)
2687 : LUIdx(LI), Imm(I), OrigReg(R) {}
2689 void print(raw_ostream &OS) const;
2695 void WorkItem::print(raw_ostream &OS) const {
2696 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2697 << " , add offset " << Imm;
2700 void WorkItem::dump() const {
2701 print(errs()); errs() << '\n';
2704 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2705 /// distance apart and try to form reuse opportunities between them.
2706 void LSRInstance::GenerateCrossUseConstantOffsets() {
2707 // Group the registers by their value without any added constant offset.
2708 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2709 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2711 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2712 SmallVector<const SCEV *, 8> Sequence;
2713 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2715 const SCEV *Reg = *I;
2716 int64_t Imm = ExtractImmediate(Reg, SE);
2717 std::pair<RegMapTy::iterator, bool> Pair =
2718 Map.insert(std::make_pair(Reg, ImmMapTy()));
2720 Sequence.push_back(Reg);
2721 Pair.first->second.insert(std::make_pair(Imm, *I));
2722 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2725 // Now examine each set of registers with the same base value. Build up
2726 // a list of work to do and do the work in a separate step so that we're
2727 // not adding formulae and register counts while we're searching.
2728 SmallVector<WorkItem, 32> WorkItems;
2729 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2730 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2731 E = Sequence.end(); I != E; ++I) {
2732 const SCEV *Reg = *I;
2733 const ImmMapTy &Imms = Map.find(Reg)->second;
2735 // It's not worthwhile looking for reuse if there's only one offset.
2736 if (Imms.size() == 1)
2739 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2740 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2742 dbgs() << ' ' << J->first;
2745 // Examine each offset.
2746 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2748 const SCEV *OrigReg = J->second;
2750 int64_t JImm = J->first;
2751 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2753 if (!isa<SCEVConstant>(OrigReg) &&
2754 UsedByIndicesMap[Reg].count() == 1) {
2755 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2759 // Conservatively examine offsets between this orig reg a few selected
2761 ImmMapTy::const_iterator OtherImms[] = {
2762 Imms.begin(), prior(Imms.end()),
2763 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2765 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2766 ImmMapTy::const_iterator M = OtherImms[i];
2767 if (M == J || M == JE) continue;
2769 // Compute the difference between the two.
2770 int64_t Imm = (uint64_t)JImm - M->first;
2771 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2772 LUIdx = UsedByIndices.find_next(LUIdx))
2773 // Make a memo of this use, offset, and register tuple.
2774 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2775 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2782 UsedByIndicesMap.clear();
2783 UniqueItems.clear();
2785 // Now iterate through the worklist and add new formulae.
2786 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2787 E = WorkItems.end(); I != E; ++I) {
2788 const WorkItem &WI = *I;
2789 size_t LUIdx = WI.LUIdx;
2790 LSRUse &LU = Uses[LUIdx];
2791 int64_t Imm = WI.Imm;
2792 const SCEV *OrigReg = WI.OrigReg;
2794 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2795 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2796 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2798 // TODO: Use a more targeted data structure.
2799 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2800 const Formula &F = LU.Formulae[L];
2801 // Use the immediate in the scaled register.
2802 if (F.ScaledReg == OrigReg) {
2803 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2804 Imm * (uint64_t)F.AM.Scale;
2805 // Don't create 50 + reg(-50).
2806 if (F.referencesReg(SE.getSCEV(
2807 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2810 NewF.AM.BaseOffs = Offs;
2811 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2812 LU.Kind, LU.AccessTy, TLI))
2814 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2816 // If the new scale is a constant in a register, and adding the constant
2817 // value to the immediate would produce a value closer to zero than the
2818 // immediate itself, then the formula isn't worthwhile.
2819 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2820 if (C->getValue()->isNegative() !=
2821 (NewF.AM.BaseOffs < 0) &&
2822 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2823 .ule(abs64(NewF.AM.BaseOffs)))
2827 (void)InsertFormula(LU, LUIdx, NewF);
2829 // Use the immediate in a base register.
2830 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2831 const SCEV *BaseReg = F.BaseRegs[N];
2832 if (BaseReg != OrigReg)
2835 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2836 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2837 LU.Kind, LU.AccessTy, TLI)) {
2839 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
2842 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
2844 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2846 // If the new formula has a constant in a register, and adding the
2847 // constant value to the immediate would produce a value closer to
2848 // zero than the immediate itself, then the formula isn't worthwhile.
2849 for (SmallVectorImpl<const SCEV *>::const_iterator
2850 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2852 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2853 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2854 abs64(NewF.AM.BaseOffs)) &&
2855 (C->getValue()->getValue() +
2856 NewF.AM.BaseOffs).countTrailingZeros() >=
2857 CountTrailingZeros_64(NewF.AM.BaseOffs))
2861 (void)InsertFormula(LU, LUIdx, NewF);
2870 /// GenerateAllReuseFormulae - Generate formulae for each use.
2872 LSRInstance::GenerateAllReuseFormulae() {
2873 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2874 // queries are more precise.
2875 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2876 LSRUse &LU = Uses[LUIdx];
2877 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2878 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2879 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2880 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2882 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2883 LSRUse &LU = Uses[LUIdx];
2884 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2885 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2886 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2887 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2888 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2889 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2890 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2891 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2893 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2894 LSRUse &LU = Uses[LUIdx];
2895 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2896 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2899 GenerateCrossUseConstantOffsets();
2901 DEBUG(dbgs() << "\n"
2902 "After generating reuse formulae:\n";
2903 print_uses(dbgs()));
2906 /// If there are multiple formulae with the same set of registers used
2907 /// by other uses, pick the best one and delete the others.
2908 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2909 DenseSet<const SCEV *> VisitedRegs;
2910 SmallPtrSet<const SCEV *, 16> Regs;
2912 bool ChangedFormulae = false;
2915 // Collect the best formula for each unique set of shared registers. This
2916 // is reset for each use.
2917 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2919 BestFormulaeTy BestFormulae;
2921 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2922 LSRUse &LU = Uses[LUIdx];
2923 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2926 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2927 FIdx != NumForms; ++FIdx) {
2928 Formula &F = LU.Formulae[FIdx];
2930 SmallVector<const SCEV *, 2> Key;
2931 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2932 JE = F.BaseRegs.end(); J != JE; ++J) {
2933 const SCEV *Reg = *J;
2934 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2938 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2939 Key.push_back(F.ScaledReg);
2940 // Unstable sort by host order ok, because this is only used for
2942 std::sort(Key.begin(), Key.end());
2944 std::pair<BestFormulaeTy::const_iterator, bool> P =
2945 BestFormulae.insert(std::make_pair(Key, FIdx));
2947 Formula &Best = LU.Formulae[P.first->second];
2950 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2953 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2955 if (CostF < CostBest)
2957 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2959 " in favor of formula "; Best.print(dbgs());
2962 ChangedFormulae = true;
2964 LU.DeleteFormula(F);
2972 // Now that we've filtered out some formulae, recompute the Regs set.
2974 LU.RecomputeRegs(LUIdx, RegUses);
2976 // Reset this to prepare for the next use.
2977 BestFormulae.clear();
2980 DEBUG(if (ChangedFormulae) {
2982 "After filtering out undesirable candidates:\n";
2987 // This is a rough guess that seems to work fairly well.
2988 static const size_t ComplexityLimit = UINT16_MAX;
2990 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2991 /// solutions the solver might have to consider. It almost never considers
2992 /// this many solutions because it prune the search space, but the pruning
2993 /// isn't always sufficient.
2994 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2996 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2997 E = Uses.end(); I != E; ++I) {
2998 size_t FSize = I->Formulae.size();
2999 if (FSize >= ComplexityLimit) {
3000 Power = ComplexityLimit;
3004 if (Power >= ComplexityLimit)
3010 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3011 /// of the registers of another formula, it won't help reduce register
3012 /// pressure (though it may not necessarily hurt register pressure); remove
3013 /// it to simplify the system.
3014 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3015 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3016 DEBUG(dbgs() << "The search space is too complex.\n");
3018 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3019 "which use a superset of registers used by other "
3022 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3023 LSRUse &LU = Uses[LUIdx];
3025 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3026 Formula &F = LU.Formulae[i];
3027 // Look for a formula with a constant or GV in a register. If the use
3028 // also has a formula with that same value in an immediate field,
3029 // delete the one that uses a register.
3030 for (SmallVectorImpl<const SCEV *>::const_iterator
3031 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3032 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3034 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3035 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3036 (I - F.BaseRegs.begin()));
3037 if (LU.HasFormulaWithSameRegs(NewF)) {
3038 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3039 LU.DeleteFormula(F);
3045 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3046 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3049 NewF.AM.BaseGV = GV;
3050 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3051 (I - F.BaseRegs.begin()));
3052 if (LU.HasFormulaWithSameRegs(NewF)) {
3053 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3055 LU.DeleteFormula(F);
3066 LU.RecomputeRegs(LUIdx, RegUses);
3069 DEBUG(dbgs() << "After pre-selection:\n";
3070 print_uses(dbgs()));
3074 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3075 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3077 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3078 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3079 DEBUG(dbgs() << "The search space is too complex.\n");
3081 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3082 "separated by a constant offset will use the same "
3085 // This is especially useful for unrolled loops.
3087 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3088 LSRUse &LU = Uses[LUIdx];
3089 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3090 E = LU.Formulae.end(); I != E; ++I) {
3091 const Formula &F = *I;
3092 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3093 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3094 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3095 /*HasBaseReg=*/false,
3096 LU.Kind, LU.AccessTy)) {
3097 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3100 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3102 // Update the relocs to reference the new use.
3103 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3104 E = Fixups.end(); I != E; ++I) {
3105 LSRFixup &Fixup = *I;
3106 if (Fixup.LUIdx == LUIdx) {
3107 Fixup.LUIdx = LUThatHas - &Uses.front();
3108 Fixup.Offset += F.AM.BaseOffs;
3109 // Add the new offset to LUThatHas' offset list.
3110 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3111 LUThatHas->Offsets.push_back(Fixup.Offset);
3112 if (Fixup.Offset > LUThatHas->MaxOffset)
3113 LUThatHas->MaxOffset = Fixup.Offset;
3114 if (Fixup.Offset < LUThatHas->MinOffset)
3115 LUThatHas->MinOffset = Fixup.Offset;
3117 DEBUG(dbgs() << "New fixup has offset "
3118 << Fixup.Offset << '\n');
3120 if (Fixup.LUIdx == NumUses-1)
3121 Fixup.LUIdx = LUIdx;
3124 // Delete formulae from the new use which are no longer legal.
3126 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3127 Formula &F = LUThatHas->Formulae[i];
3128 if (!isLegalUse(F.AM,
3129 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3130 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3131 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3133 LUThatHas->DeleteFormula(F);
3140 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3142 // Delete the old use.
3143 DeleteUse(LU, LUIdx);
3153 DEBUG(dbgs() << "After pre-selection:\n";
3154 print_uses(dbgs()));
3158 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3159 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3160 /// we've done more filtering, as it may be able to find more formulae to
3162 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3163 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3164 DEBUG(dbgs() << "The search space is too complex.\n");
3166 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3167 "undesirable dedicated registers.\n");
3169 FilterOutUndesirableDedicatedRegisters();
3171 DEBUG(dbgs() << "After pre-selection:\n";
3172 print_uses(dbgs()));
3176 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3177 /// to be profitable, and then in any use which has any reference to that
3178 /// register, delete all formulae which do not reference that register.
3179 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3180 // With all other options exhausted, loop until the system is simple
3181 // enough to handle.
3182 SmallPtrSet<const SCEV *, 4> Taken;
3183 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3184 // Ok, we have too many of formulae on our hands to conveniently handle.
3185 // Use a rough heuristic to thin out the list.
3186 DEBUG(dbgs() << "The search space is too complex.\n");
3188 // Pick the register which is used by the most LSRUses, which is likely
3189 // to be a good reuse register candidate.
3190 const SCEV *Best = 0;
3191 unsigned BestNum = 0;
3192 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3194 const SCEV *Reg = *I;
3195 if (Taken.count(Reg))
3200 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3201 if (Count > BestNum) {
3208 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3209 << " will yield profitable reuse.\n");
3212 // In any use with formulae which references this register, delete formulae
3213 // which don't reference it.
3214 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3215 LSRUse &LU = Uses[LUIdx];
3216 if (!LU.Regs.count(Best)) continue;
3219 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3220 Formula &F = LU.Formulae[i];
3221 if (!F.referencesReg(Best)) {
3222 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3223 LU.DeleteFormula(F);
3227 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3233 LU.RecomputeRegs(LUIdx, RegUses);
3236 DEBUG(dbgs() << "After pre-selection:\n";
3237 print_uses(dbgs()));
3241 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3242 /// formulae to choose from, use some rough heuristics to prune down the number
3243 /// of formulae. This keeps the main solver from taking an extraordinary amount
3244 /// of time in some worst-case scenarios.
3245 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3246 NarrowSearchSpaceByDetectingSupersets();
3247 NarrowSearchSpaceByCollapsingUnrolledCode();
3248 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3249 NarrowSearchSpaceByPickingWinnerRegs();
3252 /// SolveRecurse - This is the recursive solver.
3253 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3255 SmallVectorImpl<const Formula *> &Workspace,
3256 const Cost &CurCost,
3257 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3258 DenseSet<const SCEV *> &VisitedRegs) const {
3261 // - use more aggressive filtering
3262 // - sort the formula so that the most profitable solutions are found first
3263 // - sort the uses too
3265 // - don't compute a cost, and then compare. compare while computing a cost
3267 // - track register sets with SmallBitVector
3269 const LSRUse &LU = Uses[Workspace.size()];
3271 // If this use references any register that's already a part of the
3272 // in-progress solution, consider it a requirement that a formula must
3273 // reference that register in order to be considered. This prunes out
3274 // unprofitable searching.
3275 SmallSetVector<const SCEV *, 4> ReqRegs;
3276 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3277 E = CurRegs.end(); I != E; ++I)
3278 if (LU.Regs.count(*I))
3281 bool AnySatisfiedReqRegs = false;
3282 SmallPtrSet<const SCEV *, 16> NewRegs;
3285 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3286 E = LU.Formulae.end(); I != E; ++I) {
3287 const Formula &F = *I;
3289 // Ignore formulae which do not use any of the required registers.
3290 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3291 JE = ReqRegs.end(); J != JE; ++J) {
3292 const SCEV *Reg = *J;
3293 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3294 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3298 AnySatisfiedReqRegs = true;
3300 // Evaluate the cost of the current formula. If it's already worse than
3301 // the current best, prune the search at that point.
3304 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3305 if (NewCost < SolutionCost) {
3306 Workspace.push_back(&F);
3307 if (Workspace.size() != Uses.size()) {
3308 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3309 NewRegs, VisitedRegs);
3310 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3311 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3313 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3314 dbgs() << ". Regs:";
3315 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3316 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3317 dbgs() << ' ' << **I;
3320 SolutionCost = NewCost;
3321 Solution = Workspace;
3323 Workspace.pop_back();
3328 if (!EnableRetry && !AnySatisfiedReqRegs)
3331 // If none of the formulae had all of the required registers, relax the
3332 // constraint so that we don't exclude all formulae.
3333 if (!AnySatisfiedReqRegs) {
3334 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3340 /// Solve - Choose one formula from each use. Return the results in the given
3341 /// Solution vector.
3342 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3343 SmallVector<const Formula *, 8> Workspace;
3345 SolutionCost.Loose();
3347 SmallPtrSet<const SCEV *, 16> CurRegs;
3348 DenseSet<const SCEV *> VisitedRegs;
3349 Workspace.reserve(Uses.size());
3351 // SolveRecurse does all the work.
3352 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3353 CurRegs, VisitedRegs);
3354 if (Solution.empty()) {
3355 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
3359 // Ok, we've now made all our decisions.
3360 DEBUG(dbgs() << "\n"
3361 "The chosen solution requires "; SolutionCost.print(dbgs());
3363 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3365 Uses[i].print(dbgs());
3368 Solution[i]->print(dbgs());
3372 assert(Solution.size() == Uses.size() && "Malformed solution!");
3375 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3376 /// the dominator tree far as we can go while still being dominated by the
3377 /// input positions. This helps canonicalize the insert position, which
3378 /// encourages sharing.
3379 BasicBlock::iterator
3380 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3381 const SmallVectorImpl<Instruction *> &Inputs)
3384 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3385 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3388 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3389 if (!Rung) return IP;
3390 Rung = Rung->getIDom();
3391 if (!Rung) return IP;
3392 IDom = Rung->getBlock();
3394 // Don't climb into a loop though.
3395 const Loop *IDomLoop = LI.getLoopFor(IDom);
3396 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3397 if (IDomDepth <= IPLoopDepth &&
3398 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3402 bool AllDominate = true;
3403 Instruction *BetterPos = 0;
3404 Instruction *Tentative = IDom->getTerminator();
3405 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3406 E = Inputs.end(); I != E; ++I) {
3407 Instruction *Inst = *I;
3408 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3409 AllDominate = false;
3412 // Attempt to find an insert position in the middle of the block,
3413 // instead of at the end, so that it can be used for other expansions.
3414 if (IDom == Inst->getParent() &&
3415 (!BetterPos || DT.dominates(BetterPos, Inst)))
3416 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3429 /// AdjustInsertPositionForExpand - Determine an input position which will be
3430 /// dominated by the operands and which will dominate the result.
3431 BasicBlock::iterator
3432 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3434 const LSRUse &LU) const {
3435 // Collect some instructions which must be dominated by the
3436 // expanding replacement. These must be dominated by any operands that
3437 // will be required in the expansion.
3438 SmallVector<Instruction *, 4> Inputs;
3439 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3440 Inputs.push_back(I);
3441 if (LU.Kind == LSRUse::ICmpZero)
3442 if (Instruction *I =
3443 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3444 Inputs.push_back(I);
3445 if (LF.PostIncLoops.count(L)) {
3446 if (LF.isUseFullyOutsideLoop(L))
3447 Inputs.push_back(L->getLoopLatch()->getTerminator());
3449 Inputs.push_back(IVIncInsertPos);
3451 // The expansion must also be dominated by the increment positions of any
3452 // loops it for which it is using post-inc mode.
3453 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3454 E = LF.PostIncLoops.end(); I != E; ++I) {
3455 const Loop *PIL = *I;
3456 if (PIL == L) continue;
3458 // Be dominated by the loop exit.
3459 SmallVector<BasicBlock *, 4> ExitingBlocks;
3460 PIL->getExitingBlocks(ExitingBlocks);
3461 if (!ExitingBlocks.empty()) {
3462 BasicBlock *BB = ExitingBlocks[0];
3463 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3464 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3465 Inputs.push_back(BB->getTerminator());
3469 // Then, climb up the immediate dominator tree as far as we can go while
3470 // still being dominated by the input positions.
3471 IP = HoistInsertPosition(IP, Inputs);
3473 // Don't insert instructions before PHI nodes.
3474 while (isa<PHINode>(IP)) ++IP;
3476 // Ignore landingpad instructions.
3477 while (isa<LandingPadInst>(IP)) ++IP;
3479 // Ignore debug intrinsics.
3480 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3485 /// Expand - Emit instructions for the leading candidate expression for this
3486 /// LSRUse (this is called "expanding").
3487 Value *LSRInstance::Expand(const LSRFixup &LF,
3489 BasicBlock::iterator IP,
3490 SCEVExpander &Rewriter,
3491 SmallVectorImpl<WeakVH> &DeadInsts) const {
3492 const LSRUse &LU = Uses[LF.LUIdx];
3494 // Determine an input position which will be dominated by the operands and
3495 // which will dominate the result.
3496 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3498 // Inform the Rewriter if we have a post-increment use, so that it can
3499 // perform an advantageous expansion.
3500 Rewriter.setPostInc(LF.PostIncLoops);
3502 // This is the type that the user actually needs.
3503 Type *OpTy = LF.OperandValToReplace->getType();
3504 // This will be the type that we'll initially expand to.
3505 Type *Ty = F.getType();
3507 // No type known; just expand directly to the ultimate type.
3509 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3510 // Expand directly to the ultimate type if it's the right size.
3512 // This is the type to do integer arithmetic in.
3513 Type *IntTy = SE.getEffectiveSCEVType(Ty);
3515 // Build up a list of operands to add together to form the full base.
3516 SmallVector<const SCEV *, 8> Ops;
3518 // Expand the BaseRegs portion.
3519 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3520 E = F.BaseRegs.end(); I != E; ++I) {
3521 const SCEV *Reg = *I;
3522 assert(!Reg->isZero() && "Zero allocated in a base register!");
3524 // If we're expanding for a post-inc user, make the post-inc adjustment.
3525 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3526 Reg = TransformForPostIncUse(Denormalize, Reg,
3527 LF.UserInst, LF.OperandValToReplace,
3530 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3533 // Flush the operand list to suppress SCEVExpander hoisting.
3535 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3537 Ops.push_back(SE.getUnknown(FullV));
3540 // Expand the ScaledReg portion.
3541 Value *ICmpScaledV = 0;
3542 if (F.AM.Scale != 0) {
3543 const SCEV *ScaledS = F.ScaledReg;
3545 // If we're expanding for a post-inc user, make the post-inc adjustment.
3546 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3547 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3548 LF.UserInst, LF.OperandValToReplace,
3551 if (LU.Kind == LSRUse::ICmpZero) {
3552 // An interesting way of "folding" with an icmp is to use a negated
3553 // scale, which we'll implement by inserting it into the other operand
3555 assert(F.AM.Scale == -1 &&
3556 "The only scale supported by ICmpZero uses is -1!");
3557 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3559 // Otherwise just expand the scaled register and an explicit scale,
3560 // which is expected to be matched as part of the address.
3561 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3562 ScaledS = SE.getMulExpr(ScaledS,
3563 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3564 Ops.push_back(ScaledS);
3566 // Flush the operand list to suppress SCEVExpander hoisting.
3567 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3569 Ops.push_back(SE.getUnknown(FullV));
3573 // Expand the GV portion.
3575 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3577 // Flush the operand list to suppress SCEVExpander hoisting.
3578 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3580 Ops.push_back(SE.getUnknown(FullV));
3583 // Expand the immediate portion.
3584 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3586 if (LU.Kind == LSRUse::ICmpZero) {
3587 // The other interesting way of "folding" with an ICmpZero is to use a
3588 // negated immediate.
3590 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3592 Ops.push_back(SE.getUnknown(ICmpScaledV));
3593 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3596 // Just add the immediate values. These again are expected to be matched
3597 // as part of the address.
3598 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3602 // Expand the unfolded offset portion.
3603 int64_t UnfoldedOffset = F.UnfoldedOffset;
3604 if (UnfoldedOffset != 0) {
3605 // Just add the immediate values.
3606 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
3610 // Emit instructions summing all the operands.
3611 const SCEV *FullS = Ops.empty() ?
3612 SE.getConstant(IntTy, 0) :
3614 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3616 // We're done expanding now, so reset the rewriter.
3617 Rewriter.clearPostInc();
3619 // An ICmpZero Formula represents an ICmp which we're handling as a
3620 // comparison against zero. Now that we've expanded an expression for that
3621 // form, update the ICmp's other operand.
3622 if (LU.Kind == LSRUse::ICmpZero) {
3623 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3624 DeadInsts.push_back(CI->getOperand(1));
3625 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3626 "a scale at the same time!");
3627 if (F.AM.Scale == -1) {
3628 if (ICmpScaledV->getType() != OpTy) {
3630 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3632 ICmpScaledV, OpTy, "tmp", CI);
3635 CI->setOperand(1, ICmpScaledV);
3637 assert(F.AM.Scale == 0 &&
3638 "ICmp does not support folding a global value and "
3639 "a scale at the same time!");
3640 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3642 if (C->getType() != OpTy)
3643 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3647 CI->setOperand(1, C);
3654 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3655 /// of their operands effectively happens in their predecessor blocks, so the
3656 /// expression may need to be expanded in multiple places.
3657 void LSRInstance::RewriteForPHI(PHINode *PN,
3660 SCEVExpander &Rewriter,
3661 SmallVectorImpl<WeakVH> &DeadInsts,
3663 DenseMap<BasicBlock *, Value *> Inserted;
3664 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3665 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3666 BasicBlock *BB = PN->getIncomingBlock(i);
3668 // If this is a critical edge, split the edge so that we do not insert
3669 // the code on all predecessor/successor paths. We do this unless this
3670 // is the canonical backedge for this loop, which complicates post-inc
3672 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3673 !isa<IndirectBrInst>(BB->getTerminator())) {
3674 BasicBlock *Parent = PN->getParent();
3675 Loop *PNLoop = LI.getLoopFor(Parent);
3676 if (!PNLoop || Parent != PNLoop->getHeader()) {
3677 // Split the critical edge.
3678 BasicBlock *NewBB = 0;
3679 if (!Parent->isLandingPad()) {
3680 NewBB = SplitCriticalEdge(BB, Parent, P,
3681 /*MergeIdenticalEdges=*/true,
3682 /*DontDeleteUselessPhis=*/true);
3684 SmallVector<BasicBlock*, 2> NewBBs;
3685 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
3689 // If PN is outside of the loop and BB is in the loop, we want to
3690 // move the block to be immediately before the PHI block, not
3691 // immediately after BB.
3692 if (L->contains(BB) && !L->contains(PN))
3693 NewBB->moveBefore(PN->getParent());
3695 // Splitting the edge can reduce the number of PHI entries we have.
3696 e = PN->getNumIncomingValues();
3698 i = PN->getBasicBlockIndex(BB);
3702 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3703 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3705 PN->setIncomingValue(i, Pair.first->second);
3707 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3709 // If this is reuse-by-noop-cast, insert the noop cast.
3710 Type *OpTy = LF.OperandValToReplace->getType();
3711 if (FullV->getType() != OpTy)
3713 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3715 FullV, LF.OperandValToReplace->getType(),
3716 "tmp", BB->getTerminator());
3718 PN->setIncomingValue(i, FullV);
3719 Pair.first->second = FullV;
3724 /// Rewrite - Emit instructions for the leading candidate expression for this
3725 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3726 /// the newly expanded value.
3727 void LSRInstance::Rewrite(const LSRFixup &LF,
3729 SCEVExpander &Rewriter,
3730 SmallVectorImpl<WeakVH> &DeadInsts,
3732 // First, find an insertion point that dominates UserInst. For PHI nodes,
3733 // find the nearest block which dominates all the relevant uses.
3734 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3735 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3737 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3739 // If this is reuse-by-noop-cast, insert the noop cast.
3740 Type *OpTy = LF.OperandValToReplace->getType();
3741 if (FullV->getType() != OpTy) {
3743 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3744 FullV, OpTy, "tmp", LF.UserInst);
3748 // Update the user. ICmpZero is handled specially here (for now) because
3749 // Expand may have updated one of the operands of the icmp already, and
3750 // its new value may happen to be equal to LF.OperandValToReplace, in
3751 // which case doing replaceUsesOfWith leads to replacing both operands
3752 // with the same value. TODO: Reorganize this.
3753 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3754 LF.UserInst->setOperand(0, FullV);
3756 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3759 DeadInsts.push_back(LF.OperandValToReplace);
3762 /// ImplementSolution - Rewrite all the fixup locations with new values,
3763 /// following the chosen solution.
3765 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3767 // Keep track of instructions we may have made dead, so that
3768 // we can remove them after we are done working.
3769 SmallVector<WeakVH, 16> DeadInsts;
3771 SCEVExpander Rewriter(SE, "lsr");
3772 Rewriter.disableCanonicalMode();
3773 Rewriter.enableLSRMode();
3774 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3776 // Expand the new value definitions and update the users.
3777 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3778 E = Fixups.end(); I != E; ++I) {
3779 const LSRFixup &Fixup = *I;
3781 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3786 // Clean up after ourselves. This must be done before deleting any
3790 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3793 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3794 : IU(P->getAnalysis<IVUsers>()),
3795 SE(P->getAnalysis<ScalarEvolution>()),
3796 DT(P->getAnalysis<DominatorTree>()),
3797 LI(P->getAnalysis<LoopInfo>()),
3798 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3800 // If LoopSimplify form is not available, stay out of trouble.
3801 if (!L->isLoopSimplifyForm()) return;
3803 // If there's no interesting work to be done, bail early.
3804 if (IU.empty()) return;
3806 DEBUG(dbgs() << "\nLSR on loop ";
3807 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3810 // First, perform some low-level loop optimizations.
3812 OptimizeLoopTermCond();
3814 // If loop preparation eliminates all interesting IV users, bail.
3815 if (IU.empty()) return;
3817 // Skip nested loops until we can model them better with formulae.
3818 if (!EnableNested && !L->empty()) {
3819 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
3823 // Start collecting data and preparing for the solver.
3824 CollectInterestingTypesAndFactors();
3825 CollectFixupsAndInitialFormulae();
3826 CollectLoopInvariantFixupsAndFormulae();
3828 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3829 print_uses(dbgs()));
3831 // Now use the reuse data to generate a bunch of interesting ways
3832 // to formulate the values needed for the uses.
3833 GenerateAllReuseFormulae();
3835 FilterOutUndesirableDedicatedRegisters();
3836 NarrowSearchSpaceUsingHeuristics();
3838 SmallVector<const Formula *, 8> Solution;
3841 // Release memory that is no longer needed.
3846 if (Solution.empty())
3850 // Formulae should be legal.
3851 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3852 E = Uses.end(); I != E; ++I) {
3853 const LSRUse &LU = *I;
3854 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3855 JE = LU.Formulae.end(); J != JE; ++J)
3856 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3857 LU.Kind, LU.AccessTy, TLI) &&
3858 "Illegal formula generated!");
3862 // Now that we've decided what we want, make it so.
3863 ImplementSolution(Solution, P);
3866 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3867 if (Factors.empty() && Types.empty()) return;
3869 OS << "LSR has identified the following interesting factors and types: ";
3872 for (SmallSetVector<int64_t, 8>::const_iterator
3873 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3874 if (!First) OS << ", ";
3879 for (SmallSetVector<Type *, 4>::const_iterator
3880 I = Types.begin(), E = Types.end(); I != E; ++I) {
3881 if (!First) OS << ", ";
3883 OS << '(' << **I << ')';
3888 void LSRInstance::print_fixups(raw_ostream &OS) const {
3889 OS << "LSR is examining the following fixup sites:\n";
3890 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3891 E = Fixups.end(); I != E; ++I) {
3898 void LSRInstance::print_uses(raw_ostream &OS) const {
3899 OS << "LSR is examining the following uses:\n";
3900 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3901 E = Uses.end(); I != E; ++I) {
3902 const LSRUse &LU = *I;
3906 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3907 JE = LU.Formulae.end(); J != JE; ++J) {
3915 void LSRInstance::print(raw_ostream &OS) const {
3916 print_factors_and_types(OS);
3921 void LSRInstance::dump() const {
3922 print(errs()); errs() << '\n';
3927 class LoopStrengthReduce : public LoopPass {
3928 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3929 /// transformation profitability.
3930 const TargetLowering *const TLI;
3933 static char ID; // Pass ID, replacement for typeid
3934 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3937 bool runOnLoop(Loop *L, LPPassManager &LPM);
3938 void getAnalysisUsage(AnalysisUsage &AU) const;
3943 char LoopStrengthReduce::ID = 0;
3944 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
3945 "Loop Strength Reduction", false, false)
3946 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3947 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3948 INITIALIZE_PASS_DEPENDENCY(IVUsers)
3949 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
3950 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3951 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
3952 "Loop Strength Reduction", false, false)
3955 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3956 return new LoopStrengthReduce(TLI);
3959 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3960 : LoopPass(ID), TLI(tli) {
3961 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
3964 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3965 // We split critical edges, so we change the CFG. However, we do update
3966 // many analyses if they are around.
3967 AU.addPreservedID(LoopSimplifyID);
3969 AU.addRequired<LoopInfo>();
3970 AU.addPreserved<LoopInfo>();
3971 AU.addRequiredID(LoopSimplifyID);
3972 AU.addRequired<DominatorTree>();
3973 AU.addPreserved<DominatorTree>();
3974 AU.addRequired<ScalarEvolution>();
3975 AU.addPreserved<ScalarEvolution>();
3976 // Requiring LoopSimplify a second time here prevents IVUsers from running
3977 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
3978 AU.addRequiredID(LoopSimplifyID);
3979 AU.addRequired<IVUsers>();
3980 AU.addPreserved<IVUsers>();
3983 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3984 bool Changed = false;
3986 // Run the main LSR transformation.
3987 Changed |= LSRInstance(TLI, L, this).getChanged();
3989 // At this point, it is worth checking to see if any recurrence PHIs are also
3990 // dead, so that we can remove them as well.
3991 Changed |= DeleteDeadPHIs(L->getHeader());