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> EnableRetry(
82 "enable-lsr-retry", cl::Hidden, cl::desc("Enable LSR retry"));
87 /// RegSortData - This class holds data which is used to order reuse candidates.
90 /// UsedByIndices - This represents the set of LSRUse indices which reference
91 /// a particular register.
92 SmallBitVector UsedByIndices;
96 void print(raw_ostream &OS) const;
102 void RegSortData::print(raw_ostream &OS) const {
103 OS << "[NumUses=" << UsedByIndices.count() << ']';
106 void RegSortData::dump() const {
107 print(errs()); errs() << '\n';
112 /// RegUseTracker - Map register candidates to information about how they are
114 class RegUseTracker {
115 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
117 RegUsesTy RegUsesMap;
118 SmallVector<const SCEV *, 16> RegSequence;
121 void CountRegister(const SCEV *Reg, size_t LUIdx);
122 void DropRegister(const SCEV *Reg, size_t LUIdx);
123 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
125 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
127 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
131 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
132 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
133 iterator begin() { return RegSequence.begin(); }
134 iterator end() { return RegSequence.end(); }
135 const_iterator begin() const { return RegSequence.begin(); }
136 const_iterator end() const { return RegSequence.end(); }
142 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
143 std::pair<RegUsesTy::iterator, bool> Pair =
144 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
145 RegSortData &RSD = Pair.first->second;
147 RegSequence.push_back(Reg);
148 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
149 RSD.UsedByIndices.set(LUIdx);
153 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
154 RegUsesTy::iterator It = RegUsesMap.find(Reg);
155 assert(It != RegUsesMap.end());
156 RegSortData &RSD = It->second;
157 assert(RSD.UsedByIndices.size() > LUIdx);
158 RSD.UsedByIndices.reset(LUIdx);
162 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
163 assert(LUIdx <= LastLUIdx);
165 // Update RegUses. The data structure is not optimized for this purpose;
166 // we must iterate through it and update each of the bit vectors.
167 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
169 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
170 if (LUIdx < UsedByIndices.size())
171 UsedByIndices[LUIdx] =
172 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
173 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
178 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
179 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
180 if (I == RegUsesMap.end())
182 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
183 int i = UsedByIndices.find_first();
184 if (i == -1) return false;
185 if ((size_t)i != LUIdx) return true;
186 return UsedByIndices.find_next(i) != -1;
189 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
190 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
191 assert(I != RegUsesMap.end() && "Unknown register!");
192 return I->second.UsedByIndices;
195 void RegUseTracker::clear() {
202 /// Formula - This class holds information that describes a formula for
203 /// computing satisfying a use. It may include broken-out immediates and scaled
206 /// AM - This is used to represent complex addressing, as well as other kinds
207 /// of interesting uses.
208 TargetLowering::AddrMode AM;
210 /// BaseRegs - The list of "base" registers for this use. When this is
211 /// non-empty, AM.HasBaseReg should be set to true.
212 SmallVector<const SCEV *, 2> BaseRegs;
214 /// ScaledReg - The 'scaled' register for this use. This should be non-null
215 /// when AM.Scale is not zero.
216 const SCEV *ScaledReg;
218 /// UnfoldedOffset - An additional constant offset which added near the
219 /// use. This requires a temporary register, but the offset itself can
220 /// live in an add immediate field rather than a register.
221 int64_t UnfoldedOffset;
223 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
225 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
227 unsigned getNumRegs() const;
228 Type *getType() const;
230 void DeleteBaseReg(const SCEV *&S);
232 bool referencesReg(const SCEV *S) const;
233 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
234 const RegUseTracker &RegUses) const;
236 void print(raw_ostream &OS) const;
242 /// DoInitialMatch - Recursion helper for InitialMatch.
243 static void DoInitialMatch(const SCEV *S, Loop *L,
244 SmallVectorImpl<const SCEV *> &Good,
245 SmallVectorImpl<const SCEV *> &Bad,
246 ScalarEvolution &SE) {
247 // Collect expressions which properly dominate the loop header.
248 if (SE.properlyDominates(S, L->getHeader())) {
253 // Look at add operands.
254 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
255 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
257 DoInitialMatch(*I, L, Good, Bad, SE);
261 // Look at addrec operands.
262 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
263 if (!AR->getStart()->isZero()) {
264 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
265 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
266 AR->getStepRecurrence(SE),
267 // FIXME: AR->getNoWrapFlags()
268 AR->getLoop(), SCEV::FlagAnyWrap),
273 // Handle a multiplication by -1 (negation) if it didn't fold.
274 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
275 if (Mul->getOperand(0)->isAllOnesValue()) {
276 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
277 const SCEV *NewMul = SE.getMulExpr(Ops);
279 SmallVector<const SCEV *, 4> MyGood;
280 SmallVector<const SCEV *, 4> MyBad;
281 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
282 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
283 SE.getEffectiveSCEVType(NewMul->getType())));
284 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
285 E = MyGood.end(); I != E; ++I)
286 Good.push_back(SE.getMulExpr(NegOne, *I));
287 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
288 E = MyBad.end(); I != E; ++I)
289 Bad.push_back(SE.getMulExpr(NegOne, *I));
293 // Ok, we can't do anything interesting. Just stuff the whole thing into a
294 // register and hope for the best.
298 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
299 /// attempting to keep all loop-invariant and loop-computable values in a
300 /// single base register.
301 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
302 SmallVector<const SCEV *, 4> Good;
303 SmallVector<const SCEV *, 4> Bad;
304 DoInitialMatch(S, L, Good, Bad, SE);
306 const SCEV *Sum = SE.getAddExpr(Good);
308 BaseRegs.push_back(Sum);
309 AM.HasBaseReg = true;
312 const SCEV *Sum = SE.getAddExpr(Bad);
314 BaseRegs.push_back(Sum);
315 AM.HasBaseReg = true;
319 /// getNumRegs - Return the total number of register operands used by this
320 /// formula. This does not include register uses implied by non-constant
322 unsigned Formula::getNumRegs() const {
323 return !!ScaledReg + BaseRegs.size();
326 /// getType - Return the type of this formula, if it has one, or null
327 /// otherwise. This type is meaningless except for the bit size.
328 Type *Formula::getType() const {
329 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
330 ScaledReg ? ScaledReg->getType() :
331 AM.BaseGV ? AM.BaseGV->getType() :
335 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
336 void Formula::DeleteBaseReg(const SCEV *&S) {
337 if (&S != &BaseRegs.back())
338 std::swap(S, BaseRegs.back());
342 /// referencesReg - Test if this formula references the given register.
343 bool Formula::referencesReg(const SCEV *S) const {
344 return S == ScaledReg ||
345 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
348 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
349 /// which are used by uses other than the use with the given index.
350 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
351 const RegUseTracker &RegUses) const {
353 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
355 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
356 E = BaseRegs.end(); I != E; ++I)
357 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
362 void Formula::print(raw_ostream &OS) const {
365 if (!First) OS << " + "; else First = false;
366 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
368 if (AM.BaseOffs != 0) {
369 if (!First) OS << " + "; else First = false;
372 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
373 E = BaseRegs.end(); I != E; ++I) {
374 if (!First) OS << " + "; else First = false;
375 OS << "reg(" << **I << ')';
377 if (AM.HasBaseReg && BaseRegs.empty()) {
378 if (!First) OS << " + "; else First = false;
379 OS << "**error: HasBaseReg**";
380 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
381 if (!First) OS << " + "; else First = false;
382 OS << "**error: !HasBaseReg**";
385 if (!First) OS << " + "; else First = false;
386 OS << AM.Scale << "*reg(";
393 if (UnfoldedOffset != 0) {
394 if (!First) OS << " + "; else First = false;
395 OS << "imm(" << UnfoldedOffset << ')';
399 void Formula::dump() const {
400 print(errs()); errs() << '\n';
403 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
404 /// without changing its value.
405 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
407 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
408 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
411 /// isAddSExtable - Return true if the given add can be sign-extended
412 /// without changing its value.
413 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
415 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
416 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
419 /// isMulSExtable - Return true if the given mul can be sign-extended
420 /// without changing its value.
421 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
423 IntegerType::get(SE.getContext(),
424 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
425 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
428 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
429 /// and if the remainder is known to be zero, or null otherwise. If
430 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
431 /// to Y, ignoring that the multiplication may overflow, which is useful when
432 /// the result will be used in a context where the most significant bits are
434 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
436 bool IgnoreSignificantBits = false) {
437 // Handle the trivial case, which works for any SCEV type.
439 return SE.getConstant(LHS->getType(), 1);
441 // Handle a few RHS special cases.
442 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
444 const APInt &RA = RC->getValue()->getValue();
445 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
447 if (RA.isAllOnesValue())
448 return SE.getMulExpr(LHS, RC);
449 // Handle x /s 1 as x.
454 // Check for a division of a constant by a constant.
455 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
458 const APInt &LA = C->getValue()->getValue();
459 const APInt &RA = RC->getValue()->getValue();
460 if (LA.srem(RA) != 0)
462 return SE.getConstant(LA.sdiv(RA));
465 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
466 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
467 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
468 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
469 IgnoreSignificantBits);
471 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
472 IgnoreSignificantBits);
473 if (!Start) return 0;
474 // FlagNW is independent of the start value, step direction, and is
475 // preserved with smaller magnitude steps.
476 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
477 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
482 // Distribute the sdiv over add operands, if the add doesn't overflow.
483 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
484 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
485 SmallVector<const SCEV *, 8> Ops;
486 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
488 const SCEV *Op = getExactSDiv(*I, RHS, SE,
489 IgnoreSignificantBits);
493 return SE.getAddExpr(Ops);
498 // Check for a multiply operand that we can pull RHS out of.
499 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
500 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
501 SmallVector<const SCEV *, 4> Ops;
503 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
507 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
508 IgnoreSignificantBits)) {
514 return Found ? SE.getMulExpr(Ops) : 0;
519 // Otherwise we don't know.
523 /// ExtractImmediate - If S involves the addition of a constant integer value,
524 /// return that integer value, and mutate S to point to a new SCEV with that
526 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
527 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
528 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
529 S = SE.getConstant(C->getType(), 0);
530 return C->getValue()->getSExtValue();
532 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
533 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
534 int64_t Result = ExtractImmediate(NewOps.front(), SE);
536 S = SE.getAddExpr(NewOps);
538 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
539 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
540 int64_t Result = ExtractImmediate(NewOps.front(), SE);
542 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
543 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
550 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
551 /// return that symbol, and mutate S to point to a new SCEV with that
553 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
554 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
555 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
556 S = SE.getConstant(GV->getType(), 0);
559 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
560 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
561 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
563 S = SE.getAddExpr(NewOps);
565 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
566 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
567 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
569 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
570 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
577 /// isAddressUse - Returns true if the specified instruction is using the
578 /// specified value as an address.
579 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
580 bool isAddress = isa<LoadInst>(Inst);
581 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
582 if (SI->getOperand(1) == OperandVal)
584 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
585 // Addressing modes can also be folded into prefetches and a variety
587 switch (II->getIntrinsicID()) {
589 case Intrinsic::prefetch:
590 case Intrinsic::x86_sse_storeu_ps:
591 case Intrinsic::x86_sse2_storeu_pd:
592 case Intrinsic::x86_sse2_storeu_dq:
593 case Intrinsic::x86_sse2_storel_dq:
594 if (II->getArgOperand(0) == OperandVal)
602 /// getAccessType - Return the type of the memory being accessed.
603 static Type *getAccessType(const Instruction *Inst) {
604 Type *AccessTy = Inst->getType();
605 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
606 AccessTy = SI->getOperand(0)->getType();
607 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
608 // Addressing modes can also be folded into prefetches and a variety
610 switch (II->getIntrinsicID()) {
612 case Intrinsic::x86_sse_storeu_ps:
613 case Intrinsic::x86_sse2_storeu_pd:
614 case Intrinsic::x86_sse2_storeu_dq:
615 case Intrinsic::x86_sse2_storel_dq:
616 AccessTy = II->getArgOperand(0)->getType();
621 // All pointers have the same requirements, so canonicalize them to an
622 // arbitrary pointer type to minimize variation.
623 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
624 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
625 PTy->getAddressSpace());
630 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
631 /// specified set are trivially dead, delete them and see if this makes any of
632 /// their operands subsequently dead.
634 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
635 bool Changed = false;
637 while (!DeadInsts.empty()) {
638 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
640 if (I == 0 || !isInstructionTriviallyDead(I))
643 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
644 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
647 DeadInsts.push_back(U);
650 I->eraseFromParent();
659 /// Cost - This class is used to measure and compare candidate formulae.
661 /// TODO: Some of these could be merged. Also, a lexical ordering
662 /// isn't always optimal.
666 unsigned NumBaseAdds;
672 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
675 bool operator<(const Cost &Other) const;
680 // Once any of the metrics loses, they must all remain losers.
682 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
683 | ImmCost | SetupCost) != ~0u)
684 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
685 & ImmCost & SetupCost) == ~0u);
690 assert(isValid() && "invalid cost");
691 return NumRegs == ~0u;
694 void RateFormula(const Formula &F,
695 SmallPtrSet<const SCEV *, 16> &Regs,
696 const DenseSet<const SCEV *> &VisitedRegs,
698 const SmallVectorImpl<int64_t> &Offsets,
699 ScalarEvolution &SE, DominatorTree &DT);
701 void print(raw_ostream &OS) const;
705 void RateRegister(const SCEV *Reg,
706 SmallPtrSet<const SCEV *, 16> &Regs,
708 ScalarEvolution &SE, DominatorTree &DT);
709 void RatePrimaryRegister(const SCEV *Reg,
710 SmallPtrSet<const SCEV *, 16> &Regs,
712 ScalarEvolution &SE, DominatorTree &DT);
717 /// RateRegister - Tally up interesting quantities from the given register.
718 void Cost::RateRegister(const SCEV *Reg,
719 SmallPtrSet<const SCEV *, 16> &Regs,
721 ScalarEvolution &SE, DominatorTree &DT) {
722 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
723 if (AR->getLoop() == L)
724 AddRecCost += 1; /// TODO: This should be a function of the stride.
726 // If this is an addrec for a loop that's already been visited by LSR,
727 // don't second-guess its addrec phi nodes. LSR isn't currently smart
728 // enough to reason about more than one loop at a time. Consider these
729 // registers free and leave them alone.
730 else if (L->contains(AR->getLoop()) ||
731 (!AR->getLoop()->contains(L) &&
732 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
733 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
734 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
735 if (SE.isSCEVable(PN->getType()) &&
736 (SE.getEffectiveSCEVType(PN->getType()) ==
737 SE.getEffectiveSCEVType(AR->getType())) &&
738 SE.getSCEV(PN) == AR)
741 // If this isn't one of the addrecs that the loop already has, it
742 // would require a costly new phi and add. TODO: This isn't
743 // precisely modeled right now.
745 if (!Regs.count(AR->getStart())) {
746 RateRegister(AR->getStart(), Regs, L, SE, DT);
752 // Add the step value register, if it needs one.
753 // TODO: The non-affine case isn't precisely modeled here.
754 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
755 if (!Regs.count(AR->getOperand(1))) {
756 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
764 // Rough heuristic; favor registers which don't require extra setup
765 // instructions in the preheader.
766 if (!isa<SCEVUnknown>(Reg) &&
767 !isa<SCEVConstant>(Reg) &&
768 !(isa<SCEVAddRecExpr>(Reg) &&
769 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
770 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
773 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
774 SE.hasComputableLoopEvolution(Reg, L);
777 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
779 void Cost::RatePrimaryRegister(const SCEV *Reg,
780 SmallPtrSet<const SCEV *, 16> &Regs,
782 ScalarEvolution &SE, DominatorTree &DT) {
783 if (Regs.insert(Reg))
784 RateRegister(Reg, Regs, L, SE, DT);
787 void Cost::RateFormula(const Formula &F,
788 SmallPtrSet<const SCEV *, 16> &Regs,
789 const DenseSet<const SCEV *> &VisitedRegs,
791 const SmallVectorImpl<int64_t> &Offsets,
792 ScalarEvolution &SE, DominatorTree &DT) {
793 // Tally up the registers.
794 if (const SCEV *ScaledReg = F.ScaledReg) {
795 if (VisitedRegs.count(ScaledReg)) {
799 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
803 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
804 E = F.BaseRegs.end(); I != E; ++I) {
805 const SCEV *BaseReg = *I;
806 if (VisitedRegs.count(BaseReg)) {
810 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
815 // Determine how many (unfolded) adds we'll need inside the loop.
816 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
817 if (NumBaseParts > 1)
818 NumBaseAdds += NumBaseParts - 1;
820 // Tally up the non-zero immediates.
821 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
822 E = Offsets.end(); I != E; ++I) {
823 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
825 ImmCost += 64; // Handle symbolic values conservatively.
826 // TODO: This should probably be the pointer size.
827 else if (Offset != 0)
828 ImmCost += APInt(64, Offset, true).getMinSignedBits();
830 assert(isValid() && "invalid cost");
833 /// Loose - Set this cost to a losing value.
843 /// operator< - Choose the lower cost.
844 bool Cost::operator<(const Cost &Other) const {
845 if (NumRegs != Other.NumRegs)
846 return NumRegs < Other.NumRegs;
847 if (AddRecCost != Other.AddRecCost)
848 return AddRecCost < Other.AddRecCost;
849 if (NumIVMuls != Other.NumIVMuls)
850 return NumIVMuls < Other.NumIVMuls;
851 if (NumBaseAdds != Other.NumBaseAdds)
852 return NumBaseAdds < Other.NumBaseAdds;
853 if (ImmCost != Other.ImmCost)
854 return ImmCost < Other.ImmCost;
855 if (SetupCost != Other.SetupCost)
856 return SetupCost < Other.SetupCost;
860 void Cost::print(raw_ostream &OS) const {
861 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
863 OS << ", with addrec cost " << AddRecCost;
865 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
866 if (NumBaseAdds != 0)
867 OS << ", plus " << NumBaseAdds << " base add"
868 << (NumBaseAdds == 1 ? "" : "s");
870 OS << ", plus " << ImmCost << " imm cost";
872 OS << ", plus " << SetupCost << " setup cost";
875 void Cost::dump() const {
876 print(errs()); errs() << '\n';
881 /// LSRFixup - An operand value in an instruction which is to be replaced
882 /// with some equivalent, possibly strength-reduced, replacement.
884 /// UserInst - The instruction which will be updated.
885 Instruction *UserInst;
887 /// OperandValToReplace - The operand of the instruction which will
888 /// be replaced. The operand may be used more than once; every instance
889 /// will be replaced.
890 Value *OperandValToReplace;
892 /// PostIncLoops - If this user is to use the post-incremented value of an
893 /// induction variable, this variable is non-null and holds the loop
894 /// associated with the induction variable.
895 PostIncLoopSet PostIncLoops;
897 /// LUIdx - The index of the LSRUse describing the expression which
898 /// this fixup needs, minus an offset (below).
901 /// Offset - A constant offset to be added to the LSRUse expression.
902 /// This allows multiple fixups to share the same LSRUse with different
903 /// offsets, for example in an unrolled loop.
906 bool isUseFullyOutsideLoop(const Loop *L) const;
910 void print(raw_ostream &OS) const;
917 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
919 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
920 /// value outside of the given loop.
921 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
922 // PHI nodes use their value in their incoming blocks.
923 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
924 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
925 if (PN->getIncomingValue(i) == OperandValToReplace &&
926 L->contains(PN->getIncomingBlock(i)))
931 return !L->contains(UserInst);
934 void LSRFixup::print(raw_ostream &OS) const {
936 // Store is common and interesting enough to be worth special-casing.
937 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
939 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
940 } else if (UserInst->getType()->isVoidTy())
941 OS << UserInst->getOpcodeName();
943 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
945 OS << ", OperandValToReplace=";
946 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
948 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
949 E = PostIncLoops.end(); I != E; ++I) {
950 OS << ", PostIncLoop=";
951 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
954 if (LUIdx != ~size_t(0))
955 OS << ", LUIdx=" << LUIdx;
958 OS << ", Offset=" << Offset;
961 void LSRFixup::dump() const {
962 print(errs()); errs() << '\n';
967 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
968 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
969 struct UniquifierDenseMapInfo {
970 static SmallVector<const SCEV *, 2> getEmptyKey() {
971 SmallVector<const SCEV *, 2> V;
972 V.push_back(reinterpret_cast<const SCEV *>(-1));
976 static SmallVector<const SCEV *, 2> getTombstoneKey() {
977 SmallVector<const SCEV *, 2> V;
978 V.push_back(reinterpret_cast<const SCEV *>(-2));
982 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
984 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
985 E = V.end(); I != E; ++I)
986 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
990 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
991 const SmallVector<const SCEV *, 2> &RHS) {
996 /// LSRUse - This class holds the state that LSR keeps for each use in
997 /// IVUsers, as well as uses invented by LSR itself. It includes information
998 /// about what kinds of things can be folded into the user, information about
999 /// the user itself, and information about how the use may be satisfied.
1000 /// TODO: Represent multiple users of the same expression in common?
1002 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1005 /// KindType - An enum for a kind of use, indicating what types of
1006 /// scaled and immediate operands it might support.
1008 Basic, ///< A normal use, with no folding.
1009 Special, ///< A special case of basic, allowing -1 scales.
1010 Address, ///< An address use; folding according to TargetLowering
1011 ICmpZero ///< An equality icmp with both operands folded into one.
1012 // TODO: Add a generic icmp too?
1018 SmallVector<int64_t, 8> Offsets;
1022 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1023 /// LSRUse are outside of the loop, in which case some special-case heuristics
1025 bool AllFixupsOutsideLoop;
1027 /// WidestFixupType - This records the widest use type for any fixup using
1028 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1029 /// max fixup widths to be equivalent, because the narrower one may be relying
1030 /// on the implicit truncation to truncate away bogus bits.
1031 Type *WidestFixupType;
1033 /// Formulae - A list of ways to build a value that can satisfy this user.
1034 /// After the list is populated, one of these is selected heuristically and
1035 /// used to formulate a replacement for OperandValToReplace in UserInst.
1036 SmallVector<Formula, 12> Formulae;
1038 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1039 SmallPtrSet<const SCEV *, 4> Regs;
1041 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1042 MinOffset(INT64_MAX),
1043 MaxOffset(INT64_MIN),
1044 AllFixupsOutsideLoop(true),
1045 WidestFixupType(0) {}
1047 bool HasFormulaWithSameRegs(const Formula &F) const;
1048 bool InsertFormula(const Formula &F);
1049 void DeleteFormula(Formula &F);
1050 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1052 void print(raw_ostream &OS) const;
1058 /// HasFormula - Test whether this use as a formula which has the same
1059 /// registers as the given formula.
1060 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1061 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1062 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1063 // Unstable sort by host order ok, because this is only used for uniquifying.
1064 std::sort(Key.begin(), Key.end());
1065 return Uniquifier.count(Key);
1068 /// InsertFormula - If the given formula has not yet been inserted, add it to
1069 /// the list, and return true. Return false otherwise.
1070 bool LSRUse::InsertFormula(const Formula &F) {
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());
1076 if (!Uniquifier.insert(Key).second)
1079 // Using a register to hold the value of 0 is not profitable.
1080 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1081 "Zero allocated in a scaled register!");
1083 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1084 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1085 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1088 // Add the formula to the list.
1089 Formulae.push_back(F);
1091 // Record registers now being used by this use.
1092 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1093 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1098 /// DeleteFormula - Remove the given formula from this use's list.
1099 void LSRUse::DeleteFormula(Formula &F) {
1100 if (&F != &Formulae.back())
1101 std::swap(F, Formulae.back());
1102 Formulae.pop_back();
1103 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1106 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1107 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1108 // Now that we've filtered out some formulae, recompute the Regs set.
1109 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1111 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1112 E = Formulae.end(); I != E; ++I) {
1113 const Formula &F = *I;
1114 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1115 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1118 // Update the RegTracker.
1119 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1120 E = OldRegs.end(); I != E; ++I)
1121 if (!Regs.count(*I))
1122 RegUses.DropRegister(*I, LUIdx);
1125 void LSRUse::print(raw_ostream &OS) const {
1126 OS << "LSR Use: Kind=";
1128 case Basic: OS << "Basic"; break;
1129 case Special: OS << "Special"; break;
1130 case ICmpZero: OS << "ICmpZero"; break;
1132 OS << "Address of ";
1133 if (AccessTy->isPointerTy())
1134 OS << "pointer"; // the full pointer type could be really verbose
1139 OS << ", Offsets={";
1140 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1141 E = Offsets.end(); I != E; ++I) {
1143 if (llvm::next(I) != E)
1148 if (AllFixupsOutsideLoop)
1149 OS << ", all-fixups-outside-loop";
1151 if (WidestFixupType)
1152 OS << ", widest fixup type: " << *WidestFixupType;
1155 void LSRUse::dump() const {
1156 print(errs()); errs() << '\n';
1159 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1160 /// be completely folded into the user instruction at isel time. This includes
1161 /// address-mode folding and special icmp tricks.
1162 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1163 LSRUse::KindType Kind, Type *AccessTy,
1164 const TargetLowering *TLI) {
1166 case LSRUse::Address:
1167 // If we have low-level target information, ask the target if it can
1168 // completely fold this address.
1169 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1171 // Otherwise, just guess that reg+reg addressing is legal.
1172 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1174 case LSRUse::ICmpZero:
1175 // There's not even a target hook for querying whether it would be legal to
1176 // fold a GV into an ICmp.
1180 // ICmp only has two operands; don't allow more than two non-trivial parts.
1181 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1184 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1185 // putting the scaled register in the other operand of the icmp.
1186 if (AM.Scale != 0 && AM.Scale != -1)
1189 // If we have low-level target information, ask the target if it can fold an
1190 // integer immediate on an icmp.
1191 if (AM.BaseOffs != 0) {
1192 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1199 // Only handle single-register values.
1200 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1202 case LSRUse::Special:
1203 // Only handle -1 scales, or no scale.
1204 return AM.Scale == 0 || AM.Scale == -1;
1210 static bool isLegalUse(TargetLowering::AddrMode AM,
1211 int64_t MinOffset, int64_t MaxOffset,
1212 LSRUse::KindType Kind, Type *AccessTy,
1213 const TargetLowering *TLI) {
1214 // Check for overflow.
1215 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1218 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1219 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1220 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1221 // Check for overflow.
1222 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1225 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1226 return isLegalUse(AM, Kind, AccessTy, TLI);
1231 static bool isAlwaysFoldable(int64_t BaseOffs,
1232 GlobalValue *BaseGV,
1234 LSRUse::KindType Kind, Type *AccessTy,
1235 const TargetLowering *TLI) {
1236 // Fast-path: zero is always foldable.
1237 if (BaseOffs == 0 && !BaseGV) return true;
1239 // Conservatively, create an address with an immediate and a
1240 // base and a scale.
1241 TargetLowering::AddrMode AM;
1242 AM.BaseOffs = BaseOffs;
1244 AM.HasBaseReg = HasBaseReg;
1245 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1247 // Canonicalize a scale of 1 to a base register if the formula doesn't
1248 // already have a base register.
1249 if (!AM.HasBaseReg && AM.Scale == 1) {
1251 AM.HasBaseReg = true;
1254 return isLegalUse(AM, Kind, AccessTy, TLI);
1257 static bool isAlwaysFoldable(const SCEV *S,
1258 int64_t MinOffset, int64_t MaxOffset,
1260 LSRUse::KindType Kind, Type *AccessTy,
1261 const TargetLowering *TLI,
1262 ScalarEvolution &SE) {
1263 // Fast-path: zero is always foldable.
1264 if (S->isZero()) return true;
1266 // Conservatively, create an address with an immediate and a
1267 // base and a scale.
1268 int64_t BaseOffs = ExtractImmediate(S, SE);
1269 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1271 // If there's anything else involved, it's not foldable.
1272 if (!S->isZero()) return false;
1274 // Fast-path: zero is always foldable.
1275 if (BaseOffs == 0 && !BaseGV) return true;
1277 // Conservatively, create an address with an immediate and a
1278 // base and a scale.
1279 TargetLowering::AddrMode AM;
1280 AM.BaseOffs = BaseOffs;
1282 AM.HasBaseReg = HasBaseReg;
1283 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1285 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1290 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1291 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1292 struct UseMapDenseMapInfo {
1293 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1294 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1297 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1298 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1302 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1303 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1304 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1308 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1309 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1314 /// LSRInstance - This class holds state for the main loop strength reduction
1318 ScalarEvolution &SE;
1321 const TargetLowering *const TLI;
1325 /// IVIncInsertPos - This is the insert position that the current loop's
1326 /// induction variable increment should be placed. In simple loops, this is
1327 /// the latch block's terminator. But in more complicated cases, this is a
1328 /// position which will dominate all the in-loop post-increment users.
1329 Instruction *IVIncInsertPos;
1331 /// Factors - Interesting factors between use strides.
1332 SmallSetVector<int64_t, 8> Factors;
1334 /// Types - Interesting use types, to facilitate truncation reuse.
1335 SmallSetVector<Type *, 4> Types;
1337 /// Fixups - The list of operands which are to be replaced.
1338 SmallVector<LSRFixup, 16> Fixups;
1340 /// Uses - The list of interesting uses.
1341 SmallVector<LSRUse, 16> Uses;
1343 /// RegUses - Track which uses use which register candidates.
1344 RegUseTracker RegUses;
1346 void OptimizeShadowIV();
1347 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1348 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1349 void OptimizeLoopTermCond();
1351 void CollectInterestingTypesAndFactors();
1352 void CollectFixupsAndInitialFormulae();
1354 LSRFixup &getNewFixup() {
1355 Fixups.push_back(LSRFixup());
1356 return Fixups.back();
1359 // Support for sharing of LSRUses between LSRFixups.
1360 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1362 UseMapDenseMapInfo> UseMapTy;
1365 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1366 LSRUse::KindType Kind, Type *AccessTy);
1368 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1369 LSRUse::KindType Kind,
1372 void DeleteUse(LSRUse &LU, size_t LUIdx);
1374 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1377 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1378 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1379 void CountRegisters(const Formula &F, size_t LUIdx);
1380 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1382 void CollectLoopInvariantFixupsAndFormulae();
1384 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1385 unsigned Depth = 0);
1386 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1387 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1388 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1389 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1390 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1391 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1392 void GenerateCrossUseConstantOffsets();
1393 void GenerateAllReuseFormulae();
1395 void FilterOutUndesirableDedicatedRegisters();
1397 size_t EstimateSearchSpaceComplexity() const;
1398 void NarrowSearchSpaceByDetectingSupersets();
1399 void NarrowSearchSpaceByCollapsingUnrolledCode();
1400 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1401 void NarrowSearchSpaceByPickingWinnerRegs();
1402 void NarrowSearchSpaceUsingHeuristics();
1404 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1406 SmallVectorImpl<const Formula *> &Workspace,
1407 const Cost &CurCost,
1408 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1409 DenseSet<const SCEV *> &VisitedRegs) const;
1410 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1412 BasicBlock::iterator
1413 HoistInsertPosition(BasicBlock::iterator IP,
1414 const SmallVectorImpl<Instruction *> &Inputs) const;
1415 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1417 const LSRUse &LU) const;
1419 Value *Expand(const LSRFixup &LF,
1421 BasicBlock::iterator IP,
1422 SCEVExpander &Rewriter,
1423 SmallVectorImpl<WeakVH> &DeadInsts) const;
1424 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1426 SCEVExpander &Rewriter,
1427 SmallVectorImpl<WeakVH> &DeadInsts,
1429 void Rewrite(const LSRFixup &LF,
1431 SCEVExpander &Rewriter,
1432 SmallVectorImpl<WeakVH> &DeadInsts,
1434 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1437 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1439 bool getChanged() const { return Changed; }
1441 void print_factors_and_types(raw_ostream &OS) const;
1442 void print_fixups(raw_ostream &OS) const;
1443 void print_uses(raw_ostream &OS) const;
1444 void print(raw_ostream &OS) const;
1450 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1451 /// inside the loop then try to eliminate the cast operation.
1452 void LSRInstance::OptimizeShadowIV() {
1453 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1454 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1457 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1458 UI != E; /* empty */) {
1459 IVUsers::const_iterator CandidateUI = UI;
1461 Instruction *ShadowUse = CandidateUI->getUser();
1462 Type *DestTy = NULL;
1463 bool IsSigned = false;
1465 /* If shadow use is a int->float cast then insert a second IV
1466 to eliminate this cast.
1468 for (unsigned i = 0; i < n; ++i)
1474 for (unsigned i = 0; i < n; ++i, ++d)
1477 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1479 DestTy = UCast->getDestTy();
1481 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1483 DestTy = SCast->getDestTy();
1485 if (!DestTy) continue;
1488 // If target does not support DestTy natively then do not apply
1489 // this transformation.
1490 EVT DVT = TLI->getValueType(DestTy);
1491 if (!TLI->isTypeLegal(DVT)) continue;
1494 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1496 if (PH->getNumIncomingValues() != 2) continue;
1498 Type *SrcTy = PH->getType();
1499 int Mantissa = DestTy->getFPMantissaWidth();
1500 if (Mantissa == -1) continue;
1501 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1504 unsigned Entry, Latch;
1505 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1513 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1514 if (!Init) continue;
1515 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1516 (double)Init->getSExtValue() :
1517 (double)Init->getZExtValue());
1519 BinaryOperator *Incr =
1520 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1521 if (!Incr) continue;
1522 if (Incr->getOpcode() != Instruction::Add
1523 && Incr->getOpcode() != Instruction::Sub)
1526 /* Initialize new IV, double d = 0.0 in above example. */
1527 ConstantInt *C = NULL;
1528 if (Incr->getOperand(0) == PH)
1529 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1530 else if (Incr->getOperand(1) == PH)
1531 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1537 // Ignore negative constants, as the code below doesn't handle them
1538 // correctly. TODO: Remove this restriction.
1539 if (!C->getValue().isStrictlyPositive()) continue;
1541 /* Add new PHINode. */
1542 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1544 /* create new increment. '++d' in above example. */
1545 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1546 BinaryOperator *NewIncr =
1547 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1548 Instruction::FAdd : Instruction::FSub,
1549 NewPH, CFP, "IV.S.next.", Incr);
1551 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1552 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1554 /* Remove cast operation */
1555 ShadowUse->replaceAllUsesWith(NewPH);
1556 ShadowUse->eraseFromParent();
1562 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1563 /// set the IV user and stride information and return true, otherwise return
1565 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1566 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1567 if (UI->getUser() == Cond) {
1568 // NOTE: we could handle setcc instructions with multiple uses here, but
1569 // InstCombine does it as well for simple uses, it's not clear that it
1570 // occurs enough in real life to handle.
1577 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1578 /// a max computation.
1580 /// This is a narrow solution to a specific, but acute, problem. For loops
1586 /// } while (++i < n);
1588 /// the trip count isn't just 'n', because 'n' might not be positive. And
1589 /// unfortunately this can come up even for loops where the user didn't use
1590 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1591 /// will commonly be lowered like this:
1597 /// } while (++i < n);
1600 /// and then it's possible for subsequent optimization to obscure the if
1601 /// test in such a way that indvars can't find it.
1603 /// When indvars can't find the if test in loops like this, it creates a
1604 /// max expression, which allows it to give the loop a canonical
1605 /// induction variable:
1608 /// max = n < 1 ? 1 : n;
1611 /// } while (++i != max);
1613 /// Canonical induction variables are necessary because the loop passes
1614 /// are designed around them. The most obvious example of this is the
1615 /// LoopInfo analysis, which doesn't remember trip count values. It
1616 /// expects to be able to rediscover the trip count each time it is
1617 /// needed, and it does this using a simple analysis that only succeeds if
1618 /// the loop has a canonical induction variable.
1620 /// However, when it comes time to generate code, the maximum operation
1621 /// can be quite costly, especially if it's inside of an outer loop.
1623 /// This function solves this problem by detecting this type of loop and
1624 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1625 /// the instructions for the maximum computation.
1627 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1628 // Check that the loop matches the pattern we're looking for.
1629 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1630 Cond->getPredicate() != CmpInst::ICMP_NE)
1633 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1634 if (!Sel || !Sel->hasOneUse()) return Cond;
1636 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1637 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1639 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1641 // Add one to the backedge-taken count to get the trip count.
1642 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1643 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1645 // Check for a max calculation that matches the pattern. There's no check
1646 // for ICMP_ULE here because the comparison would be with zero, which
1647 // isn't interesting.
1648 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1649 const SCEVNAryExpr *Max = 0;
1650 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1651 Pred = ICmpInst::ICMP_SLE;
1653 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1654 Pred = ICmpInst::ICMP_SLT;
1656 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1657 Pred = ICmpInst::ICMP_ULT;
1664 // To handle a max with more than two operands, this optimization would
1665 // require additional checking and setup.
1666 if (Max->getNumOperands() != 2)
1669 const SCEV *MaxLHS = Max->getOperand(0);
1670 const SCEV *MaxRHS = Max->getOperand(1);
1672 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1673 // for a comparison with 1. For <= and >=, a comparison with zero.
1675 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1678 // Check the relevant induction variable for conformance to
1680 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1681 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1682 if (!AR || !AR->isAffine() ||
1683 AR->getStart() != One ||
1684 AR->getStepRecurrence(SE) != One)
1687 assert(AR->getLoop() == L &&
1688 "Loop condition operand is an addrec in a different loop!");
1690 // Check the right operand of the select, and remember it, as it will
1691 // be used in the new comparison instruction.
1693 if (ICmpInst::isTrueWhenEqual(Pred)) {
1694 // Look for n+1, and grab n.
1695 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1696 if (isa<ConstantInt>(BO->getOperand(1)) &&
1697 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1698 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1699 NewRHS = BO->getOperand(0);
1700 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1701 if (isa<ConstantInt>(BO->getOperand(1)) &&
1702 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1703 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1704 NewRHS = BO->getOperand(0);
1707 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1708 NewRHS = Sel->getOperand(1);
1709 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1710 NewRHS = Sel->getOperand(2);
1711 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1712 NewRHS = SU->getValue();
1714 // Max doesn't match expected pattern.
1717 // Determine the new comparison opcode. It may be signed or unsigned,
1718 // and the original comparison may be either equality or inequality.
1719 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1720 Pred = CmpInst::getInversePredicate(Pred);
1722 // Ok, everything looks ok to change the condition into an SLT or SGE and
1723 // delete the max calculation.
1725 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1727 // Delete the max calculation instructions.
1728 Cond->replaceAllUsesWith(NewCond);
1729 CondUse->setUser(NewCond);
1730 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1731 Cond->eraseFromParent();
1732 Sel->eraseFromParent();
1733 if (Cmp->use_empty())
1734 Cmp->eraseFromParent();
1738 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1739 /// postinc iv when possible.
1741 LSRInstance::OptimizeLoopTermCond() {
1742 SmallPtrSet<Instruction *, 4> PostIncs;
1744 BasicBlock *LatchBlock = L->getLoopLatch();
1745 SmallVector<BasicBlock*, 8> ExitingBlocks;
1746 L->getExitingBlocks(ExitingBlocks);
1748 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1749 BasicBlock *ExitingBlock = ExitingBlocks[i];
1751 // Get the terminating condition for the loop if possible. If we
1752 // can, we want to change it to use a post-incremented version of its
1753 // induction variable, to allow coalescing the live ranges for the IV into
1754 // one register value.
1756 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1759 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1760 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1763 // Search IVUsesByStride to find Cond's IVUse if there is one.
1764 IVStrideUse *CondUse = 0;
1765 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1766 if (!FindIVUserForCond(Cond, CondUse))
1769 // If the trip count is computed in terms of a max (due to ScalarEvolution
1770 // being unable to find a sufficient guard, for example), change the loop
1771 // comparison to use SLT or ULT instead of NE.
1772 // One consequence of doing this now is that it disrupts the count-down
1773 // optimization. That's not always a bad thing though, because in such
1774 // cases it may still be worthwhile to avoid a max.
1775 Cond = OptimizeMax(Cond, CondUse);
1777 // If this exiting block dominates the latch block, it may also use
1778 // the post-inc value if it won't be shared with other uses.
1779 // Check for dominance.
1780 if (!DT.dominates(ExitingBlock, LatchBlock))
1783 // Conservatively avoid trying to use the post-inc value in non-latch
1784 // exits if there may be pre-inc users in intervening blocks.
1785 if (LatchBlock != ExitingBlock)
1786 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1787 // Test if the use is reachable from the exiting block. This dominator
1788 // query is a conservative approximation of reachability.
1789 if (&*UI != CondUse &&
1790 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1791 // Conservatively assume there may be reuse if the quotient of their
1792 // strides could be a legal scale.
1793 const SCEV *A = IU.getStride(*CondUse, L);
1794 const SCEV *B = IU.getStride(*UI, L);
1795 if (!A || !B) continue;
1796 if (SE.getTypeSizeInBits(A->getType()) !=
1797 SE.getTypeSizeInBits(B->getType())) {
1798 if (SE.getTypeSizeInBits(A->getType()) >
1799 SE.getTypeSizeInBits(B->getType()))
1800 B = SE.getSignExtendExpr(B, A->getType());
1802 A = SE.getSignExtendExpr(A, B->getType());
1804 if (const SCEVConstant *D =
1805 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1806 const ConstantInt *C = D->getValue();
1807 // Stride of one or negative one can have reuse with non-addresses.
1808 if (C->isOne() || C->isAllOnesValue())
1809 goto decline_post_inc;
1810 // Avoid weird situations.
1811 if (C->getValue().getMinSignedBits() >= 64 ||
1812 C->getValue().isMinSignedValue())
1813 goto decline_post_inc;
1814 // Without TLI, assume that any stride might be valid, and so any
1815 // use might be shared.
1817 goto decline_post_inc;
1818 // Check for possible scaled-address reuse.
1819 Type *AccessTy = getAccessType(UI->getUser());
1820 TargetLowering::AddrMode AM;
1821 AM.Scale = C->getSExtValue();
1822 if (TLI->isLegalAddressingMode(AM, AccessTy))
1823 goto decline_post_inc;
1824 AM.Scale = -AM.Scale;
1825 if (TLI->isLegalAddressingMode(AM, AccessTy))
1826 goto decline_post_inc;
1830 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1833 // It's possible for the setcc instruction to be anywhere in the loop, and
1834 // possible for it to have multiple users. If it is not immediately before
1835 // the exiting block branch, move it.
1836 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1837 if (Cond->hasOneUse()) {
1838 Cond->moveBefore(TermBr);
1840 // Clone the terminating condition and insert into the loopend.
1841 ICmpInst *OldCond = Cond;
1842 Cond = cast<ICmpInst>(Cond->clone());
1843 Cond->setName(L->getHeader()->getName() + ".termcond");
1844 ExitingBlock->getInstList().insert(TermBr, Cond);
1846 // Clone the IVUse, as the old use still exists!
1847 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1848 TermBr->replaceUsesOfWith(OldCond, Cond);
1852 // If we get to here, we know that we can transform the setcc instruction to
1853 // use the post-incremented version of the IV, allowing us to coalesce the
1854 // live ranges for the IV correctly.
1855 CondUse->transformToPostInc(L);
1858 PostIncs.insert(Cond);
1862 // Determine an insertion point for the loop induction variable increment. It
1863 // must dominate all the post-inc comparisons we just set up, and it must
1864 // dominate the loop latch edge.
1865 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1866 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1867 E = PostIncs.end(); I != E; ++I) {
1869 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1871 if (BB == (*I)->getParent())
1872 IVIncInsertPos = *I;
1873 else if (BB != IVIncInsertPos->getParent())
1874 IVIncInsertPos = BB->getTerminator();
1878 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
1879 /// at the given offset and other details. If so, update the use and
1882 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1883 LSRUse::KindType Kind, Type *AccessTy) {
1884 int64_t NewMinOffset = LU.MinOffset;
1885 int64_t NewMaxOffset = LU.MaxOffset;
1886 Type *NewAccessTy = AccessTy;
1888 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1889 // something conservative, however this can pessimize in the case that one of
1890 // the uses will have all its uses outside the loop, for example.
1891 if (LU.Kind != Kind)
1893 // Conservatively assume HasBaseReg is true for now.
1894 if (NewOffset < LU.MinOffset) {
1895 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1896 Kind, AccessTy, TLI))
1898 NewMinOffset = NewOffset;
1899 } else if (NewOffset > LU.MaxOffset) {
1900 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1901 Kind, AccessTy, TLI))
1903 NewMaxOffset = NewOffset;
1905 // Check for a mismatched access type, and fall back conservatively as needed.
1906 // TODO: Be less conservative when the type is similar and can use the same
1907 // addressing modes.
1908 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1909 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1912 LU.MinOffset = NewMinOffset;
1913 LU.MaxOffset = NewMaxOffset;
1914 LU.AccessTy = NewAccessTy;
1915 if (NewOffset != LU.Offsets.back())
1916 LU.Offsets.push_back(NewOffset);
1920 /// getUse - Return an LSRUse index and an offset value for a fixup which
1921 /// needs the given expression, with the given kind and optional access type.
1922 /// Either reuse an existing use or create a new one, as needed.
1923 std::pair<size_t, int64_t>
1924 LSRInstance::getUse(const SCEV *&Expr,
1925 LSRUse::KindType Kind, Type *AccessTy) {
1926 const SCEV *Copy = Expr;
1927 int64_t Offset = ExtractImmediate(Expr, SE);
1929 // Basic uses can't accept any offset, for example.
1930 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1935 std::pair<UseMapTy::iterator, bool> P =
1936 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1938 // A use already existed with this base.
1939 size_t LUIdx = P.first->second;
1940 LSRUse &LU = Uses[LUIdx];
1941 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1943 return std::make_pair(LUIdx, Offset);
1946 // Create a new use.
1947 size_t LUIdx = Uses.size();
1948 P.first->second = LUIdx;
1949 Uses.push_back(LSRUse(Kind, AccessTy));
1950 LSRUse &LU = Uses[LUIdx];
1952 // We don't need to track redundant offsets, but we don't need to go out
1953 // of our way here to avoid them.
1954 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1955 LU.Offsets.push_back(Offset);
1957 LU.MinOffset = Offset;
1958 LU.MaxOffset = Offset;
1959 return std::make_pair(LUIdx, Offset);
1962 /// DeleteUse - Delete the given use from the Uses list.
1963 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
1964 if (&LU != &Uses.back())
1965 std::swap(LU, Uses.back());
1969 RegUses.SwapAndDropUse(LUIdx, Uses.size());
1972 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1973 /// a formula that has the same registers as the given formula.
1975 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1976 const LSRUse &OrigLU) {
1977 // Search all uses for the formula. This could be more clever.
1978 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1979 LSRUse &LU = Uses[LUIdx];
1980 // Check whether this use is close enough to OrigLU, to see whether it's
1981 // worthwhile looking through its formulae.
1982 // Ignore ICmpZero uses because they may contain formulae generated by
1983 // GenerateICmpZeroScales, in which case adding fixup offsets may
1985 if (&LU != &OrigLU &&
1986 LU.Kind != LSRUse::ICmpZero &&
1987 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1988 LU.WidestFixupType == OrigLU.WidestFixupType &&
1989 LU.HasFormulaWithSameRegs(OrigF)) {
1990 // Scan through this use's formulae.
1991 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1992 E = LU.Formulae.end(); I != E; ++I) {
1993 const Formula &F = *I;
1994 // Check to see if this formula has the same registers and symbols
1996 if (F.BaseRegs == OrigF.BaseRegs &&
1997 F.ScaledReg == OrigF.ScaledReg &&
1998 F.AM.BaseGV == OrigF.AM.BaseGV &&
1999 F.AM.Scale == OrigF.AM.Scale &&
2000 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2001 if (F.AM.BaseOffs == 0)
2003 // This is the formula where all the registers and symbols matched;
2004 // there aren't going to be any others. Since we declined it, we
2005 // can skip the rest of the formulae and procede to the next LSRUse.
2012 // Nothing looked good.
2016 void LSRInstance::CollectInterestingTypesAndFactors() {
2017 SmallSetVector<const SCEV *, 4> Strides;
2019 // Collect interesting types and strides.
2020 SmallVector<const SCEV *, 4> Worklist;
2021 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2022 const SCEV *Expr = IU.getExpr(*UI);
2024 // Collect interesting types.
2025 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2027 // Add strides for mentioned loops.
2028 Worklist.push_back(Expr);
2030 const SCEV *S = Worklist.pop_back_val();
2031 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2032 Strides.insert(AR->getStepRecurrence(SE));
2033 Worklist.push_back(AR->getStart());
2034 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2035 Worklist.append(Add->op_begin(), Add->op_end());
2037 } while (!Worklist.empty());
2040 // Compute interesting factors from the set of interesting strides.
2041 for (SmallSetVector<const SCEV *, 4>::const_iterator
2042 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2043 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2044 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2045 const SCEV *OldStride = *I;
2046 const SCEV *NewStride = *NewStrideIter;
2048 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2049 SE.getTypeSizeInBits(NewStride->getType())) {
2050 if (SE.getTypeSizeInBits(OldStride->getType()) >
2051 SE.getTypeSizeInBits(NewStride->getType()))
2052 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2054 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2056 if (const SCEVConstant *Factor =
2057 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2059 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2060 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2061 } else if (const SCEVConstant *Factor =
2062 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2065 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2066 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2070 // If all uses use the same type, don't bother looking for truncation-based
2072 if (Types.size() == 1)
2075 DEBUG(print_factors_and_types(dbgs()));
2078 void LSRInstance::CollectFixupsAndInitialFormulae() {
2079 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2081 LSRFixup &LF = getNewFixup();
2082 LF.UserInst = UI->getUser();
2083 LF.OperandValToReplace = UI->getOperandValToReplace();
2084 LF.PostIncLoops = UI->getPostIncLoops();
2086 LSRUse::KindType Kind = LSRUse::Basic;
2088 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2089 Kind = LSRUse::Address;
2090 AccessTy = getAccessType(LF.UserInst);
2093 const SCEV *S = IU.getExpr(*UI);
2095 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2096 // (N - i == 0), and this allows (N - i) to be the expression that we work
2097 // with rather than just N or i, so we can consider the register
2098 // requirements for both N and i at the same time. Limiting this code to
2099 // equality icmps is not a problem because all interesting loops use
2100 // equality icmps, thanks to IndVarSimplify.
2101 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2102 if (CI->isEquality()) {
2103 // Swap the operands if needed to put the OperandValToReplace on the
2104 // left, for consistency.
2105 Value *NV = CI->getOperand(1);
2106 if (NV == LF.OperandValToReplace) {
2107 CI->setOperand(1, CI->getOperand(0));
2108 CI->setOperand(0, NV);
2109 NV = CI->getOperand(1);
2113 // x == y --> x - y == 0
2114 const SCEV *N = SE.getSCEV(NV);
2115 if (SE.isLoopInvariant(N, L)) {
2116 // S is normalized, so normalize N before folding it into S
2117 // to keep the result normalized.
2118 N = TransformForPostIncUse(Normalize, N, CI, 0,
2119 LF.PostIncLoops, SE, DT);
2120 Kind = LSRUse::ICmpZero;
2121 S = SE.getMinusSCEV(N, S);
2124 // -1 and the negations of all interesting strides (except the negation
2125 // of -1) are now also interesting.
2126 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2127 if (Factors[i] != -1)
2128 Factors.insert(-(uint64_t)Factors[i]);
2132 // Set up the initial formula for this use.
2133 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2135 LF.Offset = P.second;
2136 LSRUse &LU = Uses[LF.LUIdx];
2137 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2138 if (!LU.WidestFixupType ||
2139 SE.getTypeSizeInBits(LU.WidestFixupType) <
2140 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2141 LU.WidestFixupType = LF.OperandValToReplace->getType();
2143 // If this is the first use of this LSRUse, give it a formula.
2144 if (LU.Formulae.empty()) {
2145 InsertInitialFormula(S, LU, LF.LUIdx);
2146 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2150 DEBUG(print_fixups(dbgs()));
2153 /// InsertInitialFormula - Insert a formula for the given expression into
2154 /// the given use, separating out loop-variant portions from loop-invariant
2155 /// and loop-computable portions.
2157 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2159 F.InitialMatch(S, L, SE);
2160 bool Inserted = InsertFormula(LU, LUIdx, F);
2161 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2164 /// InsertSupplementalFormula - Insert a simple single-register formula for
2165 /// the given expression into the given use.
2167 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2168 LSRUse &LU, size_t LUIdx) {
2170 F.BaseRegs.push_back(S);
2171 F.AM.HasBaseReg = true;
2172 bool Inserted = InsertFormula(LU, LUIdx, F);
2173 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2176 /// CountRegisters - Note which registers are used by the given formula,
2177 /// updating RegUses.
2178 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2180 RegUses.CountRegister(F.ScaledReg, LUIdx);
2181 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2182 E = F.BaseRegs.end(); I != E; ++I)
2183 RegUses.CountRegister(*I, LUIdx);
2186 /// InsertFormula - If the given formula has not yet been inserted, add it to
2187 /// the list, and return true. Return false otherwise.
2188 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2189 if (!LU.InsertFormula(F))
2192 CountRegisters(F, LUIdx);
2196 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2197 /// loop-invariant values which we're tracking. These other uses will pin these
2198 /// values in registers, making them less profitable for elimination.
2199 /// TODO: This currently misses non-constant addrec step registers.
2200 /// TODO: Should this give more weight to users inside the loop?
2202 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2203 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2204 SmallPtrSet<const SCEV *, 8> Inserted;
2206 while (!Worklist.empty()) {
2207 const SCEV *S = Worklist.pop_back_val();
2209 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2210 Worklist.append(N->op_begin(), N->op_end());
2211 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2212 Worklist.push_back(C->getOperand());
2213 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2214 Worklist.push_back(D->getLHS());
2215 Worklist.push_back(D->getRHS());
2216 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2217 if (!Inserted.insert(U)) continue;
2218 const Value *V = U->getValue();
2219 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2220 // Look for instructions defined outside the loop.
2221 if (L->contains(Inst)) continue;
2222 } else if (isa<UndefValue>(V))
2223 // Undef doesn't have a live range, so it doesn't matter.
2225 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2227 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2228 // Ignore non-instructions.
2231 // Ignore instructions in other functions (as can happen with
2233 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2235 // Ignore instructions not dominated by the loop.
2236 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2237 UserInst->getParent() :
2238 cast<PHINode>(UserInst)->getIncomingBlock(
2239 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2240 if (!DT.dominates(L->getHeader(), UseBB))
2242 // Ignore uses which are part of other SCEV expressions, to avoid
2243 // analyzing them multiple times.
2244 if (SE.isSCEVable(UserInst->getType())) {
2245 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2246 // If the user is a no-op, look through to its uses.
2247 if (!isa<SCEVUnknown>(UserS))
2251 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2255 // Ignore icmp instructions which are already being analyzed.
2256 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2257 unsigned OtherIdx = !UI.getOperandNo();
2258 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2259 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2263 LSRFixup &LF = getNewFixup();
2264 LF.UserInst = const_cast<Instruction *>(UserInst);
2265 LF.OperandValToReplace = UI.getUse();
2266 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2268 LF.Offset = P.second;
2269 LSRUse &LU = Uses[LF.LUIdx];
2270 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2271 if (!LU.WidestFixupType ||
2272 SE.getTypeSizeInBits(LU.WidestFixupType) <
2273 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2274 LU.WidestFixupType = LF.OperandValToReplace->getType();
2275 InsertSupplementalFormula(U, LU, LF.LUIdx);
2276 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2283 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2284 /// separate registers. If C is non-null, multiply each subexpression by C.
2285 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2286 SmallVectorImpl<const SCEV *> &Ops,
2288 ScalarEvolution &SE) {
2289 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2290 // Break out add operands.
2291 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2293 CollectSubexprs(*I, C, Ops, L, SE);
2295 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2296 // Split a non-zero base out of an addrec.
2297 if (!AR->getStart()->isZero()) {
2298 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2299 AR->getStepRecurrence(SE),
2301 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2304 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2307 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2308 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2309 if (Mul->getNumOperands() == 2)
2310 if (const SCEVConstant *Op0 =
2311 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2312 CollectSubexprs(Mul->getOperand(1),
2313 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2319 // Otherwise use the value itself, optionally with a scale applied.
2320 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2323 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2325 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2328 // Arbitrarily cap recursion to protect compile time.
2329 if (Depth >= 3) return;
2331 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2332 const SCEV *BaseReg = Base.BaseRegs[i];
2334 SmallVector<const SCEV *, 8> AddOps;
2335 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2337 if (AddOps.size() == 1) continue;
2339 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2340 JE = AddOps.end(); J != JE; ++J) {
2342 // Loop-variant "unknown" values are uninteresting; we won't be able to
2343 // do anything meaningful with them.
2344 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2347 // Don't pull a constant into a register if the constant could be folded
2348 // into an immediate field.
2349 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2350 Base.getNumRegs() > 1,
2351 LU.Kind, LU.AccessTy, TLI, SE))
2354 // Collect all operands except *J.
2355 SmallVector<const SCEV *, 8> InnerAddOps
2356 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2358 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2360 // Don't leave just a constant behind in a register if the constant could
2361 // be folded into an immediate field.
2362 if (InnerAddOps.size() == 1 &&
2363 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2364 Base.getNumRegs() > 1,
2365 LU.Kind, LU.AccessTy, TLI, SE))
2368 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2369 if (InnerSum->isZero())
2373 // Add the remaining pieces of the add back into the new formula.
2374 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
2375 if (TLI && InnerSumSC &&
2376 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
2377 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2378 InnerSumSC->getValue()->getZExtValue())) {
2379 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2380 InnerSumSC->getValue()->getZExtValue();
2381 F.BaseRegs.erase(F.BaseRegs.begin() + i);
2383 F.BaseRegs[i] = InnerSum;
2385 // Add J as its own register, or an unfolded immediate.
2386 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
2387 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
2388 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2389 SC->getValue()->getZExtValue()))
2390 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2391 SC->getValue()->getZExtValue();
2393 F.BaseRegs.push_back(*J);
2395 if (InsertFormula(LU, LUIdx, F))
2396 // If that formula hadn't been seen before, recurse to find more like
2398 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2403 /// GenerateCombinations - Generate a formula consisting of all of the
2404 /// loop-dominating registers added into a single register.
2405 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2407 // This method is only interesting on a plurality of registers.
2408 if (Base.BaseRegs.size() <= 1) return;
2412 SmallVector<const SCEV *, 4> Ops;
2413 for (SmallVectorImpl<const SCEV *>::const_iterator
2414 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2415 const SCEV *BaseReg = *I;
2416 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2417 !SE.hasComputableLoopEvolution(BaseReg, L))
2418 Ops.push_back(BaseReg);
2420 F.BaseRegs.push_back(BaseReg);
2422 if (Ops.size() > 1) {
2423 const SCEV *Sum = SE.getAddExpr(Ops);
2424 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2425 // opportunity to fold something. For now, just ignore such cases
2426 // rather than proceed with zero in a register.
2427 if (!Sum->isZero()) {
2428 F.BaseRegs.push_back(Sum);
2429 (void)InsertFormula(LU, LUIdx, F);
2434 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2435 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2437 // We can't add a symbolic offset if the address already contains one.
2438 if (Base.AM.BaseGV) return;
2440 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2441 const SCEV *G = Base.BaseRegs[i];
2442 GlobalValue *GV = ExtractSymbol(G, SE);
2443 if (G->isZero() || !GV)
2447 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2448 LU.Kind, LU.AccessTy, TLI))
2451 (void)InsertFormula(LU, LUIdx, F);
2455 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2456 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2458 // TODO: For now, just add the min and max offset, because it usually isn't
2459 // worthwhile looking at everything inbetween.
2460 SmallVector<int64_t, 2> Worklist;
2461 Worklist.push_back(LU.MinOffset);
2462 if (LU.MaxOffset != LU.MinOffset)
2463 Worklist.push_back(LU.MaxOffset);
2465 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2466 const SCEV *G = Base.BaseRegs[i];
2468 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2469 E = Worklist.end(); I != E; ++I) {
2471 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2472 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2473 LU.Kind, LU.AccessTy, TLI)) {
2474 // Add the offset to the base register.
2475 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2476 // If it cancelled out, drop the base register, otherwise update it.
2477 if (NewG->isZero()) {
2478 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2479 F.BaseRegs.pop_back();
2481 F.BaseRegs[i] = NewG;
2483 (void)InsertFormula(LU, LUIdx, F);
2487 int64_t Imm = ExtractImmediate(G, SE);
2488 if (G->isZero() || Imm == 0)
2491 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2492 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2493 LU.Kind, LU.AccessTy, TLI))
2496 (void)InsertFormula(LU, LUIdx, F);
2500 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2501 /// the comparison. For example, x == y -> x*c == y*c.
2502 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2504 if (LU.Kind != LSRUse::ICmpZero) return;
2506 // Determine the integer type for the base formula.
2507 Type *IntTy = Base.getType();
2509 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2511 // Don't do this if there is more than one offset.
2512 if (LU.MinOffset != LU.MaxOffset) return;
2514 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2516 // Check each interesting stride.
2517 for (SmallSetVector<int64_t, 8>::const_iterator
2518 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2519 int64_t Factor = *I;
2521 // Check that the multiplication doesn't overflow.
2522 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2524 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2525 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2528 // Check that multiplying with the use offset doesn't overflow.
2529 int64_t Offset = LU.MinOffset;
2530 if (Offset == INT64_MIN && Factor == -1)
2532 Offset = (uint64_t)Offset * Factor;
2533 if (Offset / Factor != LU.MinOffset)
2537 F.AM.BaseOffs = NewBaseOffs;
2539 // Check that this scale is legal.
2540 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2543 // Compensate for the use having MinOffset built into it.
2544 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2546 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2548 // Check that multiplying with each base register doesn't overflow.
2549 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2550 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2551 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2555 // Check that multiplying with the scaled register doesn't overflow.
2557 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2558 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2562 // Check that multiplying with the unfolded offset doesn't overflow.
2563 if (F.UnfoldedOffset != 0) {
2564 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
2566 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
2567 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
2571 // If we make it here and it's legal, add it.
2572 (void)InsertFormula(LU, LUIdx, F);
2577 /// GenerateScales - Generate stride factor reuse formulae by making use of
2578 /// scaled-offset address modes, for example.
2579 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2580 // Determine the integer type for the base formula.
2581 Type *IntTy = Base.getType();
2584 // If this Formula already has a scaled register, we can't add another one.
2585 if (Base.AM.Scale != 0) return;
2587 // Check each interesting stride.
2588 for (SmallSetVector<int64_t, 8>::const_iterator
2589 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2590 int64_t Factor = *I;
2592 Base.AM.Scale = Factor;
2593 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2594 // Check whether this scale is going to be legal.
2595 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2596 LU.Kind, LU.AccessTy, TLI)) {
2597 // As a special-case, handle special out-of-loop Basic users specially.
2598 // TODO: Reconsider this special case.
2599 if (LU.Kind == LSRUse::Basic &&
2600 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2601 LSRUse::Special, LU.AccessTy, TLI) &&
2602 LU.AllFixupsOutsideLoop)
2603 LU.Kind = LSRUse::Special;
2607 // For an ICmpZero, negating a solitary base register won't lead to
2609 if (LU.Kind == LSRUse::ICmpZero &&
2610 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2612 // For each addrec base reg, apply the scale, if possible.
2613 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2614 if (const SCEVAddRecExpr *AR =
2615 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2616 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2617 if (FactorS->isZero())
2619 // Divide out the factor, ignoring high bits, since we'll be
2620 // scaling the value back up in the end.
2621 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2622 // TODO: This could be optimized to avoid all the copying.
2624 F.ScaledReg = Quotient;
2625 F.DeleteBaseReg(F.BaseRegs[i]);
2626 (void)InsertFormula(LU, LUIdx, F);
2632 /// GenerateTruncates - Generate reuse formulae from different IV types.
2633 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2634 // This requires TargetLowering to tell us which truncates are free.
2637 // Don't bother truncating symbolic values.
2638 if (Base.AM.BaseGV) return;
2640 // Determine the integer type for the base formula.
2641 Type *DstTy = Base.getType();
2643 DstTy = SE.getEffectiveSCEVType(DstTy);
2645 for (SmallSetVector<Type *, 4>::const_iterator
2646 I = Types.begin(), E = Types.end(); I != E; ++I) {
2648 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2651 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2652 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2653 JE = F.BaseRegs.end(); J != JE; ++J)
2654 *J = SE.getAnyExtendExpr(*J, SrcTy);
2656 // TODO: This assumes we've done basic processing on all uses and
2657 // have an idea what the register usage is.
2658 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2661 (void)InsertFormula(LU, LUIdx, F);
2668 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2669 /// defer modifications so that the search phase doesn't have to worry about
2670 /// the data structures moving underneath it.
2674 const SCEV *OrigReg;
2676 WorkItem(size_t LI, int64_t I, const SCEV *R)
2677 : LUIdx(LI), Imm(I), OrigReg(R) {}
2679 void print(raw_ostream &OS) const;
2685 void WorkItem::print(raw_ostream &OS) const {
2686 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2687 << " , add offset " << Imm;
2690 void WorkItem::dump() const {
2691 print(errs()); errs() << '\n';
2694 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2695 /// distance apart and try to form reuse opportunities between them.
2696 void LSRInstance::GenerateCrossUseConstantOffsets() {
2697 // Group the registers by their value without any added constant offset.
2698 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2699 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2701 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2702 SmallVector<const SCEV *, 8> Sequence;
2703 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2705 const SCEV *Reg = *I;
2706 int64_t Imm = ExtractImmediate(Reg, SE);
2707 std::pair<RegMapTy::iterator, bool> Pair =
2708 Map.insert(std::make_pair(Reg, ImmMapTy()));
2710 Sequence.push_back(Reg);
2711 Pair.first->second.insert(std::make_pair(Imm, *I));
2712 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2715 // Now examine each set of registers with the same base value. Build up
2716 // a list of work to do and do the work in a separate step so that we're
2717 // not adding formulae and register counts while we're searching.
2718 SmallVector<WorkItem, 32> WorkItems;
2719 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2720 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2721 E = Sequence.end(); I != E; ++I) {
2722 const SCEV *Reg = *I;
2723 const ImmMapTy &Imms = Map.find(Reg)->second;
2725 // It's not worthwhile looking for reuse if there's only one offset.
2726 if (Imms.size() == 1)
2729 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2730 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2732 dbgs() << ' ' << J->first;
2735 // Examine each offset.
2736 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2738 const SCEV *OrigReg = J->second;
2740 int64_t JImm = J->first;
2741 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2743 if (!isa<SCEVConstant>(OrigReg) &&
2744 UsedByIndicesMap[Reg].count() == 1) {
2745 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2749 // Conservatively examine offsets between this orig reg a few selected
2751 ImmMapTy::const_iterator OtherImms[] = {
2752 Imms.begin(), prior(Imms.end()),
2753 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2755 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2756 ImmMapTy::const_iterator M = OtherImms[i];
2757 if (M == J || M == JE) continue;
2759 // Compute the difference between the two.
2760 int64_t Imm = (uint64_t)JImm - M->first;
2761 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2762 LUIdx = UsedByIndices.find_next(LUIdx))
2763 // Make a memo of this use, offset, and register tuple.
2764 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2765 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2772 UsedByIndicesMap.clear();
2773 UniqueItems.clear();
2775 // Now iterate through the worklist and add new formulae.
2776 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2777 E = WorkItems.end(); I != E; ++I) {
2778 const WorkItem &WI = *I;
2779 size_t LUIdx = WI.LUIdx;
2780 LSRUse &LU = Uses[LUIdx];
2781 int64_t Imm = WI.Imm;
2782 const SCEV *OrigReg = WI.OrigReg;
2784 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2785 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2786 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2788 // TODO: Use a more targeted data structure.
2789 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2790 const Formula &F = LU.Formulae[L];
2791 // Use the immediate in the scaled register.
2792 if (F.ScaledReg == OrigReg) {
2793 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2794 Imm * (uint64_t)F.AM.Scale;
2795 // Don't create 50 + reg(-50).
2796 if (F.referencesReg(SE.getSCEV(
2797 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2800 NewF.AM.BaseOffs = Offs;
2801 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2802 LU.Kind, LU.AccessTy, TLI))
2804 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2806 // If the new scale is a constant in a register, and adding the constant
2807 // value to the immediate would produce a value closer to zero than the
2808 // immediate itself, then the formula isn't worthwhile.
2809 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2810 if (C->getValue()->isNegative() !=
2811 (NewF.AM.BaseOffs < 0) &&
2812 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2813 .ule(abs64(NewF.AM.BaseOffs)))
2817 (void)InsertFormula(LU, LUIdx, NewF);
2819 // Use the immediate in a base register.
2820 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2821 const SCEV *BaseReg = F.BaseRegs[N];
2822 if (BaseReg != OrigReg)
2825 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2826 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2827 LU.Kind, LU.AccessTy, TLI)) {
2829 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
2832 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
2834 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2836 // If the new formula has a constant in a register, and adding the
2837 // constant value to the immediate would produce a value closer to
2838 // zero than the immediate itself, then the formula isn't worthwhile.
2839 for (SmallVectorImpl<const SCEV *>::const_iterator
2840 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2842 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2843 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2844 abs64(NewF.AM.BaseOffs)) &&
2845 (C->getValue()->getValue() +
2846 NewF.AM.BaseOffs).countTrailingZeros() >=
2847 CountTrailingZeros_64(NewF.AM.BaseOffs))
2851 (void)InsertFormula(LU, LUIdx, NewF);
2860 /// GenerateAllReuseFormulae - Generate formulae for each use.
2862 LSRInstance::GenerateAllReuseFormulae() {
2863 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2864 // queries are more precise.
2865 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2866 LSRUse &LU = Uses[LUIdx];
2867 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2868 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2869 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2870 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2872 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2873 LSRUse &LU = Uses[LUIdx];
2874 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2875 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2876 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2877 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2878 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2879 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2880 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2881 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2883 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2884 LSRUse &LU = Uses[LUIdx];
2885 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2886 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2889 GenerateCrossUseConstantOffsets();
2891 DEBUG(dbgs() << "\n"
2892 "After generating reuse formulae:\n";
2893 print_uses(dbgs()));
2896 /// If there are multiple formulae with the same set of registers used
2897 /// by other uses, pick the best one and delete the others.
2898 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2899 DenseSet<const SCEV *> VisitedRegs;
2900 SmallPtrSet<const SCEV *, 16> Regs;
2902 bool ChangedFormulae = false;
2905 // Collect the best formula for each unique set of shared registers. This
2906 // is reset for each use.
2907 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2909 BestFormulaeTy BestFormulae;
2911 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2912 LSRUse &LU = Uses[LUIdx];
2913 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2916 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2917 FIdx != NumForms; ++FIdx) {
2918 Formula &F = LU.Formulae[FIdx];
2920 SmallVector<const SCEV *, 2> Key;
2921 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2922 JE = F.BaseRegs.end(); J != JE; ++J) {
2923 const SCEV *Reg = *J;
2924 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2928 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2929 Key.push_back(F.ScaledReg);
2930 // Unstable sort by host order ok, because this is only used for
2932 std::sort(Key.begin(), Key.end());
2934 std::pair<BestFormulaeTy::const_iterator, bool> P =
2935 BestFormulae.insert(std::make_pair(Key, FIdx));
2937 Formula &Best = LU.Formulae[P.first->second];
2940 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2943 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2945 if (CostF < CostBest)
2947 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2949 " in favor of formula "; Best.print(dbgs());
2952 ChangedFormulae = true;
2954 LU.DeleteFormula(F);
2962 // Now that we've filtered out some formulae, recompute the Regs set.
2964 LU.RecomputeRegs(LUIdx, RegUses);
2966 // Reset this to prepare for the next use.
2967 BestFormulae.clear();
2970 DEBUG(if (ChangedFormulae) {
2972 "After filtering out undesirable candidates:\n";
2977 // This is a rough guess that seems to work fairly well.
2978 static const size_t ComplexityLimit = UINT16_MAX;
2980 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2981 /// solutions the solver might have to consider. It almost never considers
2982 /// this many solutions because it prune the search space, but the pruning
2983 /// isn't always sufficient.
2984 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2986 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2987 E = Uses.end(); I != E; ++I) {
2988 size_t FSize = I->Formulae.size();
2989 if (FSize >= ComplexityLimit) {
2990 Power = ComplexityLimit;
2994 if (Power >= ComplexityLimit)
3000 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3001 /// of the registers of another formula, it won't help reduce register
3002 /// pressure (though it may not necessarily hurt register pressure); remove
3003 /// it to simplify the system.
3004 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3005 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3006 DEBUG(dbgs() << "The search space is too complex.\n");
3008 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3009 "which use a superset of registers used by other "
3012 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3013 LSRUse &LU = Uses[LUIdx];
3015 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3016 Formula &F = LU.Formulae[i];
3017 // Look for a formula with a constant or GV in a register. If the use
3018 // also has a formula with that same value in an immediate field,
3019 // delete the one that uses a register.
3020 for (SmallVectorImpl<const SCEV *>::const_iterator
3021 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3022 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3024 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3025 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3026 (I - F.BaseRegs.begin()));
3027 if (LU.HasFormulaWithSameRegs(NewF)) {
3028 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3029 LU.DeleteFormula(F);
3035 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3036 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3039 NewF.AM.BaseGV = GV;
3040 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3041 (I - F.BaseRegs.begin()));
3042 if (LU.HasFormulaWithSameRegs(NewF)) {
3043 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3045 LU.DeleteFormula(F);
3056 LU.RecomputeRegs(LUIdx, RegUses);
3059 DEBUG(dbgs() << "After pre-selection:\n";
3060 print_uses(dbgs()));
3064 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3065 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3067 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3068 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3069 DEBUG(dbgs() << "The search space is too complex.\n");
3071 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3072 "separated by a constant offset will use the same "
3075 // This is especially useful for unrolled loops.
3077 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3078 LSRUse &LU = Uses[LUIdx];
3079 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3080 E = LU.Formulae.end(); I != E; ++I) {
3081 const Formula &F = *I;
3082 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3083 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3084 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3085 /*HasBaseReg=*/false,
3086 LU.Kind, LU.AccessTy)) {
3087 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3090 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3092 // Update the relocs to reference the new use.
3093 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3094 E = Fixups.end(); I != E; ++I) {
3095 LSRFixup &Fixup = *I;
3096 if (Fixup.LUIdx == LUIdx) {
3097 Fixup.LUIdx = LUThatHas - &Uses.front();
3098 Fixup.Offset += F.AM.BaseOffs;
3099 // Add the new offset to LUThatHas' offset list.
3100 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3101 LUThatHas->Offsets.push_back(Fixup.Offset);
3102 if (Fixup.Offset > LUThatHas->MaxOffset)
3103 LUThatHas->MaxOffset = Fixup.Offset;
3104 if (Fixup.Offset < LUThatHas->MinOffset)
3105 LUThatHas->MinOffset = Fixup.Offset;
3107 DEBUG(dbgs() << "New fixup has offset "
3108 << Fixup.Offset << '\n');
3110 if (Fixup.LUIdx == NumUses-1)
3111 Fixup.LUIdx = LUIdx;
3114 // Delete formulae from the new use which are no longer legal.
3116 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3117 Formula &F = LUThatHas->Formulae[i];
3118 if (!isLegalUse(F.AM,
3119 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3120 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3121 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3123 LUThatHas->DeleteFormula(F);
3130 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3132 // Delete the old use.
3133 DeleteUse(LU, LUIdx);
3143 DEBUG(dbgs() << "After pre-selection:\n";
3144 print_uses(dbgs()));
3148 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3149 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3150 /// we've done more filtering, as it may be able to find more formulae to
3152 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3153 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3154 DEBUG(dbgs() << "The search space is too complex.\n");
3156 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3157 "undesirable dedicated registers.\n");
3159 FilterOutUndesirableDedicatedRegisters();
3161 DEBUG(dbgs() << "After pre-selection:\n";
3162 print_uses(dbgs()));
3166 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3167 /// to be profitable, and then in any use which has any reference to that
3168 /// register, delete all formulae which do not reference that register.
3169 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3170 // With all other options exhausted, loop until the system is simple
3171 // enough to handle.
3172 SmallPtrSet<const SCEV *, 4> Taken;
3173 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3174 // Ok, we have too many of formulae on our hands to conveniently handle.
3175 // Use a rough heuristic to thin out the list.
3176 DEBUG(dbgs() << "The search space is too complex.\n");
3178 // Pick the register which is used by the most LSRUses, which is likely
3179 // to be a good reuse register candidate.
3180 const SCEV *Best = 0;
3181 unsigned BestNum = 0;
3182 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3184 const SCEV *Reg = *I;
3185 if (Taken.count(Reg))
3190 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3191 if (Count > BestNum) {
3198 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3199 << " will yield profitable reuse.\n");
3202 // In any use with formulae which references this register, delete formulae
3203 // which don't reference it.
3204 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3205 LSRUse &LU = Uses[LUIdx];
3206 if (!LU.Regs.count(Best)) continue;
3209 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3210 Formula &F = LU.Formulae[i];
3211 if (!F.referencesReg(Best)) {
3212 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3213 LU.DeleteFormula(F);
3217 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3223 LU.RecomputeRegs(LUIdx, RegUses);
3226 DEBUG(dbgs() << "After pre-selection:\n";
3227 print_uses(dbgs()));
3231 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3232 /// formulae to choose from, use some rough heuristics to prune down the number
3233 /// of formulae. This keeps the main solver from taking an extraordinary amount
3234 /// of time in some worst-case scenarios.
3235 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3236 NarrowSearchSpaceByDetectingSupersets();
3237 NarrowSearchSpaceByCollapsingUnrolledCode();
3238 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3239 NarrowSearchSpaceByPickingWinnerRegs();
3242 /// SolveRecurse - This is the recursive solver.
3243 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3245 SmallVectorImpl<const Formula *> &Workspace,
3246 const Cost &CurCost,
3247 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3248 DenseSet<const SCEV *> &VisitedRegs) const {
3251 // - use more aggressive filtering
3252 // - sort the formula so that the most profitable solutions are found first
3253 // - sort the uses too
3255 // - don't compute a cost, and then compare. compare while computing a cost
3257 // - track register sets with SmallBitVector
3259 const LSRUse &LU = Uses[Workspace.size()];
3261 // If this use references any register that's already a part of the
3262 // in-progress solution, consider it a requirement that a formula must
3263 // reference that register in order to be considered. This prunes out
3264 // unprofitable searching.
3265 SmallSetVector<const SCEV *, 4> ReqRegs;
3266 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3267 E = CurRegs.end(); I != E; ++I)
3268 if (LU.Regs.count(*I))
3271 bool AnySatisfiedReqRegs = false;
3272 SmallPtrSet<const SCEV *, 16> NewRegs;
3275 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3276 E = LU.Formulae.end(); I != E; ++I) {
3277 const Formula &F = *I;
3279 // Ignore formulae which do not use any of the required registers.
3280 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3281 JE = ReqRegs.end(); J != JE; ++J) {
3282 const SCEV *Reg = *J;
3283 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3284 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3288 AnySatisfiedReqRegs = true;
3290 // Evaluate the cost of the current formula. If it's already worse than
3291 // the current best, prune the search at that point.
3294 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3295 if (NewCost < SolutionCost) {
3296 Workspace.push_back(&F);
3297 if (Workspace.size() != Uses.size()) {
3298 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3299 NewRegs, VisitedRegs);
3300 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3301 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3303 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3304 dbgs() << ". Regs:";
3305 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3306 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3307 dbgs() << ' ' << **I;
3310 SolutionCost = NewCost;
3311 Solution = Workspace;
3313 Workspace.pop_back();
3318 if (!EnableRetry && !AnySatisfiedReqRegs)
3321 // If none of the formulae had all of the required registers, relax the
3322 // constraint so that we don't exclude all formulae.
3323 if (!AnySatisfiedReqRegs) {
3324 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3330 /// Solve - Choose one formula from each use. Return the results in the given
3331 /// Solution vector.
3332 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3333 SmallVector<const Formula *, 8> Workspace;
3335 SolutionCost.Loose();
3337 SmallPtrSet<const SCEV *, 16> CurRegs;
3338 DenseSet<const SCEV *> VisitedRegs;
3339 Workspace.reserve(Uses.size());
3341 // SolveRecurse does all the work.
3342 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3343 CurRegs, VisitedRegs);
3344 if (Solution.empty()) {
3345 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
3349 // Ok, we've now made all our decisions.
3350 DEBUG(dbgs() << "\n"
3351 "The chosen solution requires "; SolutionCost.print(dbgs());
3353 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3355 Uses[i].print(dbgs());
3358 Solution[i]->print(dbgs());
3362 assert(Solution.size() == Uses.size() && "Malformed solution!");
3365 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3366 /// the dominator tree far as we can go while still being dominated by the
3367 /// input positions. This helps canonicalize the insert position, which
3368 /// encourages sharing.
3369 BasicBlock::iterator
3370 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3371 const SmallVectorImpl<Instruction *> &Inputs)
3374 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3375 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3378 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3379 if (!Rung) return IP;
3380 Rung = Rung->getIDom();
3381 if (!Rung) return IP;
3382 IDom = Rung->getBlock();
3384 // Don't climb into a loop though.
3385 const Loop *IDomLoop = LI.getLoopFor(IDom);
3386 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3387 if (IDomDepth <= IPLoopDepth &&
3388 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3392 bool AllDominate = true;
3393 Instruction *BetterPos = 0;
3394 Instruction *Tentative = IDom->getTerminator();
3395 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3396 E = Inputs.end(); I != E; ++I) {
3397 Instruction *Inst = *I;
3398 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3399 AllDominate = false;
3402 // Attempt to find an insert position in the middle of the block,
3403 // instead of at the end, so that it can be used for other expansions.
3404 if (IDom == Inst->getParent() &&
3405 (!BetterPos || DT.dominates(BetterPos, Inst)))
3406 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3419 /// AdjustInsertPositionForExpand - Determine an input position which will be
3420 /// dominated by the operands and which will dominate the result.
3421 BasicBlock::iterator
3422 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3424 const LSRUse &LU) const {
3425 // Collect some instructions which must be dominated by the
3426 // expanding replacement. These must be dominated by any operands that
3427 // will be required in the expansion.
3428 SmallVector<Instruction *, 4> Inputs;
3429 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3430 Inputs.push_back(I);
3431 if (LU.Kind == LSRUse::ICmpZero)
3432 if (Instruction *I =
3433 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3434 Inputs.push_back(I);
3435 if (LF.PostIncLoops.count(L)) {
3436 if (LF.isUseFullyOutsideLoop(L))
3437 Inputs.push_back(L->getLoopLatch()->getTerminator());
3439 Inputs.push_back(IVIncInsertPos);
3441 // The expansion must also be dominated by the increment positions of any
3442 // loops it for which it is using post-inc mode.
3443 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3444 E = LF.PostIncLoops.end(); I != E; ++I) {
3445 const Loop *PIL = *I;
3446 if (PIL == L) continue;
3448 // Be dominated by the loop exit.
3449 SmallVector<BasicBlock *, 4> ExitingBlocks;
3450 PIL->getExitingBlocks(ExitingBlocks);
3451 if (!ExitingBlocks.empty()) {
3452 BasicBlock *BB = ExitingBlocks[0];
3453 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3454 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3455 Inputs.push_back(BB->getTerminator());
3459 // Then, climb up the immediate dominator tree as far as we can go while
3460 // still being dominated by the input positions.
3461 IP = HoistInsertPosition(IP, Inputs);
3463 // Don't insert instructions before PHI nodes.
3464 while (isa<PHINode>(IP)) ++IP;
3466 // Ignore landingpad instructions.
3467 while (isa<LandingPadInst>(IP)) ++IP;
3469 // Ignore debug intrinsics.
3470 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3475 /// Expand - Emit instructions for the leading candidate expression for this
3476 /// LSRUse (this is called "expanding").
3477 Value *LSRInstance::Expand(const LSRFixup &LF,
3479 BasicBlock::iterator IP,
3480 SCEVExpander &Rewriter,
3481 SmallVectorImpl<WeakVH> &DeadInsts) const {
3482 const LSRUse &LU = Uses[LF.LUIdx];
3484 // Determine an input position which will be dominated by the operands and
3485 // which will dominate the result.
3486 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3488 // Inform the Rewriter if we have a post-increment use, so that it can
3489 // perform an advantageous expansion.
3490 Rewriter.setPostInc(LF.PostIncLoops);
3492 // This is the type that the user actually needs.
3493 Type *OpTy = LF.OperandValToReplace->getType();
3494 // This will be the type that we'll initially expand to.
3495 Type *Ty = F.getType();
3497 // No type known; just expand directly to the ultimate type.
3499 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3500 // Expand directly to the ultimate type if it's the right size.
3502 // This is the type to do integer arithmetic in.
3503 Type *IntTy = SE.getEffectiveSCEVType(Ty);
3505 // Build up a list of operands to add together to form the full base.
3506 SmallVector<const SCEV *, 8> Ops;
3508 // Expand the BaseRegs portion.
3509 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3510 E = F.BaseRegs.end(); I != E; ++I) {
3511 const SCEV *Reg = *I;
3512 assert(!Reg->isZero() && "Zero allocated in a base register!");
3514 // If we're expanding for a post-inc user, make the post-inc adjustment.
3515 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3516 Reg = TransformForPostIncUse(Denormalize, Reg,
3517 LF.UserInst, LF.OperandValToReplace,
3520 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3523 // Flush the operand list to suppress SCEVExpander hoisting.
3525 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3527 Ops.push_back(SE.getUnknown(FullV));
3530 // Expand the ScaledReg portion.
3531 Value *ICmpScaledV = 0;
3532 if (F.AM.Scale != 0) {
3533 const SCEV *ScaledS = F.ScaledReg;
3535 // If we're expanding for a post-inc user, make the post-inc adjustment.
3536 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3537 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3538 LF.UserInst, LF.OperandValToReplace,
3541 if (LU.Kind == LSRUse::ICmpZero) {
3542 // An interesting way of "folding" with an icmp is to use a negated
3543 // scale, which we'll implement by inserting it into the other operand
3545 assert(F.AM.Scale == -1 &&
3546 "The only scale supported by ICmpZero uses is -1!");
3547 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3549 // Otherwise just expand the scaled register and an explicit scale,
3550 // which is expected to be matched as part of the address.
3551 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3552 ScaledS = SE.getMulExpr(ScaledS,
3553 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3554 Ops.push_back(ScaledS);
3556 // Flush the operand list to suppress SCEVExpander hoisting.
3557 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3559 Ops.push_back(SE.getUnknown(FullV));
3563 // Expand the GV portion.
3565 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3567 // Flush the operand list to suppress SCEVExpander hoisting.
3568 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3570 Ops.push_back(SE.getUnknown(FullV));
3573 // Expand the immediate portion.
3574 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3576 if (LU.Kind == LSRUse::ICmpZero) {
3577 // The other interesting way of "folding" with an ICmpZero is to use a
3578 // negated immediate.
3580 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3582 Ops.push_back(SE.getUnknown(ICmpScaledV));
3583 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3586 // Just add the immediate values. These again are expected to be matched
3587 // as part of the address.
3588 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3592 // Expand the unfolded offset portion.
3593 int64_t UnfoldedOffset = F.UnfoldedOffset;
3594 if (UnfoldedOffset != 0) {
3595 // Just add the immediate values.
3596 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
3600 // Emit instructions summing all the operands.
3601 const SCEV *FullS = Ops.empty() ?
3602 SE.getConstant(IntTy, 0) :
3604 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3606 // We're done expanding now, so reset the rewriter.
3607 Rewriter.clearPostInc();
3609 // An ICmpZero Formula represents an ICmp which we're handling as a
3610 // comparison against zero. Now that we've expanded an expression for that
3611 // form, update the ICmp's other operand.
3612 if (LU.Kind == LSRUse::ICmpZero) {
3613 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3614 DeadInsts.push_back(CI->getOperand(1));
3615 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3616 "a scale at the same time!");
3617 if (F.AM.Scale == -1) {
3618 if (ICmpScaledV->getType() != OpTy) {
3620 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3622 ICmpScaledV, OpTy, "tmp", CI);
3625 CI->setOperand(1, ICmpScaledV);
3627 assert(F.AM.Scale == 0 &&
3628 "ICmp does not support folding a global value and "
3629 "a scale at the same time!");
3630 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3632 if (C->getType() != OpTy)
3633 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3637 CI->setOperand(1, C);
3644 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3645 /// of their operands effectively happens in their predecessor blocks, so the
3646 /// expression may need to be expanded in multiple places.
3647 void LSRInstance::RewriteForPHI(PHINode *PN,
3650 SCEVExpander &Rewriter,
3651 SmallVectorImpl<WeakVH> &DeadInsts,
3653 DenseMap<BasicBlock *, Value *> Inserted;
3654 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3655 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3656 BasicBlock *BB = PN->getIncomingBlock(i);
3658 // If this is a critical edge, split the edge so that we do not insert
3659 // the code on all predecessor/successor paths. We do this unless this
3660 // is the canonical backedge for this loop, which complicates post-inc
3662 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3663 !isa<IndirectBrInst>(BB->getTerminator())) {
3664 BasicBlock *Parent = PN->getParent();
3665 Loop *PNLoop = LI.getLoopFor(Parent);
3666 if (!PNLoop || Parent != PNLoop->getHeader()) {
3667 // Split the critical edge.
3668 BasicBlock *NewBB = 0;
3669 if (!Parent->isLandingPad()) {
3670 NewBB = SplitCriticalEdge(BB, Parent, P);
3672 SmallVector<BasicBlock*, 2> NewBBs;
3673 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
3677 // If PN is outside of the loop and BB is in the loop, we want to
3678 // move the block to be immediately before the PHI block, not
3679 // immediately after BB.
3680 if (L->contains(BB) && !L->contains(PN))
3681 NewBB->moveBefore(PN->getParent());
3683 // Splitting the edge can reduce the number of PHI entries we have.
3684 e = PN->getNumIncomingValues();
3686 i = PN->getBasicBlockIndex(BB);
3690 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3691 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3693 PN->setIncomingValue(i, Pair.first->second);
3695 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3697 // If this is reuse-by-noop-cast, insert the noop cast.
3698 Type *OpTy = LF.OperandValToReplace->getType();
3699 if (FullV->getType() != OpTy)
3701 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3703 FullV, LF.OperandValToReplace->getType(),
3704 "tmp", BB->getTerminator());
3706 PN->setIncomingValue(i, FullV);
3707 Pair.first->second = FullV;
3712 /// Rewrite - Emit instructions for the leading candidate expression for this
3713 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3714 /// the newly expanded value.
3715 void LSRInstance::Rewrite(const LSRFixup &LF,
3717 SCEVExpander &Rewriter,
3718 SmallVectorImpl<WeakVH> &DeadInsts,
3720 // First, find an insertion point that dominates UserInst. For PHI nodes,
3721 // find the nearest block which dominates all the relevant uses.
3722 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3723 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3725 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3727 // If this is reuse-by-noop-cast, insert the noop cast.
3728 Type *OpTy = LF.OperandValToReplace->getType();
3729 if (FullV->getType() != OpTy) {
3731 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3732 FullV, OpTy, "tmp", LF.UserInst);
3736 // Update the user. ICmpZero is handled specially here (for now) because
3737 // Expand may have updated one of the operands of the icmp already, and
3738 // its new value may happen to be equal to LF.OperandValToReplace, in
3739 // which case doing replaceUsesOfWith leads to replacing both operands
3740 // with the same value. TODO: Reorganize this.
3741 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3742 LF.UserInst->setOperand(0, FullV);
3744 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3747 DeadInsts.push_back(LF.OperandValToReplace);
3750 /// ImplementSolution - Rewrite all the fixup locations with new values,
3751 /// following the chosen solution.
3753 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3755 // Keep track of instructions we may have made dead, so that
3756 // we can remove them after we are done working.
3757 SmallVector<WeakVH, 16> DeadInsts;
3759 SCEVExpander Rewriter(SE, "lsr");
3760 Rewriter.disableCanonicalMode();
3761 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3763 // Expand the new value definitions and update the users.
3764 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3765 E = Fixups.end(); I != E; ++I) {
3766 const LSRFixup &Fixup = *I;
3768 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3773 // Clean up after ourselves. This must be done before deleting any
3777 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3780 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3781 : IU(P->getAnalysis<IVUsers>()),
3782 SE(P->getAnalysis<ScalarEvolution>()),
3783 DT(P->getAnalysis<DominatorTree>()),
3784 LI(P->getAnalysis<LoopInfo>()),
3785 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3787 // If LoopSimplify form is not available, stay out of trouble.
3788 if (!L->isLoopSimplifyForm()) return;
3790 // If there's no interesting work to be done, bail early.
3791 if (IU.empty()) return;
3793 DEBUG(dbgs() << "\nLSR on loop ";
3794 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3797 // First, perform some low-level loop optimizations.
3799 OptimizeLoopTermCond();
3801 // If loop preparation eliminates all interesting IV users, bail.
3802 if (IU.empty()) return;
3804 // Start collecting data and preparing for the solver.
3805 CollectInterestingTypesAndFactors();
3806 CollectFixupsAndInitialFormulae();
3807 CollectLoopInvariantFixupsAndFormulae();
3809 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3810 print_uses(dbgs()));
3812 // Now use the reuse data to generate a bunch of interesting ways
3813 // to formulate the values needed for the uses.
3814 GenerateAllReuseFormulae();
3816 FilterOutUndesirableDedicatedRegisters();
3817 NarrowSearchSpaceUsingHeuristics();
3819 SmallVector<const Formula *, 8> Solution;
3822 // Release memory that is no longer needed.
3827 if (Solution.empty())
3831 // Formulae should be legal.
3832 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3833 E = Uses.end(); I != E; ++I) {
3834 const LSRUse &LU = *I;
3835 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3836 JE = LU.Formulae.end(); J != JE; ++J)
3837 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3838 LU.Kind, LU.AccessTy, TLI) &&
3839 "Illegal formula generated!");
3843 // Now that we've decided what we want, make it so.
3844 ImplementSolution(Solution, P);
3847 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3848 if (Factors.empty() && Types.empty()) return;
3850 OS << "LSR has identified the following interesting factors and types: ";
3853 for (SmallSetVector<int64_t, 8>::const_iterator
3854 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3855 if (!First) OS << ", ";
3860 for (SmallSetVector<Type *, 4>::const_iterator
3861 I = Types.begin(), E = Types.end(); I != E; ++I) {
3862 if (!First) OS << ", ";
3864 OS << '(' << **I << ')';
3869 void LSRInstance::print_fixups(raw_ostream &OS) const {
3870 OS << "LSR is examining the following fixup sites:\n";
3871 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3872 E = Fixups.end(); I != E; ++I) {
3879 void LSRInstance::print_uses(raw_ostream &OS) const {
3880 OS << "LSR is examining the following uses:\n";
3881 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3882 E = Uses.end(); I != E; ++I) {
3883 const LSRUse &LU = *I;
3887 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3888 JE = LU.Formulae.end(); J != JE; ++J) {
3896 void LSRInstance::print(raw_ostream &OS) const {
3897 print_factors_and_types(OS);
3902 void LSRInstance::dump() const {
3903 print(errs()); errs() << '\n';
3908 class LoopStrengthReduce : public LoopPass {
3909 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3910 /// transformation profitability.
3911 const TargetLowering *const TLI;
3914 static char ID; // Pass ID, replacement for typeid
3915 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3918 bool runOnLoop(Loop *L, LPPassManager &LPM);
3919 void getAnalysisUsage(AnalysisUsage &AU) const;
3924 char LoopStrengthReduce::ID = 0;
3925 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
3926 "Loop Strength Reduction", false, false)
3927 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3928 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3929 INITIALIZE_PASS_DEPENDENCY(IVUsers)
3930 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
3931 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3932 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
3933 "Loop Strength Reduction", false, false)
3936 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3937 return new LoopStrengthReduce(TLI);
3940 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3941 : LoopPass(ID), TLI(tli) {
3942 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
3945 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3946 // We split critical edges, so we change the CFG. However, we do update
3947 // many analyses if they are around.
3948 AU.addPreservedID(LoopSimplifyID);
3950 AU.addRequired<LoopInfo>();
3951 AU.addPreserved<LoopInfo>();
3952 AU.addRequiredID(LoopSimplifyID);
3953 AU.addRequired<DominatorTree>();
3954 AU.addPreserved<DominatorTree>();
3955 AU.addRequired<ScalarEvolution>();
3956 AU.addPreserved<ScalarEvolution>();
3957 // Requiring LoopSimplify a second time here prevents IVUsers from running
3958 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
3959 AU.addRequiredID(LoopSimplifyID);
3960 AU.addRequired<IVUsers>();
3961 AU.addPreserved<IVUsers>();
3964 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3965 bool Changed = false;
3967 // Run the main LSR transformation.
3968 Changed |= LSRInstance(TLI, L, this).getChanged();
3970 // At this point, it is worth checking to see if any recurrence PHIs are also
3971 // dead, so that we can remove them as well.
3972 Changed |= DeleteDeadPHIs(L->getHeader());