1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Assembly/Writer.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/ADT/SmallBitVector.h"
70 #include "llvm/ADT/SetVector.h"
71 #include "llvm/ADT/DenseSet.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/ValueHandle.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Target/TargetLowering.h"
81 /// RegSortData - This class holds data which is used to order reuse candidates.
84 /// UsedByIndices - This represents the set of LSRUse indices which reference
85 /// a particular register.
86 SmallBitVector UsedByIndices;
90 void print(raw_ostream &OS) const;
96 void RegSortData::print(raw_ostream &OS) const {
97 OS << "[NumUses=" << UsedByIndices.count() << ']';
100 void RegSortData::dump() const {
101 print(errs()); errs() << '\n';
106 /// RegUseTracker - Map register candidates to information about how they are
108 class RegUseTracker {
109 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
111 RegUsesTy RegUsesMap;
112 SmallVector<const SCEV *, 16> RegSequence;
115 void CountRegister(const SCEV *Reg, size_t LUIdx);
116 void DropRegister(const SCEV *Reg, size_t LUIdx);
117 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
119 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
121 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
125 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
126 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
127 iterator begin() { return RegSequence.begin(); }
128 iterator end() { return RegSequence.end(); }
129 const_iterator begin() const { return RegSequence.begin(); }
130 const_iterator end() const { return RegSequence.end(); }
136 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
137 std::pair<RegUsesTy::iterator, bool> Pair =
138 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
139 RegSortData &RSD = Pair.first->second;
141 RegSequence.push_back(Reg);
142 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
143 RSD.UsedByIndices.set(LUIdx);
147 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
148 RegUsesTy::iterator It = RegUsesMap.find(Reg);
149 assert(It != RegUsesMap.end());
150 RegSortData &RSD = It->second;
151 assert(RSD.UsedByIndices.size() > LUIdx);
152 RSD.UsedByIndices.reset(LUIdx);
156 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
157 assert(LUIdx <= LastLUIdx);
159 // Update RegUses. The data structure is not optimized for this purpose;
160 // we must iterate through it and update each of the bit vectors.
161 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
163 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
164 if (LUIdx < UsedByIndices.size())
165 UsedByIndices[LUIdx] =
166 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
167 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
172 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
173 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
174 if (I == RegUsesMap.end())
176 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
177 int i = UsedByIndices.find_first();
178 if (i == -1) return false;
179 if ((size_t)i != LUIdx) return true;
180 return UsedByIndices.find_next(i) != -1;
183 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
184 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
185 assert(I != RegUsesMap.end() && "Unknown register!");
186 return I->second.UsedByIndices;
189 void RegUseTracker::clear() {
196 /// Formula - This class holds information that describes a formula for
197 /// computing satisfying a use. It may include broken-out immediates and scaled
200 /// AM - This is used to represent complex addressing, as well as other kinds
201 /// of interesting uses.
202 TargetLowering::AddrMode AM;
204 /// BaseRegs - The list of "base" registers for this use. When this is
205 /// non-empty, AM.HasBaseReg should be set to true.
206 SmallVector<const SCEV *, 2> BaseRegs;
208 /// ScaledReg - The 'scaled' register for this use. This should be non-null
209 /// when AM.Scale is not zero.
210 const SCEV *ScaledReg;
212 /// UnfoldedOffset - An additional constant offset which added near the
213 /// use. This requires a temporary register, but the offset itself can
214 /// live in an add immediate field rather than a register.
215 int64_t UnfoldedOffset;
217 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
219 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
221 unsigned getNumRegs() const;
222 Type *getType() const;
224 void DeleteBaseReg(const SCEV *&S);
226 bool referencesReg(const SCEV *S) const;
227 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
228 const RegUseTracker &RegUses) const;
230 void print(raw_ostream &OS) const;
236 /// DoInitialMatch - Recursion helper for InitialMatch.
237 static void DoInitialMatch(const SCEV *S, Loop *L,
238 SmallVectorImpl<const SCEV *> &Good,
239 SmallVectorImpl<const SCEV *> &Bad,
240 ScalarEvolution &SE) {
241 // Collect expressions which properly dominate the loop header.
242 if (SE.properlyDominates(S, L->getHeader())) {
247 // Look at add operands.
248 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
249 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
251 DoInitialMatch(*I, L, Good, Bad, SE);
255 // Look at addrec operands.
256 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
257 if (!AR->getStart()->isZero()) {
258 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
259 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
260 AR->getStepRecurrence(SE),
261 // FIXME: AR->getNoWrapFlags()
262 AR->getLoop(), SCEV::FlagAnyWrap),
267 // Handle a multiplication by -1 (negation) if it didn't fold.
268 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
269 if (Mul->getOperand(0)->isAllOnesValue()) {
270 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
271 const SCEV *NewMul = SE.getMulExpr(Ops);
273 SmallVector<const SCEV *, 4> MyGood;
274 SmallVector<const SCEV *, 4> MyBad;
275 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
276 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
277 SE.getEffectiveSCEVType(NewMul->getType())));
278 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
279 E = MyGood.end(); I != E; ++I)
280 Good.push_back(SE.getMulExpr(NegOne, *I));
281 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
282 E = MyBad.end(); I != E; ++I)
283 Bad.push_back(SE.getMulExpr(NegOne, *I));
287 // Ok, we can't do anything interesting. Just stuff the whole thing into a
288 // register and hope for the best.
292 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
293 /// attempting to keep all loop-invariant and loop-computable values in a
294 /// single base register.
295 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
296 SmallVector<const SCEV *, 4> Good;
297 SmallVector<const SCEV *, 4> Bad;
298 DoInitialMatch(S, L, Good, Bad, SE);
300 const SCEV *Sum = SE.getAddExpr(Good);
302 BaseRegs.push_back(Sum);
303 AM.HasBaseReg = true;
306 const SCEV *Sum = SE.getAddExpr(Bad);
308 BaseRegs.push_back(Sum);
309 AM.HasBaseReg = true;
313 /// getNumRegs - Return the total number of register operands used by this
314 /// formula. This does not include register uses implied by non-constant
316 unsigned Formula::getNumRegs() const {
317 return !!ScaledReg + BaseRegs.size();
320 /// getType - Return the type of this formula, if it has one, or null
321 /// otherwise. This type is meaningless except for the bit size.
322 Type *Formula::getType() const {
323 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
324 ScaledReg ? ScaledReg->getType() :
325 AM.BaseGV ? AM.BaseGV->getType() :
329 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
330 void Formula::DeleteBaseReg(const SCEV *&S) {
331 if (&S != &BaseRegs.back())
332 std::swap(S, BaseRegs.back());
336 /// referencesReg - Test if this formula references the given register.
337 bool Formula::referencesReg(const SCEV *S) const {
338 return S == ScaledReg ||
339 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
342 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
343 /// which are used by uses other than the use with the given index.
344 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
345 const RegUseTracker &RegUses) const {
347 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
349 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
350 E = BaseRegs.end(); I != E; ++I)
351 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
356 void Formula::print(raw_ostream &OS) const {
359 if (!First) OS << " + "; else First = false;
360 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
362 if (AM.BaseOffs != 0) {
363 if (!First) OS << " + "; else First = false;
366 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
367 E = BaseRegs.end(); I != E; ++I) {
368 if (!First) OS << " + "; else First = false;
369 OS << "reg(" << **I << ')';
371 if (AM.HasBaseReg && BaseRegs.empty()) {
372 if (!First) OS << " + "; else First = false;
373 OS << "**error: HasBaseReg**";
374 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
375 if (!First) OS << " + "; else First = false;
376 OS << "**error: !HasBaseReg**";
379 if (!First) OS << " + "; else First = false;
380 OS << AM.Scale << "*reg(";
387 if (UnfoldedOffset != 0) {
388 if (!First) OS << " + "; else First = false;
389 OS << "imm(" << UnfoldedOffset << ')';
393 void Formula::dump() const {
394 print(errs()); errs() << '\n';
397 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
398 /// without changing its value.
399 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
401 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
402 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
405 /// isAddSExtable - Return true if the given add can be sign-extended
406 /// without changing its value.
407 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
409 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
410 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
413 /// isMulSExtable - Return true if the given mul can be sign-extended
414 /// without changing its value.
415 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
417 IntegerType::get(SE.getContext(),
418 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
419 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
422 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
423 /// and if the remainder is known to be zero, or null otherwise. If
424 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
425 /// to Y, ignoring that the multiplication may overflow, which is useful when
426 /// the result will be used in a context where the most significant bits are
428 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
430 bool IgnoreSignificantBits = false) {
431 // Handle the trivial case, which works for any SCEV type.
433 return SE.getConstant(LHS->getType(), 1);
435 // Handle a few RHS special cases.
436 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
438 const APInt &RA = RC->getValue()->getValue();
439 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
441 if (RA.isAllOnesValue())
442 return SE.getMulExpr(LHS, RC);
443 // Handle x /s 1 as x.
448 // Check for a division of a constant by a constant.
449 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
452 const APInt &LA = C->getValue()->getValue();
453 const APInt &RA = RC->getValue()->getValue();
454 if (LA.srem(RA) != 0)
456 return SE.getConstant(LA.sdiv(RA));
459 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
460 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
461 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
462 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
463 IgnoreSignificantBits);
465 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
466 IgnoreSignificantBits);
467 if (!Start) return 0;
468 // FlagNW is independent of the start value, step direction, and is
469 // preserved with smaller magnitude steps.
470 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
471 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
476 // Distribute the sdiv over add operands, if the add doesn't overflow.
477 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
478 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
479 SmallVector<const SCEV *, 8> Ops;
480 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
482 const SCEV *Op = getExactSDiv(*I, RHS, SE,
483 IgnoreSignificantBits);
487 return SE.getAddExpr(Ops);
492 // Check for a multiply operand that we can pull RHS out of.
493 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
494 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
495 SmallVector<const SCEV *, 4> Ops;
497 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
501 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
502 IgnoreSignificantBits)) {
508 return Found ? SE.getMulExpr(Ops) : 0;
513 // Otherwise we don't know.
517 /// ExtractImmediate - If S involves the addition of a constant integer value,
518 /// return that integer value, and mutate S to point to a new SCEV with that
520 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
521 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
522 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
523 S = SE.getConstant(C->getType(), 0);
524 return C->getValue()->getSExtValue();
526 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
527 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
528 int64_t Result = ExtractImmediate(NewOps.front(), SE);
530 S = SE.getAddExpr(NewOps);
532 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
533 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
534 int64_t Result = ExtractImmediate(NewOps.front(), SE);
536 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
537 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
544 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
545 /// return that symbol, and mutate S to point to a new SCEV with that
547 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
548 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
549 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
550 S = SE.getConstant(GV->getType(), 0);
553 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
554 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
555 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
557 S = SE.getAddExpr(NewOps);
559 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
560 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
561 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
563 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
564 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
571 /// isAddressUse - Returns true if the specified instruction is using the
572 /// specified value as an address.
573 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
574 bool isAddress = isa<LoadInst>(Inst);
575 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
576 if (SI->getOperand(1) == OperandVal)
578 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
579 // Addressing modes can also be folded into prefetches and a variety
581 switch (II->getIntrinsicID()) {
583 case Intrinsic::prefetch:
584 case Intrinsic::x86_sse_storeu_ps:
585 case Intrinsic::x86_sse2_storeu_pd:
586 case Intrinsic::x86_sse2_storeu_dq:
587 case Intrinsic::x86_sse2_storel_dq:
588 if (II->getArgOperand(0) == OperandVal)
596 /// getAccessType - Return the type of the memory being accessed.
597 static Type *getAccessType(const Instruction *Inst) {
598 Type *AccessTy = Inst->getType();
599 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
600 AccessTy = SI->getOperand(0)->getType();
601 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
602 // Addressing modes can also be folded into prefetches and a variety
604 switch (II->getIntrinsicID()) {
606 case Intrinsic::x86_sse_storeu_ps:
607 case Intrinsic::x86_sse2_storeu_pd:
608 case Intrinsic::x86_sse2_storeu_dq:
609 case Intrinsic::x86_sse2_storel_dq:
610 AccessTy = II->getArgOperand(0)->getType();
615 // All pointers have the same requirements, so canonicalize them to an
616 // arbitrary pointer type to minimize variation.
617 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
618 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
619 PTy->getAddressSpace());
624 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
625 /// specified set are trivially dead, delete them and see if this makes any of
626 /// their operands subsequently dead.
628 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
629 bool Changed = false;
631 while (!DeadInsts.empty()) {
632 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
634 if (I == 0 || !isInstructionTriviallyDead(I))
637 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
638 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
641 DeadInsts.push_back(U);
644 I->eraseFromParent();
653 /// Cost - This class is used to measure and compare candidate formulae.
655 /// TODO: Some of these could be merged. Also, a lexical ordering
656 /// isn't always optimal.
660 unsigned NumBaseAdds;
666 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
669 bool operator<(const Cost &Other) const;
674 // Once any of the metrics loses, they must all remain losers.
676 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
677 | ImmCost | SetupCost) != ~0u)
678 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
679 & ImmCost & SetupCost) == ~0u);
684 assert(isValid() && "invalid cost");
685 return NumRegs == ~0u;
688 void RateFormula(const Formula &F,
689 SmallPtrSet<const SCEV *, 16> &Regs,
690 const DenseSet<const SCEV *> &VisitedRegs,
692 const SmallVectorImpl<int64_t> &Offsets,
693 ScalarEvolution &SE, DominatorTree &DT);
695 void print(raw_ostream &OS) const;
699 void RateRegister(const SCEV *Reg,
700 SmallPtrSet<const SCEV *, 16> &Regs,
702 ScalarEvolution &SE, DominatorTree &DT);
703 void RatePrimaryRegister(const SCEV *Reg,
704 SmallPtrSet<const SCEV *, 16> &Regs,
706 ScalarEvolution &SE, DominatorTree &DT);
711 /// RateRegister - Tally up interesting quantities from the given register.
712 void Cost::RateRegister(const SCEV *Reg,
713 SmallPtrSet<const SCEV *, 16> &Regs,
715 ScalarEvolution &SE, DominatorTree &DT) {
716 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
717 if (AR->getLoop() == L)
718 AddRecCost += 1; /// TODO: This should be a function of the stride.
720 // If this is an addrec for a loop that's already been visited by LSR,
721 // don't second-guess its addrec phi nodes. LSR isn't currently smart
722 // enough to reason about more than one loop at a time. Consider these
723 // registers free and leave them alone.
724 else if (L->contains(AR->getLoop()) ||
725 (!AR->getLoop()->contains(L) &&
726 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
727 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
728 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
729 if (SE.isSCEVable(PN->getType()) &&
730 (SE.getEffectiveSCEVType(PN->getType()) ==
731 SE.getEffectiveSCEVType(AR->getType())) &&
732 SE.getSCEV(PN) == AR)
735 // If this isn't one of the addrecs that the loop already has, it
736 // would require a costly new phi and add. TODO: This isn't
737 // precisely modeled right now.
739 if (!Regs.count(AR->getStart())) {
740 RateRegister(AR->getStart(), Regs, L, SE, DT);
746 // Add the step value register, if it needs one.
747 // TODO: The non-affine case isn't precisely modeled here.
748 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
749 if (!Regs.count(AR->getOperand(1)))
750 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
754 // Rough heuristic; favor registers which don't require extra setup
755 // instructions in the preheader.
756 if (!isa<SCEVUnknown>(Reg) &&
757 !isa<SCEVConstant>(Reg) &&
758 !(isa<SCEVAddRecExpr>(Reg) &&
759 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
760 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
763 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
764 SE.hasComputableLoopEvolution(Reg, L);
767 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
769 void Cost::RatePrimaryRegister(const SCEV *Reg,
770 SmallPtrSet<const SCEV *, 16> &Regs,
772 ScalarEvolution &SE, DominatorTree &DT) {
773 if (Regs.insert(Reg))
774 RateRegister(Reg, Regs, L, SE, DT);
777 void Cost::RateFormula(const Formula &F,
778 SmallPtrSet<const SCEV *, 16> &Regs,
779 const DenseSet<const SCEV *> &VisitedRegs,
781 const SmallVectorImpl<int64_t> &Offsets,
782 ScalarEvolution &SE, DominatorTree &DT) {
783 // Tally up the registers.
784 if (const SCEV *ScaledReg = F.ScaledReg) {
785 if (VisitedRegs.count(ScaledReg)) {
789 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
793 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
794 E = F.BaseRegs.end(); I != E; ++I) {
795 const SCEV *BaseReg = *I;
796 if (VisitedRegs.count(BaseReg)) {
800 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
805 // Determine how many (unfolded) adds we'll need inside the loop.
806 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
807 if (NumBaseParts > 1)
808 NumBaseAdds += NumBaseParts - 1;
810 // Tally up the non-zero immediates.
811 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
812 E = Offsets.end(); I != E; ++I) {
813 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
815 ImmCost += 64; // Handle symbolic values conservatively.
816 // TODO: This should probably be the pointer size.
817 else if (Offset != 0)
818 ImmCost += APInt(64, Offset, true).getMinSignedBits();
820 assert(isValid() && "invalid cost");
823 /// Loose - Set this cost to a losing value.
833 /// operator< - Choose the lower cost.
834 bool Cost::operator<(const Cost &Other) const {
835 if (NumRegs != Other.NumRegs)
836 return NumRegs < Other.NumRegs;
837 if (AddRecCost != Other.AddRecCost)
838 return AddRecCost < Other.AddRecCost;
839 if (NumIVMuls != Other.NumIVMuls)
840 return NumIVMuls < Other.NumIVMuls;
841 if (NumBaseAdds != Other.NumBaseAdds)
842 return NumBaseAdds < Other.NumBaseAdds;
843 if (ImmCost != Other.ImmCost)
844 return ImmCost < Other.ImmCost;
845 if (SetupCost != Other.SetupCost)
846 return SetupCost < Other.SetupCost;
850 void Cost::print(raw_ostream &OS) const {
851 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
853 OS << ", with addrec cost " << AddRecCost;
855 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
856 if (NumBaseAdds != 0)
857 OS << ", plus " << NumBaseAdds << " base add"
858 << (NumBaseAdds == 1 ? "" : "s");
860 OS << ", plus " << ImmCost << " imm cost";
862 OS << ", plus " << SetupCost << " setup cost";
865 void Cost::dump() const {
866 print(errs()); errs() << '\n';
871 /// LSRFixup - An operand value in an instruction which is to be replaced
872 /// with some equivalent, possibly strength-reduced, replacement.
874 /// UserInst - The instruction which will be updated.
875 Instruction *UserInst;
877 /// OperandValToReplace - The operand of the instruction which will
878 /// be replaced. The operand may be used more than once; every instance
879 /// will be replaced.
880 Value *OperandValToReplace;
882 /// PostIncLoops - If this user is to use the post-incremented value of an
883 /// induction variable, this variable is non-null and holds the loop
884 /// associated with the induction variable.
885 PostIncLoopSet PostIncLoops;
887 /// LUIdx - The index of the LSRUse describing the expression which
888 /// this fixup needs, minus an offset (below).
891 /// Offset - A constant offset to be added to the LSRUse expression.
892 /// This allows multiple fixups to share the same LSRUse with different
893 /// offsets, for example in an unrolled loop.
896 bool isUseFullyOutsideLoop(const Loop *L) const;
900 void print(raw_ostream &OS) const;
907 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
909 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
910 /// value outside of the given loop.
911 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
912 // PHI nodes use their value in their incoming blocks.
913 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
914 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
915 if (PN->getIncomingValue(i) == OperandValToReplace &&
916 L->contains(PN->getIncomingBlock(i)))
921 return !L->contains(UserInst);
924 void LSRFixup::print(raw_ostream &OS) const {
926 // Store is common and interesting enough to be worth special-casing.
927 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
929 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
930 } else if (UserInst->getType()->isVoidTy())
931 OS << UserInst->getOpcodeName();
933 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
935 OS << ", OperandValToReplace=";
936 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
938 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
939 E = PostIncLoops.end(); I != E; ++I) {
940 OS << ", PostIncLoop=";
941 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
944 if (LUIdx != ~size_t(0))
945 OS << ", LUIdx=" << LUIdx;
948 OS << ", Offset=" << Offset;
951 void LSRFixup::dump() const {
952 print(errs()); errs() << '\n';
957 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
958 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
959 struct UniquifierDenseMapInfo {
960 static SmallVector<const SCEV *, 2> getEmptyKey() {
961 SmallVector<const SCEV *, 2> V;
962 V.push_back(reinterpret_cast<const SCEV *>(-1));
966 static SmallVector<const SCEV *, 2> getTombstoneKey() {
967 SmallVector<const SCEV *, 2> V;
968 V.push_back(reinterpret_cast<const SCEV *>(-2));
972 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
974 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
975 E = V.end(); I != E; ++I)
976 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
980 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
981 const SmallVector<const SCEV *, 2> &RHS) {
986 /// LSRUse - This class holds the state that LSR keeps for each use in
987 /// IVUsers, as well as uses invented by LSR itself. It includes information
988 /// about what kinds of things can be folded into the user, information about
989 /// the user itself, and information about how the use may be satisfied.
990 /// TODO: Represent multiple users of the same expression in common?
992 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
995 /// KindType - An enum for a kind of use, indicating what types of
996 /// scaled and immediate operands it might support.
998 Basic, ///< A normal use, with no folding.
999 Special, ///< A special case of basic, allowing -1 scales.
1000 Address, ///< An address use; folding according to TargetLowering
1001 ICmpZero ///< An equality icmp with both operands folded into one.
1002 // TODO: Add a generic icmp too?
1008 SmallVector<int64_t, 8> Offsets;
1012 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1013 /// LSRUse are outside of the loop, in which case some special-case heuristics
1015 bool AllFixupsOutsideLoop;
1017 /// WidestFixupType - This records the widest use type for any fixup using
1018 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1019 /// max fixup widths to be equivalent, because the narrower one may be relying
1020 /// on the implicit truncation to truncate away bogus bits.
1021 Type *WidestFixupType;
1023 /// Formulae - A list of ways to build a value that can satisfy this user.
1024 /// After the list is populated, one of these is selected heuristically and
1025 /// used to formulate a replacement for OperandValToReplace in UserInst.
1026 SmallVector<Formula, 12> Formulae;
1028 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1029 SmallPtrSet<const SCEV *, 4> Regs;
1031 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1032 MinOffset(INT64_MAX),
1033 MaxOffset(INT64_MIN),
1034 AllFixupsOutsideLoop(true),
1035 WidestFixupType(0) {}
1037 bool HasFormulaWithSameRegs(const Formula &F) const;
1038 bool InsertFormula(const Formula &F);
1039 void DeleteFormula(Formula &F);
1040 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1042 void print(raw_ostream &OS) const;
1048 /// HasFormula - Test whether this use as a formula which has the same
1049 /// registers as the given formula.
1050 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1051 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1052 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1053 // Unstable sort by host order ok, because this is only used for uniquifying.
1054 std::sort(Key.begin(), Key.end());
1055 return Uniquifier.count(Key);
1058 /// InsertFormula - If the given formula has not yet been inserted, add it to
1059 /// the list, and return true. Return false otherwise.
1060 bool LSRUse::InsertFormula(const Formula &F) {
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());
1066 if (!Uniquifier.insert(Key).second)
1069 // Using a register to hold the value of 0 is not profitable.
1070 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1071 "Zero allocated in a scaled register!");
1073 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1074 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1075 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1078 // Add the formula to the list.
1079 Formulae.push_back(F);
1081 // Record registers now being used by this use.
1082 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1083 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1088 /// DeleteFormula - Remove the given formula from this use's list.
1089 void LSRUse::DeleteFormula(Formula &F) {
1090 if (&F != &Formulae.back())
1091 std::swap(F, Formulae.back());
1092 Formulae.pop_back();
1093 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1096 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1097 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1098 // Now that we've filtered out some formulae, recompute the Regs set.
1099 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1101 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1102 E = Formulae.end(); I != E; ++I) {
1103 const Formula &F = *I;
1104 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1105 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1108 // Update the RegTracker.
1109 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1110 E = OldRegs.end(); I != E; ++I)
1111 if (!Regs.count(*I))
1112 RegUses.DropRegister(*I, LUIdx);
1115 void LSRUse::print(raw_ostream &OS) const {
1116 OS << "LSR Use: Kind=";
1118 case Basic: OS << "Basic"; break;
1119 case Special: OS << "Special"; break;
1120 case ICmpZero: OS << "ICmpZero"; break;
1122 OS << "Address of ";
1123 if (AccessTy->isPointerTy())
1124 OS << "pointer"; // the full pointer type could be really verbose
1129 OS << ", Offsets={";
1130 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1131 E = Offsets.end(); I != E; ++I) {
1133 if (llvm::next(I) != E)
1138 if (AllFixupsOutsideLoop)
1139 OS << ", all-fixups-outside-loop";
1141 if (WidestFixupType)
1142 OS << ", widest fixup type: " << *WidestFixupType;
1145 void LSRUse::dump() const {
1146 print(errs()); errs() << '\n';
1149 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1150 /// be completely folded into the user instruction at isel time. This includes
1151 /// address-mode folding and special icmp tricks.
1152 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1153 LSRUse::KindType Kind, Type *AccessTy,
1154 const TargetLowering *TLI) {
1156 case LSRUse::Address:
1157 // If we have low-level target information, ask the target if it can
1158 // completely fold this address.
1159 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1161 // Otherwise, just guess that reg+reg addressing is legal.
1162 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1164 case LSRUse::ICmpZero:
1165 // There's not even a target hook for querying whether it would be legal to
1166 // fold a GV into an ICmp.
1170 // ICmp only has two operands; don't allow more than two non-trivial parts.
1171 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1174 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1175 // putting the scaled register in the other operand of the icmp.
1176 if (AM.Scale != 0 && AM.Scale != -1)
1179 // If we have low-level target information, ask the target if it can fold an
1180 // integer immediate on an icmp.
1181 if (AM.BaseOffs != 0) {
1182 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1189 // Only handle single-register values.
1190 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1192 case LSRUse::Special:
1193 // Only handle -1 scales, or no scale.
1194 return AM.Scale == 0 || AM.Scale == -1;
1200 static bool isLegalUse(TargetLowering::AddrMode AM,
1201 int64_t MinOffset, int64_t MaxOffset,
1202 LSRUse::KindType Kind, Type *AccessTy,
1203 const TargetLowering *TLI) {
1204 // Check for overflow.
1205 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1208 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1209 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1210 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1211 // Check for overflow.
1212 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1215 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1216 return isLegalUse(AM, Kind, AccessTy, TLI);
1221 static bool isAlwaysFoldable(int64_t BaseOffs,
1222 GlobalValue *BaseGV,
1224 LSRUse::KindType Kind, Type *AccessTy,
1225 const TargetLowering *TLI) {
1226 // Fast-path: zero is always foldable.
1227 if (BaseOffs == 0 && !BaseGV) return true;
1229 // Conservatively, create an address with an immediate and a
1230 // base and a scale.
1231 TargetLowering::AddrMode AM;
1232 AM.BaseOffs = BaseOffs;
1234 AM.HasBaseReg = HasBaseReg;
1235 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1237 // Canonicalize a scale of 1 to a base register if the formula doesn't
1238 // already have a base register.
1239 if (!AM.HasBaseReg && AM.Scale == 1) {
1241 AM.HasBaseReg = true;
1244 return isLegalUse(AM, Kind, AccessTy, TLI);
1247 static bool isAlwaysFoldable(const SCEV *S,
1248 int64_t MinOffset, int64_t MaxOffset,
1250 LSRUse::KindType Kind, Type *AccessTy,
1251 const TargetLowering *TLI,
1252 ScalarEvolution &SE) {
1253 // Fast-path: zero is always foldable.
1254 if (S->isZero()) return true;
1256 // Conservatively, create an address with an immediate and a
1257 // base and a scale.
1258 int64_t BaseOffs = ExtractImmediate(S, SE);
1259 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1261 // If there's anything else involved, it's not foldable.
1262 if (!S->isZero()) return false;
1264 // Fast-path: zero is always foldable.
1265 if (BaseOffs == 0 && !BaseGV) return true;
1267 // Conservatively, create an address with an immediate and a
1268 // base and a scale.
1269 TargetLowering::AddrMode AM;
1270 AM.BaseOffs = BaseOffs;
1272 AM.HasBaseReg = HasBaseReg;
1273 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1275 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1280 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1281 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1282 struct UseMapDenseMapInfo {
1283 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1284 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1287 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1288 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1292 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1293 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1294 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1298 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1299 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1304 /// LSRInstance - This class holds state for the main loop strength reduction
1308 ScalarEvolution &SE;
1311 const TargetLowering *const TLI;
1315 /// IVIncInsertPos - This is the insert position that the current loop's
1316 /// induction variable increment should be placed. In simple loops, this is
1317 /// the latch block's terminator. But in more complicated cases, this is a
1318 /// position which will dominate all the in-loop post-increment users.
1319 Instruction *IVIncInsertPos;
1321 /// Factors - Interesting factors between use strides.
1322 SmallSetVector<int64_t, 8> Factors;
1324 /// Types - Interesting use types, to facilitate truncation reuse.
1325 SmallSetVector<Type *, 4> Types;
1327 /// Fixups - The list of operands which are to be replaced.
1328 SmallVector<LSRFixup, 16> Fixups;
1330 /// Uses - The list of interesting uses.
1331 SmallVector<LSRUse, 16> Uses;
1333 /// RegUses - Track which uses use which register candidates.
1334 RegUseTracker RegUses;
1336 void OptimizeShadowIV();
1337 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1338 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1339 void OptimizeLoopTermCond();
1341 void CollectInterestingTypesAndFactors();
1342 void CollectFixupsAndInitialFormulae();
1344 LSRFixup &getNewFixup() {
1345 Fixups.push_back(LSRFixup());
1346 return Fixups.back();
1349 // Support for sharing of LSRUses between LSRFixups.
1350 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1352 UseMapDenseMapInfo> UseMapTy;
1355 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1356 LSRUse::KindType Kind, Type *AccessTy);
1358 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1359 LSRUse::KindType Kind,
1362 void DeleteUse(LSRUse &LU, size_t LUIdx);
1364 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1367 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1368 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1369 void CountRegisters(const Formula &F, size_t LUIdx);
1370 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1372 void CollectLoopInvariantFixupsAndFormulae();
1374 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1375 unsigned Depth = 0);
1376 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1377 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1378 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1379 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1380 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1381 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1382 void GenerateCrossUseConstantOffsets();
1383 void GenerateAllReuseFormulae();
1385 void FilterOutUndesirableDedicatedRegisters();
1387 size_t EstimateSearchSpaceComplexity() const;
1388 void NarrowSearchSpaceByDetectingSupersets();
1389 void NarrowSearchSpaceByCollapsingUnrolledCode();
1390 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1391 void NarrowSearchSpaceByPickingWinnerRegs();
1392 void NarrowSearchSpaceUsingHeuristics();
1394 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1396 SmallVectorImpl<const Formula *> &Workspace,
1397 const Cost &CurCost,
1398 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1399 DenseSet<const SCEV *> &VisitedRegs) const;
1400 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1402 BasicBlock::iterator
1403 HoistInsertPosition(BasicBlock::iterator IP,
1404 const SmallVectorImpl<Instruction *> &Inputs) const;
1405 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1407 const LSRUse &LU) const;
1409 Value *Expand(const LSRFixup &LF,
1411 BasicBlock::iterator IP,
1412 SCEVExpander &Rewriter,
1413 SmallVectorImpl<WeakVH> &DeadInsts) const;
1414 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1416 SCEVExpander &Rewriter,
1417 SmallVectorImpl<WeakVH> &DeadInsts,
1419 void Rewrite(const LSRFixup &LF,
1421 SCEVExpander &Rewriter,
1422 SmallVectorImpl<WeakVH> &DeadInsts,
1424 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1427 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1429 bool getChanged() const { return Changed; }
1431 void print_factors_and_types(raw_ostream &OS) const;
1432 void print_fixups(raw_ostream &OS) const;
1433 void print_uses(raw_ostream &OS) const;
1434 void print(raw_ostream &OS) const;
1440 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1441 /// inside the loop then try to eliminate the cast operation.
1442 void LSRInstance::OptimizeShadowIV() {
1443 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1444 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1447 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1448 UI != E; /* empty */) {
1449 IVUsers::const_iterator CandidateUI = UI;
1451 Instruction *ShadowUse = CandidateUI->getUser();
1452 Type *DestTy = NULL;
1453 bool IsSigned = false;
1455 /* If shadow use is a int->float cast then insert a second IV
1456 to eliminate this cast.
1458 for (unsigned i = 0; i < n; ++i)
1464 for (unsigned i = 0; i < n; ++i, ++d)
1467 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1469 DestTy = UCast->getDestTy();
1471 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1473 DestTy = SCast->getDestTy();
1475 if (!DestTy) continue;
1478 // If target does not support DestTy natively then do not apply
1479 // this transformation.
1480 EVT DVT = TLI->getValueType(DestTy);
1481 if (!TLI->isTypeLegal(DVT)) continue;
1484 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1486 if (PH->getNumIncomingValues() != 2) continue;
1488 Type *SrcTy = PH->getType();
1489 int Mantissa = DestTy->getFPMantissaWidth();
1490 if (Mantissa == -1) continue;
1491 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1494 unsigned Entry, Latch;
1495 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1503 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1504 if (!Init) continue;
1505 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1506 (double)Init->getSExtValue() :
1507 (double)Init->getZExtValue());
1509 BinaryOperator *Incr =
1510 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1511 if (!Incr) continue;
1512 if (Incr->getOpcode() != Instruction::Add
1513 && Incr->getOpcode() != Instruction::Sub)
1516 /* Initialize new IV, double d = 0.0 in above example. */
1517 ConstantInt *C = NULL;
1518 if (Incr->getOperand(0) == PH)
1519 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1520 else if (Incr->getOperand(1) == PH)
1521 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1527 // Ignore negative constants, as the code below doesn't handle them
1528 // correctly. TODO: Remove this restriction.
1529 if (!C->getValue().isStrictlyPositive()) continue;
1531 /* Add new PHINode. */
1532 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1534 /* create new increment. '++d' in above example. */
1535 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1536 BinaryOperator *NewIncr =
1537 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1538 Instruction::FAdd : Instruction::FSub,
1539 NewPH, CFP, "IV.S.next.", Incr);
1541 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1542 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1544 /* Remove cast operation */
1545 ShadowUse->replaceAllUsesWith(NewPH);
1546 ShadowUse->eraseFromParent();
1552 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1553 /// set the IV user and stride information and return true, otherwise return
1555 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1556 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1557 if (UI->getUser() == Cond) {
1558 // NOTE: we could handle setcc instructions with multiple uses here, but
1559 // InstCombine does it as well for simple uses, it's not clear that it
1560 // occurs enough in real life to handle.
1567 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1568 /// a max computation.
1570 /// This is a narrow solution to a specific, but acute, problem. For loops
1576 /// } while (++i < n);
1578 /// the trip count isn't just 'n', because 'n' might not be positive. And
1579 /// unfortunately this can come up even for loops where the user didn't use
1580 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1581 /// will commonly be lowered like this:
1587 /// } while (++i < n);
1590 /// and then it's possible for subsequent optimization to obscure the if
1591 /// test in such a way that indvars can't find it.
1593 /// When indvars can't find the if test in loops like this, it creates a
1594 /// max expression, which allows it to give the loop a canonical
1595 /// induction variable:
1598 /// max = n < 1 ? 1 : n;
1601 /// } while (++i != max);
1603 /// Canonical induction variables are necessary because the loop passes
1604 /// are designed around them. The most obvious example of this is the
1605 /// LoopInfo analysis, which doesn't remember trip count values. It
1606 /// expects to be able to rediscover the trip count each time it is
1607 /// needed, and it does this using a simple analysis that only succeeds if
1608 /// the loop has a canonical induction variable.
1610 /// However, when it comes time to generate code, the maximum operation
1611 /// can be quite costly, especially if it's inside of an outer loop.
1613 /// This function solves this problem by detecting this type of loop and
1614 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1615 /// the instructions for the maximum computation.
1617 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1618 // Check that the loop matches the pattern we're looking for.
1619 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1620 Cond->getPredicate() != CmpInst::ICMP_NE)
1623 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1624 if (!Sel || !Sel->hasOneUse()) return Cond;
1626 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1627 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1629 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1631 // Add one to the backedge-taken count to get the trip count.
1632 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1633 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1635 // Check for a max calculation that matches the pattern. There's no check
1636 // for ICMP_ULE here because the comparison would be with zero, which
1637 // isn't interesting.
1638 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1639 const SCEVNAryExpr *Max = 0;
1640 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1641 Pred = ICmpInst::ICMP_SLE;
1643 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1644 Pred = ICmpInst::ICMP_SLT;
1646 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1647 Pred = ICmpInst::ICMP_ULT;
1654 // To handle a max with more than two operands, this optimization would
1655 // require additional checking and setup.
1656 if (Max->getNumOperands() != 2)
1659 const SCEV *MaxLHS = Max->getOperand(0);
1660 const SCEV *MaxRHS = Max->getOperand(1);
1662 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1663 // for a comparison with 1. For <= and >=, a comparison with zero.
1665 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1668 // Check the relevant induction variable for conformance to
1670 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1671 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1672 if (!AR || !AR->isAffine() ||
1673 AR->getStart() != One ||
1674 AR->getStepRecurrence(SE) != One)
1677 assert(AR->getLoop() == L &&
1678 "Loop condition operand is an addrec in a different loop!");
1680 // Check the right operand of the select, and remember it, as it will
1681 // be used in the new comparison instruction.
1683 if (ICmpInst::isTrueWhenEqual(Pred)) {
1684 // Look for n+1, and grab n.
1685 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1686 if (isa<ConstantInt>(BO->getOperand(1)) &&
1687 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1688 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1689 NewRHS = BO->getOperand(0);
1690 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1691 if (isa<ConstantInt>(BO->getOperand(1)) &&
1692 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1693 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1694 NewRHS = BO->getOperand(0);
1697 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1698 NewRHS = Sel->getOperand(1);
1699 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1700 NewRHS = Sel->getOperand(2);
1701 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1702 NewRHS = SU->getValue();
1704 // Max doesn't match expected pattern.
1707 // Determine the new comparison opcode. It may be signed or unsigned,
1708 // and the original comparison may be either equality or inequality.
1709 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1710 Pred = CmpInst::getInversePredicate(Pred);
1712 // Ok, everything looks ok to change the condition into an SLT or SGE and
1713 // delete the max calculation.
1715 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1717 // Delete the max calculation instructions.
1718 Cond->replaceAllUsesWith(NewCond);
1719 CondUse->setUser(NewCond);
1720 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1721 Cond->eraseFromParent();
1722 Sel->eraseFromParent();
1723 if (Cmp->use_empty())
1724 Cmp->eraseFromParent();
1728 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1729 /// postinc iv when possible.
1731 LSRInstance::OptimizeLoopTermCond() {
1732 SmallPtrSet<Instruction *, 4> PostIncs;
1734 BasicBlock *LatchBlock = L->getLoopLatch();
1735 SmallVector<BasicBlock*, 8> ExitingBlocks;
1736 L->getExitingBlocks(ExitingBlocks);
1738 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1739 BasicBlock *ExitingBlock = ExitingBlocks[i];
1741 // Get the terminating condition for the loop if possible. If we
1742 // can, we want to change it to use a post-incremented version of its
1743 // induction variable, to allow coalescing the live ranges for the IV into
1744 // one register value.
1746 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1749 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1750 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1753 // Search IVUsesByStride to find Cond's IVUse if there is one.
1754 IVStrideUse *CondUse = 0;
1755 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1756 if (!FindIVUserForCond(Cond, CondUse))
1759 // If the trip count is computed in terms of a max (due to ScalarEvolution
1760 // being unable to find a sufficient guard, for example), change the loop
1761 // comparison to use SLT or ULT instead of NE.
1762 // One consequence of doing this now is that it disrupts the count-down
1763 // optimization. That's not always a bad thing though, because in such
1764 // cases it may still be worthwhile to avoid a max.
1765 Cond = OptimizeMax(Cond, CondUse);
1767 // If this exiting block dominates the latch block, it may also use
1768 // the post-inc value if it won't be shared with other uses.
1769 // Check for dominance.
1770 if (!DT.dominates(ExitingBlock, LatchBlock))
1773 // Conservatively avoid trying to use the post-inc value in non-latch
1774 // exits if there may be pre-inc users in intervening blocks.
1775 if (LatchBlock != ExitingBlock)
1776 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1777 // Test if the use is reachable from the exiting block. This dominator
1778 // query is a conservative approximation of reachability.
1779 if (&*UI != CondUse &&
1780 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1781 // Conservatively assume there may be reuse if the quotient of their
1782 // strides could be a legal scale.
1783 const SCEV *A = IU.getStride(*CondUse, L);
1784 const SCEV *B = IU.getStride(*UI, L);
1785 if (!A || !B) continue;
1786 if (SE.getTypeSizeInBits(A->getType()) !=
1787 SE.getTypeSizeInBits(B->getType())) {
1788 if (SE.getTypeSizeInBits(A->getType()) >
1789 SE.getTypeSizeInBits(B->getType()))
1790 B = SE.getSignExtendExpr(B, A->getType());
1792 A = SE.getSignExtendExpr(A, B->getType());
1794 if (const SCEVConstant *D =
1795 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1796 const ConstantInt *C = D->getValue();
1797 // Stride of one or negative one can have reuse with non-addresses.
1798 if (C->isOne() || C->isAllOnesValue())
1799 goto decline_post_inc;
1800 // Avoid weird situations.
1801 if (C->getValue().getMinSignedBits() >= 64 ||
1802 C->getValue().isMinSignedValue())
1803 goto decline_post_inc;
1804 // Without TLI, assume that any stride might be valid, and so any
1805 // use might be shared.
1807 goto decline_post_inc;
1808 // Check for possible scaled-address reuse.
1809 Type *AccessTy = getAccessType(UI->getUser());
1810 TargetLowering::AddrMode AM;
1811 AM.Scale = C->getSExtValue();
1812 if (TLI->isLegalAddressingMode(AM, AccessTy))
1813 goto decline_post_inc;
1814 AM.Scale = -AM.Scale;
1815 if (TLI->isLegalAddressingMode(AM, AccessTy))
1816 goto decline_post_inc;
1820 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1823 // It's possible for the setcc instruction to be anywhere in the loop, and
1824 // possible for it to have multiple users. If it is not immediately before
1825 // the exiting block branch, move it.
1826 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1827 if (Cond->hasOneUse()) {
1828 Cond->moveBefore(TermBr);
1830 // Clone the terminating condition and insert into the loopend.
1831 ICmpInst *OldCond = Cond;
1832 Cond = cast<ICmpInst>(Cond->clone());
1833 Cond->setName(L->getHeader()->getName() + ".termcond");
1834 ExitingBlock->getInstList().insert(TermBr, Cond);
1836 // Clone the IVUse, as the old use still exists!
1837 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1838 TermBr->replaceUsesOfWith(OldCond, Cond);
1842 // If we get to here, we know that we can transform the setcc instruction to
1843 // use the post-incremented version of the IV, allowing us to coalesce the
1844 // live ranges for the IV correctly.
1845 CondUse->transformToPostInc(L);
1848 PostIncs.insert(Cond);
1852 // Determine an insertion point for the loop induction variable increment. It
1853 // must dominate all the post-inc comparisons we just set up, and it must
1854 // dominate the loop latch edge.
1855 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1856 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1857 E = PostIncs.end(); I != E; ++I) {
1859 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1861 if (BB == (*I)->getParent())
1862 IVIncInsertPos = *I;
1863 else if (BB != IVIncInsertPos->getParent())
1864 IVIncInsertPos = BB->getTerminator();
1868 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
1869 /// at the given offset and other details. If so, update the use and
1872 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1873 LSRUse::KindType Kind, Type *AccessTy) {
1874 int64_t NewMinOffset = LU.MinOffset;
1875 int64_t NewMaxOffset = LU.MaxOffset;
1876 Type *NewAccessTy = AccessTy;
1878 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1879 // something conservative, however this can pessimize in the case that one of
1880 // the uses will have all its uses outside the loop, for example.
1881 if (LU.Kind != Kind)
1883 // Conservatively assume HasBaseReg is true for now.
1884 if (NewOffset < LU.MinOffset) {
1885 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1886 Kind, AccessTy, TLI))
1888 NewMinOffset = NewOffset;
1889 } else if (NewOffset > LU.MaxOffset) {
1890 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1891 Kind, AccessTy, TLI))
1893 NewMaxOffset = NewOffset;
1895 // Check for a mismatched access type, and fall back conservatively as needed.
1896 // TODO: Be less conservative when the type is similar and can use the same
1897 // addressing modes.
1898 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1899 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1902 LU.MinOffset = NewMinOffset;
1903 LU.MaxOffset = NewMaxOffset;
1904 LU.AccessTy = NewAccessTy;
1905 if (NewOffset != LU.Offsets.back())
1906 LU.Offsets.push_back(NewOffset);
1910 /// getUse - Return an LSRUse index and an offset value for a fixup which
1911 /// needs the given expression, with the given kind and optional access type.
1912 /// Either reuse an existing use or create a new one, as needed.
1913 std::pair<size_t, int64_t>
1914 LSRInstance::getUse(const SCEV *&Expr,
1915 LSRUse::KindType Kind, Type *AccessTy) {
1916 const SCEV *Copy = Expr;
1917 int64_t Offset = ExtractImmediate(Expr, SE);
1919 // Basic uses can't accept any offset, for example.
1920 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1925 std::pair<UseMapTy::iterator, bool> P =
1926 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1928 // A use already existed with this base.
1929 size_t LUIdx = P.first->second;
1930 LSRUse &LU = Uses[LUIdx];
1931 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1933 return std::make_pair(LUIdx, Offset);
1936 // Create a new use.
1937 size_t LUIdx = Uses.size();
1938 P.first->second = LUIdx;
1939 Uses.push_back(LSRUse(Kind, AccessTy));
1940 LSRUse &LU = Uses[LUIdx];
1942 // We don't need to track redundant offsets, but we don't need to go out
1943 // of our way here to avoid them.
1944 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1945 LU.Offsets.push_back(Offset);
1947 LU.MinOffset = Offset;
1948 LU.MaxOffset = Offset;
1949 return std::make_pair(LUIdx, Offset);
1952 /// DeleteUse - Delete the given use from the Uses list.
1953 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
1954 if (&LU != &Uses.back())
1955 std::swap(LU, Uses.back());
1959 RegUses.SwapAndDropUse(LUIdx, Uses.size());
1962 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1963 /// a formula that has the same registers as the given formula.
1965 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1966 const LSRUse &OrigLU) {
1967 // Search all uses for the formula. This could be more clever.
1968 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1969 LSRUse &LU = Uses[LUIdx];
1970 // Check whether this use is close enough to OrigLU, to see whether it's
1971 // worthwhile looking through its formulae.
1972 // Ignore ICmpZero uses because they may contain formulae generated by
1973 // GenerateICmpZeroScales, in which case adding fixup offsets may
1975 if (&LU != &OrigLU &&
1976 LU.Kind != LSRUse::ICmpZero &&
1977 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
1978 LU.WidestFixupType == OrigLU.WidestFixupType &&
1979 LU.HasFormulaWithSameRegs(OrigF)) {
1980 // Scan through this use's formulae.
1981 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
1982 E = LU.Formulae.end(); I != E; ++I) {
1983 const Formula &F = *I;
1984 // Check to see if this formula has the same registers and symbols
1986 if (F.BaseRegs == OrigF.BaseRegs &&
1987 F.ScaledReg == OrigF.ScaledReg &&
1988 F.AM.BaseGV == OrigF.AM.BaseGV &&
1989 F.AM.Scale == OrigF.AM.Scale &&
1990 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
1991 if (F.AM.BaseOffs == 0)
1993 // This is the formula where all the registers and symbols matched;
1994 // there aren't going to be any others. Since we declined it, we
1995 // can skip the rest of the formulae and procede to the next LSRUse.
2002 // Nothing looked good.
2006 void LSRInstance::CollectInterestingTypesAndFactors() {
2007 SmallSetVector<const SCEV *, 4> Strides;
2009 // Collect interesting types and strides.
2010 SmallVector<const SCEV *, 4> Worklist;
2011 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2012 const SCEV *Expr = IU.getExpr(*UI);
2014 // Collect interesting types.
2015 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2017 // Add strides for mentioned loops.
2018 Worklist.push_back(Expr);
2020 const SCEV *S = Worklist.pop_back_val();
2021 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2022 Strides.insert(AR->getStepRecurrence(SE));
2023 Worklist.push_back(AR->getStart());
2024 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2025 Worklist.append(Add->op_begin(), Add->op_end());
2027 } while (!Worklist.empty());
2030 // Compute interesting factors from the set of interesting strides.
2031 for (SmallSetVector<const SCEV *, 4>::const_iterator
2032 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2033 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2034 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2035 const SCEV *OldStride = *I;
2036 const SCEV *NewStride = *NewStrideIter;
2038 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2039 SE.getTypeSizeInBits(NewStride->getType())) {
2040 if (SE.getTypeSizeInBits(OldStride->getType()) >
2041 SE.getTypeSizeInBits(NewStride->getType()))
2042 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2044 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2046 if (const SCEVConstant *Factor =
2047 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2049 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2050 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2051 } else if (const SCEVConstant *Factor =
2052 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2055 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2056 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2060 // If all uses use the same type, don't bother looking for truncation-based
2062 if (Types.size() == 1)
2065 DEBUG(print_factors_and_types(dbgs()));
2068 void LSRInstance::CollectFixupsAndInitialFormulae() {
2069 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2071 LSRFixup &LF = getNewFixup();
2072 LF.UserInst = UI->getUser();
2073 LF.OperandValToReplace = UI->getOperandValToReplace();
2074 LF.PostIncLoops = UI->getPostIncLoops();
2076 LSRUse::KindType Kind = LSRUse::Basic;
2078 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2079 Kind = LSRUse::Address;
2080 AccessTy = getAccessType(LF.UserInst);
2083 const SCEV *S = IU.getExpr(*UI);
2085 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2086 // (N - i == 0), and this allows (N - i) to be the expression that we work
2087 // with rather than just N or i, so we can consider the register
2088 // requirements for both N and i at the same time. Limiting this code to
2089 // equality icmps is not a problem because all interesting loops use
2090 // equality icmps, thanks to IndVarSimplify.
2091 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2092 if (CI->isEquality()) {
2093 // Swap the operands if needed to put the OperandValToReplace on the
2094 // left, for consistency.
2095 Value *NV = CI->getOperand(1);
2096 if (NV == LF.OperandValToReplace) {
2097 CI->setOperand(1, CI->getOperand(0));
2098 CI->setOperand(0, NV);
2099 NV = CI->getOperand(1);
2103 // x == y --> x - y == 0
2104 const SCEV *N = SE.getSCEV(NV);
2105 if (SE.isLoopInvariant(N, L)) {
2106 // S is normalized, so normalize N before folding it into S
2107 // to keep the result normalized.
2108 N = TransformForPostIncUse(Normalize, N, CI, 0,
2109 LF.PostIncLoops, SE, DT);
2110 Kind = LSRUse::ICmpZero;
2111 S = SE.getMinusSCEV(N, S);
2114 // -1 and the negations of all interesting strides (except the negation
2115 // of -1) are now also interesting.
2116 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2117 if (Factors[i] != -1)
2118 Factors.insert(-(uint64_t)Factors[i]);
2122 // Set up the initial formula for this use.
2123 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2125 LF.Offset = P.second;
2126 LSRUse &LU = Uses[LF.LUIdx];
2127 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2128 if (!LU.WidestFixupType ||
2129 SE.getTypeSizeInBits(LU.WidestFixupType) <
2130 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2131 LU.WidestFixupType = LF.OperandValToReplace->getType();
2133 // If this is the first use of this LSRUse, give it a formula.
2134 if (LU.Formulae.empty()) {
2135 InsertInitialFormula(S, LU, LF.LUIdx);
2136 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2140 DEBUG(print_fixups(dbgs()));
2143 /// InsertInitialFormula - Insert a formula for the given expression into
2144 /// the given use, separating out loop-variant portions from loop-invariant
2145 /// and loop-computable portions.
2147 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2149 F.InitialMatch(S, L, SE);
2150 bool Inserted = InsertFormula(LU, LUIdx, F);
2151 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2154 /// InsertSupplementalFormula - Insert a simple single-register formula for
2155 /// the given expression into the given use.
2157 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2158 LSRUse &LU, size_t LUIdx) {
2160 F.BaseRegs.push_back(S);
2161 F.AM.HasBaseReg = true;
2162 bool Inserted = InsertFormula(LU, LUIdx, F);
2163 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2166 /// CountRegisters - Note which registers are used by the given formula,
2167 /// updating RegUses.
2168 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2170 RegUses.CountRegister(F.ScaledReg, LUIdx);
2171 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2172 E = F.BaseRegs.end(); I != E; ++I)
2173 RegUses.CountRegister(*I, LUIdx);
2176 /// InsertFormula - If the given formula has not yet been inserted, add it to
2177 /// the list, and return true. Return false otherwise.
2178 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2179 if (!LU.InsertFormula(F))
2182 CountRegisters(F, LUIdx);
2186 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2187 /// loop-invariant values which we're tracking. These other uses will pin these
2188 /// values in registers, making them less profitable for elimination.
2189 /// TODO: This currently misses non-constant addrec step registers.
2190 /// TODO: Should this give more weight to users inside the loop?
2192 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2193 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2194 SmallPtrSet<const SCEV *, 8> Inserted;
2196 while (!Worklist.empty()) {
2197 const SCEV *S = Worklist.pop_back_val();
2199 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2200 Worklist.append(N->op_begin(), N->op_end());
2201 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2202 Worklist.push_back(C->getOperand());
2203 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2204 Worklist.push_back(D->getLHS());
2205 Worklist.push_back(D->getRHS());
2206 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2207 if (!Inserted.insert(U)) continue;
2208 const Value *V = U->getValue();
2209 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2210 // Look for instructions defined outside the loop.
2211 if (L->contains(Inst)) continue;
2212 } else if (isa<UndefValue>(V))
2213 // Undef doesn't have a live range, so it doesn't matter.
2215 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2217 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2218 // Ignore non-instructions.
2221 // Ignore instructions in other functions (as can happen with
2223 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2225 // Ignore instructions not dominated by the loop.
2226 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2227 UserInst->getParent() :
2228 cast<PHINode>(UserInst)->getIncomingBlock(
2229 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2230 if (!DT.dominates(L->getHeader(), UseBB))
2232 // Ignore uses which are part of other SCEV expressions, to avoid
2233 // analyzing them multiple times.
2234 if (SE.isSCEVable(UserInst->getType())) {
2235 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2236 // If the user is a no-op, look through to its uses.
2237 if (!isa<SCEVUnknown>(UserS))
2241 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2245 // Ignore icmp instructions which are already being analyzed.
2246 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2247 unsigned OtherIdx = !UI.getOperandNo();
2248 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2249 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2253 LSRFixup &LF = getNewFixup();
2254 LF.UserInst = const_cast<Instruction *>(UserInst);
2255 LF.OperandValToReplace = UI.getUse();
2256 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2258 LF.Offset = P.second;
2259 LSRUse &LU = Uses[LF.LUIdx];
2260 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2261 if (!LU.WidestFixupType ||
2262 SE.getTypeSizeInBits(LU.WidestFixupType) <
2263 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2264 LU.WidestFixupType = LF.OperandValToReplace->getType();
2265 InsertSupplementalFormula(U, LU, LF.LUIdx);
2266 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2273 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2274 /// separate registers. If C is non-null, multiply each subexpression by C.
2275 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2276 SmallVectorImpl<const SCEV *> &Ops,
2278 ScalarEvolution &SE) {
2279 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2280 // Break out add operands.
2281 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2283 CollectSubexprs(*I, C, Ops, L, SE);
2285 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2286 // Split a non-zero base out of an addrec.
2287 if (!AR->getStart()->isZero()) {
2288 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2289 AR->getStepRecurrence(SE),
2291 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2294 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2297 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2298 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2299 if (Mul->getNumOperands() == 2)
2300 if (const SCEVConstant *Op0 =
2301 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2302 CollectSubexprs(Mul->getOperand(1),
2303 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2309 // Otherwise use the value itself, optionally with a scale applied.
2310 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2313 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2315 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2318 // Arbitrarily cap recursion to protect compile time.
2319 if (Depth >= 3) return;
2321 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2322 const SCEV *BaseReg = Base.BaseRegs[i];
2324 SmallVector<const SCEV *, 8> AddOps;
2325 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2327 if (AddOps.size() == 1) continue;
2329 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2330 JE = AddOps.end(); J != JE; ++J) {
2332 // Loop-variant "unknown" values are uninteresting; we won't be able to
2333 // do anything meaningful with them.
2334 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2337 // Don't pull a constant into a register if the constant could be folded
2338 // into an immediate field.
2339 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2340 Base.getNumRegs() > 1,
2341 LU.Kind, LU.AccessTy, TLI, SE))
2344 // Collect all operands except *J.
2345 SmallVector<const SCEV *, 8> InnerAddOps
2346 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2348 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2350 // Don't leave just a constant behind in a register if the constant could
2351 // be folded into an immediate field.
2352 if (InnerAddOps.size() == 1 &&
2353 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2354 Base.getNumRegs() > 1,
2355 LU.Kind, LU.AccessTy, TLI, SE))
2358 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2359 if (InnerSum->isZero())
2363 // Add the remaining pieces of the add back into the new formula.
2364 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
2365 if (TLI && InnerSumSC &&
2366 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
2367 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2368 InnerSumSC->getValue()->getZExtValue())) {
2369 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2370 InnerSumSC->getValue()->getZExtValue();
2371 F.BaseRegs.erase(F.BaseRegs.begin() + i);
2373 F.BaseRegs[i] = InnerSum;
2375 // Add J as its own register, or an unfolded immediate.
2376 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
2377 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
2378 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2379 SC->getValue()->getZExtValue()))
2380 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2381 SC->getValue()->getZExtValue();
2383 F.BaseRegs.push_back(*J);
2385 if (InsertFormula(LU, LUIdx, F))
2386 // If that formula hadn't been seen before, recurse to find more like
2388 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2393 /// GenerateCombinations - Generate a formula consisting of all of the
2394 /// loop-dominating registers added into a single register.
2395 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2397 // This method is only interesting on a plurality of registers.
2398 if (Base.BaseRegs.size() <= 1) return;
2402 SmallVector<const SCEV *, 4> Ops;
2403 for (SmallVectorImpl<const SCEV *>::const_iterator
2404 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2405 const SCEV *BaseReg = *I;
2406 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2407 !SE.hasComputableLoopEvolution(BaseReg, L))
2408 Ops.push_back(BaseReg);
2410 F.BaseRegs.push_back(BaseReg);
2412 if (Ops.size() > 1) {
2413 const SCEV *Sum = SE.getAddExpr(Ops);
2414 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2415 // opportunity to fold something. For now, just ignore such cases
2416 // rather than proceed with zero in a register.
2417 if (!Sum->isZero()) {
2418 F.BaseRegs.push_back(Sum);
2419 (void)InsertFormula(LU, LUIdx, F);
2424 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2425 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2427 // We can't add a symbolic offset if the address already contains one.
2428 if (Base.AM.BaseGV) return;
2430 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2431 const SCEV *G = Base.BaseRegs[i];
2432 GlobalValue *GV = ExtractSymbol(G, SE);
2433 if (G->isZero() || !GV)
2437 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2438 LU.Kind, LU.AccessTy, TLI))
2441 (void)InsertFormula(LU, LUIdx, F);
2445 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2446 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2448 // TODO: For now, just add the min and max offset, because it usually isn't
2449 // worthwhile looking at everything inbetween.
2450 SmallVector<int64_t, 2> Worklist;
2451 Worklist.push_back(LU.MinOffset);
2452 if (LU.MaxOffset != LU.MinOffset)
2453 Worklist.push_back(LU.MaxOffset);
2455 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2456 const SCEV *G = Base.BaseRegs[i];
2458 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2459 E = Worklist.end(); I != E; ++I) {
2461 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2462 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2463 LU.Kind, LU.AccessTy, TLI)) {
2464 // Add the offset to the base register.
2465 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2466 // If it cancelled out, drop the base register, otherwise update it.
2467 if (NewG->isZero()) {
2468 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2469 F.BaseRegs.pop_back();
2471 F.BaseRegs[i] = NewG;
2473 (void)InsertFormula(LU, LUIdx, F);
2477 int64_t Imm = ExtractImmediate(G, SE);
2478 if (G->isZero() || Imm == 0)
2481 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2482 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2483 LU.Kind, LU.AccessTy, TLI))
2486 (void)InsertFormula(LU, LUIdx, F);
2490 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2491 /// the comparison. For example, x == y -> x*c == y*c.
2492 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2494 if (LU.Kind != LSRUse::ICmpZero) return;
2496 // Determine the integer type for the base formula.
2497 Type *IntTy = Base.getType();
2499 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2501 // Don't do this if there is more than one offset.
2502 if (LU.MinOffset != LU.MaxOffset) return;
2504 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2506 // Check each interesting stride.
2507 for (SmallSetVector<int64_t, 8>::const_iterator
2508 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2509 int64_t Factor = *I;
2511 // Check that the multiplication doesn't overflow.
2512 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2514 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2515 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2518 // Check that multiplying with the use offset doesn't overflow.
2519 int64_t Offset = LU.MinOffset;
2520 if (Offset == INT64_MIN && Factor == -1)
2522 Offset = (uint64_t)Offset * Factor;
2523 if (Offset / Factor != LU.MinOffset)
2527 F.AM.BaseOffs = NewBaseOffs;
2529 // Check that this scale is legal.
2530 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2533 // Compensate for the use having MinOffset built into it.
2534 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2536 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2538 // Check that multiplying with each base register doesn't overflow.
2539 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2540 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2541 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2545 // Check that multiplying with the scaled register doesn't overflow.
2547 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2548 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2552 // Check that multiplying with the unfolded offset doesn't overflow.
2553 if (F.UnfoldedOffset != 0) {
2554 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
2556 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
2557 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
2561 // If we make it here and it's legal, add it.
2562 (void)InsertFormula(LU, LUIdx, F);
2567 /// GenerateScales - Generate stride factor reuse formulae by making use of
2568 /// scaled-offset address modes, for example.
2569 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2570 // Determine the integer type for the base formula.
2571 Type *IntTy = Base.getType();
2574 // If this Formula already has a scaled register, we can't add another one.
2575 if (Base.AM.Scale != 0) return;
2577 // Check each interesting stride.
2578 for (SmallSetVector<int64_t, 8>::const_iterator
2579 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2580 int64_t Factor = *I;
2582 Base.AM.Scale = Factor;
2583 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2584 // Check whether this scale is going to be legal.
2585 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2586 LU.Kind, LU.AccessTy, TLI)) {
2587 // As a special-case, handle special out-of-loop Basic users specially.
2588 // TODO: Reconsider this special case.
2589 if (LU.Kind == LSRUse::Basic &&
2590 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2591 LSRUse::Special, LU.AccessTy, TLI) &&
2592 LU.AllFixupsOutsideLoop)
2593 LU.Kind = LSRUse::Special;
2597 // For an ICmpZero, negating a solitary base register won't lead to
2599 if (LU.Kind == LSRUse::ICmpZero &&
2600 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2602 // For each addrec base reg, apply the scale, if possible.
2603 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2604 if (const SCEVAddRecExpr *AR =
2605 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2606 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2607 if (FactorS->isZero())
2609 // Divide out the factor, ignoring high bits, since we'll be
2610 // scaling the value back up in the end.
2611 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2612 // TODO: This could be optimized to avoid all the copying.
2614 F.ScaledReg = Quotient;
2615 F.DeleteBaseReg(F.BaseRegs[i]);
2616 (void)InsertFormula(LU, LUIdx, F);
2622 /// GenerateTruncates - Generate reuse formulae from different IV types.
2623 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2624 // This requires TargetLowering to tell us which truncates are free.
2627 // Don't bother truncating symbolic values.
2628 if (Base.AM.BaseGV) return;
2630 // Determine the integer type for the base formula.
2631 Type *DstTy = Base.getType();
2633 DstTy = SE.getEffectiveSCEVType(DstTy);
2635 for (SmallSetVector<Type *, 4>::const_iterator
2636 I = Types.begin(), E = Types.end(); I != E; ++I) {
2638 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2641 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2642 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2643 JE = F.BaseRegs.end(); J != JE; ++J)
2644 *J = SE.getAnyExtendExpr(*J, SrcTy);
2646 // TODO: This assumes we've done basic processing on all uses and
2647 // have an idea what the register usage is.
2648 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2651 (void)InsertFormula(LU, LUIdx, F);
2658 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2659 /// defer modifications so that the search phase doesn't have to worry about
2660 /// the data structures moving underneath it.
2664 const SCEV *OrigReg;
2666 WorkItem(size_t LI, int64_t I, const SCEV *R)
2667 : LUIdx(LI), Imm(I), OrigReg(R) {}
2669 void print(raw_ostream &OS) const;
2675 void WorkItem::print(raw_ostream &OS) const {
2676 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2677 << " , add offset " << Imm;
2680 void WorkItem::dump() const {
2681 print(errs()); errs() << '\n';
2684 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2685 /// distance apart and try to form reuse opportunities between them.
2686 void LSRInstance::GenerateCrossUseConstantOffsets() {
2687 // Group the registers by their value without any added constant offset.
2688 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2689 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2691 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2692 SmallVector<const SCEV *, 8> Sequence;
2693 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2695 const SCEV *Reg = *I;
2696 int64_t Imm = ExtractImmediate(Reg, SE);
2697 std::pair<RegMapTy::iterator, bool> Pair =
2698 Map.insert(std::make_pair(Reg, ImmMapTy()));
2700 Sequence.push_back(Reg);
2701 Pair.first->second.insert(std::make_pair(Imm, *I));
2702 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2705 // Now examine each set of registers with the same base value. Build up
2706 // a list of work to do and do the work in a separate step so that we're
2707 // not adding formulae and register counts while we're searching.
2708 SmallVector<WorkItem, 32> WorkItems;
2709 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2710 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2711 E = Sequence.end(); I != E; ++I) {
2712 const SCEV *Reg = *I;
2713 const ImmMapTy &Imms = Map.find(Reg)->second;
2715 // It's not worthwhile looking for reuse if there's only one offset.
2716 if (Imms.size() == 1)
2719 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2720 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2722 dbgs() << ' ' << J->first;
2725 // Examine each offset.
2726 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2728 const SCEV *OrigReg = J->second;
2730 int64_t JImm = J->first;
2731 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2733 if (!isa<SCEVConstant>(OrigReg) &&
2734 UsedByIndicesMap[Reg].count() == 1) {
2735 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2739 // Conservatively examine offsets between this orig reg a few selected
2741 ImmMapTy::const_iterator OtherImms[] = {
2742 Imms.begin(), prior(Imms.end()),
2743 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2745 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2746 ImmMapTy::const_iterator M = OtherImms[i];
2747 if (M == J || M == JE) continue;
2749 // Compute the difference between the two.
2750 int64_t Imm = (uint64_t)JImm - M->first;
2751 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2752 LUIdx = UsedByIndices.find_next(LUIdx))
2753 // Make a memo of this use, offset, and register tuple.
2754 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2755 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2762 UsedByIndicesMap.clear();
2763 UniqueItems.clear();
2765 // Now iterate through the worklist and add new formulae.
2766 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2767 E = WorkItems.end(); I != E; ++I) {
2768 const WorkItem &WI = *I;
2769 size_t LUIdx = WI.LUIdx;
2770 LSRUse &LU = Uses[LUIdx];
2771 int64_t Imm = WI.Imm;
2772 const SCEV *OrigReg = WI.OrigReg;
2774 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2775 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2776 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2778 // TODO: Use a more targeted data structure.
2779 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2780 const Formula &F = LU.Formulae[L];
2781 // Use the immediate in the scaled register.
2782 if (F.ScaledReg == OrigReg) {
2783 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2784 Imm * (uint64_t)F.AM.Scale;
2785 // Don't create 50 + reg(-50).
2786 if (F.referencesReg(SE.getSCEV(
2787 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2790 NewF.AM.BaseOffs = Offs;
2791 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2792 LU.Kind, LU.AccessTy, TLI))
2794 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2796 // If the new scale is a constant in a register, and adding the constant
2797 // value to the immediate would produce a value closer to zero than the
2798 // immediate itself, then the formula isn't worthwhile.
2799 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2800 if (C->getValue()->isNegative() !=
2801 (NewF.AM.BaseOffs < 0) &&
2802 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2803 .ule(abs64(NewF.AM.BaseOffs)))
2807 (void)InsertFormula(LU, LUIdx, NewF);
2809 // Use the immediate in a base register.
2810 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2811 const SCEV *BaseReg = F.BaseRegs[N];
2812 if (BaseReg != OrigReg)
2815 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2816 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2817 LU.Kind, LU.AccessTy, TLI)) {
2819 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
2822 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
2824 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2826 // If the new formula has a constant in a register, and adding the
2827 // constant value to the immediate would produce a value closer to
2828 // zero than the immediate itself, then the formula isn't worthwhile.
2829 for (SmallVectorImpl<const SCEV *>::const_iterator
2830 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2832 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2833 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2834 abs64(NewF.AM.BaseOffs)) &&
2835 (C->getValue()->getValue() +
2836 NewF.AM.BaseOffs).countTrailingZeros() >=
2837 CountTrailingZeros_64(NewF.AM.BaseOffs))
2841 (void)InsertFormula(LU, LUIdx, NewF);
2850 /// GenerateAllReuseFormulae - Generate formulae for each use.
2852 LSRInstance::GenerateAllReuseFormulae() {
2853 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2854 // queries are more precise.
2855 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2856 LSRUse &LU = Uses[LUIdx];
2857 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2858 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2859 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2860 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2862 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2863 LSRUse &LU = Uses[LUIdx];
2864 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2865 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2866 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2867 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2868 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2869 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2870 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2871 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2873 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2874 LSRUse &LU = Uses[LUIdx];
2875 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2876 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2879 GenerateCrossUseConstantOffsets();
2881 DEBUG(dbgs() << "\n"
2882 "After generating reuse formulae:\n";
2883 print_uses(dbgs()));
2886 /// If there are multiple formulae with the same set of registers used
2887 /// by other uses, pick the best one and delete the others.
2888 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2889 DenseSet<const SCEV *> VisitedRegs;
2890 SmallPtrSet<const SCEV *, 16> Regs;
2892 bool ChangedFormulae = false;
2895 // Collect the best formula for each unique set of shared registers. This
2896 // is reset for each use.
2897 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2899 BestFormulaeTy BestFormulae;
2901 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2902 LSRUse &LU = Uses[LUIdx];
2903 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2906 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2907 FIdx != NumForms; ++FIdx) {
2908 Formula &F = LU.Formulae[FIdx];
2910 SmallVector<const SCEV *, 2> Key;
2911 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2912 JE = F.BaseRegs.end(); J != JE; ++J) {
2913 const SCEV *Reg = *J;
2914 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2918 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2919 Key.push_back(F.ScaledReg);
2920 // Unstable sort by host order ok, because this is only used for
2922 std::sort(Key.begin(), Key.end());
2924 std::pair<BestFormulaeTy::const_iterator, bool> P =
2925 BestFormulae.insert(std::make_pair(Key, FIdx));
2927 Formula &Best = LU.Formulae[P.first->second];
2930 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2933 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2935 if (CostF < CostBest)
2937 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2939 " in favor of formula "; Best.print(dbgs());
2942 ChangedFormulae = true;
2944 LU.DeleteFormula(F);
2952 // Now that we've filtered out some formulae, recompute the Regs set.
2954 LU.RecomputeRegs(LUIdx, RegUses);
2956 // Reset this to prepare for the next use.
2957 BestFormulae.clear();
2960 DEBUG(if (ChangedFormulae) {
2962 "After filtering out undesirable candidates:\n";
2967 // This is a rough guess that seems to work fairly well.
2968 static const size_t ComplexityLimit = UINT16_MAX;
2970 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2971 /// solutions the solver might have to consider. It almost never considers
2972 /// this many solutions because it prune the search space, but the pruning
2973 /// isn't always sufficient.
2974 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
2976 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2977 E = Uses.end(); I != E; ++I) {
2978 size_t FSize = I->Formulae.size();
2979 if (FSize >= ComplexityLimit) {
2980 Power = ComplexityLimit;
2984 if (Power >= ComplexityLimit)
2990 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
2991 /// of the registers of another formula, it won't help reduce register
2992 /// pressure (though it may not necessarily hurt register pressure); remove
2993 /// it to simplify the system.
2994 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
2995 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
2996 DEBUG(dbgs() << "The search space is too complex.\n");
2998 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
2999 "which use a superset of registers used by other "
3002 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3003 LSRUse &LU = Uses[LUIdx];
3005 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3006 Formula &F = LU.Formulae[i];
3007 // Look for a formula with a constant or GV in a register. If the use
3008 // also has a formula with that same value in an immediate field,
3009 // delete the one that uses a register.
3010 for (SmallVectorImpl<const SCEV *>::const_iterator
3011 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3012 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3014 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3015 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3016 (I - F.BaseRegs.begin()));
3017 if (LU.HasFormulaWithSameRegs(NewF)) {
3018 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3019 LU.DeleteFormula(F);
3025 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3026 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3029 NewF.AM.BaseGV = GV;
3030 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3031 (I - F.BaseRegs.begin()));
3032 if (LU.HasFormulaWithSameRegs(NewF)) {
3033 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3035 LU.DeleteFormula(F);
3046 LU.RecomputeRegs(LUIdx, RegUses);
3049 DEBUG(dbgs() << "After pre-selection:\n";
3050 print_uses(dbgs()));
3054 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3055 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3057 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3058 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3059 DEBUG(dbgs() << "The search space is too complex.\n");
3061 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3062 "separated by a constant offset will use the same "
3065 // This is especially useful for unrolled loops.
3067 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3068 LSRUse &LU = Uses[LUIdx];
3069 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3070 E = LU.Formulae.end(); I != E; ++I) {
3071 const Formula &F = *I;
3072 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3073 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3074 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3075 /*HasBaseReg=*/false,
3076 LU.Kind, LU.AccessTy)) {
3077 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3080 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3082 // Update the relocs to reference the new use.
3083 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3084 E = Fixups.end(); I != E; ++I) {
3085 LSRFixup &Fixup = *I;
3086 if (Fixup.LUIdx == LUIdx) {
3087 Fixup.LUIdx = LUThatHas - &Uses.front();
3088 Fixup.Offset += F.AM.BaseOffs;
3089 // Add the new offset to LUThatHas' offset list.
3090 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3091 LUThatHas->Offsets.push_back(Fixup.Offset);
3092 if (Fixup.Offset > LUThatHas->MaxOffset)
3093 LUThatHas->MaxOffset = Fixup.Offset;
3094 if (Fixup.Offset < LUThatHas->MinOffset)
3095 LUThatHas->MinOffset = Fixup.Offset;
3097 DEBUG(dbgs() << "New fixup has offset "
3098 << Fixup.Offset << '\n');
3100 if (Fixup.LUIdx == NumUses-1)
3101 Fixup.LUIdx = LUIdx;
3104 // Delete formulae from the new use which are no longer legal.
3106 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3107 Formula &F = LUThatHas->Formulae[i];
3108 if (!isLegalUse(F.AM,
3109 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3110 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3111 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3113 LUThatHas->DeleteFormula(F);
3120 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3122 // Delete the old use.
3123 DeleteUse(LU, LUIdx);
3133 DEBUG(dbgs() << "After pre-selection:\n";
3134 print_uses(dbgs()));
3138 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3139 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3140 /// we've done more filtering, as it may be able to find more formulae to
3142 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3143 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3144 DEBUG(dbgs() << "The search space is too complex.\n");
3146 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3147 "undesirable dedicated registers.\n");
3149 FilterOutUndesirableDedicatedRegisters();
3151 DEBUG(dbgs() << "After pre-selection:\n";
3152 print_uses(dbgs()));
3156 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3157 /// to be profitable, and then in any use which has any reference to that
3158 /// register, delete all formulae which do not reference that register.
3159 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3160 // With all other options exhausted, loop until the system is simple
3161 // enough to handle.
3162 SmallPtrSet<const SCEV *, 4> Taken;
3163 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3164 // Ok, we have too many of formulae on our hands to conveniently handle.
3165 // Use a rough heuristic to thin out the list.
3166 DEBUG(dbgs() << "The search space is too complex.\n");
3168 // Pick the register which is used by the most LSRUses, which is likely
3169 // to be a good reuse register candidate.
3170 const SCEV *Best = 0;
3171 unsigned BestNum = 0;
3172 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3174 const SCEV *Reg = *I;
3175 if (Taken.count(Reg))
3180 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3181 if (Count > BestNum) {
3188 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3189 << " will yield profitable reuse.\n");
3192 // In any use with formulae which references this register, delete formulae
3193 // which don't reference it.
3194 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3195 LSRUse &LU = Uses[LUIdx];
3196 if (!LU.Regs.count(Best)) continue;
3199 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3200 Formula &F = LU.Formulae[i];
3201 if (!F.referencesReg(Best)) {
3202 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3203 LU.DeleteFormula(F);
3207 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3213 LU.RecomputeRegs(LUIdx, RegUses);
3216 DEBUG(dbgs() << "After pre-selection:\n";
3217 print_uses(dbgs()));
3221 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3222 /// formulae to choose from, use some rough heuristics to prune down the number
3223 /// of formulae. This keeps the main solver from taking an extraordinary amount
3224 /// of time in some worst-case scenarios.
3225 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3226 NarrowSearchSpaceByDetectingSupersets();
3227 NarrowSearchSpaceByCollapsingUnrolledCode();
3228 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3229 NarrowSearchSpaceByPickingWinnerRegs();
3232 /// SolveRecurse - This is the recursive solver.
3233 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3235 SmallVectorImpl<const Formula *> &Workspace,
3236 const Cost &CurCost,
3237 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3238 DenseSet<const SCEV *> &VisitedRegs) const {
3241 // - use more aggressive filtering
3242 // - sort the formula so that the most profitable solutions are found first
3243 // - sort the uses too
3245 // - don't compute a cost, and then compare. compare while computing a cost
3247 // - track register sets with SmallBitVector
3249 const LSRUse &LU = Uses[Workspace.size()];
3251 // If this use references any register that's already a part of the
3252 // in-progress solution, consider it a requirement that a formula must
3253 // reference that register in order to be considered. This prunes out
3254 // unprofitable searching.
3255 SmallSetVector<const SCEV *, 4> ReqRegs;
3256 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3257 E = CurRegs.end(); I != E; ++I)
3258 if (LU.Regs.count(*I))
3261 bool AnySatisfiedReqRegs = false;
3262 SmallPtrSet<const SCEV *, 16> NewRegs;
3265 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3266 E = LU.Formulae.end(); I != E; ++I) {
3267 const Formula &F = *I;
3269 // Ignore formulae which do not use any of the required registers.
3270 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3271 JE = ReqRegs.end(); J != JE; ++J) {
3272 const SCEV *Reg = *J;
3273 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3274 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3278 AnySatisfiedReqRegs = true;
3280 // Evaluate the cost of the current formula. If it's already worse than
3281 // the current best, prune the search at that point.
3284 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3285 if (NewCost < SolutionCost) {
3286 Workspace.push_back(&F);
3287 if (Workspace.size() != Uses.size()) {
3288 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3289 NewRegs, VisitedRegs);
3290 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3291 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3293 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3294 dbgs() << ". Regs:";
3295 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3296 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3297 dbgs() << ' ' << **I;
3300 SolutionCost = NewCost;
3301 Solution = Workspace;
3303 Workspace.pop_back();
3308 // If none of the formulae had all of the required registers, relax the
3309 // constraint so that we don't exclude all formulae.
3310 if (!AnySatisfiedReqRegs) {
3311 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3317 /// Solve - Choose one formula from each use. Return the results in the given
3318 /// Solution vector.
3319 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3320 SmallVector<const Formula *, 8> Workspace;
3322 SolutionCost.Loose();
3324 SmallPtrSet<const SCEV *, 16> CurRegs;
3325 DenseSet<const SCEV *> VisitedRegs;
3326 Workspace.reserve(Uses.size());
3328 // SolveRecurse does all the work.
3329 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3330 CurRegs, VisitedRegs);
3332 // Ok, we've now made all our decisions.
3333 DEBUG(dbgs() << "\n"
3334 "The chosen solution requires "; SolutionCost.print(dbgs());
3336 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3338 Uses[i].print(dbgs());
3341 Solution[i]->print(dbgs());
3345 assert(Solution.size() == Uses.size() && "Malformed solution!");
3348 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3349 /// the dominator tree far as we can go while still being dominated by the
3350 /// input positions. This helps canonicalize the insert position, which
3351 /// encourages sharing.
3352 BasicBlock::iterator
3353 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3354 const SmallVectorImpl<Instruction *> &Inputs)
3357 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3358 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3361 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3362 if (!Rung) return IP;
3363 Rung = Rung->getIDom();
3364 if (!Rung) return IP;
3365 IDom = Rung->getBlock();
3367 // Don't climb into a loop though.
3368 const Loop *IDomLoop = LI.getLoopFor(IDom);
3369 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3370 if (IDomDepth <= IPLoopDepth &&
3371 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3375 bool AllDominate = true;
3376 Instruction *BetterPos = 0;
3377 Instruction *Tentative = IDom->getTerminator();
3378 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3379 E = Inputs.end(); I != E; ++I) {
3380 Instruction *Inst = *I;
3381 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3382 AllDominate = false;
3385 // Attempt to find an insert position in the middle of the block,
3386 // instead of at the end, so that it can be used for other expansions.
3387 if (IDom == Inst->getParent() &&
3388 (!BetterPos || DT.dominates(BetterPos, Inst)))
3389 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3402 /// AdjustInsertPositionForExpand - Determine an input position which will be
3403 /// dominated by the operands and which will dominate the result.
3404 BasicBlock::iterator
3405 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3407 const LSRUse &LU) const {
3408 // Collect some instructions which must be dominated by the
3409 // expanding replacement. These must be dominated by any operands that
3410 // will be required in the expansion.
3411 SmallVector<Instruction *, 4> Inputs;
3412 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3413 Inputs.push_back(I);
3414 if (LU.Kind == LSRUse::ICmpZero)
3415 if (Instruction *I =
3416 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3417 Inputs.push_back(I);
3418 if (LF.PostIncLoops.count(L)) {
3419 if (LF.isUseFullyOutsideLoop(L))
3420 Inputs.push_back(L->getLoopLatch()->getTerminator());
3422 Inputs.push_back(IVIncInsertPos);
3424 // The expansion must also be dominated by the increment positions of any
3425 // loops it for which it is using post-inc mode.
3426 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3427 E = LF.PostIncLoops.end(); I != E; ++I) {
3428 const Loop *PIL = *I;
3429 if (PIL == L) continue;
3431 // Be dominated by the loop exit.
3432 SmallVector<BasicBlock *, 4> ExitingBlocks;
3433 PIL->getExitingBlocks(ExitingBlocks);
3434 if (!ExitingBlocks.empty()) {
3435 BasicBlock *BB = ExitingBlocks[0];
3436 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3437 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3438 Inputs.push_back(BB->getTerminator());
3442 // Then, climb up the immediate dominator tree as far as we can go while
3443 // still being dominated by the input positions.
3444 IP = HoistInsertPosition(IP, Inputs);
3446 // Don't insert instructions before PHI nodes.
3447 while (isa<PHINode>(IP)) ++IP;
3449 // Ignore landingpad instructions.
3450 while (isa<LandingPadInst>(IP)) ++IP;
3452 // Ignore debug intrinsics.
3453 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3458 /// Expand - Emit instructions for the leading candidate expression for this
3459 /// LSRUse (this is called "expanding").
3460 Value *LSRInstance::Expand(const LSRFixup &LF,
3462 BasicBlock::iterator IP,
3463 SCEVExpander &Rewriter,
3464 SmallVectorImpl<WeakVH> &DeadInsts) const {
3465 const LSRUse &LU = Uses[LF.LUIdx];
3467 // Determine an input position which will be dominated by the operands and
3468 // which will dominate the result.
3469 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3471 // Inform the Rewriter if we have a post-increment use, so that it can
3472 // perform an advantageous expansion.
3473 Rewriter.setPostInc(LF.PostIncLoops);
3475 // This is the type that the user actually needs.
3476 Type *OpTy = LF.OperandValToReplace->getType();
3477 // This will be the type that we'll initially expand to.
3478 Type *Ty = F.getType();
3480 // No type known; just expand directly to the ultimate type.
3482 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3483 // Expand directly to the ultimate type if it's the right size.
3485 // This is the type to do integer arithmetic in.
3486 Type *IntTy = SE.getEffectiveSCEVType(Ty);
3488 // Build up a list of operands to add together to form the full base.
3489 SmallVector<const SCEV *, 8> Ops;
3491 // Expand the BaseRegs portion.
3492 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3493 E = F.BaseRegs.end(); I != E; ++I) {
3494 const SCEV *Reg = *I;
3495 assert(!Reg->isZero() && "Zero allocated in a base register!");
3497 // If we're expanding for a post-inc user, make the post-inc adjustment.
3498 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3499 Reg = TransformForPostIncUse(Denormalize, Reg,
3500 LF.UserInst, LF.OperandValToReplace,
3503 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3506 // Flush the operand list to suppress SCEVExpander hoisting.
3508 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3510 Ops.push_back(SE.getUnknown(FullV));
3513 // Expand the ScaledReg portion.
3514 Value *ICmpScaledV = 0;
3515 if (F.AM.Scale != 0) {
3516 const SCEV *ScaledS = F.ScaledReg;
3518 // If we're expanding for a post-inc user, make the post-inc adjustment.
3519 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3520 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3521 LF.UserInst, LF.OperandValToReplace,
3524 if (LU.Kind == LSRUse::ICmpZero) {
3525 // An interesting way of "folding" with an icmp is to use a negated
3526 // scale, which we'll implement by inserting it into the other operand
3528 assert(F.AM.Scale == -1 &&
3529 "The only scale supported by ICmpZero uses is -1!");
3530 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3532 // Otherwise just expand the scaled register and an explicit scale,
3533 // which is expected to be matched as part of the address.
3534 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3535 ScaledS = SE.getMulExpr(ScaledS,
3536 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3537 Ops.push_back(ScaledS);
3539 // Flush the operand list to suppress SCEVExpander hoisting.
3540 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3542 Ops.push_back(SE.getUnknown(FullV));
3546 // Expand the GV portion.
3548 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3550 // Flush the operand list to suppress SCEVExpander hoisting.
3551 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3553 Ops.push_back(SE.getUnknown(FullV));
3556 // Expand the immediate portion.
3557 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3559 if (LU.Kind == LSRUse::ICmpZero) {
3560 // The other interesting way of "folding" with an ICmpZero is to use a
3561 // negated immediate.
3563 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3565 Ops.push_back(SE.getUnknown(ICmpScaledV));
3566 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3569 // Just add the immediate values. These again are expected to be matched
3570 // as part of the address.
3571 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3575 // Expand the unfolded offset portion.
3576 int64_t UnfoldedOffset = F.UnfoldedOffset;
3577 if (UnfoldedOffset != 0) {
3578 // Just add the immediate values.
3579 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
3583 // Emit instructions summing all the operands.
3584 const SCEV *FullS = Ops.empty() ?
3585 SE.getConstant(IntTy, 0) :
3587 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3589 // We're done expanding now, so reset the rewriter.
3590 Rewriter.clearPostInc();
3592 // An ICmpZero Formula represents an ICmp which we're handling as a
3593 // comparison against zero. Now that we've expanded an expression for that
3594 // form, update the ICmp's other operand.
3595 if (LU.Kind == LSRUse::ICmpZero) {
3596 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3597 DeadInsts.push_back(CI->getOperand(1));
3598 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3599 "a scale at the same time!");
3600 if (F.AM.Scale == -1) {
3601 if (ICmpScaledV->getType() != OpTy) {
3603 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3605 ICmpScaledV, OpTy, "tmp", CI);
3608 CI->setOperand(1, ICmpScaledV);
3610 assert(F.AM.Scale == 0 &&
3611 "ICmp does not support folding a global value and "
3612 "a scale at the same time!");
3613 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3615 if (C->getType() != OpTy)
3616 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3620 CI->setOperand(1, C);
3627 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3628 /// of their operands effectively happens in their predecessor blocks, so the
3629 /// expression may need to be expanded in multiple places.
3630 void LSRInstance::RewriteForPHI(PHINode *PN,
3633 SCEVExpander &Rewriter,
3634 SmallVectorImpl<WeakVH> &DeadInsts,
3636 DenseMap<BasicBlock *, Value *> Inserted;
3637 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3638 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3639 BasicBlock *BB = PN->getIncomingBlock(i);
3641 // If this is a critical edge, split the edge so that we do not insert
3642 // the code on all predecessor/successor paths. We do this unless this
3643 // is the canonical backedge for this loop, which complicates post-inc
3645 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3646 !isa<IndirectBrInst>(BB->getTerminator())) {
3647 BasicBlock *Parent = PN->getParent();
3648 Loop *PNLoop = LI.getLoopFor(Parent);
3649 if (!PNLoop || Parent != PNLoop->getHeader()) {
3650 // Split the critical edge.
3651 BasicBlock *NewBB = 0;
3652 if (!Parent->isLandingPad()) {
3653 NewBB = SplitCriticalEdge(BB, Parent, P);
3655 SmallVector<BasicBlock*, 2> NewBBs;
3656 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
3660 // If PN is outside of the loop and BB is in the loop, we want to
3661 // move the block to be immediately before the PHI block, not
3662 // immediately after BB.
3663 if (L->contains(BB) && !L->contains(PN))
3664 NewBB->moveBefore(PN->getParent());
3666 // Splitting the edge can reduce the number of PHI entries we have.
3667 e = PN->getNumIncomingValues();
3669 i = PN->getBasicBlockIndex(BB);
3673 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3674 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3676 PN->setIncomingValue(i, Pair.first->second);
3678 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3680 // If this is reuse-by-noop-cast, insert the noop cast.
3681 Type *OpTy = LF.OperandValToReplace->getType();
3682 if (FullV->getType() != OpTy)
3684 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3686 FullV, LF.OperandValToReplace->getType(),
3687 "tmp", BB->getTerminator());
3689 PN->setIncomingValue(i, FullV);
3690 Pair.first->second = FullV;
3695 /// Rewrite - Emit instructions for the leading candidate expression for this
3696 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3697 /// the newly expanded value.
3698 void LSRInstance::Rewrite(const LSRFixup &LF,
3700 SCEVExpander &Rewriter,
3701 SmallVectorImpl<WeakVH> &DeadInsts,
3703 // First, find an insertion point that dominates UserInst. For PHI nodes,
3704 // find the nearest block which dominates all the relevant uses.
3705 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3706 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3708 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3710 // If this is reuse-by-noop-cast, insert the noop cast.
3711 Type *OpTy = LF.OperandValToReplace->getType();
3712 if (FullV->getType() != OpTy) {
3714 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3715 FullV, OpTy, "tmp", LF.UserInst);
3719 // Update the user. ICmpZero is handled specially here (for now) because
3720 // Expand may have updated one of the operands of the icmp already, and
3721 // its new value may happen to be equal to LF.OperandValToReplace, in
3722 // which case doing replaceUsesOfWith leads to replacing both operands
3723 // with the same value. TODO: Reorganize this.
3724 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3725 LF.UserInst->setOperand(0, FullV);
3727 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3730 DeadInsts.push_back(LF.OperandValToReplace);
3733 /// ImplementSolution - Rewrite all the fixup locations with new values,
3734 /// following the chosen solution.
3736 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3738 // Keep track of instructions we may have made dead, so that
3739 // we can remove them after we are done working.
3740 SmallVector<WeakVH, 16> DeadInsts;
3742 SCEVExpander Rewriter(SE, "lsr");
3743 Rewriter.disableCanonicalMode();
3744 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3746 // Expand the new value definitions and update the users.
3747 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3748 E = Fixups.end(); I != E; ++I) {
3749 const LSRFixup &Fixup = *I;
3751 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3756 // Clean up after ourselves. This must be done before deleting any
3760 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3763 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3764 : IU(P->getAnalysis<IVUsers>()),
3765 SE(P->getAnalysis<ScalarEvolution>()),
3766 DT(P->getAnalysis<DominatorTree>()),
3767 LI(P->getAnalysis<LoopInfo>()),
3768 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3770 // If LoopSimplify form is not available, stay out of trouble.
3771 if (!L->isLoopSimplifyForm()) return;
3773 // If there's no interesting work to be done, bail early.
3774 if (IU.empty()) return;
3776 DEBUG(dbgs() << "\nLSR on loop ";
3777 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3780 // First, perform some low-level loop optimizations.
3782 OptimizeLoopTermCond();
3784 // If loop preparation eliminates all interesting IV users, bail.
3785 if (IU.empty()) return;
3787 // Start collecting data and preparing for the solver.
3788 CollectInterestingTypesAndFactors();
3789 CollectFixupsAndInitialFormulae();
3790 CollectLoopInvariantFixupsAndFormulae();
3792 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3793 print_uses(dbgs()));
3795 // Now use the reuse data to generate a bunch of interesting ways
3796 // to formulate the values needed for the uses.
3797 GenerateAllReuseFormulae();
3799 FilterOutUndesirableDedicatedRegisters();
3800 NarrowSearchSpaceUsingHeuristics();
3802 SmallVector<const Formula *, 8> Solution;
3805 // Release memory that is no longer needed.
3811 // Formulae should be legal.
3812 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3813 E = Uses.end(); I != E; ++I) {
3814 const LSRUse &LU = *I;
3815 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3816 JE = LU.Formulae.end(); J != JE; ++J)
3817 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3818 LU.Kind, LU.AccessTy, TLI) &&
3819 "Illegal formula generated!");
3823 // Now that we've decided what we want, make it so.
3824 ImplementSolution(Solution, P);
3827 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3828 if (Factors.empty() && Types.empty()) return;
3830 OS << "LSR has identified the following interesting factors and types: ";
3833 for (SmallSetVector<int64_t, 8>::const_iterator
3834 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3835 if (!First) OS << ", ";
3840 for (SmallSetVector<Type *, 4>::const_iterator
3841 I = Types.begin(), E = Types.end(); I != E; ++I) {
3842 if (!First) OS << ", ";
3844 OS << '(' << **I << ')';
3849 void LSRInstance::print_fixups(raw_ostream &OS) const {
3850 OS << "LSR is examining the following fixup sites:\n";
3851 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3852 E = Fixups.end(); I != E; ++I) {
3859 void LSRInstance::print_uses(raw_ostream &OS) const {
3860 OS << "LSR is examining the following uses:\n";
3861 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3862 E = Uses.end(); I != E; ++I) {
3863 const LSRUse &LU = *I;
3867 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3868 JE = LU.Formulae.end(); J != JE; ++J) {
3876 void LSRInstance::print(raw_ostream &OS) const {
3877 print_factors_and_types(OS);
3882 void LSRInstance::dump() const {
3883 print(errs()); errs() << '\n';
3888 class LoopStrengthReduce : public LoopPass {
3889 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3890 /// transformation profitability.
3891 const TargetLowering *const TLI;
3894 static char ID; // Pass ID, replacement for typeid
3895 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3898 bool runOnLoop(Loop *L, LPPassManager &LPM);
3899 void getAnalysisUsage(AnalysisUsage &AU) const;
3904 char LoopStrengthReduce::ID = 0;
3905 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
3906 "Loop Strength Reduction", false, false)
3907 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3908 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3909 INITIALIZE_PASS_DEPENDENCY(IVUsers)
3910 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
3911 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3912 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
3913 "Loop Strength Reduction", false, false)
3916 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3917 return new LoopStrengthReduce(TLI);
3920 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3921 : LoopPass(ID), TLI(tli) {
3922 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
3925 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3926 // We split critical edges, so we change the CFG. However, we do update
3927 // many analyses if they are around.
3928 AU.addPreservedID(LoopSimplifyID);
3930 AU.addRequired<LoopInfo>();
3931 AU.addPreserved<LoopInfo>();
3932 AU.addRequiredID(LoopSimplifyID);
3933 AU.addRequired<DominatorTree>();
3934 AU.addPreserved<DominatorTree>();
3935 AU.addRequired<ScalarEvolution>();
3936 AU.addPreserved<ScalarEvolution>();
3937 // Requiring LoopSimplify a second time here prevents IVUsers from running
3938 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
3939 AU.addRequiredID(LoopSimplifyID);
3940 AU.addRequired<IVUsers>();
3941 AU.addPreserved<IVUsers>();
3944 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3945 bool Changed = false;
3947 // Run the main LSR transformation.
3948 Changed |= LSRInstance(TLI, L, this).getChanged();
3950 // At this point, it is worth checking to see if any recurrence PHIs are also
3951 // dead, so that we can remove them as well.
3952 Changed |= DeleteDeadPHIs(L->getHeader());