1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Assembly/Writer.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/ADT/SmallBitVector.h"
70 #include "llvm/ADT/SetVector.h"
71 #include "llvm/ADT/DenseSet.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Target/TargetLowering.h"
80 static cl::opt<bool> EnableNested(
81 "enable-lsr-nested", cl::Hidden, cl::desc("Enable LSR on nested loops"));
83 static cl::opt<bool> EnableRetry(
84 "enable-lsr-retry", cl::Hidden, cl::desc("Enable LSR retry"));
86 // Temporary flag to cleanup congruent phis after LSR phi expansion.
87 // It's currently disabled until we can determine whether it's truly useful or
88 // not. The flag should be removed after the v3.0 release.
89 static cl::opt<bool> EnablePhiElim(
90 "enable-lsr-phielim", cl::Hidden, cl::desc("Enable LSR phi elimination"));
94 /// RegSortData - This class holds data which is used to order reuse candidates.
97 /// UsedByIndices - This represents the set of LSRUse indices which reference
98 /// a particular register.
99 SmallBitVector UsedByIndices;
103 void print(raw_ostream &OS) const;
109 void RegSortData::print(raw_ostream &OS) const {
110 OS << "[NumUses=" << UsedByIndices.count() << ']';
113 void RegSortData::dump() const {
114 print(errs()); errs() << '\n';
119 /// RegUseTracker - Map register candidates to information about how they are
121 class RegUseTracker {
122 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
124 RegUsesTy RegUsesMap;
125 SmallVector<const SCEV *, 16> RegSequence;
128 void CountRegister(const SCEV *Reg, size_t LUIdx);
129 void DropRegister(const SCEV *Reg, size_t LUIdx);
130 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
132 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
134 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
138 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
139 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
140 iterator begin() { return RegSequence.begin(); }
141 iterator end() { return RegSequence.end(); }
142 const_iterator begin() const { return RegSequence.begin(); }
143 const_iterator end() const { return RegSequence.end(); }
149 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
150 std::pair<RegUsesTy::iterator, bool> Pair =
151 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
152 RegSortData &RSD = Pair.first->second;
154 RegSequence.push_back(Reg);
155 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
156 RSD.UsedByIndices.set(LUIdx);
160 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
161 RegUsesTy::iterator It = RegUsesMap.find(Reg);
162 assert(It != RegUsesMap.end());
163 RegSortData &RSD = It->second;
164 assert(RSD.UsedByIndices.size() > LUIdx);
165 RSD.UsedByIndices.reset(LUIdx);
169 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
170 assert(LUIdx <= LastLUIdx);
172 // Update RegUses. The data structure is not optimized for this purpose;
173 // we must iterate through it and update each of the bit vectors.
174 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
176 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
177 if (LUIdx < UsedByIndices.size())
178 UsedByIndices[LUIdx] =
179 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
180 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
185 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
186 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
187 if (I == RegUsesMap.end())
189 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
190 int i = UsedByIndices.find_first();
191 if (i == -1) return false;
192 if ((size_t)i != LUIdx) return true;
193 return UsedByIndices.find_next(i) != -1;
196 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
197 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
198 assert(I != RegUsesMap.end() && "Unknown register!");
199 return I->second.UsedByIndices;
202 void RegUseTracker::clear() {
209 /// Formula - This class holds information that describes a formula for
210 /// computing satisfying a use. It may include broken-out immediates and scaled
213 /// AM - This is used to represent complex addressing, as well as other kinds
214 /// of interesting uses.
215 TargetLowering::AddrMode AM;
217 /// BaseRegs - The list of "base" registers for this use. When this is
218 /// non-empty, AM.HasBaseReg should be set to true.
219 SmallVector<const SCEV *, 2> BaseRegs;
221 /// ScaledReg - The 'scaled' register for this use. This should be non-null
222 /// when AM.Scale is not zero.
223 const SCEV *ScaledReg;
225 /// UnfoldedOffset - An additional constant offset which added near the
226 /// use. This requires a temporary register, but the offset itself can
227 /// live in an add immediate field rather than a register.
228 int64_t UnfoldedOffset;
230 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
232 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
234 unsigned getNumRegs() const;
235 Type *getType() const;
237 void DeleteBaseReg(const SCEV *&S);
239 bool referencesReg(const SCEV *S) const;
240 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
241 const RegUseTracker &RegUses) const;
243 void print(raw_ostream &OS) const;
249 /// DoInitialMatch - Recursion helper for InitialMatch.
250 static void DoInitialMatch(const SCEV *S, Loop *L,
251 SmallVectorImpl<const SCEV *> &Good,
252 SmallVectorImpl<const SCEV *> &Bad,
253 ScalarEvolution &SE) {
254 // Collect expressions which properly dominate the loop header.
255 if (SE.properlyDominates(S, L->getHeader())) {
260 // Look at add operands.
261 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
262 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
264 DoInitialMatch(*I, L, Good, Bad, SE);
268 // Look at addrec operands.
269 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
270 if (!AR->getStart()->isZero()) {
271 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
272 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
273 AR->getStepRecurrence(SE),
274 // FIXME: AR->getNoWrapFlags()
275 AR->getLoop(), SCEV::FlagAnyWrap),
280 // Handle a multiplication by -1 (negation) if it didn't fold.
281 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
282 if (Mul->getOperand(0)->isAllOnesValue()) {
283 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
284 const SCEV *NewMul = SE.getMulExpr(Ops);
286 SmallVector<const SCEV *, 4> MyGood;
287 SmallVector<const SCEV *, 4> MyBad;
288 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
289 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
290 SE.getEffectiveSCEVType(NewMul->getType())));
291 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
292 E = MyGood.end(); I != E; ++I)
293 Good.push_back(SE.getMulExpr(NegOne, *I));
294 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
295 E = MyBad.end(); I != E; ++I)
296 Bad.push_back(SE.getMulExpr(NegOne, *I));
300 // Ok, we can't do anything interesting. Just stuff the whole thing into a
301 // register and hope for the best.
305 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
306 /// attempting to keep all loop-invariant and loop-computable values in a
307 /// single base register.
308 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
309 SmallVector<const SCEV *, 4> Good;
310 SmallVector<const SCEV *, 4> Bad;
311 DoInitialMatch(S, L, Good, Bad, SE);
313 const SCEV *Sum = SE.getAddExpr(Good);
315 BaseRegs.push_back(Sum);
316 AM.HasBaseReg = true;
319 const SCEV *Sum = SE.getAddExpr(Bad);
321 BaseRegs.push_back(Sum);
322 AM.HasBaseReg = true;
326 /// getNumRegs - Return the total number of register operands used by this
327 /// formula. This does not include register uses implied by non-constant
329 unsigned Formula::getNumRegs() const {
330 return !!ScaledReg + BaseRegs.size();
333 /// getType - Return the type of this formula, if it has one, or null
334 /// otherwise. This type is meaningless except for the bit size.
335 Type *Formula::getType() const {
336 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
337 ScaledReg ? ScaledReg->getType() :
338 AM.BaseGV ? AM.BaseGV->getType() :
342 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
343 void Formula::DeleteBaseReg(const SCEV *&S) {
344 if (&S != &BaseRegs.back())
345 std::swap(S, BaseRegs.back());
349 /// referencesReg - Test if this formula references the given register.
350 bool Formula::referencesReg(const SCEV *S) const {
351 return S == ScaledReg ||
352 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
355 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
356 /// which are used by uses other than the use with the given index.
357 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
358 const RegUseTracker &RegUses) const {
360 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
362 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
363 E = BaseRegs.end(); I != E; ++I)
364 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
369 void Formula::print(raw_ostream &OS) const {
372 if (!First) OS << " + "; else First = false;
373 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
375 if (AM.BaseOffs != 0) {
376 if (!First) OS << " + "; else First = false;
379 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
380 E = BaseRegs.end(); I != E; ++I) {
381 if (!First) OS << " + "; else First = false;
382 OS << "reg(" << **I << ')';
384 if (AM.HasBaseReg && BaseRegs.empty()) {
385 if (!First) OS << " + "; else First = false;
386 OS << "**error: HasBaseReg**";
387 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
388 if (!First) OS << " + "; else First = false;
389 OS << "**error: !HasBaseReg**";
392 if (!First) OS << " + "; else First = false;
393 OS << AM.Scale << "*reg(";
400 if (UnfoldedOffset != 0) {
401 if (!First) OS << " + "; else First = false;
402 OS << "imm(" << UnfoldedOffset << ')';
406 void Formula::dump() const {
407 print(errs()); errs() << '\n';
410 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
411 /// without changing its value.
412 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
414 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
415 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
418 /// isAddSExtable - Return true if the given add can be sign-extended
419 /// without changing its value.
420 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
422 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
423 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
426 /// isMulSExtable - Return true if the given mul can be sign-extended
427 /// without changing its value.
428 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
430 IntegerType::get(SE.getContext(),
431 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
432 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
435 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
436 /// and if the remainder is known to be zero, or null otherwise. If
437 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
438 /// to Y, ignoring that the multiplication may overflow, which is useful when
439 /// the result will be used in a context where the most significant bits are
441 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
443 bool IgnoreSignificantBits = false) {
444 // Handle the trivial case, which works for any SCEV type.
446 return SE.getConstant(LHS->getType(), 1);
448 // Handle a few RHS special cases.
449 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
451 const APInt &RA = RC->getValue()->getValue();
452 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
454 if (RA.isAllOnesValue())
455 return SE.getMulExpr(LHS, RC);
456 // Handle x /s 1 as x.
461 // Check for a division of a constant by a constant.
462 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
465 const APInt &LA = C->getValue()->getValue();
466 const APInt &RA = RC->getValue()->getValue();
467 if (LA.srem(RA) != 0)
469 return SE.getConstant(LA.sdiv(RA));
472 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
473 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
474 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
475 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
476 IgnoreSignificantBits);
478 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
479 IgnoreSignificantBits);
480 if (!Start) return 0;
481 // FlagNW is independent of the start value, step direction, and is
482 // preserved with smaller magnitude steps.
483 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
484 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
489 // Distribute the sdiv over add operands, if the add doesn't overflow.
490 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
491 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
492 SmallVector<const SCEV *, 8> Ops;
493 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
495 const SCEV *Op = getExactSDiv(*I, RHS, SE,
496 IgnoreSignificantBits);
500 return SE.getAddExpr(Ops);
505 // Check for a multiply operand that we can pull RHS out of.
506 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
507 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
508 SmallVector<const SCEV *, 4> Ops;
510 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
514 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
515 IgnoreSignificantBits)) {
521 return Found ? SE.getMulExpr(Ops) : 0;
526 // Otherwise we don't know.
530 /// ExtractImmediate - If S involves the addition of a constant integer value,
531 /// return that integer value, and mutate S to point to a new SCEV with that
533 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
534 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
535 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
536 S = SE.getConstant(C->getType(), 0);
537 return C->getValue()->getSExtValue();
539 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
540 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
541 int64_t Result = ExtractImmediate(NewOps.front(), SE);
543 S = SE.getAddExpr(NewOps);
545 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
546 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
547 int64_t Result = ExtractImmediate(NewOps.front(), SE);
549 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
550 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
557 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
558 /// return that symbol, and mutate S to point to a new SCEV with that
560 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
561 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
562 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
563 S = SE.getConstant(GV->getType(), 0);
566 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
567 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
568 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
570 S = SE.getAddExpr(NewOps);
572 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
573 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
574 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
576 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
577 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
584 /// isAddressUse - Returns true if the specified instruction is using the
585 /// specified value as an address.
586 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
587 bool isAddress = isa<LoadInst>(Inst);
588 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
589 if (SI->getOperand(1) == OperandVal)
591 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
592 // Addressing modes can also be folded into prefetches and a variety
594 switch (II->getIntrinsicID()) {
596 case Intrinsic::prefetch:
597 case Intrinsic::x86_sse_storeu_ps:
598 case Intrinsic::x86_sse2_storeu_pd:
599 case Intrinsic::x86_sse2_storeu_dq:
600 case Intrinsic::x86_sse2_storel_dq:
601 if (II->getArgOperand(0) == OperandVal)
609 /// getAccessType - Return the type of the memory being accessed.
610 static Type *getAccessType(const Instruction *Inst) {
611 Type *AccessTy = Inst->getType();
612 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
613 AccessTy = SI->getOperand(0)->getType();
614 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
615 // Addressing modes can also be folded into prefetches and a variety
617 switch (II->getIntrinsicID()) {
619 case Intrinsic::x86_sse_storeu_ps:
620 case Intrinsic::x86_sse2_storeu_pd:
621 case Intrinsic::x86_sse2_storeu_dq:
622 case Intrinsic::x86_sse2_storel_dq:
623 AccessTy = II->getArgOperand(0)->getType();
628 // All pointers have the same requirements, so canonicalize them to an
629 // arbitrary pointer type to minimize variation.
630 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
631 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
632 PTy->getAddressSpace());
637 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
638 /// specified set are trivially dead, delete them and see if this makes any of
639 /// their operands subsequently dead.
641 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
642 bool Changed = false;
644 while (!DeadInsts.empty()) {
645 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
647 if (I == 0 || !isInstructionTriviallyDead(I))
650 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
651 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
654 DeadInsts.push_back(U);
657 I->eraseFromParent();
666 /// Cost - This class is used to measure and compare candidate formulae.
668 /// TODO: Some of these could be merged. Also, a lexical ordering
669 /// isn't always optimal.
673 unsigned NumBaseAdds;
679 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
682 bool operator<(const Cost &Other) const;
687 // Once any of the metrics loses, they must all remain losers.
689 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
690 | ImmCost | SetupCost) != ~0u)
691 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
692 & ImmCost & SetupCost) == ~0u);
697 assert(isValid() && "invalid cost");
698 return NumRegs == ~0u;
701 void RateFormula(const Formula &F,
702 SmallPtrSet<const SCEV *, 16> &Regs,
703 const DenseSet<const SCEV *> &VisitedRegs,
705 const SmallVectorImpl<int64_t> &Offsets,
706 ScalarEvolution &SE, DominatorTree &DT);
708 void print(raw_ostream &OS) const;
712 void RateRegister(const SCEV *Reg,
713 SmallPtrSet<const SCEV *, 16> &Regs,
715 ScalarEvolution &SE, DominatorTree &DT);
716 void RatePrimaryRegister(const SCEV *Reg,
717 SmallPtrSet<const SCEV *, 16> &Regs,
719 ScalarEvolution &SE, DominatorTree &DT);
724 /// RateRegister - Tally up interesting quantities from the given register.
725 void Cost::RateRegister(const SCEV *Reg,
726 SmallPtrSet<const SCEV *, 16> &Regs,
728 ScalarEvolution &SE, DominatorTree &DT) {
729 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
730 if (AR->getLoop() == L)
731 AddRecCost += 1; /// TODO: This should be a function of the stride.
733 // If this is an addrec for another loop, don't second-guess its addrec phi
734 // nodes. LSR isn't currently smart enough to reason about more than one
735 // loop at a time. LSR has either already run on inner loops, will not run
736 // on other loops, and cannot be expected to change sibling loops. If the
737 // AddRec exists, consider it's register free and leave it alone. Otherwise,
738 // do not consider this formula at all.
739 // FIXME: why do we need to generate such fomulae?
740 else if (!EnableNested || L->contains(AR->getLoop()) ||
741 (!AR->getLoop()->contains(L) &&
742 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
743 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
744 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
745 if (SE.isSCEVable(PN->getType()) &&
746 (SE.getEffectiveSCEVType(PN->getType()) ==
747 SE.getEffectiveSCEVType(AR->getType())) &&
748 SE.getSCEV(PN) == AR)
755 // If this isn't one of the addrecs that the loop already has, it
756 // would require a costly new phi and add. TODO: This isn't
757 // precisely modeled right now.
759 if (!Regs.count(AR->getStart())) {
760 RateRegister(AR->getStart(), Regs, L, SE, DT);
766 // Add the step value register, if it needs one.
767 // TODO: The non-affine case isn't precisely modeled here.
768 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
769 if (!Regs.count(AR->getOperand(1))) {
770 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
778 // Rough heuristic; favor registers which don't require extra setup
779 // instructions in the preheader.
780 if (!isa<SCEVUnknown>(Reg) &&
781 !isa<SCEVConstant>(Reg) &&
782 !(isa<SCEVAddRecExpr>(Reg) &&
783 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
784 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
787 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
788 SE.hasComputableLoopEvolution(Reg, L);
791 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
793 void Cost::RatePrimaryRegister(const SCEV *Reg,
794 SmallPtrSet<const SCEV *, 16> &Regs,
796 ScalarEvolution &SE, DominatorTree &DT) {
797 if (Regs.insert(Reg))
798 RateRegister(Reg, Regs, L, SE, DT);
801 void Cost::RateFormula(const Formula &F,
802 SmallPtrSet<const SCEV *, 16> &Regs,
803 const DenseSet<const SCEV *> &VisitedRegs,
805 const SmallVectorImpl<int64_t> &Offsets,
806 ScalarEvolution &SE, DominatorTree &DT) {
807 // Tally up the registers.
808 if (const SCEV *ScaledReg = F.ScaledReg) {
809 if (VisitedRegs.count(ScaledReg)) {
813 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
817 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
818 E = F.BaseRegs.end(); I != E; ++I) {
819 const SCEV *BaseReg = *I;
820 if (VisitedRegs.count(BaseReg)) {
824 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
829 // Determine how many (unfolded) adds we'll need inside the loop.
830 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
831 if (NumBaseParts > 1)
832 NumBaseAdds += NumBaseParts - 1;
834 // Tally up the non-zero immediates.
835 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
836 E = Offsets.end(); I != E; ++I) {
837 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
839 ImmCost += 64; // Handle symbolic values conservatively.
840 // TODO: This should probably be the pointer size.
841 else if (Offset != 0)
842 ImmCost += APInt(64, Offset, true).getMinSignedBits();
844 assert(isValid() && "invalid cost");
847 /// Loose - Set this cost to a losing value.
857 /// operator< - Choose the lower cost.
858 bool Cost::operator<(const Cost &Other) const {
859 if (NumRegs != Other.NumRegs)
860 return NumRegs < Other.NumRegs;
861 if (AddRecCost != Other.AddRecCost)
862 return AddRecCost < Other.AddRecCost;
863 if (NumIVMuls != Other.NumIVMuls)
864 return NumIVMuls < Other.NumIVMuls;
865 if (NumBaseAdds != Other.NumBaseAdds)
866 return NumBaseAdds < Other.NumBaseAdds;
867 if (ImmCost != Other.ImmCost)
868 return ImmCost < Other.ImmCost;
869 if (SetupCost != Other.SetupCost)
870 return SetupCost < Other.SetupCost;
874 void Cost::print(raw_ostream &OS) const {
875 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
877 OS << ", with addrec cost " << AddRecCost;
879 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
880 if (NumBaseAdds != 0)
881 OS << ", plus " << NumBaseAdds << " base add"
882 << (NumBaseAdds == 1 ? "" : "s");
884 OS << ", plus " << ImmCost << " imm cost";
886 OS << ", plus " << SetupCost << " setup cost";
889 void Cost::dump() const {
890 print(errs()); errs() << '\n';
895 /// LSRFixup - An operand value in an instruction which is to be replaced
896 /// with some equivalent, possibly strength-reduced, replacement.
898 /// UserInst - The instruction which will be updated.
899 Instruction *UserInst;
901 /// OperandValToReplace - The operand of the instruction which will
902 /// be replaced. The operand may be used more than once; every instance
903 /// will be replaced.
904 Value *OperandValToReplace;
906 /// PostIncLoops - If this user is to use the post-incremented value of an
907 /// induction variable, this variable is non-null and holds the loop
908 /// associated with the induction variable.
909 PostIncLoopSet PostIncLoops;
911 /// LUIdx - The index of the LSRUse describing the expression which
912 /// this fixup needs, minus an offset (below).
915 /// Offset - A constant offset to be added to the LSRUse expression.
916 /// This allows multiple fixups to share the same LSRUse with different
917 /// offsets, for example in an unrolled loop.
920 bool isUseFullyOutsideLoop(const Loop *L) const;
924 void print(raw_ostream &OS) const;
931 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
933 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
934 /// value outside of the given loop.
935 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
936 // PHI nodes use their value in their incoming blocks.
937 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
938 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
939 if (PN->getIncomingValue(i) == OperandValToReplace &&
940 L->contains(PN->getIncomingBlock(i)))
945 return !L->contains(UserInst);
948 void LSRFixup::print(raw_ostream &OS) const {
950 // Store is common and interesting enough to be worth special-casing.
951 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
953 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
954 } else if (UserInst->getType()->isVoidTy())
955 OS << UserInst->getOpcodeName();
957 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
959 OS << ", OperandValToReplace=";
960 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
962 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
963 E = PostIncLoops.end(); I != E; ++I) {
964 OS << ", PostIncLoop=";
965 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
968 if (LUIdx != ~size_t(0))
969 OS << ", LUIdx=" << LUIdx;
972 OS << ", Offset=" << Offset;
975 void LSRFixup::dump() const {
976 print(errs()); errs() << '\n';
981 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
982 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
983 struct UniquifierDenseMapInfo {
984 static SmallVector<const SCEV *, 2> getEmptyKey() {
985 SmallVector<const SCEV *, 2> V;
986 V.push_back(reinterpret_cast<const SCEV *>(-1));
990 static SmallVector<const SCEV *, 2> getTombstoneKey() {
991 SmallVector<const SCEV *, 2> V;
992 V.push_back(reinterpret_cast<const SCEV *>(-2));
996 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
998 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
999 E = V.end(); I != E; ++I)
1000 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1004 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1005 const SmallVector<const SCEV *, 2> &RHS) {
1010 /// LSRUse - This class holds the state that LSR keeps for each use in
1011 /// IVUsers, as well as uses invented by LSR itself. It includes information
1012 /// about what kinds of things can be folded into the user, information about
1013 /// the user itself, and information about how the use may be satisfied.
1014 /// TODO: Represent multiple users of the same expression in common?
1016 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1019 /// KindType - An enum for a kind of use, indicating what types of
1020 /// scaled and immediate operands it might support.
1022 Basic, ///< A normal use, with no folding.
1023 Special, ///< A special case of basic, allowing -1 scales.
1024 Address, ///< An address use; folding according to TargetLowering
1025 ICmpZero ///< An equality icmp with both operands folded into one.
1026 // TODO: Add a generic icmp too?
1032 SmallVector<int64_t, 8> Offsets;
1036 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1037 /// LSRUse are outside of the loop, in which case some special-case heuristics
1039 bool AllFixupsOutsideLoop;
1041 /// WidestFixupType - This records the widest use type for any fixup using
1042 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1043 /// max fixup widths to be equivalent, because the narrower one may be relying
1044 /// on the implicit truncation to truncate away bogus bits.
1045 Type *WidestFixupType;
1047 /// Formulae - A list of ways to build a value that can satisfy this user.
1048 /// After the list is populated, one of these is selected heuristically and
1049 /// used to formulate a replacement for OperandValToReplace in UserInst.
1050 SmallVector<Formula, 12> Formulae;
1052 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1053 SmallPtrSet<const SCEV *, 4> Regs;
1055 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1056 MinOffset(INT64_MAX),
1057 MaxOffset(INT64_MIN),
1058 AllFixupsOutsideLoop(true),
1059 WidestFixupType(0) {}
1061 bool HasFormulaWithSameRegs(const Formula &F) const;
1062 bool InsertFormula(const Formula &F);
1063 void DeleteFormula(Formula &F);
1064 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1066 void print(raw_ostream &OS) const;
1072 /// HasFormula - Test whether this use as a formula which has the same
1073 /// registers as the given formula.
1074 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1075 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1076 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1077 // Unstable sort by host order ok, because this is only used for uniquifying.
1078 std::sort(Key.begin(), Key.end());
1079 return Uniquifier.count(Key);
1082 /// InsertFormula - If the given formula has not yet been inserted, add it to
1083 /// the list, and return true. Return false otherwise.
1084 bool LSRUse::InsertFormula(const Formula &F) {
1085 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1086 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1087 // Unstable sort by host order ok, because this is only used for uniquifying.
1088 std::sort(Key.begin(), Key.end());
1090 if (!Uniquifier.insert(Key).second)
1093 // Using a register to hold the value of 0 is not profitable.
1094 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1095 "Zero allocated in a scaled register!");
1097 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1098 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1099 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1102 // Add the formula to the list.
1103 Formulae.push_back(F);
1105 // Record registers now being used by this use.
1106 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1107 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1112 /// DeleteFormula - Remove the given formula from this use's list.
1113 void LSRUse::DeleteFormula(Formula &F) {
1114 if (&F != &Formulae.back())
1115 std::swap(F, Formulae.back());
1116 Formulae.pop_back();
1117 assert(!Formulae.empty() && "LSRUse has no formulae left!");
1120 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1121 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1122 // Now that we've filtered out some formulae, recompute the Regs set.
1123 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1125 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1126 E = Formulae.end(); I != E; ++I) {
1127 const Formula &F = *I;
1128 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1129 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1132 // Update the RegTracker.
1133 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1134 E = OldRegs.end(); I != E; ++I)
1135 if (!Regs.count(*I))
1136 RegUses.DropRegister(*I, LUIdx);
1139 void LSRUse::print(raw_ostream &OS) const {
1140 OS << "LSR Use: Kind=";
1142 case Basic: OS << "Basic"; break;
1143 case Special: OS << "Special"; break;
1144 case ICmpZero: OS << "ICmpZero"; break;
1146 OS << "Address of ";
1147 if (AccessTy->isPointerTy())
1148 OS << "pointer"; // the full pointer type could be really verbose
1153 OS << ", Offsets={";
1154 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1155 E = Offsets.end(); I != E; ++I) {
1157 if (llvm::next(I) != E)
1162 if (AllFixupsOutsideLoop)
1163 OS << ", all-fixups-outside-loop";
1165 if (WidestFixupType)
1166 OS << ", widest fixup type: " << *WidestFixupType;
1169 void LSRUse::dump() const {
1170 print(errs()); errs() << '\n';
1173 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1174 /// be completely folded into the user instruction at isel time. This includes
1175 /// address-mode folding and special icmp tricks.
1176 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1177 LSRUse::KindType Kind, Type *AccessTy,
1178 const TargetLowering *TLI) {
1180 case LSRUse::Address:
1181 // If we have low-level target information, ask the target if it can
1182 // completely fold this address.
1183 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1185 // Otherwise, just guess that reg+reg addressing is legal.
1186 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1188 case LSRUse::ICmpZero:
1189 // There's not even a target hook for querying whether it would be legal to
1190 // fold a GV into an ICmp.
1194 // ICmp only has two operands; don't allow more than two non-trivial parts.
1195 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1198 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1199 // putting the scaled register in the other operand of the icmp.
1200 if (AM.Scale != 0 && AM.Scale != -1)
1203 // If we have low-level target information, ask the target if it can fold an
1204 // integer immediate on an icmp.
1205 if (AM.BaseOffs != 0) {
1206 if (TLI) return TLI->isLegalICmpImmediate(-(uint64_t)AM.BaseOffs);
1213 // Only handle single-register values.
1214 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1216 case LSRUse::Special:
1217 // Only handle -1 scales, or no scale.
1218 return AM.Scale == 0 || AM.Scale == -1;
1224 static bool isLegalUse(TargetLowering::AddrMode AM,
1225 int64_t MinOffset, int64_t MaxOffset,
1226 LSRUse::KindType Kind, Type *AccessTy,
1227 const TargetLowering *TLI) {
1228 // Check for overflow.
1229 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1232 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1233 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1234 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1235 // Check for overflow.
1236 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1239 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1240 return isLegalUse(AM, Kind, AccessTy, TLI);
1245 static bool isAlwaysFoldable(int64_t BaseOffs,
1246 GlobalValue *BaseGV,
1248 LSRUse::KindType Kind, Type *AccessTy,
1249 const TargetLowering *TLI) {
1250 // Fast-path: zero is always foldable.
1251 if (BaseOffs == 0 && !BaseGV) return true;
1253 // Conservatively, create an address with an immediate and a
1254 // base and a scale.
1255 TargetLowering::AddrMode AM;
1256 AM.BaseOffs = BaseOffs;
1258 AM.HasBaseReg = HasBaseReg;
1259 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1261 // Canonicalize a scale of 1 to a base register if the formula doesn't
1262 // already have a base register.
1263 if (!AM.HasBaseReg && AM.Scale == 1) {
1265 AM.HasBaseReg = true;
1268 return isLegalUse(AM, Kind, AccessTy, TLI);
1271 static bool isAlwaysFoldable(const SCEV *S,
1272 int64_t MinOffset, int64_t MaxOffset,
1274 LSRUse::KindType Kind, Type *AccessTy,
1275 const TargetLowering *TLI,
1276 ScalarEvolution &SE) {
1277 // Fast-path: zero is always foldable.
1278 if (S->isZero()) return true;
1280 // Conservatively, create an address with an immediate and a
1281 // base and a scale.
1282 int64_t BaseOffs = ExtractImmediate(S, SE);
1283 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1285 // If there's anything else involved, it's not foldable.
1286 if (!S->isZero()) return false;
1288 // Fast-path: zero is always foldable.
1289 if (BaseOffs == 0 && !BaseGV) return true;
1291 // Conservatively, create an address with an immediate and a
1292 // base and a scale.
1293 TargetLowering::AddrMode AM;
1294 AM.BaseOffs = BaseOffs;
1296 AM.HasBaseReg = HasBaseReg;
1297 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1299 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1304 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1305 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1306 struct UseMapDenseMapInfo {
1307 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1308 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1311 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1312 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1316 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1317 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1318 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1322 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1323 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1328 /// LSRInstance - This class holds state for the main loop strength reduction
1332 ScalarEvolution &SE;
1335 const TargetLowering *const TLI;
1339 /// IVIncInsertPos - This is the insert position that the current loop's
1340 /// induction variable increment should be placed. In simple loops, this is
1341 /// the latch block's terminator. But in more complicated cases, this is a
1342 /// position which will dominate all the in-loop post-increment users.
1343 Instruction *IVIncInsertPos;
1345 /// Factors - Interesting factors between use strides.
1346 SmallSetVector<int64_t, 8> Factors;
1348 /// Types - Interesting use types, to facilitate truncation reuse.
1349 SmallSetVector<Type *, 4> Types;
1351 /// Fixups - The list of operands which are to be replaced.
1352 SmallVector<LSRFixup, 16> Fixups;
1354 /// Uses - The list of interesting uses.
1355 SmallVector<LSRUse, 16> Uses;
1357 /// RegUses - Track which uses use which register candidates.
1358 RegUseTracker RegUses;
1360 void OptimizeShadowIV();
1361 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1362 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1363 void OptimizeLoopTermCond();
1365 void CollectInterestingTypesAndFactors();
1366 void CollectFixupsAndInitialFormulae();
1368 LSRFixup &getNewFixup() {
1369 Fixups.push_back(LSRFixup());
1370 return Fixups.back();
1373 // Support for sharing of LSRUses between LSRFixups.
1374 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1376 UseMapDenseMapInfo> UseMapTy;
1379 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1380 LSRUse::KindType Kind, Type *AccessTy);
1382 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1383 LSRUse::KindType Kind,
1386 void DeleteUse(LSRUse &LU, size_t LUIdx);
1388 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1391 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1392 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1393 void CountRegisters(const Formula &F, size_t LUIdx);
1394 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1396 void CollectLoopInvariantFixupsAndFormulae();
1398 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1399 unsigned Depth = 0);
1400 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1401 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1402 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1403 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1404 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1405 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1406 void GenerateCrossUseConstantOffsets();
1407 void GenerateAllReuseFormulae();
1409 void FilterOutUndesirableDedicatedRegisters();
1411 size_t EstimateSearchSpaceComplexity() const;
1412 void NarrowSearchSpaceByDetectingSupersets();
1413 void NarrowSearchSpaceByCollapsingUnrolledCode();
1414 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1415 void NarrowSearchSpaceByPickingWinnerRegs();
1416 void NarrowSearchSpaceUsingHeuristics();
1418 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1420 SmallVectorImpl<const Formula *> &Workspace,
1421 const Cost &CurCost,
1422 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1423 DenseSet<const SCEV *> &VisitedRegs) const;
1424 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1426 BasicBlock::iterator
1427 HoistInsertPosition(BasicBlock::iterator IP,
1428 const SmallVectorImpl<Instruction *> &Inputs) const;
1429 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1431 const LSRUse &LU) const;
1433 Value *Expand(const LSRFixup &LF,
1435 BasicBlock::iterator IP,
1436 SCEVExpander &Rewriter,
1437 SmallVectorImpl<WeakVH> &DeadInsts) const;
1438 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1440 SCEVExpander &Rewriter,
1441 SmallVectorImpl<WeakVH> &DeadInsts,
1443 void Rewrite(const LSRFixup &LF,
1445 SCEVExpander &Rewriter,
1446 SmallVectorImpl<WeakVH> &DeadInsts,
1448 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1451 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1453 bool getChanged() const { return Changed; }
1455 void print_factors_and_types(raw_ostream &OS) const;
1456 void print_fixups(raw_ostream &OS) const;
1457 void print_uses(raw_ostream &OS) const;
1458 void print(raw_ostream &OS) const;
1464 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1465 /// inside the loop then try to eliminate the cast operation.
1466 void LSRInstance::OptimizeShadowIV() {
1467 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1468 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1471 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1472 UI != E; /* empty */) {
1473 IVUsers::const_iterator CandidateUI = UI;
1475 Instruction *ShadowUse = CandidateUI->getUser();
1476 Type *DestTy = NULL;
1477 bool IsSigned = false;
1479 /* If shadow use is a int->float cast then insert a second IV
1480 to eliminate this cast.
1482 for (unsigned i = 0; i < n; ++i)
1488 for (unsigned i = 0; i < n; ++i, ++d)
1491 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1493 DestTy = UCast->getDestTy();
1495 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1497 DestTy = SCast->getDestTy();
1499 if (!DestTy) continue;
1502 // If target does not support DestTy natively then do not apply
1503 // this transformation.
1504 EVT DVT = TLI->getValueType(DestTy);
1505 if (!TLI->isTypeLegal(DVT)) continue;
1508 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1510 if (PH->getNumIncomingValues() != 2) continue;
1512 Type *SrcTy = PH->getType();
1513 int Mantissa = DestTy->getFPMantissaWidth();
1514 if (Mantissa == -1) continue;
1515 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1518 unsigned Entry, Latch;
1519 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1527 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1528 if (!Init) continue;
1529 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1530 (double)Init->getSExtValue() :
1531 (double)Init->getZExtValue());
1533 BinaryOperator *Incr =
1534 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1535 if (!Incr) continue;
1536 if (Incr->getOpcode() != Instruction::Add
1537 && Incr->getOpcode() != Instruction::Sub)
1540 /* Initialize new IV, double d = 0.0 in above example. */
1541 ConstantInt *C = NULL;
1542 if (Incr->getOperand(0) == PH)
1543 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1544 else if (Incr->getOperand(1) == PH)
1545 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1551 // Ignore negative constants, as the code below doesn't handle them
1552 // correctly. TODO: Remove this restriction.
1553 if (!C->getValue().isStrictlyPositive()) continue;
1555 /* Add new PHINode. */
1556 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1558 /* create new increment. '++d' in above example. */
1559 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1560 BinaryOperator *NewIncr =
1561 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1562 Instruction::FAdd : Instruction::FSub,
1563 NewPH, CFP, "IV.S.next.", Incr);
1565 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1566 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1568 /* Remove cast operation */
1569 ShadowUse->replaceAllUsesWith(NewPH);
1570 ShadowUse->eraseFromParent();
1576 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1577 /// set the IV user and stride information and return true, otherwise return
1579 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1580 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1581 if (UI->getUser() == Cond) {
1582 // NOTE: we could handle setcc instructions with multiple uses here, but
1583 // InstCombine does it as well for simple uses, it's not clear that it
1584 // occurs enough in real life to handle.
1591 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1592 /// a max computation.
1594 /// This is a narrow solution to a specific, but acute, problem. For loops
1600 /// } while (++i < n);
1602 /// the trip count isn't just 'n', because 'n' might not be positive. And
1603 /// unfortunately this can come up even for loops where the user didn't use
1604 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1605 /// will commonly be lowered like this:
1611 /// } while (++i < n);
1614 /// and then it's possible for subsequent optimization to obscure the if
1615 /// test in such a way that indvars can't find it.
1617 /// When indvars can't find the if test in loops like this, it creates a
1618 /// max expression, which allows it to give the loop a canonical
1619 /// induction variable:
1622 /// max = n < 1 ? 1 : n;
1625 /// } while (++i != max);
1627 /// Canonical induction variables are necessary because the loop passes
1628 /// are designed around them. The most obvious example of this is the
1629 /// LoopInfo analysis, which doesn't remember trip count values. It
1630 /// expects to be able to rediscover the trip count each time it is
1631 /// needed, and it does this using a simple analysis that only succeeds if
1632 /// the loop has a canonical induction variable.
1634 /// However, when it comes time to generate code, the maximum operation
1635 /// can be quite costly, especially if it's inside of an outer loop.
1637 /// This function solves this problem by detecting this type of loop and
1638 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1639 /// the instructions for the maximum computation.
1641 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1642 // Check that the loop matches the pattern we're looking for.
1643 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1644 Cond->getPredicate() != CmpInst::ICMP_NE)
1647 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1648 if (!Sel || !Sel->hasOneUse()) return Cond;
1650 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1651 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1653 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1655 // Add one to the backedge-taken count to get the trip count.
1656 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1657 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1659 // Check for a max calculation that matches the pattern. There's no check
1660 // for ICMP_ULE here because the comparison would be with zero, which
1661 // isn't interesting.
1662 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1663 const SCEVNAryExpr *Max = 0;
1664 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1665 Pred = ICmpInst::ICMP_SLE;
1667 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1668 Pred = ICmpInst::ICMP_SLT;
1670 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1671 Pred = ICmpInst::ICMP_ULT;
1678 // To handle a max with more than two operands, this optimization would
1679 // require additional checking and setup.
1680 if (Max->getNumOperands() != 2)
1683 const SCEV *MaxLHS = Max->getOperand(0);
1684 const SCEV *MaxRHS = Max->getOperand(1);
1686 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1687 // for a comparison with 1. For <= and >=, a comparison with zero.
1689 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1692 // Check the relevant induction variable for conformance to
1694 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1695 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1696 if (!AR || !AR->isAffine() ||
1697 AR->getStart() != One ||
1698 AR->getStepRecurrence(SE) != One)
1701 assert(AR->getLoop() == L &&
1702 "Loop condition operand is an addrec in a different loop!");
1704 // Check the right operand of the select, and remember it, as it will
1705 // be used in the new comparison instruction.
1707 if (ICmpInst::isTrueWhenEqual(Pred)) {
1708 // Look for n+1, and grab n.
1709 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1710 if (isa<ConstantInt>(BO->getOperand(1)) &&
1711 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1712 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1713 NewRHS = BO->getOperand(0);
1714 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1715 if (isa<ConstantInt>(BO->getOperand(1)) &&
1716 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1717 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1718 NewRHS = BO->getOperand(0);
1721 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1722 NewRHS = Sel->getOperand(1);
1723 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1724 NewRHS = Sel->getOperand(2);
1725 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1726 NewRHS = SU->getValue();
1728 // Max doesn't match expected pattern.
1731 // Determine the new comparison opcode. It may be signed or unsigned,
1732 // and the original comparison may be either equality or inequality.
1733 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1734 Pred = CmpInst::getInversePredicate(Pred);
1736 // Ok, everything looks ok to change the condition into an SLT or SGE and
1737 // delete the max calculation.
1739 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1741 // Delete the max calculation instructions.
1742 Cond->replaceAllUsesWith(NewCond);
1743 CondUse->setUser(NewCond);
1744 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1745 Cond->eraseFromParent();
1746 Sel->eraseFromParent();
1747 if (Cmp->use_empty())
1748 Cmp->eraseFromParent();
1752 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1753 /// postinc iv when possible.
1755 LSRInstance::OptimizeLoopTermCond() {
1756 SmallPtrSet<Instruction *, 4> PostIncs;
1758 BasicBlock *LatchBlock = L->getLoopLatch();
1759 SmallVector<BasicBlock*, 8> ExitingBlocks;
1760 L->getExitingBlocks(ExitingBlocks);
1762 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1763 BasicBlock *ExitingBlock = ExitingBlocks[i];
1765 // Get the terminating condition for the loop if possible. If we
1766 // can, we want to change it to use a post-incremented version of its
1767 // induction variable, to allow coalescing the live ranges for the IV into
1768 // one register value.
1770 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1773 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1774 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1777 // Search IVUsesByStride to find Cond's IVUse if there is one.
1778 IVStrideUse *CondUse = 0;
1779 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1780 if (!FindIVUserForCond(Cond, CondUse))
1783 // If the trip count is computed in terms of a max (due to ScalarEvolution
1784 // being unable to find a sufficient guard, for example), change the loop
1785 // comparison to use SLT or ULT instead of NE.
1786 // One consequence of doing this now is that it disrupts the count-down
1787 // optimization. That's not always a bad thing though, because in such
1788 // cases it may still be worthwhile to avoid a max.
1789 Cond = OptimizeMax(Cond, CondUse);
1791 // If this exiting block dominates the latch block, it may also use
1792 // the post-inc value if it won't be shared with other uses.
1793 // Check for dominance.
1794 if (!DT.dominates(ExitingBlock, LatchBlock))
1797 // Conservatively avoid trying to use the post-inc value in non-latch
1798 // exits if there may be pre-inc users in intervening blocks.
1799 if (LatchBlock != ExitingBlock)
1800 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1801 // Test if the use is reachable from the exiting block. This dominator
1802 // query is a conservative approximation of reachability.
1803 if (&*UI != CondUse &&
1804 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1805 // Conservatively assume there may be reuse if the quotient of their
1806 // strides could be a legal scale.
1807 const SCEV *A = IU.getStride(*CondUse, L);
1808 const SCEV *B = IU.getStride(*UI, L);
1809 if (!A || !B) continue;
1810 if (SE.getTypeSizeInBits(A->getType()) !=
1811 SE.getTypeSizeInBits(B->getType())) {
1812 if (SE.getTypeSizeInBits(A->getType()) >
1813 SE.getTypeSizeInBits(B->getType()))
1814 B = SE.getSignExtendExpr(B, A->getType());
1816 A = SE.getSignExtendExpr(A, B->getType());
1818 if (const SCEVConstant *D =
1819 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1820 const ConstantInt *C = D->getValue();
1821 // Stride of one or negative one can have reuse with non-addresses.
1822 if (C->isOne() || C->isAllOnesValue())
1823 goto decline_post_inc;
1824 // Avoid weird situations.
1825 if (C->getValue().getMinSignedBits() >= 64 ||
1826 C->getValue().isMinSignedValue())
1827 goto decline_post_inc;
1828 // Without TLI, assume that any stride might be valid, and so any
1829 // use might be shared.
1831 goto decline_post_inc;
1832 // Check for possible scaled-address reuse.
1833 Type *AccessTy = getAccessType(UI->getUser());
1834 TargetLowering::AddrMode AM;
1835 AM.Scale = C->getSExtValue();
1836 if (TLI->isLegalAddressingMode(AM, AccessTy))
1837 goto decline_post_inc;
1838 AM.Scale = -AM.Scale;
1839 if (TLI->isLegalAddressingMode(AM, AccessTy))
1840 goto decline_post_inc;
1844 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1847 // It's possible for the setcc instruction to be anywhere in the loop, and
1848 // possible for it to have multiple users. If it is not immediately before
1849 // the exiting block branch, move it.
1850 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1851 if (Cond->hasOneUse()) {
1852 Cond->moveBefore(TermBr);
1854 // Clone the terminating condition and insert into the loopend.
1855 ICmpInst *OldCond = Cond;
1856 Cond = cast<ICmpInst>(Cond->clone());
1857 Cond->setName(L->getHeader()->getName() + ".termcond");
1858 ExitingBlock->getInstList().insert(TermBr, Cond);
1860 // Clone the IVUse, as the old use still exists!
1861 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1862 TermBr->replaceUsesOfWith(OldCond, Cond);
1866 // If we get to here, we know that we can transform the setcc instruction to
1867 // use the post-incremented version of the IV, allowing us to coalesce the
1868 // live ranges for the IV correctly.
1869 CondUse->transformToPostInc(L);
1872 PostIncs.insert(Cond);
1876 // Determine an insertion point for the loop induction variable increment. It
1877 // must dominate all the post-inc comparisons we just set up, and it must
1878 // dominate the loop latch edge.
1879 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1880 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1881 E = PostIncs.end(); I != E; ++I) {
1883 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1885 if (BB == (*I)->getParent())
1886 IVIncInsertPos = *I;
1887 else if (BB != IVIncInsertPos->getParent())
1888 IVIncInsertPos = BB->getTerminator();
1892 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
1893 /// at the given offset and other details. If so, update the use and
1896 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1897 LSRUse::KindType Kind, Type *AccessTy) {
1898 int64_t NewMinOffset = LU.MinOffset;
1899 int64_t NewMaxOffset = LU.MaxOffset;
1900 Type *NewAccessTy = AccessTy;
1902 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1903 // something conservative, however this can pessimize in the case that one of
1904 // the uses will have all its uses outside the loop, for example.
1905 if (LU.Kind != Kind)
1907 // Conservatively assume HasBaseReg is true for now.
1908 if (NewOffset < LU.MinOffset) {
1909 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1910 Kind, AccessTy, TLI))
1912 NewMinOffset = NewOffset;
1913 } else if (NewOffset > LU.MaxOffset) {
1914 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1915 Kind, AccessTy, TLI))
1917 NewMaxOffset = NewOffset;
1919 // Check for a mismatched access type, and fall back conservatively as needed.
1920 // TODO: Be less conservative when the type is similar and can use the same
1921 // addressing modes.
1922 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1923 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1926 LU.MinOffset = NewMinOffset;
1927 LU.MaxOffset = NewMaxOffset;
1928 LU.AccessTy = NewAccessTy;
1929 if (NewOffset != LU.Offsets.back())
1930 LU.Offsets.push_back(NewOffset);
1934 /// getUse - Return an LSRUse index and an offset value for a fixup which
1935 /// needs the given expression, with the given kind and optional access type.
1936 /// Either reuse an existing use or create a new one, as needed.
1937 std::pair<size_t, int64_t>
1938 LSRInstance::getUse(const SCEV *&Expr,
1939 LSRUse::KindType Kind, Type *AccessTy) {
1940 const SCEV *Copy = Expr;
1941 int64_t Offset = ExtractImmediate(Expr, SE);
1943 // Basic uses can't accept any offset, for example.
1944 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1949 std::pair<UseMapTy::iterator, bool> P =
1950 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
1952 // A use already existed with this base.
1953 size_t LUIdx = P.first->second;
1954 LSRUse &LU = Uses[LUIdx];
1955 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
1957 return std::make_pair(LUIdx, Offset);
1960 // Create a new use.
1961 size_t LUIdx = Uses.size();
1962 P.first->second = LUIdx;
1963 Uses.push_back(LSRUse(Kind, AccessTy));
1964 LSRUse &LU = Uses[LUIdx];
1966 // We don't need to track redundant offsets, but we don't need to go out
1967 // of our way here to avoid them.
1968 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1969 LU.Offsets.push_back(Offset);
1971 LU.MinOffset = Offset;
1972 LU.MaxOffset = Offset;
1973 return std::make_pair(LUIdx, Offset);
1976 /// DeleteUse - Delete the given use from the Uses list.
1977 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
1978 if (&LU != &Uses.back())
1979 std::swap(LU, Uses.back());
1983 RegUses.SwapAndDropUse(LUIdx, Uses.size());
1986 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
1987 /// a formula that has the same registers as the given formula.
1989 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1990 const LSRUse &OrigLU) {
1991 // Search all uses for the formula. This could be more clever.
1992 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
1993 LSRUse &LU = Uses[LUIdx];
1994 // Check whether this use is close enough to OrigLU, to see whether it's
1995 // worthwhile looking through its formulae.
1996 // Ignore ICmpZero uses because they may contain formulae generated by
1997 // GenerateICmpZeroScales, in which case adding fixup offsets may
1999 if (&LU != &OrigLU &&
2000 LU.Kind != LSRUse::ICmpZero &&
2001 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2002 LU.WidestFixupType == OrigLU.WidestFixupType &&
2003 LU.HasFormulaWithSameRegs(OrigF)) {
2004 // Scan through this use's formulae.
2005 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2006 E = LU.Formulae.end(); I != E; ++I) {
2007 const Formula &F = *I;
2008 // Check to see if this formula has the same registers and symbols
2010 if (F.BaseRegs == OrigF.BaseRegs &&
2011 F.ScaledReg == OrigF.ScaledReg &&
2012 F.AM.BaseGV == OrigF.AM.BaseGV &&
2013 F.AM.Scale == OrigF.AM.Scale &&
2014 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2015 if (F.AM.BaseOffs == 0)
2017 // This is the formula where all the registers and symbols matched;
2018 // there aren't going to be any others. Since we declined it, we
2019 // can skip the rest of the formulae and procede to the next LSRUse.
2026 // Nothing looked good.
2030 void LSRInstance::CollectInterestingTypesAndFactors() {
2031 SmallSetVector<const SCEV *, 4> Strides;
2033 // Collect interesting types and strides.
2034 SmallVector<const SCEV *, 4> Worklist;
2035 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2036 const SCEV *Expr = IU.getExpr(*UI);
2038 // Collect interesting types.
2039 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2041 // Add strides for mentioned loops.
2042 Worklist.push_back(Expr);
2044 const SCEV *S = Worklist.pop_back_val();
2045 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2046 Strides.insert(AR->getStepRecurrence(SE));
2047 Worklist.push_back(AR->getStart());
2048 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2049 Worklist.append(Add->op_begin(), Add->op_end());
2051 } while (!Worklist.empty());
2054 // Compute interesting factors from the set of interesting strides.
2055 for (SmallSetVector<const SCEV *, 4>::const_iterator
2056 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2057 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2058 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2059 const SCEV *OldStride = *I;
2060 const SCEV *NewStride = *NewStrideIter;
2062 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2063 SE.getTypeSizeInBits(NewStride->getType())) {
2064 if (SE.getTypeSizeInBits(OldStride->getType()) >
2065 SE.getTypeSizeInBits(NewStride->getType()))
2066 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2068 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2070 if (const SCEVConstant *Factor =
2071 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2073 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2074 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2075 } else if (const SCEVConstant *Factor =
2076 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2079 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2080 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2084 // If all uses use the same type, don't bother looking for truncation-based
2086 if (Types.size() == 1)
2089 DEBUG(print_factors_and_types(dbgs()));
2092 void LSRInstance::CollectFixupsAndInitialFormulae() {
2093 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2095 LSRFixup &LF = getNewFixup();
2096 LF.UserInst = UI->getUser();
2097 LF.OperandValToReplace = UI->getOperandValToReplace();
2098 LF.PostIncLoops = UI->getPostIncLoops();
2100 LSRUse::KindType Kind = LSRUse::Basic;
2102 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2103 Kind = LSRUse::Address;
2104 AccessTy = getAccessType(LF.UserInst);
2107 const SCEV *S = IU.getExpr(*UI);
2109 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2110 // (N - i == 0), and this allows (N - i) to be the expression that we work
2111 // with rather than just N or i, so we can consider the register
2112 // requirements for both N and i at the same time. Limiting this code to
2113 // equality icmps is not a problem because all interesting loops use
2114 // equality icmps, thanks to IndVarSimplify.
2115 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2116 if (CI->isEquality()) {
2117 // Swap the operands if needed to put the OperandValToReplace on the
2118 // left, for consistency.
2119 Value *NV = CI->getOperand(1);
2120 if (NV == LF.OperandValToReplace) {
2121 CI->setOperand(1, CI->getOperand(0));
2122 CI->setOperand(0, NV);
2123 NV = CI->getOperand(1);
2127 // x == y --> x - y == 0
2128 const SCEV *N = SE.getSCEV(NV);
2129 if (SE.isLoopInvariant(N, L)) {
2130 // S is normalized, so normalize N before folding it into S
2131 // to keep the result normalized.
2132 N = TransformForPostIncUse(Normalize, N, CI, 0,
2133 LF.PostIncLoops, SE, DT);
2134 Kind = LSRUse::ICmpZero;
2135 S = SE.getMinusSCEV(N, S);
2138 // -1 and the negations of all interesting strides (except the negation
2139 // of -1) are now also interesting.
2140 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2141 if (Factors[i] != -1)
2142 Factors.insert(-(uint64_t)Factors[i]);
2146 // Set up the initial formula for this use.
2147 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2149 LF.Offset = P.second;
2150 LSRUse &LU = Uses[LF.LUIdx];
2151 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2152 if (!LU.WidestFixupType ||
2153 SE.getTypeSizeInBits(LU.WidestFixupType) <
2154 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2155 LU.WidestFixupType = LF.OperandValToReplace->getType();
2157 // If this is the first use of this LSRUse, give it a formula.
2158 if (LU.Formulae.empty()) {
2159 InsertInitialFormula(S, LU, LF.LUIdx);
2160 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2164 DEBUG(print_fixups(dbgs()));
2167 /// InsertInitialFormula - Insert a formula for the given expression into
2168 /// the given use, separating out loop-variant portions from loop-invariant
2169 /// and loop-computable portions.
2171 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2173 F.InitialMatch(S, L, SE);
2174 bool Inserted = InsertFormula(LU, LUIdx, F);
2175 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2178 /// InsertSupplementalFormula - Insert a simple single-register formula for
2179 /// the given expression into the given use.
2181 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2182 LSRUse &LU, size_t LUIdx) {
2184 F.BaseRegs.push_back(S);
2185 F.AM.HasBaseReg = true;
2186 bool Inserted = InsertFormula(LU, LUIdx, F);
2187 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2190 /// CountRegisters - Note which registers are used by the given formula,
2191 /// updating RegUses.
2192 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2194 RegUses.CountRegister(F.ScaledReg, LUIdx);
2195 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2196 E = F.BaseRegs.end(); I != E; ++I)
2197 RegUses.CountRegister(*I, LUIdx);
2200 /// InsertFormula - If the given formula has not yet been inserted, add it to
2201 /// the list, and return true. Return false otherwise.
2202 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2203 if (!LU.InsertFormula(F))
2206 CountRegisters(F, LUIdx);
2210 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2211 /// loop-invariant values which we're tracking. These other uses will pin these
2212 /// values in registers, making them less profitable for elimination.
2213 /// TODO: This currently misses non-constant addrec step registers.
2214 /// TODO: Should this give more weight to users inside the loop?
2216 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2217 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2218 SmallPtrSet<const SCEV *, 8> Inserted;
2220 while (!Worklist.empty()) {
2221 const SCEV *S = Worklist.pop_back_val();
2223 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2224 Worklist.append(N->op_begin(), N->op_end());
2225 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2226 Worklist.push_back(C->getOperand());
2227 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2228 Worklist.push_back(D->getLHS());
2229 Worklist.push_back(D->getRHS());
2230 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2231 if (!Inserted.insert(U)) continue;
2232 const Value *V = U->getValue();
2233 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2234 // Look for instructions defined outside the loop.
2235 if (L->contains(Inst)) continue;
2236 } else if (isa<UndefValue>(V))
2237 // Undef doesn't have a live range, so it doesn't matter.
2239 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2241 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2242 // Ignore non-instructions.
2245 // Ignore instructions in other functions (as can happen with
2247 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2249 // Ignore instructions not dominated by the loop.
2250 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2251 UserInst->getParent() :
2252 cast<PHINode>(UserInst)->getIncomingBlock(
2253 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2254 if (!DT.dominates(L->getHeader(), UseBB))
2256 // Ignore uses which are part of other SCEV expressions, to avoid
2257 // analyzing them multiple times.
2258 if (SE.isSCEVable(UserInst->getType())) {
2259 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2260 // If the user is a no-op, look through to its uses.
2261 if (!isa<SCEVUnknown>(UserS))
2265 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2269 // Ignore icmp instructions which are already being analyzed.
2270 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2271 unsigned OtherIdx = !UI.getOperandNo();
2272 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2273 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2277 LSRFixup &LF = getNewFixup();
2278 LF.UserInst = const_cast<Instruction *>(UserInst);
2279 LF.OperandValToReplace = UI.getUse();
2280 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2282 LF.Offset = P.second;
2283 LSRUse &LU = Uses[LF.LUIdx];
2284 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2285 if (!LU.WidestFixupType ||
2286 SE.getTypeSizeInBits(LU.WidestFixupType) <
2287 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2288 LU.WidestFixupType = LF.OperandValToReplace->getType();
2289 InsertSupplementalFormula(U, LU, LF.LUIdx);
2290 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2297 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2298 /// separate registers. If C is non-null, multiply each subexpression by C.
2299 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2300 SmallVectorImpl<const SCEV *> &Ops,
2302 ScalarEvolution &SE) {
2303 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2304 // Break out add operands.
2305 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2307 CollectSubexprs(*I, C, Ops, L, SE);
2309 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2310 // Split a non-zero base out of an addrec.
2311 if (!AR->getStart()->isZero()) {
2312 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2313 AR->getStepRecurrence(SE),
2315 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2318 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2321 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2322 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2323 if (Mul->getNumOperands() == 2)
2324 if (const SCEVConstant *Op0 =
2325 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2326 CollectSubexprs(Mul->getOperand(1),
2327 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2333 // Otherwise use the value itself, optionally with a scale applied.
2334 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2337 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2339 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2342 // Arbitrarily cap recursion to protect compile time.
2343 if (Depth >= 3) return;
2345 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2346 const SCEV *BaseReg = Base.BaseRegs[i];
2348 SmallVector<const SCEV *, 8> AddOps;
2349 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2351 if (AddOps.size() == 1) continue;
2353 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2354 JE = AddOps.end(); J != JE; ++J) {
2356 // Loop-variant "unknown" values are uninteresting; we won't be able to
2357 // do anything meaningful with them.
2358 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2361 // Don't pull a constant into a register if the constant could be folded
2362 // into an immediate field.
2363 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2364 Base.getNumRegs() > 1,
2365 LU.Kind, LU.AccessTy, TLI, SE))
2368 // Collect all operands except *J.
2369 SmallVector<const SCEV *, 8> InnerAddOps
2370 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2372 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2374 // Don't leave just a constant behind in a register if the constant could
2375 // be folded into an immediate field.
2376 if (InnerAddOps.size() == 1 &&
2377 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2378 Base.getNumRegs() > 1,
2379 LU.Kind, LU.AccessTy, TLI, SE))
2382 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2383 if (InnerSum->isZero())
2387 // Add the remaining pieces of the add back into the new formula.
2388 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
2389 if (TLI && InnerSumSC &&
2390 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
2391 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2392 InnerSumSC->getValue()->getZExtValue())) {
2393 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2394 InnerSumSC->getValue()->getZExtValue();
2395 F.BaseRegs.erase(F.BaseRegs.begin() + i);
2397 F.BaseRegs[i] = InnerSum;
2399 // Add J as its own register, or an unfolded immediate.
2400 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
2401 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
2402 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2403 SC->getValue()->getZExtValue()))
2404 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2405 SC->getValue()->getZExtValue();
2407 F.BaseRegs.push_back(*J);
2409 if (InsertFormula(LU, LUIdx, F))
2410 // If that formula hadn't been seen before, recurse to find more like
2412 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2417 /// GenerateCombinations - Generate a formula consisting of all of the
2418 /// loop-dominating registers added into a single register.
2419 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2421 // This method is only interesting on a plurality of registers.
2422 if (Base.BaseRegs.size() <= 1) return;
2426 SmallVector<const SCEV *, 4> Ops;
2427 for (SmallVectorImpl<const SCEV *>::const_iterator
2428 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2429 const SCEV *BaseReg = *I;
2430 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2431 !SE.hasComputableLoopEvolution(BaseReg, L))
2432 Ops.push_back(BaseReg);
2434 F.BaseRegs.push_back(BaseReg);
2436 if (Ops.size() > 1) {
2437 const SCEV *Sum = SE.getAddExpr(Ops);
2438 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2439 // opportunity to fold something. For now, just ignore such cases
2440 // rather than proceed with zero in a register.
2441 if (!Sum->isZero()) {
2442 F.BaseRegs.push_back(Sum);
2443 (void)InsertFormula(LU, LUIdx, F);
2448 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2449 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2451 // We can't add a symbolic offset if the address already contains one.
2452 if (Base.AM.BaseGV) return;
2454 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2455 const SCEV *G = Base.BaseRegs[i];
2456 GlobalValue *GV = ExtractSymbol(G, SE);
2457 if (G->isZero() || !GV)
2461 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2462 LU.Kind, LU.AccessTy, TLI))
2465 (void)InsertFormula(LU, LUIdx, F);
2469 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2470 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2472 // TODO: For now, just add the min and max offset, because it usually isn't
2473 // worthwhile looking at everything inbetween.
2474 SmallVector<int64_t, 2> Worklist;
2475 Worklist.push_back(LU.MinOffset);
2476 if (LU.MaxOffset != LU.MinOffset)
2477 Worklist.push_back(LU.MaxOffset);
2479 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2480 const SCEV *G = Base.BaseRegs[i];
2482 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2483 E = Worklist.end(); I != E; ++I) {
2485 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2486 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2487 LU.Kind, LU.AccessTy, TLI)) {
2488 // Add the offset to the base register.
2489 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2490 // If it cancelled out, drop the base register, otherwise update it.
2491 if (NewG->isZero()) {
2492 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2493 F.BaseRegs.pop_back();
2495 F.BaseRegs[i] = NewG;
2497 (void)InsertFormula(LU, LUIdx, F);
2501 int64_t Imm = ExtractImmediate(G, SE);
2502 if (G->isZero() || Imm == 0)
2505 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2506 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2507 LU.Kind, LU.AccessTy, TLI))
2510 (void)InsertFormula(LU, LUIdx, F);
2514 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2515 /// the comparison. For example, x == y -> x*c == y*c.
2516 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2518 if (LU.Kind != LSRUse::ICmpZero) return;
2520 // Determine the integer type for the base formula.
2521 Type *IntTy = Base.getType();
2523 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2525 // Don't do this if there is more than one offset.
2526 if (LU.MinOffset != LU.MaxOffset) return;
2528 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2530 // Check each interesting stride.
2531 for (SmallSetVector<int64_t, 8>::const_iterator
2532 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2533 int64_t Factor = *I;
2535 // Check that the multiplication doesn't overflow.
2536 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
2538 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2539 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
2542 // Check that multiplying with the use offset doesn't overflow.
2543 int64_t Offset = LU.MinOffset;
2544 if (Offset == INT64_MIN && Factor == -1)
2546 Offset = (uint64_t)Offset * Factor;
2547 if (Offset / Factor != LU.MinOffset)
2551 F.AM.BaseOffs = NewBaseOffs;
2553 // Check that this scale is legal.
2554 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2557 // Compensate for the use having MinOffset built into it.
2558 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2560 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2562 // Check that multiplying with each base register doesn't overflow.
2563 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2564 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2565 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2569 // Check that multiplying with the scaled register doesn't overflow.
2571 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2572 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2576 // Check that multiplying with the unfolded offset doesn't overflow.
2577 if (F.UnfoldedOffset != 0) {
2578 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
2580 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
2581 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
2585 // If we make it here and it's legal, add it.
2586 (void)InsertFormula(LU, LUIdx, F);
2591 /// GenerateScales - Generate stride factor reuse formulae by making use of
2592 /// scaled-offset address modes, for example.
2593 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
2594 // Determine the integer type for the base formula.
2595 Type *IntTy = Base.getType();
2598 // If this Formula already has a scaled register, we can't add another one.
2599 if (Base.AM.Scale != 0) return;
2601 // Check each interesting stride.
2602 for (SmallSetVector<int64_t, 8>::const_iterator
2603 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2604 int64_t Factor = *I;
2606 Base.AM.Scale = Factor;
2607 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2608 // Check whether this scale is going to be legal.
2609 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2610 LU.Kind, LU.AccessTy, TLI)) {
2611 // As a special-case, handle special out-of-loop Basic users specially.
2612 // TODO: Reconsider this special case.
2613 if (LU.Kind == LSRUse::Basic &&
2614 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2615 LSRUse::Special, LU.AccessTy, TLI) &&
2616 LU.AllFixupsOutsideLoop)
2617 LU.Kind = LSRUse::Special;
2621 // For an ICmpZero, negating a solitary base register won't lead to
2623 if (LU.Kind == LSRUse::ICmpZero &&
2624 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2626 // For each addrec base reg, apply the scale, if possible.
2627 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2628 if (const SCEVAddRecExpr *AR =
2629 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2630 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
2631 if (FactorS->isZero())
2633 // Divide out the factor, ignoring high bits, since we'll be
2634 // scaling the value back up in the end.
2635 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2636 // TODO: This could be optimized to avoid all the copying.
2638 F.ScaledReg = Quotient;
2639 F.DeleteBaseReg(F.BaseRegs[i]);
2640 (void)InsertFormula(LU, LUIdx, F);
2646 /// GenerateTruncates - Generate reuse formulae from different IV types.
2647 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
2648 // This requires TargetLowering to tell us which truncates are free.
2651 // Don't bother truncating symbolic values.
2652 if (Base.AM.BaseGV) return;
2654 // Determine the integer type for the base formula.
2655 Type *DstTy = Base.getType();
2657 DstTy = SE.getEffectiveSCEVType(DstTy);
2659 for (SmallSetVector<Type *, 4>::const_iterator
2660 I = Types.begin(), E = Types.end(); I != E; ++I) {
2662 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2665 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2666 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2667 JE = F.BaseRegs.end(); J != JE; ++J)
2668 *J = SE.getAnyExtendExpr(*J, SrcTy);
2670 // TODO: This assumes we've done basic processing on all uses and
2671 // have an idea what the register usage is.
2672 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2675 (void)InsertFormula(LU, LUIdx, F);
2682 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2683 /// defer modifications so that the search phase doesn't have to worry about
2684 /// the data structures moving underneath it.
2688 const SCEV *OrigReg;
2690 WorkItem(size_t LI, int64_t I, const SCEV *R)
2691 : LUIdx(LI), Imm(I), OrigReg(R) {}
2693 void print(raw_ostream &OS) const;
2699 void WorkItem::print(raw_ostream &OS) const {
2700 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2701 << " , add offset " << Imm;
2704 void WorkItem::dump() const {
2705 print(errs()); errs() << '\n';
2708 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2709 /// distance apart and try to form reuse opportunities between them.
2710 void LSRInstance::GenerateCrossUseConstantOffsets() {
2711 // Group the registers by their value without any added constant offset.
2712 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2713 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2715 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2716 SmallVector<const SCEV *, 8> Sequence;
2717 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2719 const SCEV *Reg = *I;
2720 int64_t Imm = ExtractImmediate(Reg, SE);
2721 std::pair<RegMapTy::iterator, bool> Pair =
2722 Map.insert(std::make_pair(Reg, ImmMapTy()));
2724 Sequence.push_back(Reg);
2725 Pair.first->second.insert(std::make_pair(Imm, *I));
2726 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2729 // Now examine each set of registers with the same base value. Build up
2730 // a list of work to do and do the work in a separate step so that we're
2731 // not adding formulae and register counts while we're searching.
2732 SmallVector<WorkItem, 32> WorkItems;
2733 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2734 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2735 E = Sequence.end(); I != E; ++I) {
2736 const SCEV *Reg = *I;
2737 const ImmMapTy &Imms = Map.find(Reg)->second;
2739 // It's not worthwhile looking for reuse if there's only one offset.
2740 if (Imms.size() == 1)
2743 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2744 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2746 dbgs() << ' ' << J->first;
2749 // Examine each offset.
2750 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2752 const SCEV *OrigReg = J->second;
2754 int64_t JImm = J->first;
2755 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2757 if (!isa<SCEVConstant>(OrigReg) &&
2758 UsedByIndicesMap[Reg].count() == 1) {
2759 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2763 // Conservatively examine offsets between this orig reg a few selected
2765 ImmMapTy::const_iterator OtherImms[] = {
2766 Imms.begin(), prior(Imms.end()),
2767 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2769 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2770 ImmMapTy::const_iterator M = OtherImms[i];
2771 if (M == J || M == JE) continue;
2773 // Compute the difference between the two.
2774 int64_t Imm = (uint64_t)JImm - M->first;
2775 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2776 LUIdx = UsedByIndices.find_next(LUIdx))
2777 // Make a memo of this use, offset, and register tuple.
2778 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2779 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2786 UsedByIndicesMap.clear();
2787 UniqueItems.clear();
2789 // Now iterate through the worklist and add new formulae.
2790 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2791 E = WorkItems.end(); I != E; ++I) {
2792 const WorkItem &WI = *I;
2793 size_t LUIdx = WI.LUIdx;
2794 LSRUse &LU = Uses[LUIdx];
2795 int64_t Imm = WI.Imm;
2796 const SCEV *OrigReg = WI.OrigReg;
2798 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2799 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2800 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2802 // TODO: Use a more targeted data structure.
2803 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2804 const Formula &F = LU.Formulae[L];
2805 // Use the immediate in the scaled register.
2806 if (F.ScaledReg == OrigReg) {
2807 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2808 Imm * (uint64_t)F.AM.Scale;
2809 // Don't create 50 + reg(-50).
2810 if (F.referencesReg(SE.getSCEV(
2811 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2814 NewF.AM.BaseOffs = Offs;
2815 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2816 LU.Kind, LU.AccessTy, TLI))
2818 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2820 // If the new scale is a constant in a register, and adding the constant
2821 // value to the immediate would produce a value closer to zero than the
2822 // immediate itself, then the formula isn't worthwhile.
2823 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2824 if (C->getValue()->isNegative() !=
2825 (NewF.AM.BaseOffs < 0) &&
2826 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2827 .ule(abs64(NewF.AM.BaseOffs)))
2831 (void)InsertFormula(LU, LUIdx, NewF);
2833 // Use the immediate in a base register.
2834 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2835 const SCEV *BaseReg = F.BaseRegs[N];
2836 if (BaseReg != OrigReg)
2839 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2840 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2841 LU.Kind, LU.AccessTy, TLI)) {
2843 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
2846 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
2848 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2850 // If the new formula has a constant in a register, and adding the
2851 // constant value to the immediate would produce a value closer to
2852 // zero than the immediate itself, then the formula isn't worthwhile.
2853 for (SmallVectorImpl<const SCEV *>::const_iterator
2854 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2856 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2857 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
2858 abs64(NewF.AM.BaseOffs)) &&
2859 (C->getValue()->getValue() +
2860 NewF.AM.BaseOffs).countTrailingZeros() >=
2861 CountTrailingZeros_64(NewF.AM.BaseOffs))
2865 (void)InsertFormula(LU, LUIdx, NewF);
2874 /// GenerateAllReuseFormulae - Generate formulae for each use.
2876 LSRInstance::GenerateAllReuseFormulae() {
2877 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2878 // queries are more precise.
2879 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2880 LSRUse &LU = Uses[LUIdx];
2881 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2882 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2883 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2884 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2886 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2887 LSRUse &LU = Uses[LUIdx];
2888 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2889 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2890 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2891 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2892 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2893 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2894 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2895 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2897 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2898 LSRUse &LU = Uses[LUIdx];
2899 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2900 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2903 GenerateCrossUseConstantOffsets();
2905 DEBUG(dbgs() << "\n"
2906 "After generating reuse formulae:\n";
2907 print_uses(dbgs()));
2910 /// If there are multiple formulae with the same set of registers used
2911 /// by other uses, pick the best one and delete the others.
2912 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2913 DenseSet<const SCEV *> VisitedRegs;
2914 SmallPtrSet<const SCEV *, 16> Regs;
2916 bool ChangedFormulae = false;
2919 // Collect the best formula for each unique set of shared registers. This
2920 // is reset for each use.
2921 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2923 BestFormulaeTy BestFormulae;
2925 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2926 LSRUse &LU = Uses[LUIdx];
2927 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
2930 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2931 FIdx != NumForms; ++FIdx) {
2932 Formula &F = LU.Formulae[FIdx];
2934 SmallVector<const SCEV *, 2> Key;
2935 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2936 JE = F.BaseRegs.end(); J != JE; ++J) {
2937 const SCEV *Reg = *J;
2938 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2942 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2943 Key.push_back(F.ScaledReg);
2944 // Unstable sort by host order ok, because this is only used for
2946 std::sort(Key.begin(), Key.end());
2948 std::pair<BestFormulaeTy::const_iterator, bool> P =
2949 BestFormulae.insert(std::make_pair(Key, FIdx));
2951 Formula &Best = LU.Formulae[P.first->second];
2954 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2957 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
2959 if (CostF < CostBest)
2961 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
2963 " in favor of formula "; Best.print(dbgs());
2966 ChangedFormulae = true;
2968 LU.DeleteFormula(F);
2976 // Now that we've filtered out some formulae, recompute the Regs set.
2978 LU.RecomputeRegs(LUIdx, RegUses);
2980 // Reset this to prepare for the next use.
2981 BestFormulae.clear();
2984 DEBUG(if (ChangedFormulae) {
2986 "After filtering out undesirable candidates:\n";
2991 // This is a rough guess that seems to work fairly well.
2992 static const size_t ComplexityLimit = UINT16_MAX;
2994 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
2995 /// solutions the solver might have to consider. It almost never considers
2996 /// this many solutions because it prune the search space, but the pruning
2997 /// isn't always sufficient.
2998 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3000 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3001 E = Uses.end(); I != E; ++I) {
3002 size_t FSize = I->Formulae.size();
3003 if (FSize >= ComplexityLimit) {
3004 Power = ComplexityLimit;
3008 if (Power >= ComplexityLimit)
3014 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3015 /// of the registers of another formula, it won't help reduce register
3016 /// pressure (though it may not necessarily hurt register pressure); remove
3017 /// it to simplify the system.
3018 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3019 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3020 DEBUG(dbgs() << "The search space is too complex.\n");
3022 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3023 "which use a superset of registers used by other "
3026 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3027 LSRUse &LU = Uses[LUIdx];
3029 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3030 Formula &F = LU.Formulae[i];
3031 // Look for a formula with a constant or GV in a register. If the use
3032 // also has a formula with that same value in an immediate field,
3033 // delete the one that uses a register.
3034 for (SmallVectorImpl<const SCEV *>::const_iterator
3035 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3036 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3038 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3039 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3040 (I - F.BaseRegs.begin()));
3041 if (LU.HasFormulaWithSameRegs(NewF)) {
3042 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3043 LU.DeleteFormula(F);
3049 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3050 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3053 NewF.AM.BaseGV = GV;
3054 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3055 (I - F.BaseRegs.begin()));
3056 if (LU.HasFormulaWithSameRegs(NewF)) {
3057 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3059 LU.DeleteFormula(F);
3070 LU.RecomputeRegs(LUIdx, RegUses);
3073 DEBUG(dbgs() << "After pre-selection:\n";
3074 print_uses(dbgs()));
3078 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3079 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3081 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3082 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3083 DEBUG(dbgs() << "The search space is too complex.\n");
3085 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3086 "separated by a constant offset will use the same "
3089 // This is especially useful for unrolled loops.
3091 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3092 LSRUse &LU = Uses[LUIdx];
3093 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3094 E = LU.Formulae.end(); I != E; ++I) {
3095 const Formula &F = *I;
3096 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3097 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3098 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3099 /*HasBaseReg=*/false,
3100 LU.Kind, LU.AccessTy)) {
3101 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3104 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3106 // Update the relocs to reference the new use.
3107 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3108 E = Fixups.end(); I != E; ++I) {
3109 LSRFixup &Fixup = *I;
3110 if (Fixup.LUIdx == LUIdx) {
3111 Fixup.LUIdx = LUThatHas - &Uses.front();
3112 Fixup.Offset += F.AM.BaseOffs;
3113 // Add the new offset to LUThatHas' offset list.
3114 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3115 LUThatHas->Offsets.push_back(Fixup.Offset);
3116 if (Fixup.Offset > LUThatHas->MaxOffset)
3117 LUThatHas->MaxOffset = Fixup.Offset;
3118 if (Fixup.Offset < LUThatHas->MinOffset)
3119 LUThatHas->MinOffset = Fixup.Offset;
3121 DEBUG(dbgs() << "New fixup has offset "
3122 << Fixup.Offset << '\n');
3124 if (Fixup.LUIdx == NumUses-1)
3125 Fixup.LUIdx = LUIdx;
3128 // Delete formulae from the new use which are no longer legal.
3130 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3131 Formula &F = LUThatHas->Formulae[i];
3132 if (!isLegalUse(F.AM,
3133 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3134 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3135 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3137 LUThatHas->DeleteFormula(F);
3144 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3146 // Delete the old use.
3147 DeleteUse(LU, LUIdx);
3157 DEBUG(dbgs() << "After pre-selection:\n";
3158 print_uses(dbgs()));
3162 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3163 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3164 /// we've done more filtering, as it may be able to find more formulae to
3166 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3167 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3168 DEBUG(dbgs() << "The search space is too complex.\n");
3170 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3171 "undesirable dedicated registers.\n");
3173 FilterOutUndesirableDedicatedRegisters();
3175 DEBUG(dbgs() << "After pre-selection:\n";
3176 print_uses(dbgs()));
3180 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3181 /// to be profitable, and then in any use which has any reference to that
3182 /// register, delete all formulae which do not reference that register.
3183 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3184 // With all other options exhausted, loop until the system is simple
3185 // enough to handle.
3186 SmallPtrSet<const SCEV *, 4> Taken;
3187 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3188 // Ok, we have too many of formulae on our hands to conveniently handle.
3189 // Use a rough heuristic to thin out the list.
3190 DEBUG(dbgs() << "The search space is too complex.\n");
3192 // Pick the register which is used by the most LSRUses, which is likely
3193 // to be a good reuse register candidate.
3194 const SCEV *Best = 0;
3195 unsigned BestNum = 0;
3196 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3198 const SCEV *Reg = *I;
3199 if (Taken.count(Reg))
3204 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3205 if (Count > BestNum) {
3212 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3213 << " will yield profitable reuse.\n");
3216 // In any use with formulae which references this register, delete formulae
3217 // which don't reference it.
3218 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3219 LSRUse &LU = Uses[LUIdx];
3220 if (!LU.Regs.count(Best)) continue;
3223 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3224 Formula &F = LU.Formulae[i];
3225 if (!F.referencesReg(Best)) {
3226 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3227 LU.DeleteFormula(F);
3231 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3237 LU.RecomputeRegs(LUIdx, RegUses);
3240 DEBUG(dbgs() << "After pre-selection:\n";
3241 print_uses(dbgs()));
3245 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3246 /// formulae to choose from, use some rough heuristics to prune down the number
3247 /// of formulae. This keeps the main solver from taking an extraordinary amount
3248 /// of time in some worst-case scenarios.
3249 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3250 NarrowSearchSpaceByDetectingSupersets();
3251 NarrowSearchSpaceByCollapsingUnrolledCode();
3252 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3253 NarrowSearchSpaceByPickingWinnerRegs();
3256 /// SolveRecurse - This is the recursive solver.
3257 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3259 SmallVectorImpl<const Formula *> &Workspace,
3260 const Cost &CurCost,
3261 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3262 DenseSet<const SCEV *> &VisitedRegs) const {
3265 // - use more aggressive filtering
3266 // - sort the formula so that the most profitable solutions are found first
3267 // - sort the uses too
3269 // - don't compute a cost, and then compare. compare while computing a cost
3271 // - track register sets with SmallBitVector
3273 const LSRUse &LU = Uses[Workspace.size()];
3275 // If this use references any register that's already a part of the
3276 // in-progress solution, consider it a requirement that a formula must
3277 // reference that register in order to be considered. This prunes out
3278 // unprofitable searching.
3279 SmallSetVector<const SCEV *, 4> ReqRegs;
3280 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3281 E = CurRegs.end(); I != E; ++I)
3282 if (LU.Regs.count(*I))
3285 bool AnySatisfiedReqRegs = false;
3286 SmallPtrSet<const SCEV *, 16> NewRegs;
3289 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3290 E = LU.Formulae.end(); I != E; ++I) {
3291 const Formula &F = *I;
3293 // Ignore formulae which do not use any of the required registers.
3294 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3295 JE = ReqRegs.end(); J != JE; ++J) {
3296 const SCEV *Reg = *J;
3297 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3298 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3302 AnySatisfiedReqRegs = true;
3304 // Evaluate the cost of the current formula. If it's already worse than
3305 // the current best, prune the search at that point.
3308 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3309 if (NewCost < SolutionCost) {
3310 Workspace.push_back(&F);
3311 if (Workspace.size() != Uses.size()) {
3312 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3313 NewRegs, VisitedRegs);
3314 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3315 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3317 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3318 dbgs() << ". Regs:";
3319 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3320 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3321 dbgs() << ' ' << **I;
3324 SolutionCost = NewCost;
3325 Solution = Workspace;
3327 Workspace.pop_back();
3332 if (!EnableRetry && !AnySatisfiedReqRegs)
3335 // If none of the formulae had all of the required registers, relax the
3336 // constraint so that we don't exclude all formulae.
3337 if (!AnySatisfiedReqRegs) {
3338 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3344 /// Solve - Choose one formula from each use. Return the results in the given
3345 /// Solution vector.
3346 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3347 SmallVector<const Formula *, 8> Workspace;
3349 SolutionCost.Loose();
3351 SmallPtrSet<const SCEV *, 16> CurRegs;
3352 DenseSet<const SCEV *> VisitedRegs;
3353 Workspace.reserve(Uses.size());
3355 // SolveRecurse does all the work.
3356 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3357 CurRegs, VisitedRegs);
3358 if (Solution.empty()) {
3359 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
3363 // Ok, we've now made all our decisions.
3364 DEBUG(dbgs() << "\n"
3365 "The chosen solution requires "; SolutionCost.print(dbgs());
3367 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3369 Uses[i].print(dbgs());
3372 Solution[i]->print(dbgs());
3376 assert(Solution.size() == Uses.size() && "Malformed solution!");
3379 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3380 /// the dominator tree far as we can go while still being dominated by the
3381 /// input positions. This helps canonicalize the insert position, which
3382 /// encourages sharing.
3383 BasicBlock::iterator
3384 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3385 const SmallVectorImpl<Instruction *> &Inputs)
3388 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3389 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3392 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3393 if (!Rung) return IP;
3394 Rung = Rung->getIDom();
3395 if (!Rung) return IP;
3396 IDom = Rung->getBlock();
3398 // Don't climb into a loop though.
3399 const Loop *IDomLoop = LI.getLoopFor(IDom);
3400 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3401 if (IDomDepth <= IPLoopDepth &&
3402 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3406 bool AllDominate = true;
3407 Instruction *BetterPos = 0;
3408 Instruction *Tentative = IDom->getTerminator();
3409 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3410 E = Inputs.end(); I != E; ++I) {
3411 Instruction *Inst = *I;
3412 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3413 AllDominate = false;
3416 // Attempt to find an insert position in the middle of the block,
3417 // instead of at the end, so that it can be used for other expansions.
3418 if (IDom == Inst->getParent() &&
3419 (!BetterPos || DT.dominates(BetterPos, Inst)))
3420 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3433 /// AdjustInsertPositionForExpand - Determine an input position which will be
3434 /// dominated by the operands and which will dominate the result.
3435 BasicBlock::iterator
3436 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3438 const LSRUse &LU) const {
3439 // Collect some instructions which must be dominated by the
3440 // expanding replacement. These must be dominated by any operands that
3441 // will be required in the expansion.
3442 SmallVector<Instruction *, 4> Inputs;
3443 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3444 Inputs.push_back(I);
3445 if (LU.Kind == LSRUse::ICmpZero)
3446 if (Instruction *I =
3447 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3448 Inputs.push_back(I);
3449 if (LF.PostIncLoops.count(L)) {
3450 if (LF.isUseFullyOutsideLoop(L))
3451 Inputs.push_back(L->getLoopLatch()->getTerminator());
3453 Inputs.push_back(IVIncInsertPos);
3455 // The expansion must also be dominated by the increment positions of any
3456 // loops it for which it is using post-inc mode.
3457 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3458 E = LF.PostIncLoops.end(); I != E; ++I) {
3459 const Loop *PIL = *I;
3460 if (PIL == L) continue;
3462 // Be dominated by the loop exit.
3463 SmallVector<BasicBlock *, 4> ExitingBlocks;
3464 PIL->getExitingBlocks(ExitingBlocks);
3465 if (!ExitingBlocks.empty()) {
3466 BasicBlock *BB = ExitingBlocks[0];
3467 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3468 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3469 Inputs.push_back(BB->getTerminator());
3473 // Then, climb up the immediate dominator tree as far as we can go while
3474 // still being dominated by the input positions.
3475 IP = HoistInsertPosition(IP, Inputs);
3477 // Don't insert instructions before PHI nodes.
3478 while (isa<PHINode>(IP)) ++IP;
3480 // Ignore landingpad instructions.
3481 while (isa<LandingPadInst>(IP)) ++IP;
3483 // Ignore debug intrinsics.
3484 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3489 /// Expand - Emit instructions for the leading candidate expression for this
3490 /// LSRUse (this is called "expanding").
3491 Value *LSRInstance::Expand(const LSRFixup &LF,
3493 BasicBlock::iterator IP,
3494 SCEVExpander &Rewriter,
3495 SmallVectorImpl<WeakVH> &DeadInsts) const {
3496 const LSRUse &LU = Uses[LF.LUIdx];
3498 // Determine an input position which will be dominated by the operands and
3499 // which will dominate the result.
3500 IP = AdjustInsertPositionForExpand(IP, LF, LU);
3502 // Inform the Rewriter if we have a post-increment use, so that it can
3503 // perform an advantageous expansion.
3504 Rewriter.setPostInc(LF.PostIncLoops);
3506 // This is the type that the user actually needs.
3507 Type *OpTy = LF.OperandValToReplace->getType();
3508 // This will be the type that we'll initially expand to.
3509 Type *Ty = F.getType();
3511 // No type known; just expand directly to the ultimate type.
3513 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
3514 // Expand directly to the ultimate type if it's the right size.
3516 // This is the type to do integer arithmetic in.
3517 Type *IntTy = SE.getEffectiveSCEVType(Ty);
3519 // Build up a list of operands to add together to form the full base.
3520 SmallVector<const SCEV *, 8> Ops;
3522 // Expand the BaseRegs portion.
3523 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3524 E = F.BaseRegs.end(); I != E; ++I) {
3525 const SCEV *Reg = *I;
3526 assert(!Reg->isZero() && "Zero allocated in a base register!");
3528 // If we're expanding for a post-inc user, make the post-inc adjustment.
3529 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3530 Reg = TransformForPostIncUse(Denormalize, Reg,
3531 LF.UserInst, LF.OperandValToReplace,
3534 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
3537 // Flush the operand list to suppress SCEVExpander hoisting.
3539 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3541 Ops.push_back(SE.getUnknown(FullV));
3544 // Expand the ScaledReg portion.
3545 Value *ICmpScaledV = 0;
3546 if (F.AM.Scale != 0) {
3547 const SCEV *ScaledS = F.ScaledReg;
3549 // If we're expanding for a post-inc user, make the post-inc adjustment.
3550 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
3551 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
3552 LF.UserInst, LF.OperandValToReplace,
3555 if (LU.Kind == LSRUse::ICmpZero) {
3556 // An interesting way of "folding" with an icmp is to use a negated
3557 // scale, which we'll implement by inserting it into the other operand
3559 assert(F.AM.Scale == -1 &&
3560 "The only scale supported by ICmpZero uses is -1!");
3561 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
3563 // Otherwise just expand the scaled register and an explicit scale,
3564 // which is expected to be matched as part of the address.
3565 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3566 ScaledS = SE.getMulExpr(ScaledS,
3567 SE.getConstant(ScaledS->getType(), F.AM.Scale));
3568 Ops.push_back(ScaledS);
3570 // Flush the operand list to suppress SCEVExpander hoisting.
3571 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3573 Ops.push_back(SE.getUnknown(FullV));
3577 // Expand the GV portion.
3579 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3581 // Flush the operand list to suppress SCEVExpander hoisting.
3582 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3584 Ops.push_back(SE.getUnknown(FullV));
3587 // Expand the immediate portion.
3588 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3590 if (LU.Kind == LSRUse::ICmpZero) {
3591 // The other interesting way of "folding" with an ICmpZero is to use a
3592 // negated immediate.
3594 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
3596 Ops.push_back(SE.getUnknown(ICmpScaledV));
3597 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3600 // Just add the immediate values. These again are expected to be matched
3601 // as part of the address.
3602 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3606 // Expand the unfolded offset portion.
3607 int64_t UnfoldedOffset = F.UnfoldedOffset;
3608 if (UnfoldedOffset != 0) {
3609 // Just add the immediate values.
3610 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
3614 // Emit instructions summing all the operands.
3615 const SCEV *FullS = Ops.empty() ?
3616 SE.getConstant(IntTy, 0) :
3618 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3620 // We're done expanding now, so reset the rewriter.
3621 Rewriter.clearPostInc();
3623 // An ICmpZero Formula represents an ICmp which we're handling as a
3624 // comparison against zero. Now that we've expanded an expression for that
3625 // form, update the ICmp's other operand.
3626 if (LU.Kind == LSRUse::ICmpZero) {
3627 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3628 DeadInsts.push_back(CI->getOperand(1));
3629 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3630 "a scale at the same time!");
3631 if (F.AM.Scale == -1) {
3632 if (ICmpScaledV->getType() != OpTy) {
3634 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3636 ICmpScaledV, OpTy, "tmp", CI);
3639 CI->setOperand(1, ICmpScaledV);
3641 assert(F.AM.Scale == 0 &&
3642 "ICmp does not support folding a global value and "
3643 "a scale at the same time!");
3644 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3646 if (C->getType() != OpTy)
3647 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3651 CI->setOperand(1, C);
3658 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3659 /// of their operands effectively happens in their predecessor blocks, so the
3660 /// expression may need to be expanded in multiple places.
3661 void LSRInstance::RewriteForPHI(PHINode *PN,
3664 SCEVExpander &Rewriter,
3665 SmallVectorImpl<WeakVH> &DeadInsts,
3667 DenseMap<BasicBlock *, Value *> Inserted;
3668 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3669 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3670 BasicBlock *BB = PN->getIncomingBlock(i);
3672 // If this is a critical edge, split the edge so that we do not insert
3673 // the code on all predecessor/successor paths. We do this unless this
3674 // is the canonical backedge for this loop, which complicates post-inc
3676 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3677 !isa<IndirectBrInst>(BB->getTerminator())) {
3678 BasicBlock *Parent = PN->getParent();
3679 Loop *PNLoop = LI.getLoopFor(Parent);
3680 if (!PNLoop || Parent != PNLoop->getHeader()) {
3681 // Split the critical edge.
3682 BasicBlock *NewBB = 0;
3683 if (!Parent->isLandingPad()) {
3684 NewBB = SplitCriticalEdge(BB, Parent, P,
3685 /*MergeIdenticalEdges=*/true,
3686 /*DontDeleteUselessPhis=*/true);
3688 SmallVector<BasicBlock*, 2> NewBBs;
3689 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
3693 // If PN is outside of the loop and BB is in the loop, we want to
3694 // move the block to be immediately before the PHI block, not
3695 // immediately after BB.
3696 if (L->contains(BB) && !L->contains(PN))
3697 NewBB->moveBefore(PN->getParent());
3699 // Splitting the edge can reduce the number of PHI entries we have.
3700 e = PN->getNumIncomingValues();
3702 i = PN->getBasicBlockIndex(BB);
3706 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3707 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3709 PN->setIncomingValue(i, Pair.first->second);
3711 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3713 // If this is reuse-by-noop-cast, insert the noop cast.
3714 Type *OpTy = LF.OperandValToReplace->getType();
3715 if (FullV->getType() != OpTy)
3717 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3719 FullV, LF.OperandValToReplace->getType(),
3720 "tmp", BB->getTerminator());
3722 PN->setIncomingValue(i, FullV);
3723 Pair.first->second = FullV;
3728 /// Rewrite - Emit instructions for the leading candidate expression for this
3729 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3730 /// the newly expanded value.
3731 void LSRInstance::Rewrite(const LSRFixup &LF,
3733 SCEVExpander &Rewriter,
3734 SmallVectorImpl<WeakVH> &DeadInsts,
3736 // First, find an insertion point that dominates UserInst. For PHI nodes,
3737 // find the nearest block which dominates all the relevant uses.
3738 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3739 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3741 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3743 // If this is reuse-by-noop-cast, insert the noop cast.
3744 Type *OpTy = LF.OperandValToReplace->getType();
3745 if (FullV->getType() != OpTy) {
3747 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3748 FullV, OpTy, "tmp", LF.UserInst);
3752 // Update the user. ICmpZero is handled specially here (for now) because
3753 // Expand may have updated one of the operands of the icmp already, and
3754 // its new value may happen to be equal to LF.OperandValToReplace, in
3755 // which case doing replaceUsesOfWith leads to replacing both operands
3756 // with the same value. TODO: Reorganize this.
3757 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3758 LF.UserInst->setOperand(0, FullV);
3760 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3763 DeadInsts.push_back(LF.OperandValToReplace);
3766 /// ImplementSolution - Rewrite all the fixup locations with new values,
3767 /// following the chosen solution.
3769 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3771 // Keep track of instructions we may have made dead, so that
3772 // we can remove them after we are done working.
3773 SmallVector<WeakVH, 16> DeadInsts;
3775 SCEVExpander Rewriter(SE, "lsr");
3776 Rewriter.disableCanonicalMode();
3777 Rewriter.enableLSRMode();
3778 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3780 // Expand the new value definitions and update the users.
3781 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3782 E = Fixups.end(); I != E; ++I) {
3783 const LSRFixup &Fixup = *I;
3785 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
3790 // Clean up after ourselves. This must be done before deleting any
3794 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3797 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3798 : IU(P->getAnalysis<IVUsers>()),
3799 SE(P->getAnalysis<ScalarEvolution>()),
3800 DT(P->getAnalysis<DominatorTree>()),
3801 LI(P->getAnalysis<LoopInfo>()),
3802 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3804 // If LoopSimplify form is not available, stay out of trouble.
3805 if (!L->isLoopSimplifyForm()) return;
3807 // If there's no interesting work to be done, bail early.
3808 if (IU.empty()) return;
3810 DEBUG(dbgs() << "\nLSR on loop ";
3811 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3814 // First, perform some low-level loop optimizations.
3816 OptimizeLoopTermCond();
3818 // If loop preparation eliminates all interesting IV users, bail.
3819 if (IU.empty()) return;
3821 // Skip nested loops until we can model them better with formulae.
3822 if (!EnableNested && !L->empty()) {
3824 if (EnablePhiElim) {
3825 // Remove any extra phis created by processing inner loops.
3826 SmallVector<WeakVH, 16> DeadInsts;
3827 SCEVExpander Rewriter(SE, "lsr");
3828 Changed |= (bool)Rewriter.replaceCongruentIVs(L, &DT, DeadInsts);
3829 Changed |= (bool)DeleteTriviallyDeadInstructions(DeadInsts);
3831 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
3835 // Start collecting data and preparing for the solver.
3836 CollectInterestingTypesAndFactors();
3837 CollectFixupsAndInitialFormulae();
3838 CollectLoopInvariantFixupsAndFormulae();
3840 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3841 print_uses(dbgs()));
3843 // Now use the reuse data to generate a bunch of interesting ways
3844 // to formulate the values needed for the uses.
3845 GenerateAllReuseFormulae();
3847 FilterOutUndesirableDedicatedRegisters();
3848 NarrowSearchSpaceUsingHeuristics();
3850 SmallVector<const Formula *, 8> Solution;
3853 // Release memory that is no longer needed.
3858 if (Solution.empty())
3862 // Formulae should be legal.
3863 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3864 E = Uses.end(); I != E; ++I) {
3865 const LSRUse &LU = *I;
3866 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3867 JE = LU.Formulae.end(); J != JE; ++J)
3868 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3869 LU.Kind, LU.AccessTy, TLI) &&
3870 "Illegal formula generated!");
3874 // Now that we've decided what we want, make it so.
3875 ImplementSolution(Solution, P);
3877 if (EnablePhiElim) {
3878 // Remove any extra phis created by processing inner loops.
3879 SmallVector<WeakVH, 16> DeadInsts;
3880 SCEVExpander Rewriter(SE, "lsr");
3881 Changed |= (bool)Rewriter.replaceCongruentIVs(L, &DT, DeadInsts);
3882 Changed |= (bool)DeleteTriviallyDeadInstructions(DeadInsts);
3886 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3887 if (Factors.empty() && Types.empty()) return;
3889 OS << "LSR has identified the following interesting factors and types: ";
3892 for (SmallSetVector<int64_t, 8>::const_iterator
3893 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3894 if (!First) OS << ", ";
3899 for (SmallSetVector<Type *, 4>::const_iterator
3900 I = Types.begin(), E = Types.end(); I != E; ++I) {
3901 if (!First) OS << ", ";
3903 OS << '(' << **I << ')';
3908 void LSRInstance::print_fixups(raw_ostream &OS) const {
3909 OS << "LSR is examining the following fixup sites:\n";
3910 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3911 E = Fixups.end(); I != E; ++I) {
3918 void LSRInstance::print_uses(raw_ostream &OS) const {
3919 OS << "LSR is examining the following uses:\n";
3920 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3921 E = Uses.end(); I != E; ++I) {
3922 const LSRUse &LU = *I;
3926 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3927 JE = LU.Formulae.end(); J != JE; ++J) {
3935 void LSRInstance::print(raw_ostream &OS) const {
3936 print_factors_and_types(OS);
3941 void LSRInstance::dump() const {
3942 print(errs()); errs() << '\n';
3947 class LoopStrengthReduce : public LoopPass {
3948 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3949 /// transformation profitability.
3950 const TargetLowering *const TLI;
3953 static char ID; // Pass ID, replacement for typeid
3954 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3957 bool runOnLoop(Loop *L, LPPassManager &LPM);
3958 void getAnalysisUsage(AnalysisUsage &AU) const;
3963 char LoopStrengthReduce::ID = 0;
3964 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
3965 "Loop Strength Reduction", false, false)
3966 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
3967 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3968 INITIALIZE_PASS_DEPENDENCY(IVUsers)
3969 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
3970 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3971 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
3972 "Loop Strength Reduction", false, false)
3975 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3976 return new LoopStrengthReduce(TLI);
3979 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3980 : LoopPass(ID), TLI(tli) {
3981 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
3984 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3985 // We split critical edges, so we change the CFG. However, we do update
3986 // many analyses if they are around.
3987 AU.addPreservedID(LoopSimplifyID);
3989 AU.addRequired<LoopInfo>();
3990 AU.addPreserved<LoopInfo>();
3991 AU.addRequiredID(LoopSimplifyID);
3992 AU.addRequired<DominatorTree>();
3993 AU.addPreserved<DominatorTree>();
3994 AU.addRequired<ScalarEvolution>();
3995 AU.addPreserved<ScalarEvolution>();
3996 // Requiring LoopSimplify a second time here prevents IVUsers from running
3997 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
3998 AU.addRequiredID(LoopSimplifyID);
3999 AU.addRequired<IVUsers>();
4000 AU.addPreserved<IVUsers>();
4003 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4004 bool Changed = false;
4006 // Run the main LSR transformation.
4007 Changed |= LSRInstance(TLI, L, this).getChanged();
4009 // At this point, it is worth checking to see if any recurrence PHIs are also
4010 // dead, so that we can remove them as well.
4011 Changed |= DeleteDeadPHIs(L->getHeader());