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 the addressing mode BaseGV be changed to a ConstantExpr instead
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 #include "llvm/Transforms/Scalar.h"
57 #include "llvm/ADT/DenseSet.h"
58 #include "llvm/ADT/Hashing.h"
59 #include "llvm/ADT/STLExtras.h"
60 #include "llvm/ADT/SetVector.h"
61 #include "llvm/ADT/SmallBitVector.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/LoopPass.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/IR/Constants.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/Instructions.h"
70 #include "llvm/IR/IntrinsicInst.h"
71 #include "llvm/IR/Module.h"
72 #include "llvm/IR/ValueHandle.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/Debug.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
81 #define DEBUG_TYPE "loop-reduce"
83 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
84 /// bail out. This threshold is far beyond the number of users that LSR can
85 /// conceivably solve, so it should not affect generated code, but catches the
86 /// worst cases before LSR burns too much compile time and stack space.
87 static const unsigned MaxIVUsers = 200;
89 // Temporary flag to cleanup congruent phis after LSR phi expansion.
90 // It's currently disabled until we can determine whether it's truly useful or
91 // not. The flag should be removed after the v3.0 release.
92 // This is now needed for ivchains.
93 static cl::opt<bool> EnablePhiElim(
94 "enable-lsr-phielim", cl::Hidden, cl::init(true),
95 cl::desc("Enable LSR phi elimination"));
98 // Stress test IV chain generation.
99 static cl::opt<bool> StressIVChain(
100 "stress-ivchain", cl::Hidden, cl::init(false),
101 cl::desc("Stress test LSR IV chains"));
103 static bool StressIVChain = false;
109 /// Used in situations where the accessed memory type is unknown.
110 static const unsigned UnknownAddressSpace = ~0u;
115 MemAccessTy() : MemTy(nullptr), AddrSpace(UnknownAddressSpace) {}
117 MemAccessTy(Type *Ty, unsigned AS) :
118 MemTy(Ty), AddrSpace(AS) {}
120 bool operator==(MemAccessTy Other) const {
121 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
124 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
126 static MemAccessTy getUnknown(LLVMContext &Ctx) {
127 return MemAccessTy(Type::getVoidTy(Ctx), UnknownAddressSpace);
131 /// This class holds data which is used to order reuse candidates.
134 /// This represents the set of LSRUse indices which reference
135 /// a particular register.
136 SmallBitVector UsedByIndices;
138 void print(raw_ostream &OS) const;
144 void RegSortData::print(raw_ostream &OS) const {
145 OS << "[NumUses=" << UsedByIndices.count() << ']';
149 void RegSortData::dump() const {
150 print(errs()); errs() << '\n';
155 /// Map register candidates to information about how they are used.
156 class RegUseTracker {
157 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
159 RegUsesTy RegUsesMap;
160 SmallVector<const SCEV *, 16> RegSequence;
163 void countRegister(const SCEV *Reg, size_t LUIdx);
164 void dropRegister(const SCEV *Reg, size_t LUIdx);
165 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
167 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
169 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
173 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
174 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
175 iterator begin() { return RegSequence.begin(); }
176 iterator end() { return RegSequence.end(); }
177 const_iterator begin() const { return RegSequence.begin(); }
178 const_iterator end() const { return RegSequence.end(); }
184 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
185 std::pair<RegUsesTy::iterator, bool> Pair =
186 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
187 RegSortData &RSD = Pair.first->second;
189 RegSequence.push_back(Reg);
190 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
191 RSD.UsedByIndices.set(LUIdx);
195 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
196 RegUsesTy::iterator It = RegUsesMap.find(Reg);
197 assert(It != RegUsesMap.end());
198 RegSortData &RSD = It->second;
199 assert(RSD.UsedByIndices.size() > LUIdx);
200 RSD.UsedByIndices.reset(LUIdx);
204 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
205 assert(LUIdx <= LastLUIdx);
207 // Update RegUses. The data structure is not optimized for this purpose;
208 // we must iterate through it and update each of the bit vectors.
209 for (auto &Pair : RegUsesMap) {
210 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
211 if (LUIdx < UsedByIndices.size())
212 UsedByIndices[LUIdx] =
213 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
214 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
219 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
220 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
221 if (I == RegUsesMap.end())
223 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
224 int i = UsedByIndices.find_first();
225 if (i == -1) return false;
226 if ((size_t)i != LUIdx) return true;
227 return UsedByIndices.find_next(i) != -1;
230 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
231 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
232 assert(I != RegUsesMap.end() && "Unknown register!");
233 return I->second.UsedByIndices;
236 void RegUseTracker::clear() {
243 /// This class holds information that describes a formula for computing
244 /// satisfying a use. It may include broken-out immediates and scaled registers.
246 /// Global base address used for complex addressing.
249 /// Base offset for complex addressing.
252 /// Whether any complex addressing has a base register.
255 /// The scale of any complex addressing.
258 /// The list of "base" registers for this use. When this is non-empty. The
259 /// canonical representation of a formula is
260 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
261 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
262 /// #1 enforces that the scaled register is always used when at least two
263 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
264 /// #2 enforces that 1 * reg is reg.
265 /// This invariant can be temporarly broken while building a formula.
266 /// However, every formula inserted into the LSRInstance must be in canonical
268 SmallVector<const SCEV *, 4> BaseRegs;
270 /// The 'scaled' register for this use. This should be non-null when Scale is
272 const SCEV *ScaledReg;
274 /// An additional constant offset which added near the use. This requires a
275 /// temporary register, but the offset itself can live in an add immediate
276 /// field rather than a register.
277 int64_t UnfoldedOffset;
280 : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
281 ScaledReg(nullptr), UnfoldedOffset(0) {}
283 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
285 bool isCanonical() const;
291 size_t getNumRegs() const;
292 Type *getType() const;
294 void deleteBaseReg(const SCEV *&S);
296 bool referencesReg(const SCEV *S) const;
297 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
298 const RegUseTracker &RegUses) const;
300 void print(raw_ostream &OS) const;
306 /// Recursion helper for initialMatch.
307 static void DoInitialMatch(const SCEV *S, Loop *L,
308 SmallVectorImpl<const SCEV *> &Good,
309 SmallVectorImpl<const SCEV *> &Bad,
310 ScalarEvolution &SE) {
311 // Collect expressions which properly dominate the loop header.
312 if (SE.properlyDominates(S, L->getHeader())) {
317 // Look at add operands.
318 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
319 for (const SCEV *S : Add->operands())
320 DoInitialMatch(S, L, Good, Bad, SE);
324 // Look at addrec operands.
325 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
326 if (!AR->getStart()->isZero()) {
327 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
328 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
329 AR->getStepRecurrence(SE),
330 // FIXME: AR->getNoWrapFlags()
331 AR->getLoop(), SCEV::FlagAnyWrap),
336 // Handle a multiplication by -1 (negation) if it didn't fold.
337 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
338 if (Mul->getOperand(0)->isAllOnesValue()) {
339 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
340 const SCEV *NewMul = SE.getMulExpr(Ops);
342 SmallVector<const SCEV *, 4> MyGood;
343 SmallVector<const SCEV *, 4> MyBad;
344 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
345 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
346 SE.getEffectiveSCEVType(NewMul->getType())));
347 for (const SCEV *S : MyGood)
348 Good.push_back(SE.getMulExpr(NegOne, S));
349 for (const SCEV *S : MyBad)
350 Bad.push_back(SE.getMulExpr(NegOne, S));
354 // Ok, we can't do anything interesting. Just stuff the whole thing into a
355 // register and hope for the best.
359 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
360 /// all loop-invariant and loop-computable values in a single base register.
361 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
362 SmallVector<const SCEV *, 4> Good;
363 SmallVector<const SCEV *, 4> Bad;
364 DoInitialMatch(S, L, Good, Bad, SE);
366 const SCEV *Sum = SE.getAddExpr(Good);
368 BaseRegs.push_back(Sum);
372 const SCEV *Sum = SE.getAddExpr(Bad);
374 BaseRegs.push_back(Sum);
380 /// \brief Check whether or not this formula statisfies the canonical
382 /// \see Formula::BaseRegs.
383 bool Formula::isCanonical() const {
385 return Scale != 1 || !BaseRegs.empty();
386 return BaseRegs.size() <= 1;
389 /// \brief Helper method to morph a formula into its canonical representation.
390 /// \see Formula::BaseRegs.
391 /// Every formula having more than one base register, must use the ScaledReg
392 /// field. Otherwise, we would have to do special cases everywhere in LSR
393 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
394 /// On the other hand, 1*reg should be canonicalized into reg.
395 void Formula::canonicalize() {
398 // So far we did not need this case. This is easy to implement but it is
399 // useless to maintain dead code. Beside it could hurt compile time.
400 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
401 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
402 ScaledReg = BaseRegs.back();
405 size_t BaseRegsSize = BaseRegs.size();
407 // If ScaledReg is an invariant, try to find a variant expression.
408 while (Try < BaseRegsSize && !isa<SCEVAddRecExpr>(ScaledReg))
409 std::swap(ScaledReg, BaseRegs[Try++]);
412 /// \brief Get rid of the scale in the formula.
413 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
414 /// \return true if it was possible to get rid of the scale, false otherwise.
415 /// \note After this operation the formula may not be in the canonical form.
416 bool Formula::unscale() {
420 BaseRegs.push_back(ScaledReg);
425 /// Return the total number of register operands used by this formula. This does
426 /// not include register uses implied by non-constant addrec strides.
427 size_t Formula::getNumRegs() const {
428 return !!ScaledReg + BaseRegs.size();
431 /// Return the type of this formula, if it has one, or null otherwise. This type
432 /// is meaningless except for the bit size.
433 Type *Formula::getType() const {
434 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
435 ScaledReg ? ScaledReg->getType() :
436 BaseGV ? BaseGV->getType() :
440 /// Delete the given base reg from the BaseRegs list.
441 void Formula::deleteBaseReg(const SCEV *&S) {
442 if (&S != &BaseRegs.back())
443 std::swap(S, BaseRegs.back());
447 /// Test if this formula references the given register.
448 bool Formula::referencesReg(const SCEV *S) const {
449 return S == ScaledReg ||
450 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
453 /// Test whether this formula uses registers which are used by uses other than
454 /// the use with the given index.
455 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
456 const RegUseTracker &RegUses) const {
458 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
460 for (const SCEV *BaseReg : BaseRegs)
461 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
466 void Formula::print(raw_ostream &OS) const {
469 if (!First) OS << " + "; else First = false;
470 BaseGV->printAsOperand(OS, /*PrintType=*/false);
472 if (BaseOffset != 0) {
473 if (!First) OS << " + "; else First = false;
476 for (const SCEV *BaseReg : BaseRegs) {
477 if (!First) OS << " + "; else First = false;
478 OS << "reg(" << *BaseReg << ')';
480 if (HasBaseReg && BaseRegs.empty()) {
481 if (!First) OS << " + "; else First = false;
482 OS << "**error: HasBaseReg**";
483 } else if (!HasBaseReg && !BaseRegs.empty()) {
484 if (!First) OS << " + "; else First = false;
485 OS << "**error: !HasBaseReg**";
488 if (!First) OS << " + "; else First = false;
489 OS << Scale << "*reg(";
496 if (UnfoldedOffset != 0) {
497 if (!First) OS << " + ";
498 OS << "imm(" << UnfoldedOffset << ')';
503 void Formula::dump() const {
504 print(errs()); errs() << '\n';
507 /// Return true if the given addrec can be sign-extended without changing its
509 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
511 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
512 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
515 /// Return true if the given add can be sign-extended without changing its
517 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
519 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
520 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
523 /// Return true if the given mul can be sign-extended without changing its
525 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
527 IntegerType::get(SE.getContext(),
528 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
529 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
532 /// Return an expression for LHS /s RHS, if it can be determined and if the
533 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
534 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
535 /// the multiplication may overflow, which is useful when the result will be
536 /// used in a context where the most significant bits are ignored.
537 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
539 bool IgnoreSignificantBits = false) {
540 // Handle the trivial case, which works for any SCEV type.
542 return SE.getConstant(LHS->getType(), 1);
544 // Handle a few RHS special cases.
545 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
547 const APInt &RA = RC->getAPInt();
548 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
550 if (RA.isAllOnesValue())
551 return SE.getMulExpr(LHS, RC);
552 // Handle x /s 1 as x.
557 // Check for a division of a constant by a constant.
558 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
561 const APInt &LA = C->getAPInt();
562 const APInt &RA = RC->getAPInt();
563 if (LA.srem(RA) != 0)
565 return SE.getConstant(LA.sdiv(RA));
568 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
569 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
570 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
571 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
572 IgnoreSignificantBits);
573 if (!Step) return nullptr;
574 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
575 IgnoreSignificantBits);
576 if (!Start) return nullptr;
577 // FlagNW is independent of the start value, step direction, and is
578 // preserved with smaller magnitude steps.
579 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
580 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
585 // Distribute the sdiv over add operands, if the add doesn't overflow.
586 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
587 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
588 SmallVector<const SCEV *, 8> Ops;
589 for (const SCEV *S : Add->operands()) {
590 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
591 if (!Op) return nullptr;
594 return SE.getAddExpr(Ops);
599 // Check for a multiply operand that we can pull RHS out of.
600 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
601 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
602 SmallVector<const SCEV *, 4> Ops;
604 for (const SCEV *S : Mul->operands()) {
606 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
607 IgnoreSignificantBits)) {
613 return Found ? SE.getMulExpr(Ops) : nullptr;
618 // Otherwise we don't know.
622 /// If S involves the addition of a constant integer value, return that integer
623 /// value, and mutate S to point to a new SCEV with that value excluded.
624 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
625 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
626 if (C->getAPInt().getMinSignedBits() <= 64) {
627 S = SE.getConstant(C->getType(), 0);
628 return C->getValue()->getSExtValue();
630 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
631 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
632 int64_t Result = ExtractImmediate(NewOps.front(), SE);
634 S = SE.getAddExpr(NewOps);
636 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
637 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
638 int64_t Result = ExtractImmediate(NewOps.front(), SE);
640 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
641 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
648 /// If S involves the addition of a GlobalValue address, return that symbol, and
649 /// mutate S to point to a new SCEV with that value excluded.
650 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
651 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
652 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
653 S = SE.getConstant(GV->getType(), 0);
656 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
657 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
658 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
660 S = SE.getAddExpr(NewOps);
662 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
663 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
664 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
666 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
667 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
674 /// Returns true if the specified instruction is using the specified value as an
676 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
677 bool isAddress = isa<LoadInst>(Inst);
678 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
679 if (SI->getOperand(1) == OperandVal)
681 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
682 // Addressing modes can also be folded into prefetches and a variety
684 switch (II->getIntrinsicID()) {
686 case Intrinsic::prefetch:
687 if (II->getArgOperand(0) == OperandVal)
695 /// Return the type of the memory being accessed.
696 static MemAccessTy getAccessType(const Instruction *Inst) {
697 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
698 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
699 AccessTy.MemTy = SI->getOperand(0)->getType();
700 AccessTy.AddrSpace = SI->getPointerAddressSpace();
701 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
702 AccessTy.AddrSpace = LI->getPointerAddressSpace();
705 // All pointers have the same requirements, so canonicalize them to an
706 // arbitrary pointer type to minimize variation.
707 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
708 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
709 PTy->getAddressSpace());
714 /// Return true if this AddRec is already a phi in its loop.
715 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
716 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
717 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
718 if (SE.isSCEVable(PN->getType()) &&
719 (SE.getEffectiveSCEVType(PN->getType()) ==
720 SE.getEffectiveSCEVType(AR->getType())) &&
721 SE.getSCEV(PN) == AR)
727 /// Check if expanding this expression is likely to incur significant cost. This
728 /// is tricky because SCEV doesn't track which expressions are actually computed
729 /// by the current IR.
731 /// We currently allow expansion of IV increments that involve adds,
732 /// multiplication by constants, and AddRecs from existing phis.
734 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
735 /// obvious multiple of the UDivExpr.
736 static bool isHighCostExpansion(const SCEV *S,
737 SmallPtrSetImpl<const SCEV*> &Processed,
738 ScalarEvolution &SE) {
739 // Zero/One operand expressions
740 switch (S->getSCEVType()) {
745 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
748 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
751 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
755 if (!Processed.insert(S).second)
758 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
759 for (const SCEV *S : Add->operands()) {
760 if (isHighCostExpansion(S, Processed, SE))
766 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
767 if (Mul->getNumOperands() == 2) {
768 // Multiplication by a constant is ok
769 if (isa<SCEVConstant>(Mul->getOperand(0)))
770 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
772 // If we have the value of one operand, check if an existing
773 // multiplication already generates this expression.
774 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
775 Value *UVal = U->getValue();
776 for (User *UR : UVal->users()) {
777 // If U is a constant, it may be used by a ConstantExpr.
778 Instruction *UI = dyn_cast<Instruction>(UR);
779 if (UI && UI->getOpcode() == Instruction::Mul &&
780 SE.isSCEVable(UI->getType())) {
781 return SE.getSCEV(UI) == Mul;
788 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
789 if (isExistingPhi(AR, SE))
793 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
797 /// If any of the instructions is the specified set are trivially dead, delete
798 /// them and see if this makes any of their operands subsequently dead.
800 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
801 bool Changed = false;
803 while (!DeadInsts.empty()) {
804 Value *V = DeadInsts.pop_back_val();
805 Instruction *I = dyn_cast_or_null<Instruction>(V);
807 if (!I || !isInstructionTriviallyDead(I))
810 for (Use &O : I->operands())
811 if (Instruction *U = dyn_cast<Instruction>(O)) {
814 DeadInsts.emplace_back(U);
817 I->eraseFromParent();
828 /// \brief Check if the addressing mode defined by \p F is completely
829 /// folded in \p LU at isel time.
830 /// This includes address-mode folding and special icmp tricks.
831 /// This function returns true if \p LU can accommodate what \p F
832 /// defines and up to 1 base + 1 scaled + offset.
833 /// In other words, if \p F has several base registers, this function may
834 /// still return true. Therefore, users still need to account for
835 /// additional base registers and/or unfolded offsets to derive an
836 /// accurate cost model.
837 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
838 const LSRUse &LU, const Formula &F);
839 // Get the cost of the scaling factor used in F for LU.
840 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
841 const LSRUse &LU, const Formula &F);
845 /// This class is used to measure and compare candidate formulae.
847 /// TODO: Some of these could be merged. Also, a lexical ordering
848 /// isn't always optimal.
852 unsigned NumBaseAdds;
859 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
860 SetupCost(0), ScaleCost(0) {}
862 bool operator<(const Cost &Other) const;
867 // Once any of the metrics loses, they must all remain losers.
869 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
870 | ImmCost | SetupCost | ScaleCost) != ~0u)
871 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
872 & ImmCost & SetupCost & ScaleCost) == ~0u);
877 assert(isValid() && "invalid cost");
878 return NumRegs == ~0u;
881 void RateFormula(const TargetTransformInfo &TTI,
883 SmallPtrSetImpl<const SCEV *> &Regs,
884 const DenseSet<const SCEV *> &VisitedRegs,
886 const SmallVectorImpl<int64_t> &Offsets,
887 ScalarEvolution &SE, DominatorTree &DT,
889 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
891 void print(raw_ostream &OS) const;
895 void RateRegister(const SCEV *Reg,
896 SmallPtrSetImpl<const SCEV *> &Regs,
898 ScalarEvolution &SE, DominatorTree &DT);
899 void RatePrimaryRegister(const SCEV *Reg,
900 SmallPtrSetImpl<const SCEV *> &Regs,
902 ScalarEvolution &SE, DominatorTree &DT,
903 SmallPtrSetImpl<const SCEV *> *LoserRegs);
908 /// Tally up interesting quantities from the given register.
909 void Cost::RateRegister(const SCEV *Reg,
910 SmallPtrSetImpl<const SCEV *> &Regs,
912 ScalarEvolution &SE, DominatorTree &DT) {
913 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
914 // If this is an addrec for another loop, don't second-guess its addrec phi
915 // nodes. LSR isn't currently smart enough to reason about more than one
916 // loop at a time. LSR has already run on inner loops, will not run on outer
917 // loops, and cannot be expected to change sibling loops.
918 if (AR->getLoop() != L) {
919 // If the AddRec exists, consider it's register free and leave it alone.
920 if (isExistingPhi(AR, SE))
923 // Otherwise, do not consider this formula at all.
927 AddRecCost += 1; /// TODO: This should be a function of the stride.
929 // Add the step value register, if it needs one.
930 // TODO: The non-affine case isn't precisely modeled here.
931 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
932 if (!Regs.count(AR->getOperand(1))) {
933 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
941 // Rough heuristic; favor registers which don't require extra setup
942 // instructions in the preheader.
943 if (!isa<SCEVUnknown>(Reg) &&
944 !isa<SCEVConstant>(Reg) &&
945 !(isa<SCEVAddRecExpr>(Reg) &&
946 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
947 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
950 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
951 SE.hasComputableLoopEvolution(Reg, L);
954 /// Record this register in the set. If we haven't seen it before, rate
955 /// it. Optional LoserRegs provides a way to declare any formula that refers to
956 /// one of those regs an instant loser.
957 void Cost::RatePrimaryRegister(const SCEV *Reg,
958 SmallPtrSetImpl<const SCEV *> &Regs,
960 ScalarEvolution &SE, DominatorTree &DT,
961 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
962 if (LoserRegs && LoserRegs->count(Reg)) {
966 if (Regs.insert(Reg).second) {
967 RateRegister(Reg, Regs, L, SE, DT);
968 if (LoserRegs && isLoser())
969 LoserRegs->insert(Reg);
973 void Cost::RateFormula(const TargetTransformInfo &TTI,
975 SmallPtrSetImpl<const SCEV *> &Regs,
976 const DenseSet<const SCEV *> &VisitedRegs,
978 const SmallVectorImpl<int64_t> &Offsets,
979 ScalarEvolution &SE, DominatorTree &DT,
981 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
982 assert(F.isCanonical() && "Cost is accurate only for canonical formula");
983 // Tally up the registers.
984 if (const SCEV *ScaledReg = F.ScaledReg) {
985 if (VisitedRegs.count(ScaledReg)) {
989 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
993 for (const SCEV *BaseReg : F.BaseRegs) {
994 if (VisitedRegs.count(BaseReg)) {
998 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1003 // Determine how many (unfolded) adds we'll need inside the loop.
1004 size_t NumBaseParts = F.getNumRegs();
1005 if (NumBaseParts > 1)
1006 // Do not count the base and a possible second register if the target
1007 // allows to fold 2 registers.
1009 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1010 NumBaseAdds += (F.UnfoldedOffset != 0);
1012 // Accumulate non-free scaling amounts.
1013 ScaleCost += getScalingFactorCost(TTI, LU, F);
1015 // Tally up the non-zero immediates.
1016 for (int64_t O : Offsets) {
1017 int64_t Offset = (uint64_t)O + F.BaseOffset;
1019 ImmCost += 64; // Handle symbolic values conservatively.
1020 // TODO: This should probably be the pointer size.
1021 else if (Offset != 0)
1022 ImmCost += APInt(64, Offset, true).getMinSignedBits();
1024 assert(isValid() && "invalid cost");
1027 /// Set this cost to a losing value.
1038 /// Choose the lower cost.
1039 bool Cost::operator<(const Cost &Other) const {
1040 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
1041 ImmCost, SetupCost) <
1042 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
1043 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
1047 void Cost::print(raw_ostream &OS) const {
1048 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
1049 if (AddRecCost != 0)
1050 OS << ", with addrec cost " << AddRecCost;
1052 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
1053 if (NumBaseAdds != 0)
1054 OS << ", plus " << NumBaseAdds << " base add"
1055 << (NumBaseAdds == 1 ? "" : "s");
1057 OS << ", plus " << ScaleCost << " scale cost";
1059 OS << ", plus " << ImmCost << " imm cost";
1061 OS << ", plus " << SetupCost << " setup cost";
1065 void Cost::dump() const {
1066 print(errs()); errs() << '\n';
1071 /// An operand value in an instruction which is to be replaced with some
1072 /// equivalent, possibly strength-reduced, replacement.
1074 /// The instruction which will be updated.
1075 Instruction *UserInst;
1077 /// The operand of the instruction which will be replaced. The operand may be
1078 /// used more than once; every instance will be replaced.
1079 Value *OperandValToReplace;
1081 /// If this user is to use the post-incremented value of an induction
1082 /// variable, this variable is non-null and holds the loop associated with the
1083 /// induction variable.
1084 PostIncLoopSet PostIncLoops;
1086 /// The index of the LSRUse describing the expression which this fixup needs,
1087 /// minus an offset (below).
1090 /// A constant offset to be added to the LSRUse expression. This allows
1091 /// multiple fixups to share the same LSRUse with different offsets, for
1092 /// example in an unrolled loop.
1095 bool isUseFullyOutsideLoop(const Loop *L) const;
1099 void print(raw_ostream &OS) const;
1105 LSRFixup::LSRFixup()
1106 : UserInst(nullptr), OperandValToReplace(nullptr), LUIdx(~size_t(0)),
1109 /// Test whether this fixup always uses its value outside of the given loop.
1110 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1111 // PHI nodes use their value in their incoming blocks.
1112 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1113 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1114 if (PN->getIncomingValue(i) == OperandValToReplace &&
1115 L->contains(PN->getIncomingBlock(i)))
1120 return !L->contains(UserInst);
1123 void LSRFixup::print(raw_ostream &OS) const {
1125 // Store is common and interesting enough to be worth special-casing.
1126 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1128 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1129 } else if (UserInst->getType()->isVoidTy())
1130 OS << UserInst->getOpcodeName();
1132 UserInst->printAsOperand(OS, /*PrintType=*/false);
1134 OS << ", OperandValToReplace=";
1135 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1137 for (const Loop *PIL : PostIncLoops) {
1138 OS << ", PostIncLoop=";
1139 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1142 if (LUIdx != ~size_t(0))
1143 OS << ", LUIdx=" << LUIdx;
1146 OS << ", Offset=" << Offset;
1150 void LSRFixup::dump() const {
1151 print(errs()); errs() << '\n';
1156 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1157 /// SmallVectors of const SCEV*.
1158 struct UniquifierDenseMapInfo {
1159 static SmallVector<const SCEV *, 4> getEmptyKey() {
1160 SmallVector<const SCEV *, 4> V;
1161 V.push_back(reinterpret_cast<const SCEV *>(-1));
1165 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1166 SmallVector<const SCEV *, 4> V;
1167 V.push_back(reinterpret_cast<const SCEV *>(-2));
1171 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1172 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1175 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1176 const SmallVector<const SCEV *, 4> &RHS) {
1181 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1182 /// as uses invented by LSR itself. It includes information about what kinds of
1183 /// things can be folded into the user, information about the user itself, and
1184 /// information about how the use may be satisfied. TODO: Represent multiple
1185 /// users of the same expression in common?
1187 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1190 /// An enum for a kind of use, indicating what types of scaled and immediate
1191 /// operands it might support.
1193 Basic, ///< A normal use, with no folding.
1194 Special, ///< A special case of basic, allowing -1 scales.
1195 Address, ///< An address use; folding according to TargetLowering
1196 ICmpZero ///< An equality icmp with both operands folded into one.
1197 // TODO: Add a generic icmp too?
1200 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1203 MemAccessTy AccessTy;
1205 SmallVector<int64_t, 8> Offsets;
1209 /// This records whether all of the fixups using this LSRUse are outside of
1210 /// the loop, in which case some special-case heuristics may be used.
1211 bool AllFixupsOutsideLoop;
1213 /// RigidFormula is set to true to guarantee that this use will be associated
1214 /// with a single formula--the one that initially matched. Some SCEV
1215 /// expressions cannot be expanded. This allows LSR to consider the registers
1216 /// used by those expressions without the need to expand them later after
1217 /// changing the formula.
1220 /// This records the widest use type for any fixup using this
1221 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1222 /// fixup widths to be equivalent, because the narrower one may be relying on
1223 /// the implicit truncation to truncate away bogus bits.
1224 Type *WidestFixupType;
1226 /// A list of ways to build a value that can satisfy this user. After the
1227 /// list is populated, one of these is selected heuristically and used to
1228 /// formulate a replacement for OperandValToReplace in UserInst.
1229 SmallVector<Formula, 12> Formulae;
1231 /// The set of register candidates used by all formulae in this LSRUse.
1232 SmallPtrSet<const SCEV *, 4> Regs;
1234 LSRUse(KindType K, MemAccessTy AT)
1235 : Kind(K), AccessTy(AT), MinOffset(INT64_MAX), MaxOffset(INT64_MIN),
1236 AllFixupsOutsideLoop(true), RigidFormula(false),
1237 WidestFixupType(nullptr) {}
1239 bool HasFormulaWithSameRegs(const Formula &F) const;
1240 bool InsertFormula(const Formula &F);
1241 void DeleteFormula(Formula &F);
1242 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1244 void print(raw_ostream &OS) const;
1250 /// Test whether this use as a formula which has the same registers as the given
1252 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1253 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1254 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1255 // Unstable sort by host order ok, because this is only used for uniquifying.
1256 std::sort(Key.begin(), Key.end());
1257 return Uniquifier.count(Key);
1260 /// If the given formula has not yet been inserted, add it to the list, and
1261 /// return true. Return false otherwise. The formula must be in canonical form.
1262 bool LSRUse::InsertFormula(const Formula &F) {
1263 assert(F.isCanonical() && "Invalid canonical representation");
1265 if (!Formulae.empty() && RigidFormula)
1268 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1269 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1270 // Unstable sort by host order ok, because this is only used for uniquifying.
1271 std::sort(Key.begin(), Key.end());
1273 if (!Uniquifier.insert(Key).second)
1276 // Using a register to hold the value of 0 is not profitable.
1277 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1278 "Zero allocated in a scaled register!");
1280 for (const SCEV *BaseReg : F.BaseRegs)
1281 assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1284 // Add the formula to the list.
1285 Formulae.push_back(F);
1287 // Record registers now being used by this use.
1288 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1290 Regs.insert(F.ScaledReg);
1295 /// Remove the given formula from this use's list.
1296 void LSRUse::DeleteFormula(Formula &F) {
1297 if (&F != &Formulae.back())
1298 std::swap(F, Formulae.back());
1299 Formulae.pop_back();
1302 /// Recompute the Regs field, and update RegUses.
1303 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1304 // Now that we've filtered out some formulae, recompute the Regs set.
1305 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1307 for (const Formula &F : Formulae) {
1308 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1309 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1312 // Update the RegTracker.
1313 for (const SCEV *S : OldRegs)
1315 RegUses.dropRegister(S, LUIdx);
1318 void LSRUse::print(raw_ostream &OS) const {
1319 OS << "LSR Use: Kind=";
1321 case Basic: OS << "Basic"; break;
1322 case Special: OS << "Special"; break;
1323 case ICmpZero: OS << "ICmpZero"; break;
1325 OS << "Address of ";
1326 if (AccessTy.MemTy->isPointerTy())
1327 OS << "pointer"; // the full pointer type could be really verbose
1329 OS << *AccessTy.MemTy;
1332 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1335 OS << ", Offsets={";
1336 bool NeedComma = false;
1337 for (int64_t O : Offsets) {
1338 if (NeedComma) OS << ',';
1344 if (AllFixupsOutsideLoop)
1345 OS << ", all-fixups-outside-loop";
1347 if (WidestFixupType)
1348 OS << ", widest fixup type: " << *WidestFixupType;
1352 void LSRUse::dump() const {
1353 print(errs()); errs() << '\n';
1356 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1357 LSRUse::KindType Kind, MemAccessTy AccessTy,
1358 GlobalValue *BaseGV, int64_t BaseOffset,
1359 bool HasBaseReg, int64_t Scale) {
1361 case LSRUse::Address:
1362 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1363 HasBaseReg, Scale, AccessTy.AddrSpace);
1365 case LSRUse::ICmpZero:
1366 // There's not even a target hook for querying whether it would be legal to
1367 // fold a GV into an ICmp.
1371 // ICmp only has two operands; don't allow more than two non-trivial parts.
1372 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1375 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1376 // putting the scaled register in the other operand of the icmp.
1377 if (Scale != 0 && Scale != -1)
1380 // If we have low-level target information, ask the target if it can fold an
1381 // integer immediate on an icmp.
1382 if (BaseOffset != 0) {
1384 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1385 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1386 // Offs is the ICmp immediate.
1388 // The cast does the right thing with INT64_MIN.
1389 BaseOffset = -(uint64_t)BaseOffset;
1390 return TTI.isLegalICmpImmediate(BaseOffset);
1393 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1397 // Only handle single-register values.
1398 return !BaseGV && Scale == 0 && BaseOffset == 0;
1400 case LSRUse::Special:
1401 // Special case Basic to handle -1 scales.
1402 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1405 llvm_unreachable("Invalid LSRUse Kind!");
1408 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1409 int64_t MinOffset, int64_t MaxOffset,
1410 LSRUse::KindType Kind, MemAccessTy AccessTy,
1411 GlobalValue *BaseGV, int64_t BaseOffset,
1412 bool HasBaseReg, int64_t Scale) {
1413 // Check for overflow.
1414 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1417 MinOffset = (uint64_t)BaseOffset + MinOffset;
1418 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1421 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1423 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1424 HasBaseReg, Scale) &&
1425 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1429 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1430 int64_t MinOffset, int64_t MaxOffset,
1431 LSRUse::KindType Kind, MemAccessTy AccessTy,
1433 // For the purpose of isAMCompletelyFolded either having a canonical formula
1434 // or a scale not equal to zero is correct.
1435 // Problems may arise from non canonical formulae having a scale == 0.
1436 // Strictly speaking it would best to just rely on canonical formulae.
1437 // However, when we generate the scaled formulae, we first check that the
1438 // scaling factor is profitable before computing the actual ScaledReg for
1439 // compile time sake.
1440 assert((F.isCanonical() || F.Scale != 0));
1441 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1442 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1445 /// Test whether we know how to expand the current formula.
1446 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1447 int64_t MaxOffset, LSRUse::KindType Kind,
1448 MemAccessTy AccessTy, GlobalValue *BaseGV,
1449 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1450 // We know how to expand completely foldable formulae.
1451 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1452 BaseOffset, HasBaseReg, Scale) ||
1453 // Or formulae that use a base register produced by a sum of base
1456 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1457 BaseGV, BaseOffset, true, 0));
1460 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1461 int64_t MaxOffset, LSRUse::KindType Kind,
1462 MemAccessTy AccessTy, const Formula &F) {
1463 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1464 F.BaseOffset, F.HasBaseReg, F.Scale);
1467 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1468 const LSRUse &LU, const Formula &F) {
1469 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1470 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1474 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1475 const LSRUse &LU, const Formula &F) {
1479 // If the use is not completely folded in that instruction, we will have to
1480 // pay an extra cost only for scale != 1.
1481 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1483 return F.Scale != 1;
1486 case LSRUse::Address: {
1487 // Check the scaling factor cost with both the min and max offsets.
1488 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1489 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1490 F.Scale, LU.AccessTy.AddrSpace);
1491 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1492 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1493 F.Scale, LU.AccessTy.AddrSpace);
1495 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1496 "Legal addressing mode has an illegal cost!");
1497 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1499 case LSRUse::ICmpZero:
1501 case LSRUse::Special:
1502 // The use is completely folded, i.e., everything is folded into the
1507 llvm_unreachable("Invalid LSRUse Kind!");
1510 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1511 LSRUse::KindType Kind, MemAccessTy AccessTy,
1512 GlobalValue *BaseGV, int64_t BaseOffset,
1514 // Fast-path: zero is always foldable.
1515 if (BaseOffset == 0 && !BaseGV) return true;
1517 // Conservatively, create an address with an immediate and a
1518 // base and a scale.
1519 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1521 // Canonicalize a scale of 1 to a base register if the formula doesn't
1522 // already have a base register.
1523 if (!HasBaseReg && Scale == 1) {
1528 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1532 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1533 ScalarEvolution &SE, int64_t MinOffset,
1534 int64_t MaxOffset, LSRUse::KindType Kind,
1535 MemAccessTy AccessTy, const SCEV *S,
1537 // Fast-path: zero is always foldable.
1538 if (S->isZero()) return true;
1540 // Conservatively, create an address with an immediate and a
1541 // base and a scale.
1542 int64_t BaseOffset = ExtractImmediate(S, SE);
1543 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1545 // If there's anything else involved, it's not foldable.
1546 if (!S->isZero()) return false;
1548 // Fast-path: zero is always foldable.
1549 if (BaseOffset == 0 && !BaseGV) return true;
1551 // Conservatively, create an address with an immediate and a
1552 // base and a scale.
1553 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1555 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1556 BaseOffset, HasBaseReg, Scale);
1561 /// An individual increment in a Chain of IV increments. Relate an IV user to
1562 /// an expression that computes the IV it uses from the IV used by the previous
1563 /// link in the Chain.
1565 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1566 /// original IVOperand. The head of the chain's IVOperand is only valid during
1567 /// chain collection, before LSR replaces IV users. During chain generation,
1568 /// IncExpr can be used to find the new IVOperand that computes the same
1571 Instruction *UserInst;
1573 const SCEV *IncExpr;
1575 IVInc(Instruction *U, Value *O, const SCEV *E):
1576 UserInst(U), IVOperand(O), IncExpr(E) {}
1579 // The list of IV increments in program order. We typically add the head of a
1580 // chain without finding subsequent links.
1582 SmallVector<IVInc,1> Incs;
1583 const SCEV *ExprBase;
1585 IVChain() : ExprBase(nullptr) {}
1587 IVChain(const IVInc &Head, const SCEV *Base)
1588 : Incs(1, Head), ExprBase(Base) {}
1590 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1592 // Return the first increment in the chain.
1593 const_iterator begin() const {
1594 assert(!Incs.empty());
1595 return std::next(Incs.begin());
1597 const_iterator end() const {
1601 // Returns true if this chain contains any increments.
1602 bool hasIncs() const { return Incs.size() >= 2; }
1604 // Add an IVInc to the end of this chain.
1605 void add(const IVInc &X) { Incs.push_back(X); }
1607 // Returns the last UserInst in the chain.
1608 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1610 // Returns true if IncExpr can be profitably added to this chain.
1611 bool isProfitableIncrement(const SCEV *OperExpr,
1612 const SCEV *IncExpr,
1616 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1617 /// between FarUsers that definitely cross IV increments and NearUsers that may
1618 /// be used between IV increments.
1620 SmallPtrSet<Instruction*, 4> FarUsers;
1621 SmallPtrSet<Instruction*, 4> NearUsers;
1624 /// This class holds state for the main loop strength reduction logic.
1627 ScalarEvolution &SE;
1630 const TargetTransformInfo &TTI;
1634 /// This is the insert position that the current loop's induction variable
1635 /// increment should be placed. In simple loops, this is the latch block's
1636 /// terminator. But in more complicated cases, this is a position which will
1637 /// dominate all the in-loop post-increment users.
1638 Instruction *IVIncInsertPos;
1640 /// Interesting factors between use strides.
1641 SmallSetVector<int64_t, 8> Factors;
1643 /// Interesting use types, to facilitate truncation reuse.
1644 SmallSetVector<Type *, 4> Types;
1646 /// The list of operands which are to be replaced.
1647 SmallVector<LSRFixup, 16> Fixups;
1649 /// The list of interesting uses.
1650 SmallVector<LSRUse, 16> Uses;
1652 /// Track which uses use which register candidates.
1653 RegUseTracker RegUses;
1655 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1656 // have more than a few IV increment chains in a loop. Missing a Chain falls
1657 // back to normal LSR behavior for those uses.
1658 static const unsigned MaxChains = 8;
1660 /// IV users can form a chain of IV increments.
1661 SmallVector<IVChain, MaxChains> IVChainVec;
1663 /// IV users that belong to profitable IVChains.
1664 SmallPtrSet<Use*, MaxChains> IVIncSet;
1666 void OptimizeShadowIV();
1667 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1668 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1669 void OptimizeLoopTermCond();
1671 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1672 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1673 void FinalizeChain(IVChain &Chain);
1674 void CollectChains();
1675 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1676 SmallVectorImpl<WeakVH> &DeadInsts);
1678 void CollectInterestingTypesAndFactors();
1679 void CollectFixupsAndInitialFormulae();
1681 LSRFixup &getNewFixup() {
1682 Fixups.push_back(LSRFixup());
1683 return Fixups.back();
1686 // Support for sharing of LSRUses between LSRFixups.
1687 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1690 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1691 LSRUse::KindType Kind, MemAccessTy AccessTy);
1693 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1694 MemAccessTy AccessTy);
1696 void DeleteUse(LSRUse &LU, size_t LUIdx);
1698 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1700 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1701 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1702 void CountRegisters(const Formula &F, size_t LUIdx);
1703 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1705 void CollectLoopInvariantFixupsAndFormulae();
1707 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1708 unsigned Depth = 0);
1710 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1711 const Formula &Base, unsigned Depth,
1712 size_t Idx, bool IsScaledReg = false);
1713 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1714 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1715 const Formula &Base, size_t Idx,
1716 bool IsScaledReg = false);
1717 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1718 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1719 const Formula &Base,
1720 const SmallVectorImpl<int64_t> &Worklist,
1721 size_t Idx, bool IsScaledReg = false);
1722 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1723 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1724 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1725 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1726 void GenerateCrossUseConstantOffsets();
1727 void GenerateAllReuseFormulae();
1729 void FilterOutUndesirableDedicatedRegisters();
1731 size_t EstimateSearchSpaceComplexity() const;
1732 void NarrowSearchSpaceByDetectingSupersets();
1733 void NarrowSearchSpaceByCollapsingUnrolledCode();
1734 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1735 void NarrowSearchSpaceByPickingWinnerRegs();
1736 void NarrowSearchSpaceUsingHeuristics();
1738 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1740 SmallVectorImpl<const Formula *> &Workspace,
1741 const Cost &CurCost,
1742 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1743 DenseSet<const SCEV *> &VisitedRegs) const;
1744 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1746 BasicBlock::iterator
1747 HoistInsertPosition(BasicBlock::iterator IP,
1748 const SmallVectorImpl<Instruction *> &Inputs) const;
1749 BasicBlock::iterator
1750 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1753 SCEVExpander &Rewriter) const;
1755 Value *Expand(const LSRFixup &LF,
1757 BasicBlock::iterator IP,
1758 SCEVExpander &Rewriter,
1759 SmallVectorImpl<WeakVH> &DeadInsts) const;
1760 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1762 SCEVExpander &Rewriter,
1763 SmallVectorImpl<WeakVH> &DeadInsts) const;
1764 void Rewrite(const LSRFixup &LF,
1766 SCEVExpander &Rewriter,
1767 SmallVectorImpl<WeakVH> &DeadInsts) const;
1768 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
1771 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
1772 LoopInfo &LI, const TargetTransformInfo &TTI);
1774 bool getChanged() const { return Changed; }
1776 void print_factors_and_types(raw_ostream &OS) const;
1777 void print_fixups(raw_ostream &OS) const;
1778 void print_uses(raw_ostream &OS) const;
1779 void print(raw_ostream &OS) const;
1785 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
1786 /// the cast operation.
1787 void LSRInstance::OptimizeShadowIV() {
1788 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1789 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1792 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1793 UI != E; /* empty */) {
1794 IVUsers::const_iterator CandidateUI = UI;
1796 Instruction *ShadowUse = CandidateUI->getUser();
1797 Type *DestTy = nullptr;
1798 bool IsSigned = false;
1800 /* If shadow use is a int->float cast then insert a second IV
1801 to eliminate this cast.
1803 for (unsigned i = 0; i < n; ++i)
1809 for (unsigned i = 0; i < n; ++i, ++d)
1812 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1814 DestTy = UCast->getDestTy();
1816 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1818 DestTy = SCast->getDestTy();
1820 if (!DestTy) continue;
1822 // If target does not support DestTy natively then do not apply
1823 // this transformation.
1824 if (!TTI.isTypeLegal(DestTy)) continue;
1826 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1828 if (PH->getNumIncomingValues() != 2) continue;
1830 Type *SrcTy = PH->getType();
1831 int Mantissa = DestTy->getFPMantissaWidth();
1832 if (Mantissa == -1) continue;
1833 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1836 unsigned Entry, Latch;
1837 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1845 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1846 if (!Init) continue;
1847 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1848 (double)Init->getSExtValue() :
1849 (double)Init->getZExtValue());
1851 BinaryOperator *Incr =
1852 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1853 if (!Incr) continue;
1854 if (Incr->getOpcode() != Instruction::Add
1855 && Incr->getOpcode() != Instruction::Sub)
1858 /* Initialize new IV, double d = 0.0 in above example. */
1859 ConstantInt *C = nullptr;
1860 if (Incr->getOperand(0) == PH)
1861 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1862 else if (Incr->getOperand(1) == PH)
1863 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1869 // Ignore negative constants, as the code below doesn't handle them
1870 // correctly. TODO: Remove this restriction.
1871 if (!C->getValue().isStrictlyPositive()) continue;
1873 /* Add new PHINode. */
1874 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1876 /* create new increment. '++d' in above example. */
1877 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1878 BinaryOperator *NewIncr =
1879 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1880 Instruction::FAdd : Instruction::FSub,
1881 NewPH, CFP, "IV.S.next.", Incr);
1883 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1884 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1886 /* Remove cast operation */
1887 ShadowUse->replaceAllUsesWith(NewPH);
1888 ShadowUse->eraseFromParent();
1894 /// If Cond has an operand that is an expression of an IV, set the IV user and
1895 /// stride information and return true, otherwise return false.
1896 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1897 for (IVStrideUse &U : IU)
1898 if (U.getUser() == Cond) {
1899 // NOTE: we could handle setcc instructions with multiple uses here, but
1900 // InstCombine does it as well for simple uses, it's not clear that it
1901 // occurs enough in real life to handle.
1908 /// Rewrite the loop's terminating condition if it uses a max computation.
1910 /// This is a narrow solution to a specific, but acute, problem. For loops
1916 /// } while (++i < n);
1918 /// the trip count isn't just 'n', because 'n' might not be positive. And
1919 /// unfortunately this can come up even for loops where the user didn't use
1920 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1921 /// will commonly be lowered like this:
1927 /// } while (++i < n);
1930 /// and then it's possible for subsequent optimization to obscure the if
1931 /// test in such a way that indvars can't find it.
1933 /// When indvars can't find the if test in loops like this, it creates a
1934 /// max expression, which allows it to give the loop a canonical
1935 /// induction variable:
1938 /// max = n < 1 ? 1 : n;
1941 /// } while (++i != max);
1943 /// Canonical induction variables are necessary because the loop passes
1944 /// are designed around them. The most obvious example of this is the
1945 /// LoopInfo analysis, which doesn't remember trip count values. It
1946 /// expects to be able to rediscover the trip count each time it is
1947 /// needed, and it does this using a simple analysis that only succeeds if
1948 /// the loop has a canonical induction variable.
1950 /// However, when it comes time to generate code, the maximum operation
1951 /// can be quite costly, especially if it's inside of an outer loop.
1953 /// This function solves this problem by detecting this type of loop and
1954 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1955 /// the instructions for the maximum computation.
1957 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1958 // Check that the loop matches the pattern we're looking for.
1959 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1960 Cond->getPredicate() != CmpInst::ICMP_NE)
1963 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1964 if (!Sel || !Sel->hasOneUse()) return Cond;
1966 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1967 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1969 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1971 // Add one to the backedge-taken count to get the trip count.
1972 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1973 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1975 // Check for a max calculation that matches the pattern. There's no check
1976 // for ICMP_ULE here because the comparison would be with zero, which
1977 // isn't interesting.
1978 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1979 const SCEVNAryExpr *Max = nullptr;
1980 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1981 Pred = ICmpInst::ICMP_SLE;
1983 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1984 Pred = ICmpInst::ICMP_SLT;
1986 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1987 Pred = ICmpInst::ICMP_ULT;
1994 // To handle a max with more than two operands, this optimization would
1995 // require additional checking and setup.
1996 if (Max->getNumOperands() != 2)
1999 const SCEV *MaxLHS = Max->getOperand(0);
2000 const SCEV *MaxRHS = Max->getOperand(1);
2002 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2003 // for a comparison with 1. For <= and >=, a comparison with zero.
2005 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2008 // Check the relevant induction variable for conformance to
2010 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2011 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2012 if (!AR || !AR->isAffine() ||
2013 AR->getStart() != One ||
2014 AR->getStepRecurrence(SE) != One)
2017 assert(AR->getLoop() == L &&
2018 "Loop condition operand is an addrec in a different loop!");
2020 // Check the right operand of the select, and remember it, as it will
2021 // be used in the new comparison instruction.
2022 Value *NewRHS = nullptr;
2023 if (ICmpInst::isTrueWhenEqual(Pred)) {
2024 // Look for n+1, and grab n.
2025 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2026 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2027 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2028 NewRHS = BO->getOperand(0);
2029 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2030 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2031 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2032 NewRHS = BO->getOperand(0);
2035 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2036 NewRHS = Sel->getOperand(1);
2037 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2038 NewRHS = Sel->getOperand(2);
2039 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2040 NewRHS = SU->getValue();
2042 // Max doesn't match expected pattern.
2045 // Determine the new comparison opcode. It may be signed or unsigned,
2046 // and the original comparison may be either equality or inequality.
2047 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2048 Pred = CmpInst::getInversePredicate(Pred);
2050 // Ok, everything looks ok to change the condition into an SLT or SGE and
2051 // delete the max calculation.
2053 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2055 // Delete the max calculation instructions.
2056 Cond->replaceAllUsesWith(NewCond);
2057 CondUse->setUser(NewCond);
2058 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2059 Cond->eraseFromParent();
2060 Sel->eraseFromParent();
2061 if (Cmp->use_empty())
2062 Cmp->eraseFromParent();
2066 /// Change loop terminating condition to use the postinc iv when possible.
2068 LSRInstance::OptimizeLoopTermCond() {
2069 SmallPtrSet<Instruction *, 4> PostIncs;
2071 BasicBlock *LatchBlock = L->getLoopLatch();
2072 SmallVector<BasicBlock*, 8> ExitingBlocks;
2073 L->getExitingBlocks(ExitingBlocks);
2075 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2077 // Get the terminating condition for the loop if possible. If we
2078 // can, we want to change it to use a post-incremented version of its
2079 // induction variable, to allow coalescing the live ranges for the IV into
2080 // one register value.
2082 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2085 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2086 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2089 // Search IVUsesByStride to find Cond's IVUse if there is one.
2090 IVStrideUse *CondUse = nullptr;
2091 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2092 if (!FindIVUserForCond(Cond, CondUse))
2095 // If the trip count is computed in terms of a max (due to ScalarEvolution
2096 // being unable to find a sufficient guard, for example), change the loop
2097 // comparison to use SLT or ULT instead of NE.
2098 // One consequence of doing this now is that it disrupts the count-down
2099 // optimization. That's not always a bad thing though, because in such
2100 // cases it may still be worthwhile to avoid a max.
2101 Cond = OptimizeMax(Cond, CondUse);
2103 // If this exiting block dominates the latch block, it may also use
2104 // the post-inc value if it won't be shared with other uses.
2105 // Check for dominance.
2106 if (!DT.dominates(ExitingBlock, LatchBlock))
2109 // Conservatively avoid trying to use the post-inc value in non-latch
2110 // exits if there may be pre-inc users in intervening blocks.
2111 if (LatchBlock != ExitingBlock)
2112 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2113 // Test if the use is reachable from the exiting block. This dominator
2114 // query is a conservative approximation of reachability.
2115 if (&*UI != CondUse &&
2116 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2117 // Conservatively assume there may be reuse if the quotient of their
2118 // strides could be a legal scale.
2119 const SCEV *A = IU.getStride(*CondUse, L);
2120 const SCEV *B = IU.getStride(*UI, L);
2121 if (!A || !B) continue;
2122 if (SE.getTypeSizeInBits(A->getType()) !=
2123 SE.getTypeSizeInBits(B->getType())) {
2124 if (SE.getTypeSizeInBits(A->getType()) >
2125 SE.getTypeSizeInBits(B->getType()))
2126 B = SE.getSignExtendExpr(B, A->getType());
2128 A = SE.getSignExtendExpr(A, B->getType());
2130 if (const SCEVConstant *D =
2131 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2132 const ConstantInt *C = D->getValue();
2133 // Stride of one or negative one can have reuse with non-addresses.
2134 if (C->isOne() || C->isAllOnesValue())
2135 goto decline_post_inc;
2136 // Avoid weird situations.
2137 if (C->getValue().getMinSignedBits() >= 64 ||
2138 C->getValue().isMinSignedValue())
2139 goto decline_post_inc;
2140 // Check for possible scaled-address reuse.
2141 MemAccessTy AccessTy = getAccessType(UI->getUser());
2142 int64_t Scale = C->getSExtValue();
2143 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2145 /*HasBaseReg=*/false, Scale,
2146 AccessTy.AddrSpace))
2147 goto decline_post_inc;
2149 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2151 /*HasBaseReg=*/false, Scale,
2152 AccessTy.AddrSpace))
2153 goto decline_post_inc;
2157 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2160 // It's possible for the setcc instruction to be anywhere in the loop, and
2161 // possible for it to have multiple users. If it is not immediately before
2162 // the exiting block branch, move it.
2163 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2164 if (Cond->hasOneUse()) {
2165 Cond->moveBefore(TermBr);
2167 // Clone the terminating condition and insert into the loopend.
2168 ICmpInst *OldCond = Cond;
2169 Cond = cast<ICmpInst>(Cond->clone());
2170 Cond->setName(L->getHeader()->getName() + ".termcond");
2171 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2173 // Clone the IVUse, as the old use still exists!
2174 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2175 TermBr->replaceUsesOfWith(OldCond, Cond);
2179 // If we get to here, we know that we can transform the setcc instruction to
2180 // use the post-incremented version of the IV, allowing us to coalesce the
2181 // live ranges for the IV correctly.
2182 CondUse->transformToPostInc(L);
2185 PostIncs.insert(Cond);
2189 // Determine an insertion point for the loop induction variable increment. It
2190 // must dominate all the post-inc comparisons we just set up, and it must
2191 // dominate the loop latch edge.
2192 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2193 for (Instruction *Inst : PostIncs) {
2195 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2197 if (BB == Inst->getParent())
2198 IVIncInsertPos = Inst;
2199 else if (BB != IVIncInsertPos->getParent())
2200 IVIncInsertPos = BB->getTerminator();
2204 /// Determine if the given use can accommodate a fixup at the given offset and
2205 /// other details. If so, update the use and return true.
2206 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2207 bool HasBaseReg, LSRUse::KindType Kind,
2208 MemAccessTy AccessTy) {
2209 int64_t NewMinOffset = LU.MinOffset;
2210 int64_t NewMaxOffset = LU.MaxOffset;
2211 MemAccessTy NewAccessTy = AccessTy;
2213 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2214 // something conservative, however this can pessimize in the case that one of
2215 // the uses will have all its uses outside the loop, for example.
2216 if (LU.Kind != Kind)
2219 // Check for a mismatched access type, and fall back conservatively as needed.
2220 // TODO: Be less conservative when the type is similar and can use the same
2221 // addressing modes.
2222 if (Kind == LSRUse::Address) {
2223 if (AccessTy != LU.AccessTy)
2224 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext());
2227 // Conservatively assume HasBaseReg is true for now.
2228 if (NewOffset < LU.MinOffset) {
2229 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2230 LU.MaxOffset - NewOffset, HasBaseReg))
2232 NewMinOffset = NewOffset;
2233 } else if (NewOffset > LU.MaxOffset) {
2234 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2235 NewOffset - LU.MinOffset, HasBaseReg))
2237 NewMaxOffset = NewOffset;
2241 LU.MinOffset = NewMinOffset;
2242 LU.MaxOffset = NewMaxOffset;
2243 LU.AccessTy = NewAccessTy;
2244 if (NewOffset != LU.Offsets.back())
2245 LU.Offsets.push_back(NewOffset);
2249 /// Return an LSRUse index and an offset value for a fixup which needs the given
2250 /// expression, with the given kind and optional access type. Either reuse an
2251 /// existing use or create a new one, as needed.
2252 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2253 LSRUse::KindType Kind,
2254 MemAccessTy AccessTy) {
2255 const SCEV *Copy = Expr;
2256 int64_t Offset = ExtractImmediate(Expr, SE);
2258 // Basic uses can't accept any offset, for example.
2259 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2260 Offset, /*HasBaseReg=*/ true)) {
2265 std::pair<UseMapTy::iterator, bool> P =
2266 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2268 // A use already existed with this base.
2269 size_t LUIdx = P.first->second;
2270 LSRUse &LU = Uses[LUIdx];
2271 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2273 return std::make_pair(LUIdx, Offset);
2276 // Create a new use.
2277 size_t LUIdx = Uses.size();
2278 P.first->second = LUIdx;
2279 Uses.push_back(LSRUse(Kind, AccessTy));
2280 LSRUse &LU = Uses[LUIdx];
2282 // We don't need to track redundant offsets, but we don't need to go out
2283 // of our way here to avoid them.
2284 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2285 LU.Offsets.push_back(Offset);
2287 LU.MinOffset = Offset;
2288 LU.MaxOffset = Offset;
2289 return std::make_pair(LUIdx, Offset);
2292 /// Delete the given use from the Uses list.
2293 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2294 if (&LU != &Uses.back())
2295 std::swap(LU, Uses.back());
2299 RegUses.swapAndDropUse(LUIdx, Uses.size());
2302 /// Look for a use distinct from OrigLU which is has a formula that has the same
2303 /// registers as the given formula.
2305 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2306 const LSRUse &OrigLU) {
2307 // Search all uses for the formula. This could be more clever.
2308 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2309 LSRUse &LU = Uses[LUIdx];
2310 // Check whether this use is close enough to OrigLU, to see whether it's
2311 // worthwhile looking through its formulae.
2312 // Ignore ICmpZero uses because they may contain formulae generated by
2313 // GenerateICmpZeroScales, in which case adding fixup offsets may
2315 if (&LU != &OrigLU &&
2316 LU.Kind != LSRUse::ICmpZero &&
2317 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2318 LU.WidestFixupType == OrigLU.WidestFixupType &&
2319 LU.HasFormulaWithSameRegs(OrigF)) {
2320 // Scan through this use's formulae.
2321 for (const Formula &F : LU.Formulae) {
2322 // Check to see if this formula has the same registers and symbols
2324 if (F.BaseRegs == OrigF.BaseRegs &&
2325 F.ScaledReg == OrigF.ScaledReg &&
2326 F.BaseGV == OrigF.BaseGV &&
2327 F.Scale == OrigF.Scale &&
2328 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2329 if (F.BaseOffset == 0)
2331 // This is the formula where all the registers and symbols matched;
2332 // there aren't going to be any others. Since we declined it, we
2333 // can skip the rest of the formulae and proceed to the next LSRUse.
2340 // Nothing looked good.
2344 void LSRInstance::CollectInterestingTypesAndFactors() {
2345 SmallSetVector<const SCEV *, 4> Strides;
2347 // Collect interesting types and strides.
2348 SmallVector<const SCEV *, 4> Worklist;
2349 for (const IVStrideUse &U : IU) {
2350 const SCEV *Expr = IU.getExpr(U);
2352 // Collect interesting types.
2353 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2355 // Add strides for mentioned loops.
2356 Worklist.push_back(Expr);
2358 const SCEV *S = Worklist.pop_back_val();
2359 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2360 if (AR->getLoop() == L)
2361 Strides.insert(AR->getStepRecurrence(SE));
2362 Worklist.push_back(AR->getStart());
2363 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2364 Worklist.append(Add->op_begin(), Add->op_end());
2366 } while (!Worklist.empty());
2369 // Compute interesting factors from the set of interesting strides.
2370 for (SmallSetVector<const SCEV *, 4>::const_iterator
2371 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2372 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2373 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2374 const SCEV *OldStride = *I;
2375 const SCEV *NewStride = *NewStrideIter;
2377 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2378 SE.getTypeSizeInBits(NewStride->getType())) {
2379 if (SE.getTypeSizeInBits(OldStride->getType()) >
2380 SE.getTypeSizeInBits(NewStride->getType()))
2381 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2383 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2385 if (const SCEVConstant *Factor =
2386 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2388 if (Factor->getAPInt().getMinSignedBits() <= 64)
2389 Factors.insert(Factor->getAPInt().getSExtValue());
2390 } else if (const SCEVConstant *Factor =
2391 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2394 if (Factor->getAPInt().getMinSignedBits() <= 64)
2395 Factors.insert(Factor->getAPInt().getSExtValue());
2399 // If all uses use the same type, don't bother looking for truncation-based
2401 if (Types.size() == 1)
2404 DEBUG(print_factors_and_types(dbgs()));
2407 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2408 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2409 /// IVStrideUses, we could partially skip this.
2410 static User::op_iterator
2411 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2412 Loop *L, ScalarEvolution &SE) {
2413 for(; OI != OE; ++OI) {
2414 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2415 if (!SE.isSCEVable(Oper->getType()))
2418 if (const SCEVAddRecExpr *AR =
2419 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2420 if (AR->getLoop() == L)
2428 /// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2429 /// a convenient helper.
2430 static Value *getWideOperand(Value *Oper) {
2431 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2432 return Trunc->getOperand(0);
2436 /// Return true if we allow an IV chain to include both types.
2437 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2438 Type *LType = LVal->getType();
2439 Type *RType = RVal->getType();
2440 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2443 /// Return an approximation of this SCEV expression's "base", or NULL for any
2444 /// constant. Returning the expression itself is conservative. Returning a
2445 /// deeper subexpression is more precise and valid as long as it isn't less
2446 /// complex than another subexpression. For expressions involving multiple
2447 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2448 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2451 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2452 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2453 static const SCEV *getExprBase(const SCEV *S) {
2454 switch (S->getSCEVType()) {
2455 default: // uncluding scUnknown.
2460 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2462 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2464 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2466 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2467 // there's nothing more complex.
2468 // FIXME: not sure if we want to recognize negation.
2469 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2470 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2471 E(Add->op_begin()); I != E; ++I) {
2472 const SCEV *SubExpr = *I;
2473 if (SubExpr->getSCEVType() == scAddExpr)
2474 return getExprBase(SubExpr);
2476 if (SubExpr->getSCEVType() != scMulExpr)
2479 return S; // all operands are scaled, be conservative.
2482 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2486 /// Return true if the chain increment is profitable to expand into a loop
2487 /// invariant value, which may require its own register. A profitable chain
2488 /// increment will be an offset relative to the same base. We allow such offsets
2489 /// to potentially be used as chain increment as long as it's not obviously
2490 /// expensive to expand using real instructions.
2491 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2492 const SCEV *IncExpr,
2493 ScalarEvolution &SE) {
2494 // Aggressively form chains when -stress-ivchain.
2498 // Do not replace a constant offset from IV head with a nonconstant IV
2500 if (!isa<SCEVConstant>(IncExpr)) {
2501 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2502 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2506 SmallPtrSet<const SCEV*, 8> Processed;
2507 return !isHighCostExpansion(IncExpr, Processed, SE);
2510 /// Return true if the number of registers needed for the chain is estimated to
2511 /// be less than the number required for the individual IV users. First prohibit
2512 /// any IV users that keep the IV live across increments (the Users set should
2513 /// be empty). Next count the number and type of increments in the chain.
2515 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2516 /// effectively use postinc addressing modes. Only consider it profitable it the
2517 /// increments can be computed in fewer registers when chained.
2519 /// TODO: Consider IVInc free if it's already used in another chains.
2521 isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2522 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2526 if (!Chain.hasIncs())
2529 if (!Users.empty()) {
2530 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2531 for (Instruction *Inst : Users) {
2532 dbgs() << " " << *Inst << "\n";
2536 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2538 // The chain itself may require a register, so intialize cost to 1.
2541 // A complete chain likely eliminates the need for keeping the original IV in
2542 // a register. LSR does not currently know how to form a complete chain unless
2543 // the header phi already exists.
2544 if (isa<PHINode>(Chain.tailUserInst())
2545 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2548 const SCEV *LastIncExpr = nullptr;
2549 unsigned NumConstIncrements = 0;
2550 unsigned NumVarIncrements = 0;
2551 unsigned NumReusedIncrements = 0;
2552 for (const IVInc &Inc : Chain) {
2553 if (Inc.IncExpr->isZero())
2556 // Incrementing by zero or some constant is neutral. We assume constants can
2557 // be folded into an addressing mode or an add's immediate operand.
2558 if (isa<SCEVConstant>(Inc.IncExpr)) {
2559 ++NumConstIncrements;
2563 if (Inc.IncExpr == LastIncExpr)
2564 ++NumReusedIncrements;
2568 LastIncExpr = Inc.IncExpr;
2570 // An IV chain with a single increment is handled by LSR's postinc
2571 // uses. However, a chain with multiple increments requires keeping the IV's
2572 // value live longer than it needs to be if chained.
2573 if (NumConstIncrements > 1)
2576 // Materializing increment expressions in the preheader that didn't exist in
2577 // the original code may cost a register. For example, sign-extended array
2578 // indices can produce ridiculous increments like this:
2579 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2580 cost += NumVarIncrements;
2582 // Reusing variable increments likely saves a register to hold the multiple of
2584 cost -= NumReusedIncrements;
2586 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2592 /// Add this IV user to an existing chain or make it the head of a new chain.
2593 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2594 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2595 // When IVs are used as types of varying widths, they are generally converted
2596 // to a wider type with some uses remaining narrow under a (free) trunc.
2597 Value *const NextIV = getWideOperand(IVOper);
2598 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2599 const SCEV *const OperExprBase = getExprBase(OperExpr);
2601 // Visit all existing chains. Check if its IVOper can be computed as a
2602 // profitable loop invariant increment from the last link in the Chain.
2603 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2604 const SCEV *LastIncExpr = nullptr;
2605 for (; ChainIdx < NChains; ++ChainIdx) {
2606 IVChain &Chain = IVChainVec[ChainIdx];
2608 // Prune the solution space aggressively by checking that both IV operands
2609 // are expressions that operate on the same unscaled SCEVUnknown. This
2610 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2611 // first avoids creating extra SCEV expressions.
2612 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2615 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2616 if (!isCompatibleIVType(PrevIV, NextIV))
2619 // A phi node terminates a chain.
2620 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2623 // The increment must be loop-invariant so it can be kept in a register.
2624 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2625 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2626 if (!SE.isLoopInvariant(IncExpr, L))
2629 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2630 LastIncExpr = IncExpr;
2634 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2635 // bother for phi nodes, because they must be last in the chain.
2636 if (ChainIdx == NChains) {
2637 if (isa<PHINode>(UserInst))
2639 if (NChains >= MaxChains && !StressIVChain) {
2640 DEBUG(dbgs() << "IV Chain Limit\n");
2643 LastIncExpr = OperExpr;
2644 // IVUsers may have skipped over sign/zero extensions. We don't currently
2645 // attempt to form chains involving extensions unless they can be hoisted
2646 // into this loop's AddRec.
2647 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2650 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2652 ChainUsersVec.resize(NChains);
2653 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2654 << ") IV=" << *LastIncExpr << "\n");
2656 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2657 << ") IV+" << *LastIncExpr << "\n");
2658 // Add this IV user to the end of the chain.
2659 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2661 IVChain &Chain = IVChainVec[ChainIdx];
2663 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2664 // This chain's NearUsers become FarUsers.
2665 if (!LastIncExpr->isZero()) {
2666 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2671 // All other uses of IVOperand become near uses of the chain.
2672 // We currently ignore intermediate values within SCEV expressions, assuming
2673 // they will eventually be used be the current chain, or can be computed
2674 // from one of the chain increments. To be more precise we could
2675 // transitively follow its user and only add leaf IV users to the set.
2676 for (User *U : IVOper->users()) {
2677 Instruction *OtherUse = dyn_cast<Instruction>(U);
2680 // Uses in the chain will no longer be uses if the chain is formed.
2681 // Include the head of the chain in this iteration (not Chain.begin()).
2682 IVChain::const_iterator IncIter = Chain.Incs.begin();
2683 IVChain::const_iterator IncEnd = Chain.Incs.end();
2684 for( ; IncIter != IncEnd; ++IncIter) {
2685 if (IncIter->UserInst == OtherUse)
2688 if (IncIter != IncEnd)
2691 if (SE.isSCEVable(OtherUse->getType())
2692 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2693 && IU.isIVUserOrOperand(OtherUse)) {
2696 NearUsers.insert(OtherUse);
2699 // Since this user is part of the chain, it's no longer considered a use
2701 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2704 /// Populate the vector of Chains.
2706 /// This decreases ILP at the architecture level. Targets with ample registers,
2707 /// multiple memory ports, and no register renaming probably don't want
2708 /// this. However, such targets should probably disable LSR altogether.
2710 /// The job of LSR is to make a reasonable choice of induction variables across
2711 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2712 /// ILP *within the loop* if the target wants it.
2714 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2715 /// will not reorder memory operations, it will recognize this as a chain, but
2716 /// will generate redundant IV increments. Ideally this would be corrected later
2717 /// by a smart scheduler:
2723 /// TODO: Walk the entire domtree within this loop, not just the path to the
2724 /// loop latch. This will discover chains on side paths, but requires
2725 /// maintaining multiple copies of the Chains state.
2726 void LSRInstance::CollectChains() {
2727 DEBUG(dbgs() << "Collecting IV Chains.\n");
2728 SmallVector<ChainUsers, 8> ChainUsersVec;
2730 SmallVector<BasicBlock *,8> LatchPath;
2731 BasicBlock *LoopHeader = L->getHeader();
2732 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2733 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2734 LatchPath.push_back(Rung->getBlock());
2736 LatchPath.push_back(LoopHeader);
2738 // Walk the instruction stream from the loop header to the loop latch.
2739 for (BasicBlock *BB : reverse(LatchPath)) {
2740 for (Instruction &I : *BB) {
2741 // Skip instructions that weren't seen by IVUsers analysis.
2742 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
2745 // Ignore users that are part of a SCEV expression. This way we only
2746 // consider leaf IV Users. This effectively rediscovers a portion of
2747 // IVUsers analysis but in program order this time.
2748 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
2751 // Remove this instruction from any NearUsers set it may be in.
2752 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2753 ChainIdx < NChains; ++ChainIdx) {
2754 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
2756 // Search for operands that can be chained.
2757 SmallPtrSet<Instruction*, 4> UniqueOperands;
2758 User::op_iterator IVOpEnd = I.op_end();
2759 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
2760 while (IVOpIter != IVOpEnd) {
2761 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2762 if (UniqueOperands.insert(IVOpInst).second)
2763 ChainInstruction(&I, IVOpInst, ChainUsersVec);
2764 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2766 } // Continue walking down the instructions.
2767 } // Continue walking down the domtree.
2768 // Visit phi backedges to determine if the chain can generate the IV postinc.
2769 for (BasicBlock::iterator I = L->getHeader()->begin();
2770 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2771 if (!SE.isSCEVable(PN->getType()))
2775 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2777 ChainInstruction(PN, IncV, ChainUsersVec);
2779 // Remove any unprofitable chains.
2780 unsigned ChainIdx = 0;
2781 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2782 UsersIdx < NChains; ++UsersIdx) {
2783 if (!isProfitableChain(IVChainVec[UsersIdx],
2784 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2786 // Preserve the chain at UsesIdx.
2787 if (ChainIdx != UsersIdx)
2788 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2789 FinalizeChain(IVChainVec[ChainIdx]);
2792 IVChainVec.resize(ChainIdx);
2795 void LSRInstance::FinalizeChain(IVChain &Chain) {
2796 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2797 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2799 for (const IVInc &Inc : Chain) {
2800 DEBUG(dbgs() << " Inc: " << Inc.UserInst << "\n");
2801 auto UseI = std::find(Inc.UserInst->op_begin(), Inc.UserInst->op_end(),
2803 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
2804 IVIncSet.insert(UseI);
2808 /// Return true if the IVInc can be folded into an addressing mode.
2809 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2810 Value *Operand, const TargetTransformInfo &TTI) {
2811 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2812 if (!IncConst || !isAddressUse(UserInst, Operand))
2815 if (IncConst->getAPInt().getMinSignedBits() > 64)
2818 MemAccessTy AccessTy = getAccessType(UserInst);
2819 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2820 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
2821 IncOffset, /*HaseBaseReg=*/false))
2827 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
2828 /// user's operand from the previous IV user's operand.
2829 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2830 SmallVectorImpl<WeakVH> &DeadInsts) {
2831 // Find the new IVOperand for the head of the chain. It may have been replaced
2833 const IVInc &Head = Chain.Incs[0];
2834 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2835 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2836 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2838 Value *IVSrc = nullptr;
2839 while (IVOpIter != IVOpEnd) {
2840 IVSrc = getWideOperand(*IVOpIter);
2842 // If this operand computes the expression that the chain needs, we may use
2843 // it. (Check this after setting IVSrc which is used below.)
2845 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2846 // narrow for the chain, so we can no longer use it. We do allow using a
2847 // wider phi, assuming the LSR checked for free truncation. In that case we
2848 // should already have a truncate on this operand such that
2849 // getSCEV(IVSrc) == IncExpr.
2850 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2851 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2854 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2856 if (IVOpIter == IVOpEnd) {
2857 // Gracefully give up on this chain.
2858 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2862 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2863 Type *IVTy = IVSrc->getType();
2864 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2865 const SCEV *LeftOverExpr = nullptr;
2866 for (const IVInc &Inc : Chain) {
2867 Instruction *InsertPt = Inc.UserInst;
2868 if (isa<PHINode>(InsertPt))
2869 InsertPt = L->getLoopLatch()->getTerminator();
2871 // IVOper will replace the current IV User's operand. IVSrc is the IV
2872 // value currently held in a register.
2873 Value *IVOper = IVSrc;
2874 if (!Inc.IncExpr->isZero()) {
2875 // IncExpr was the result of subtraction of two narrow values, so must
2877 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
2878 LeftOverExpr = LeftOverExpr ?
2879 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2881 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2882 // Expand the IV increment.
2883 Rewriter.clearPostInc();
2884 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2885 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2886 SE.getUnknown(IncV));
2887 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2889 // If an IV increment can't be folded, use it as the next IV value.
2890 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
2891 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2893 LeftOverExpr = nullptr;
2896 Type *OperTy = Inc.IVOperand->getType();
2897 if (IVTy != OperTy) {
2898 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2899 "cannot extend a chained IV");
2900 IRBuilder<> Builder(InsertPt);
2901 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2903 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
2904 DeadInsts.emplace_back(Inc.IVOperand);
2906 // If LSR created a new, wider phi, we may also replace its postinc. We only
2907 // do this if we also found a wide value for the head of the chain.
2908 if (isa<PHINode>(Chain.tailUserInst())) {
2909 for (BasicBlock::iterator I = L->getHeader()->begin();
2910 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2911 if (!isCompatibleIVType(Phi, IVSrc))
2913 Instruction *PostIncV = dyn_cast<Instruction>(
2914 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2915 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2917 Value *IVOper = IVSrc;
2918 Type *PostIncTy = PostIncV->getType();
2919 if (IVTy != PostIncTy) {
2920 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2921 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2922 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2923 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2925 Phi->replaceUsesOfWith(PostIncV, IVOper);
2926 DeadInsts.emplace_back(PostIncV);
2931 void LSRInstance::CollectFixupsAndInitialFormulae() {
2932 for (const IVStrideUse &U : IU) {
2933 Instruction *UserInst = U.getUser();
2934 // Skip IV users that are part of profitable IV Chains.
2935 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2936 U.getOperandValToReplace());
2937 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2938 if (IVIncSet.count(UseI))
2942 LSRFixup &LF = getNewFixup();
2943 LF.UserInst = UserInst;
2944 LF.OperandValToReplace = U.getOperandValToReplace();
2945 LF.PostIncLoops = U.getPostIncLoops();
2947 LSRUse::KindType Kind = LSRUse::Basic;
2948 MemAccessTy AccessTy;
2949 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2950 Kind = LSRUse::Address;
2951 AccessTy = getAccessType(LF.UserInst);
2954 const SCEV *S = IU.getExpr(U);
2956 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2957 // (N - i == 0), and this allows (N - i) to be the expression that we work
2958 // with rather than just N or i, so we can consider the register
2959 // requirements for both N and i at the same time. Limiting this code to
2960 // equality icmps is not a problem because all interesting loops use
2961 // equality icmps, thanks to IndVarSimplify.
2962 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2963 if (CI->isEquality()) {
2964 // Swap the operands if needed to put the OperandValToReplace on the
2965 // left, for consistency.
2966 Value *NV = CI->getOperand(1);
2967 if (NV == LF.OperandValToReplace) {
2968 CI->setOperand(1, CI->getOperand(0));
2969 CI->setOperand(0, NV);
2970 NV = CI->getOperand(1);
2974 // x == y --> x - y == 0
2975 const SCEV *N = SE.getSCEV(NV);
2976 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2977 // S is normalized, so normalize N before folding it into S
2978 // to keep the result normalized.
2979 N = TransformForPostIncUse(Normalize, N, CI, nullptr,
2980 LF.PostIncLoops, SE, DT);
2981 Kind = LSRUse::ICmpZero;
2982 S = SE.getMinusSCEV(N, S);
2985 // -1 and the negations of all interesting strides (except the negation
2986 // of -1) are now also interesting.
2987 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2988 if (Factors[i] != -1)
2989 Factors.insert(-(uint64_t)Factors[i]);
2993 // Set up the initial formula for this use.
2994 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2996 LF.Offset = P.second;
2997 LSRUse &LU = Uses[LF.LUIdx];
2998 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2999 if (!LU.WidestFixupType ||
3000 SE.getTypeSizeInBits(LU.WidestFixupType) <
3001 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3002 LU.WidestFixupType = LF.OperandValToReplace->getType();
3004 // If this is the first use of this LSRUse, give it a formula.
3005 if (LU.Formulae.empty()) {
3006 InsertInitialFormula(S, LU, LF.LUIdx);
3007 CountRegisters(LU.Formulae.back(), LF.LUIdx);
3011 DEBUG(print_fixups(dbgs()));
3014 /// Insert a formula for the given expression into the given use, separating out
3015 /// loop-variant portions from loop-invariant and loop-computable portions.
3017 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3018 // Mark uses whose expressions cannot be expanded.
3019 if (!isSafeToExpand(S, SE))
3020 LU.RigidFormula = true;
3023 F.initialMatch(S, L, SE);
3024 bool Inserted = InsertFormula(LU, LUIdx, F);
3025 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3028 /// Insert a simple single-register formula for the given expression into the
3031 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3032 LSRUse &LU, size_t LUIdx) {
3034 F.BaseRegs.push_back(S);
3035 F.HasBaseReg = true;
3036 bool Inserted = InsertFormula(LU, LUIdx, F);
3037 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3040 /// Note which registers are used by the given formula, updating RegUses.
3041 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3043 RegUses.countRegister(F.ScaledReg, LUIdx);
3044 for (const SCEV *BaseReg : F.BaseRegs)
3045 RegUses.countRegister(BaseReg, LUIdx);
3048 /// If the given formula has not yet been inserted, add it to the list, and
3049 /// return true. Return false otherwise.
3050 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3051 // Do not insert formula that we will not be able to expand.
3052 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3053 "Formula is illegal");
3054 if (!LU.InsertFormula(F))
3057 CountRegisters(F, LUIdx);
3061 /// Check for other uses of loop-invariant values which we're tracking. These
3062 /// other uses will pin these values in registers, making them less profitable
3063 /// for elimination.
3064 /// TODO: This currently misses non-constant addrec step registers.
3065 /// TODO: Should this give more weight to users inside the loop?
3067 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3068 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3069 SmallPtrSet<const SCEV *, 32> Visited;
3071 while (!Worklist.empty()) {
3072 const SCEV *S = Worklist.pop_back_val();
3074 // Don't process the same SCEV twice
3075 if (!Visited.insert(S).second)
3078 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3079 Worklist.append(N->op_begin(), N->op_end());
3080 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3081 Worklist.push_back(C->getOperand());
3082 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3083 Worklist.push_back(D->getLHS());
3084 Worklist.push_back(D->getRHS());
3085 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3086 const Value *V = US->getValue();
3087 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3088 // Look for instructions defined outside the loop.
3089 if (L->contains(Inst)) continue;
3090 } else if (isa<UndefValue>(V))
3091 // Undef doesn't have a live range, so it doesn't matter.
3093 for (const Use &U : V->uses()) {
3094 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3095 // Ignore non-instructions.
3098 // Ignore instructions in other functions (as can happen with
3100 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3102 // Ignore instructions not dominated by the loop.
3103 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3104 UserInst->getParent() :
3105 cast<PHINode>(UserInst)->getIncomingBlock(
3106 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3107 if (!DT.dominates(L->getHeader(), UseBB))
3109 // Don't bother if the instruction is in a BB which ends in an EHPad.
3110 if (UseBB->getTerminator()->isEHPad())
3112 // Ignore uses which are part of other SCEV expressions, to avoid
3113 // analyzing them multiple times.
3114 if (SE.isSCEVable(UserInst->getType())) {
3115 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3116 // If the user is a no-op, look through to its uses.
3117 if (!isa<SCEVUnknown>(UserS))
3121 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3125 // Ignore icmp instructions which are already being analyzed.
3126 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3127 unsigned OtherIdx = !U.getOperandNo();
3128 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3129 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3133 LSRFixup &LF = getNewFixup();
3134 LF.UserInst = const_cast<Instruction *>(UserInst);
3135 LF.OperandValToReplace = U;
3136 std::pair<size_t, int64_t> P = getUse(
3137 S, LSRUse::Basic, MemAccessTy());
3139 LF.Offset = P.second;
3140 LSRUse &LU = Uses[LF.LUIdx];
3141 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3142 if (!LU.WidestFixupType ||
3143 SE.getTypeSizeInBits(LU.WidestFixupType) <
3144 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3145 LU.WidestFixupType = LF.OperandValToReplace->getType();
3146 InsertSupplementalFormula(US, LU, LF.LUIdx);
3147 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3154 /// Split S into subexpressions which can be pulled out into separate
3155 /// registers. If C is non-null, multiply each subexpression by C.
3157 /// Return remainder expression after factoring the subexpressions captured by
3158 /// Ops. If Ops is complete, return NULL.
3159 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3160 SmallVectorImpl<const SCEV *> &Ops,
3162 ScalarEvolution &SE,
3163 unsigned Depth = 0) {
3164 // Arbitrarily cap recursion to protect compile time.
3168 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3169 // Break out add operands.
3170 for (const SCEV *S : Add->operands()) {
3171 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3173 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3176 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3177 // Split a non-zero base out of an addrec.
3178 if (AR->getStart()->isZero())
3181 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3182 C, Ops, L, SE, Depth+1);
3183 // Split the non-zero AddRec unless it is part of a nested recurrence that
3184 // does not pertain to this loop.
3185 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3186 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3187 Remainder = nullptr;
3189 if (Remainder != AR->getStart()) {
3191 Remainder = SE.getConstant(AR->getType(), 0);
3192 return SE.getAddRecExpr(Remainder,
3193 AR->getStepRecurrence(SE),
3195 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3198 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3199 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3200 if (Mul->getNumOperands() != 2)
3202 if (const SCEVConstant *Op0 =
3203 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3204 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3205 const SCEV *Remainder =
3206 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3208 Ops.push_back(SE.getMulExpr(C, Remainder));
3215 /// \brief Helper function for LSRInstance::GenerateReassociations.
3216 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3217 const Formula &Base,
3218 unsigned Depth, size_t Idx,
3220 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3221 SmallVector<const SCEV *, 8> AddOps;
3222 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3224 AddOps.push_back(Remainder);
3226 if (AddOps.size() == 1)
3229 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3233 // Loop-variant "unknown" values are uninteresting; we won't be able to
3234 // do anything meaningful with them.
3235 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3238 // Don't pull a constant into a register if the constant could be folded
3239 // into an immediate field.
3240 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3241 LU.AccessTy, *J, Base.getNumRegs() > 1))
3244 // Collect all operands except *J.
3245 SmallVector<const SCEV *, 8> InnerAddOps(
3246 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3247 InnerAddOps.append(std::next(J),
3248 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3250 // Don't leave just a constant behind in a register if the constant could
3251 // be folded into an immediate field.
3252 if (InnerAddOps.size() == 1 &&
3253 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3254 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3257 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3258 if (InnerSum->isZero())
3262 // Add the remaining pieces of the add back into the new formula.
3263 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3264 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3265 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3266 InnerSumSC->getValue()->getZExtValue())) {
3268 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3270 F.ScaledReg = nullptr;
3272 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3273 } else if (IsScaledReg)
3274 F.ScaledReg = InnerSum;
3276 F.BaseRegs[Idx] = InnerSum;
3278 // Add J as its own register, or an unfolded immediate.
3279 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3280 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3281 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3282 SC->getValue()->getZExtValue()))
3284 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3286 F.BaseRegs.push_back(*J);
3287 // We may have changed the number of register in base regs, adjust the
3288 // formula accordingly.
3291 if (InsertFormula(LU, LUIdx, F))
3292 // If that formula hadn't been seen before, recurse to find more like
3294 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3298 /// Split out subexpressions from adds and the bases of addrecs.
3299 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3300 Formula Base, unsigned Depth) {
3301 assert(Base.isCanonical() && "Input must be in the canonical form");
3302 // Arbitrarily cap recursion to protect compile time.
3306 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3307 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3309 if (Base.Scale == 1)
3310 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3311 /* Idx */ -1, /* IsScaledReg */ true);
3314 /// Generate a formula consisting of all of the loop-dominating registers added
3315 /// into a single register.
3316 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3318 // This method is only interesting on a plurality of registers.
3319 if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3322 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3323 // processing the formula.
3327 SmallVector<const SCEV *, 4> Ops;
3328 for (const SCEV *BaseReg : Base.BaseRegs) {
3329 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3330 !SE.hasComputableLoopEvolution(BaseReg, L))
3331 Ops.push_back(BaseReg);
3333 F.BaseRegs.push_back(BaseReg);
3335 if (Ops.size() > 1) {
3336 const SCEV *Sum = SE.getAddExpr(Ops);
3337 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3338 // opportunity to fold something. For now, just ignore such cases
3339 // rather than proceed with zero in a register.
3340 if (!Sum->isZero()) {
3341 F.BaseRegs.push_back(Sum);
3343 (void)InsertFormula(LU, LUIdx, F);
3348 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3349 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3350 const Formula &Base, size_t Idx,
3352 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3353 GlobalValue *GV = ExtractSymbol(G, SE);
3354 if (G->isZero() || !GV)
3358 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3363 F.BaseRegs[Idx] = G;
3364 (void)InsertFormula(LU, LUIdx, F);
3367 /// Generate reuse formulae using symbolic offsets.
3368 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3370 // We can't add a symbolic offset if the address already contains one.
3371 if (Base.BaseGV) return;
3373 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3374 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3375 if (Base.Scale == 1)
3376 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3377 /* IsScaledReg */ true);
3380 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3381 void LSRInstance::GenerateConstantOffsetsImpl(
3382 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3383 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3384 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3385 for (int64_t Offset : Worklist) {
3387 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3388 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3390 // Add the offset to the base register.
3391 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3392 // If it cancelled out, drop the base register, otherwise update it.
3393 if (NewG->isZero()) {
3396 F.ScaledReg = nullptr;
3398 F.deleteBaseReg(F.BaseRegs[Idx]);
3400 } else if (IsScaledReg)
3403 F.BaseRegs[Idx] = NewG;
3405 (void)InsertFormula(LU, LUIdx, F);
3409 int64_t Imm = ExtractImmediate(G, SE);
3410 if (G->isZero() || Imm == 0)
3413 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3414 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3419 F.BaseRegs[Idx] = G;
3420 (void)InsertFormula(LU, LUIdx, F);
3423 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3424 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3426 // TODO: For now, just add the min and max offset, because it usually isn't
3427 // worthwhile looking at everything inbetween.
3428 SmallVector<int64_t, 2> Worklist;
3429 Worklist.push_back(LU.MinOffset);
3430 if (LU.MaxOffset != LU.MinOffset)
3431 Worklist.push_back(LU.MaxOffset);
3433 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3434 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3435 if (Base.Scale == 1)
3436 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3437 /* IsScaledReg */ true);
3440 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3441 /// == y -> x*c == y*c.
3442 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3444 if (LU.Kind != LSRUse::ICmpZero) return;
3446 // Determine the integer type for the base formula.
3447 Type *IntTy = Base.getType();
3449 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3451 // Don't do this if there is more than one offset.
3452 if (LU.MinOffset != LU.MaxOffset) return;
3454 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3456 // Check each interesting stride.
3457 for (int64_t Factor : Factors) {
3458 // Check that the multiplication doesn't overflow.
3459 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3461 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3462 if (NewBaseOffset / Factor != Base.BaseOffset)
3464 // If the offset will be truncated at this use, check that it is in bounds.
3465 if (!IntTy->isPointerTy() &&
3466 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3469 // Check that multiplying with the use offset doesn't overflow.
3470 int64_t Offset = LU.MinOffset;
3471 if (Offset == INT64_MIN && Factor == -1)
3473 Offset = (uint64_t)Offset * Factor;
3474 if (Offset / Factor != LU.MinOffset)
3476 // If the offset will be truncated at this use, check that it is in bounds.
3477 if (!IntTy->isPointerTy() &&
3478 !ConstantInt::isValueValidForType(IntTy, Offset))
3482 F.BaseOffset = NewBaseOffset;
3484 // Check that this scale is legal.
3485 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3488 // Compensate for the use having MinOffset built into it.
3489 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3491 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3493 // Check that multiplying with each base register doesn't overflow.
3494 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3495 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3496 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3500 // Check that multiplying with the scaled register doesn't overflow.
3502 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3503 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3507 // Check that multiplying with the unfolded offset doesn't overflow.
3508 if (F.UnfoldedOffset != 0) {
3509 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3511 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3512 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3514 // If the offset will be truncated, check that it is in bounds.
3515 if (!IntTy->isPointerTy() &&
3516 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3520 // If we make it here and it's legal, add it.
3521 (void)InsertFormula(LU, LUIdx, F);
3526 /// Generate stride factor reuse formulae by making use of scaled-offset address
3527 /// modes, for example.
3528 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3529 // Determine the integer type for the base formula.
3530 Type *IntTy = Base.getType();
3533 // If this Formula already has a scaled register, we can't add another one.
3534 // Try to unscale the formula to generate a better scale.
3535 if (Base.Scale != 0 && !Base.unscale())
3538 assert(Base.Scale == 0 && "unscale did not did its job!");
3540 // Check each interesting stride.
3541 for (int64_t Factor : Factors) {
3542 Base.Scale = Factor;
3543 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3544 // Check whether this scale is going to be legal.
3545 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3547 // As a special-case, handle special out-of-loop Basic users specially.
3548 // TODO: Reconsider this special case.
3549 if (LU.Kind == LSRUse::Basic &&
3550 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3551 LU.AccessTy, Base) &&
3552 LU.AllFixupsOutsideLoop)
3553 LU.Kind = LSRUse::Special;
3557 // For an ICmpZero, negating a solitary base register won't lead to
3559 if (LU.Kind == LSRUse::ICmpZero &&
3560 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3562 // For each addrec base reg, apply the scale, if possible.
3563 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3564 if (const SCEVAddRecExpr *AR =
3565 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3566 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3567 if (FactorS->isZero())
3569 // Divide out the factor, ignoring high bits, since we'll be
3570 // scaling the value back up in the end.
3571 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3572 // TODO: This could be optimized to avoid all the copying.
3574 F.ScaledReg = Quotient;
3575 F.deleteBaseReg(F.BaseRegs[i]);
3576 // The canonical representation of 1*reg is reg, which is already in
3577 // Base. In that case, do not try to insert the formula, it will be
3579 if (F.Scale == 1 && F.BaseRegs.empty())
3581 (void)InsertFormula(LU, LUIdx, F);
3587 /// Generate reuse formulae from different IV types.
3588 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3589 // Don't bother truncating symbolic values.
3590 if (Base.BaseGV) return;
3592 // Determine the integer type for the base formula.
3593 Type *DstTy = Base.getType();
3595 DstTy = SE.getEffectiveSCEVType(DstTy);
3597 for (Type *SrcTy : Types) {
3598 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3601 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3602 for (const SCEV *&BaseReg : F.BaseRegs)
3603 BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3605 // TODO: This assumes we've done basic processing on all uses and
3606 // have an idea what the register usage is.
3607 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3610 (void)InsertFormula(LU, LUIdx, F);
3617 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3618 /// modifications so that the search phase doesn't have to worry about the data
3619 /// structures moving underneath it.
3623 const SCEV *OrigReg;
3625 WorkItem(size_t LI, int64_t I, const SCEV *R)
3626 : LUIdx(LI), Imm(I), OrigReg(R) {}
3628 void print(raw_ostream &OS) const;
3634 void WorkItem::print(raw_ostream &OS) const {
3635 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3636 << " , add offset " << Imm;
3640 void WorkItem::dump() const {
3641 print(errs()); errs() << '\n';
3644 /// Look for registers which are a constant distance apart and try to form reuse
3645 /// opportunities between them.
3646 void LSRInstance::GenerateCrossUseConstantOffsets() {
3647 // Group the registers by their value without any added constant offset.
3648 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3649 DenseMap<const SCEV *, ImmMapTy> Map;
3650 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3651 SmallVector<const SCEV *, 8> Sequence;
3652 for (const SCEV *Use : RegUses) {
3653 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
3654 int64_t Imm = ExtractImmediate(Reg, SE);
3655 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
3657 Sequence.push_back(Reg);
3658 Pair.first->second.insert(std::make_pair(Imm, Use));
3659 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
3662 // Now examine each set of registers with the same base value. Build up
3663 // a list of work to do and do the work in a separate step so that we're
3664 // not adding formulae and register counts while we're searching.
3665 SmallVector<WorkItem, 32> WorkItems;
3666 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3667 for (const SCEV *Reg : Sequence) {
3668 const ImmMapTy &Imms = Map.find(Reg)->second;
3670 // It's not worthwhile looking for reuse if there's only one offset.
3671 if (Imms.size() == 1)
3674 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3675 for (const auto &Entry : Imms)
3676 dbgs() << ' ' << Entry.first;
3679 // Examine each offset.
3680 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3682 const SCEV *OrigReg = J->second;
3684 int64_t JImm = J->first;
3685 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3687 if (!isa<SCEVConstant>(OrigReg) &&
3688 UsedByIndicesMap[Reg].count() == 1) {
3689 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3693 // Conservatively examine offsets between this orig reg a few selected
3695 ImmMapTy::const_iterator OtherImms[] = {
3696 Imms.begin(), std::prev(Imms.end()),
3697 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3700 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3701 ImmMapTy::const_iterator M = OtherImms[i];
3702 if (M == J || M == JE) continue;
3704 // Compute the difference between the two.
3705 int64_t Imm = (uint64_t)JImm - M->first;
3706 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3707 LUIdx = UsedByIndices.find_next(LUIdx))
3708 // Make a memo of this use, offset, and register tuple.
3709 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3710 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3717 UsedByIndicesMap.clear();
3718 UniqueItems.clear();
3720 // Now iterate through the worklist and add new formulae.
3721 for (const WorkItem &WI : WorkItems) {
3722 size_t LUIdx = WI.LUIdx;
3723 LSRUse &LU = Uses[LUIdx];
3724 int64_t Imm = WI.Imm;
3725 const SCEV *OrigReg = WI.OrigReg;
3727 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3728 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3729 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3731 // TODO: Use a more targeted data structure.
3732 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3733 Formula F = LU.Formulae[L];
3734 // FIXME: The code for the scaled and unscaled registers looks
3735 // very similar but slightly different. Investigate if they
3736 // could be merged. That way, we would not have to unscale the
3739 // Use the immediate in the scaled register.
3740 if (F.ScaledReg == OrigReg) {
3741 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3742 // Don't create 50 + reg(-50).
3743 if (F.referencesReg(SE.getSCEV(
3744 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3747 NewF.BaseOffset = Offset;
3748 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3751 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3753 // If the new scale is a constant in a register, and adding the constant
3754 // value to the immediate would produce a value closer to zero than the
3755 // immediate itself, then the formula isn't worthwhile.
3756 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3757 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
3758 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
3759 .ule(std::abs(NewF.BaseOffset)))
3763 NewF.canonicalize();
3764 (void)InsertFormula(LU, LUIdx, NewF);
3766 // Use the immediate in a base register.
3767 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3768 const SCEV *BaseReg = F.BaseRegs[N];
3769 if (BaseReg != OrigReg)
3772 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3773 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3774 LU.Kind, LU.AccessTy, NewF)) {
3775 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3778 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3780 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3782 // If the new formula has a constant in a register, and adding the
3783 // constant value to the immediate would produce a value closer to
3784 // zero than the immediate itself, then the formula isn't worthwhile.
3785 for (const SCEV *NewReg : NewF.BaseRegs)
3786 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
3787 if ((C->getAPInt() + NewF.BaseOffset)
3789 .slt(std::abs(NewF.BaseOffset)) &&
3790 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
3791 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3795 NewF.canonicalize();
3796 (void)InsertFormula(LU, LUIdx, NewF);
3805 /// Generate formulae for each use.
3807 LSRInstance::GenerateAllReuseFormulae() {
3808 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3809 // queries are more precise.
3810 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3811 LSRUse &LU = Uses[LUIdx];
3812 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3813 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3814 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3815 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3817 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3818 LSRUse &LU = Uses[LUIdx];
3819 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3820 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3821 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3822 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3823 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3824 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3825 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3826 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3828 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3829 LSRUse &LU = Uses[LUIdx];
3830 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3831 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3834 GenerateCrossUseConstantOffsets();
3836 DEBUG(dbgs() << "\n"
3837 "After generating reuse formulae:\n";
3838 print_uses(dbgs()));
3841 /// If there are multiple formulae with the same set of registers used
3842 /// by other uses, pick the best one and delete the others.
3843 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3844 DenseSet<const SCEV *> VisitedRegs;
3845 SmallPtrSet<const SCEV *, 16> Regs;
3846 SmallPtrSet<const SCEV *, 16> LoserRegs;
3848 bool ChangedFormulae = false;
3851 // Collect the best formula for each unique set of shared registers. This
3852 // is reset for each use.
3853 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3855 BestFormulaeTy BestFormulae;
3857 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3858 LSRUse &LU = Uses[LUIdx];
3859 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3862 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3863 FIdx != NumForms; ++FIdx) {
3864 Formula &F = LU.Formulae[FIdx];
3866 // Some formulas are instant losers. For example, they may depend on
3867 // nonexistent AddRecs from other loops. These need to be filtered
3868 // immediately, otherwise heuristics could choose them over others leading
3869 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3870 // avoids the need to recompute this information across formulae using the
3871 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3872 // the corresponding bad register from the Regs set.
3875 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3877 if (CostF.isLoser()) {
3878 // During initial formula generation, undesirable formulae are generated
3879 // by uses within other loops that have some non-trivial address mode or
3880 // use the postinc form of the IV. LSR needs to provide these formulae
3881 // as the basis of rediscovering the desired formula that uses an AddRec
3882 // corresponding to the existing phi. Once all formulae have been
3883 // generated, these initial losers may be pruned.
3884 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3888 SmallVector<const SCEV *, 4> Key;
3889 for (const SCEV *Reg : F.BaseRegs) {
3890 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3894 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3895 Key.push_back(F.ScaledReg);
3896 // Unstable sort by host order ok, because this is only used for
3898 std::sort(Key.begin(), Key.end());
3900 std::pair<BestFormulaeTy::const_iterator, bool> P =
3901 BestFormulae.insert(std::make_pair(Key, FIdx));
3905 Formula &Best = LU.Formulae[P.first->second];
3909 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3911 if (CostF < CostBest)
3913 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3915 " in favor of formula "; Best.print(dbgs());
3919 ChangedFormulae = true;
3921 LU.DeleteFormula(F);
3927 // Now that we've filtered out some formulae, recompute the Regs set.
3929 LU.RecomputeRegs(LUIdx, RegUses);
3931 // Reset this to prepare for the next use.
3932 BestFormulae.clear();
3935 DEBUG(if (ChangedFormulae) {
3937 "After filtering out undesirable candidates:\n";
3942 // This is a rough guess that seems to work fairly well.
3943 static const size_t ComplexityLimit = UINT16_MAX;
3945 /// Estimate the worst-case number of solutions the solver might have to
3946 /// consider. It almost never considers this many solutions because it prune the
3947 /// search space, but the pruning isn't always sufficient.
3948 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3950 for (const LSRUse &LU : Uses) {
3951 size_t FSize = LU.Formulae.size();
3952 if (FSize >= ComplexityLimit) {
3953 Power = ComplexityLimit;
3957 if (Power >= ComplexityLimit)
3963 /// When one formula uses a superset of the registers of another formula, it
3964 /// won't help reduce register pressure (though it may not necessarily hurt
3965 /// register pressure); remove it to simplify the system.
3966 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3967 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3968 DEBUG(dbgs() << "The search space is too complex.\n");
3970 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3971 "which use a superset of registers used by other "
3974 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3975 LSRUse &LU = Uses[LUIdx];
3977 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3978 Formula &F = LU.Formulae[i];
3979 // Look for a formula with a constant or GV in a register. If the use
3980 // also has a formula with that same value in an immediate field,
3981 // delete the one that uses a register.
3982 for (SmallVectorImpl<const SCEV *>::const_iterator
3983 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3984 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3986 NewF.BaseOffset += C->getValue()->getSExtValue();
3987 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3988 (I - F.BaseRegs.begin()));
3989 if (LU.HasFormulaWithSameRegs(NewF)) {
3990 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3991 LU.DeleteFormula(F);
3997 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3998 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4002 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4003 (I - F.BaseRegs.begin()));
4004 if (LU.HasFormulaWithSameRegs(NewF)) {
4005 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4007 LU.DeleteFormula(F);
4018 LU.RecomputeRegs(LUIdx, RegUses);
4021 DEBUG(dbgs() << "After pre-selection:\n";
4022 print_uses(dbgs()));
4026 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4027 /// allocate a single register for them.
4028 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4029 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4032 DEBUG(dbgs() << "The search space is too complex.\n"
4033 "Narrowing the search space by assuming that uses separated "
4034 "by a constant offset will use the same registers.\n");
4036 // This is especially useful for unrolled loops.
4038 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4039 LSRUse &LU = Uses[LUIdx];
4040 for (const Formula &F : LU.Formulae) {
4041 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4044 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4048 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4049 LU.Kind, LU.AccessTy))
4052 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4054 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4056 // Update the relocs to reference the new use.
4057 for (LSRFixup &Fixup : Fixups) {
4058 if (Fixup.LUIdx == LUIdx) {
4059 Fixup.LUIdx = LUThatHas - &Uses.front();
4060 Fixup.Offset += F.BaseOffset;
4061 // Add the new offset to LUThatHas' offset list.
4062 if (LUThatHas->Offsets.back() != Fixup.Offset) {
4063 LUThatHas->Offsets.push_back(Fixup.Offset);
4064 if (Fixup.Offset > LUThatHas->MaxOffset)
4065 LUThatHas->MaxOffset = Fixup.Offset;
4066 if (Fixup.Offset < LUThatHas->MinOffset)
4067 LUThatHas->MinOffset = Fixup.Offset;
4069 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4071 if (Fixup.LUIdx == NumUses-1)
4072 Fixup.LUIdx = LUIdx;
4075 // Delete formulae from the new use which are no longer legal.
4077 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4078 Formula &F = LUThatHas->Formulae[i];
4079 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4080 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4081 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4083 LUThatHas->DeleteFormula(F);
4091 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4093 // Delete the old use.
4094 DeleteUse(LU, LUIdx);
4101 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4104 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4105 /// we've done more filtering, as it may be able to find more formulae to
4107 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4108 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4109 DEBUG(dbgs() << "The search space is too complex.\n");
4111 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4112 "undesirable dedicated registers.\n");
4114 FilterOutUndesirableDedicatedRegisters();
4116 DEBUG(dbgs() << "After pre-selection:\n";
4117 print_uses(dbgs()));
4121 /// Pick a register which seems likely to be profitable, and then in any use
4122 /// which has any reference to that register, delete all formulae which do not
4123 /// reference that register.
4124 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4125 // With all other options exhausted, loop until the system is simple
4126 // enough to handle.
4127 SmallPtrSet<const SCEV *, 4> Taken;
4128 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4129 // Ok, we have too many of formulae on our hands to conveniently handle.
4130 // Use a rough heuristic to thin out the list.
4131 DEBUG(dbgs() << "The search space is too complex.\n");
4133 // Pick the register which is used by the most LSRUses, which is likely
4134 // to be a good reuse register candidate.
4135 const SCEV *Best = nullptr;
4136 unsigned BestNum = 0;
4137 for (const SCEV *Reg : RegUses) {
4138 if (Taken.count(Reg))
4143 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4144 if (Count > BestNum) {
4151 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4152 << " will yield profitable reuse.\n");
4155 // In any use with formulae which references this register, delete formulae
4156 // which don't reference it.
4157 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4158 LSRUse &LU = Uses[LUIdx];
4159 if (!LU.Regs.count(Best)) continue;
4162 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4163 Formula &F = LU.Formulae[i];
4164 if (!F.referencesReg(Best)) {
4165 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4166 LU.DeleteFormula(F);
4170 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4176 LU.RecomputeRegs(LUIdx, RegUses);
4179 DEBUG(dbgs() << "After pre-selection:\n";
4180 print_uses(dbgs()));
4184 /// If there are an extraordinary number of formulae to choose from, use some
4185 /// rough heuristics to prune down the number of formulae. This keeps the main
4186 /// solver from taking an extraordinary amount of time in some worst-case
4188 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4189 NarrowSearchSpaceByDetectingSupersets();
4190 NarrowSearchSpaceByCollapsingUnrolledCode();
4191 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4192 NarrowSearchSpaceByPickingWinnerRegs();
4195 /// This is the recursive solver.
4196 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4198 SmallVectorImpl<const Formula *> &Workspace,
4199 const Cost &CurCost,
4200 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4201 DenseSet<const SCEV *> &VisitedRegs) const {
4204 // - use more aggressive filtering
4205 // - sort the formula so that the most profitable solutions are found first
4206 // - sort the uses too
4208 // - don't compute a cost, and then compare. compare while computing a cost
4210 // - track register sets with SmallBitVector
4212 const LSRUse &LU = Uses[Workspace.size()];
4214 // If this use references any register that's already a part of the
4215 // in-progress solution, consider it a requirement that a formula must
4216 // reference that register in order to be considered. This prunes out
4217 // unprofitable searching.
4218 SmallSetVector<const SCEV *, 4> ReqRegs;
4219 for (const SCEV *S : CurRegs)
4220 if (LU.Regs.count(S))
4223 SmallPtrSet<const SCEV *, 16> NewRegs;
4225 for (const Formula &F : LU.Formulae) {
4226 // Ignore formulae which may not be ideal in terms of register reuse of
4227 // ReqRegs. The formula should use all required registers before
4228 // introducing new ones.
4229 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4230 for (const SCEV *Reg : ReqRegs) {
4231 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4232 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) !=
4235 if (NumReqRegsToFind == 0)
4239 if (NumReqRegsToFind != 0) {
4240 // If none of the formulae satisfied the required registers, then we could
4241 // clear ReqRegs and try again. Currently, we simply give up in this case.
4245 // Evaluate the cost of the current formula. If it's already worse than
4246 // the current best, prune the search at that point.
4249 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4251 if (NewCost < SolutionCost) {
4252 Workspace.push_back(&F);
4253 if (Workspace.size() != Uses.size()) {
4254 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4255 NewRegs, VisitedRegs);
4256 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4257 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4259 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4260 dbgs() << ".\n Regs:";
4261 for (const SCEV *S : NewRegs)
4262 dbgs() << ' ' << *S;
4265 SolutionCost = NewCost;
4266 Solution = Workspace;
4268 Workspace.pop_back();
4273 /// Choose one formula from each use. Return the results in the given Solution
4275 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4276 SmallVector<const Formula *, 8> Workspace;
4278 SolutionCost.Lose();
4280 SmallPtrSet<const SCEV *, 16> CurRegs;
4281 DenseSet<const SCEV *> VisitedRegs;
4282 Workspace.reserve(Uses.size());
4284 // SolveRecurse does all the work.
4285 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4286 CurRegs, VisitedRegs);
4287 if (Solution.empty()) {
4288 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4292 // Ok, we've now made all our decisions.
4293 DEBUG(dbgs() << "\n"
4294 "The chosen solution requires "; SolutionCost.print(dbgs());
4296 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4298 Uses[i].print(dbgs());
4301 Solution[i]->print(dbgs());
4305 assert(Solution.size() == Uses.size() && "Malformed solution!");
4308 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4309 /// we can go while still being dominated by the input positions. This helps
4310 /// canonicalize the insert position, which encourages sharing.
4311 BasicBlock::iterator
4312 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4313 const SmallVectorImpl<Instruction *> &Inputs)
4315 Instruction *Tentative = &*IP;
4317 bool AllDominate = true;
4318 Instruction *BetterPos = nullptr;
4319 // Don't bother attempting to insert before a catchswitch, their basic block
4320 // cannot have other non-PHI instructions.
4321 if (isa<CatchSwitchInst>(Tentative))
4324 for (Instruction *Inst : Inputs) {
4325 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4326 AllDominate = false;
4329 // Attempt to find an insert position in the middle of the block,
4330 // instead of at the end, so that it can be used for other expansions.
4331 if (Tentative->getParent() == Inst->getParent() &&
4332 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4333 BetterPos = &*std::next(BasicBlock::iterator(Inst));
4338 IP = BetterPos->getIterator();
4340 IP = Tentative->getIterator();
4342 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4343 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4346 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4347 if (!Rung) return IP;
4348 Rung = Rung->getIDom();
4349 if (!Rung) return IP;
4350 IDom = Rung->getBlock();
4352 // Don't climb into a loop though.
4353 const Loop *IDomLoop = LI.getLoopFor(IDom);
4354 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4355 if (IDomDepth <= IPLoopDepth &&
4356 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4360 Tentative = IDom->getTerminator();
4366 /// Determine an input position which will be dominated by the operands and
4367 /// which will dominate the result.
4368 BasicBlock::iterator
4369 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4372 SCEVExpander &Rewriter) const {
4373 // Collect some instructions which must be dominated by the
4374 // expanding replacement. These must be dominated by any operands that
4375 // will be required in the expansion.
4376 SmallVector<Instruction *, 4> Inputs;
4377 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4378 Inputs.push_back(I);
4379 if (LU.Kind == LSRUse::ICmpZero)
4380 if (Instruction *I =
4381 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4382 Inputs.push_back(I);
4383 if (LF.PostIncLoops.count(L)) {
4384 if (LF.isUseFullyOutsideLoop(L))
4385 Inputs.push_back(L->getLoopLatch()->getTerminator());
4387 Inputs.push_back(IVIncInsertPos);
4389 // The expansion must also be dominated by the increment positions of any
4390 // loops it for which it is using post-inc mode.
4391 for (const Loop *PIL : LF.PostIncLoops) {
4392 if (PIL == L) continue;
4394 // Be dominated by the loop exit.
4395 SmallVector<BasicBlock *, 4> ExitingBlocks;
4396 PIL->getExitingBlocks(ExitingBlocks);
4397 if (!ExitingBlocks.empty()) {
4398 BasicBlock *BB = ExitingBlocks[0];
4399 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4400 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4401 Inputs.push_back(BB->getTerminator());
4405 assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
4406 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4407 "Insertion point must be a normal instruction");
4409 // Then, climb up the immediate dominator tree as far as we can go while
4410 // still being dominated by the input positions.
4411 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4413 // Don't insert instructions before PHI nodes.
4414 while (isa<PHINode>(IP)) ++IP;
4416 // Ignore landingpad instructions.
4417 while (IP->isEHPad()) ++IP;
4419 // Ignore debug intrinsics.
4420 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4422 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4423 // IP consistent across expansions and allows the previously inserted
4424 // instructions to be reused by subsequent expansion.
4425 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
4431 /// Emit instructions for the leading candidate expression for this LSRUse (this
4432 /// is called "expanding").
4433 Value *LSRInstance::Expand(const LSRFixup &LF,
4435 BasicBlock::iterator IP,
4436 SCEVExpander &Rewriter,
4437 SmallVectorImpl<WeakVH> &DeadInsts) const {
4438 const LSRUse &LU = Uses[LF.LUIdx];
4439 if (LU.RigidFormula)
4440 return LF.OperandValToReplace;
4442 // Determine an input position which will be dominated by the operands and
4443 // which will dominate the result.
4444 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4445 Rewriter.setInsertPoint(&*IP);
4447 // Inform the Rewriter if we have a post-increment use, so that it can
4448 // perform an advantageous expansion.
4449 Rewriter.setPostInc(LF.PostIncLoops);
4451 // This is the type that the user actually needs.
4452 Type *OpTy = LF.OperandValToReplace->getType();
4453 // This will be the type that we'll initially expand to.
4454 Type *Ty = F.getType();
4456 // No type known; just expand directly to the ultimate type.
4458 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4459 // Expand directly to the ultimate type if it's the right size.
4461 // This is the type to do integer arithmetic in.
4462 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4464 // Build up a list of operands to add together to form the full base.
4465 SmallVector<const SCEV *, 8> Ops;
4467 // Expand the BaseRegs portion.
4468 for (const SCEV *Reg : F.BaseRegs) {
4469 assert(!Reg->isZero() && "Zero allocated in a base register!");
4471 // If we're expanding for a post-inc user, make the post-inc adjustment.
4472 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4473 Reg = TransformForPostIncUse(Denormalize, Reg,
4474 LF.UserInst, LF.OperandValToReplace,
4477 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
4480 // Expand the ScaledReg portion.
4481 Value *ICmpScaledV = nullptr;
4483 const SCEV *ScaledS = F.ScaledReg;
4485 // If we're expanding for a post-inc user, make the post-inc adjustment.
4486 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4487 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4488 LF.UserInst, LF.OperandValToReplace,
4491 if (LU.Kind == LSRUse::ICmpZero) {
4492 // Expand ScaleReg as if it was part of the base regs.
4495 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
4497 // An interesting way of "folding" with an icmp is to use a negated
4498 // scale, which we'll implement by inserting it into the other operand
4500 assert(F.Scale == -1 &&
4501 "The only scale supported by ICmpZero uses is -1!");
4502 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
4505 // Otherwise just expand the scaled register and an explicit scale,
4506 // which is expected to be matched as part of the address.
4508 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4509 // Unless the addressing mode will not be folded.
4510 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4511 isAMCompletelyFolded(TTI, LU, F)) {
4512 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4514 Ops.push_back(SE.getUnknown(FullV));
4516 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
4519 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4520 Ops.push_back(ScaledS);
4524 // Expand the GV portion.
4526 // Flush the operand list to suppress SCEVExpander hoisting.
4528 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4530 Ops.push_back(SE.getUnknown(FullV));
4532 Ops.push_back(SE.getUnknown(F.BaseGV));
4535 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4536 // unfolded offsets. LSR assumes they both live next to their uses.
4538 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4540 Ops.push_back(SE.getUnknown(FullV));
4543 // Expand the immediate portion.
4544 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4546 if (LU.Kind == LSRUse::ICmpZero) {
4547 // The other interesting way of "folding" with an ICmpZero is to use a
4548 // negated immediate.
4550 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4552 Ops.push_back(SE.getUnknown(ICmpScaledV));
4553 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4556 // Just add the immediate values. These again are expected to be matched
4557 // as part of the address.
4558 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4562 // Expand the unfolded offset portion.
4563 int64_t UnfoldedOffset = F.UnfoldedOffset;
4564 if (UnfoldedOffset != 0) {
4565 // Just add the immediate values.
4566 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4570 // Emit instructions summing all the operands.
4571 const SCEV *FullS = Ops.empty() ?
4572 SE.getConstant(IntTy, 0) :
4574 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
4576 // We're done expanding now, so reset the rewriter.
4577 Rewriter.clearPostInc();
4579 // An ICmpZero Formula represents an ICmp which we're handling as a
4580 // comparison against zero. Now that we've expanded an expression for that
4581 // form, update the ICmp's other operand.
4582 if (LU.Kind == LSRUse::ICmpZero) {
4583 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4584 DeadInsts.emplace_back(CI->getOperand(1));
4585 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4586 "a scale at the same time!");
4587 if (F.Scale == -1) {
4588 if (ICmpScaledV->getType() != OpTy) {
4590 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4592 ICmpScaledV, OpTy, "tmp", CI);
4595 CI->setOperand(1, ICmpScaledV);
4597 // A scale of 1 means that the scale has been expanded as part of the
4599 assert((F.Scale == 0 || F.Scale == 1) &&
4600 "ICmp does not support folding a global value and "
4601 "a scale at the same time!");
4602 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4604 if (C->getType() != OpTy)
4605 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4609 CI->setOperand(1, C);
4616 /// Helper for Rewrite. PHI nodes are special because the use of their operands
4617 /// effectively happens in their predecessor blocks, so the expression may need
4618 /// to be expanded in multiple places.
4619 void LSRInstance::RewriteForPHI(PHINode *PN,
4622 SCEVExpander &Rewriter,
4623 SmallVectorImpl<WeakVH> &DeadInsts) const {
4624 DenseMap<BasicBlock *, Value *> Inserted;
4625 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4626 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4627 BasicBlock *BB = PN->getIncomingBlock(i);
4629 // If this is a critical edge, split the edge so that we do not insert
4630 // the code on all predecessor/successor paths. We do this unless this
4631 // is the canonical backedge for this loop, which complicates post-inc
4633 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4634 !isa<IndirectBrInst>(BB->getTerminator())) {
4635 BasicBlock *Parent = PN->getParent();
4636 Loop *PNLoop = LI.getLoopFor(Parent);
4637 if (!PNLoop || Parent != PNLoop->getHeader()) {
4638 // Split the critical edge.
4639 BasicBlock *NewBB = nullptr;
4640 if (!Parent->isLandingPad()) {
4641 NewBB = SplitCriticalEdge(BB, Parent,
4642 CriticalEdgeSplittingOptions(&DT, &LI)
4643 .setMergeIdenticalEdges()
4644 .setDontDeleteUselessPHIs());
4646 SmallVector<BasicBlock*, 2> NewBBs;
4647 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
4650 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4651 // phi predecessors are identical. The simple thing to do is skip
4652 // splitting in this case rather than complicate the API.
4654 // If PN is outside of the loop and BB is in the loop, we want to
4655 // move the block to be immediately before the PHI block, not
4656 // immediately after BB.
4657 if (L->contains(BB) && !L->contains(PN))
4658 NewBB->moveBefore(PN->getParent());
4660 // Splitting the edge can reduce the number of PHI entries we have.
4661 e = PN->getNumIncomingValues();
4663 i = PN->getBasicBlockIndex(BB);
4668 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4669 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
4671 PN->setIncomingValue(i, Pair.first->second);
4673 Value *FullV = Expand(LF, F, BB->getTerminator()->getIterator(),
4674 Rewriter, DeadInsts);
4676 // If this is reuse-by-noop-cast, insert the noop cast.
4677 Type *OpTy = LF.OperandValToReplace->getType();
4678 if (FullV->getType() != OpTy)
4680 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4682 FullV, LF.OperandValToReplace->getType(),
4683 "tmp", BB->getTerminator());
4685 PN->setIncomingValue(i, FullV);
4686 Pair.first->second = FullV;
4691 /// Emit instructions for the leading candidate expression for this LSRUse (this
4692 /// is called "expanding"), and update the UserInst to reference the newly
4694 void LSRInstance::Rewrite(const LSRFixup &LF,
4696 SCEVExpander &Rewriter,
4697 SmallVectorImpl<WeakVH> &DeadInsts) const {
4698 // First, find an insertion point that dominates UserInst. For PHI nodes,
4699 // find the nearest block which dominates all the relevant uses.
4700 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4701 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts);
4704 Expand(LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
4706 // If this is reuse-by-noop-cast, insert the noop cast.
4707 Type *OpTy = LF.OperandValToReplace->getType();
4708 if (FullV->getType() != OpTy) {
4710 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4711 FullV, OpTy, "tmp", LF.UserInst);
4715 // Update the user. ICmpZero is handled specially here (for now) because
4716 // Expand may have updated one of the operands of the icmp already, and
4717 // its new value may happen to be equal to LF.OperandValToReplace, in
4718 // which case doing replaceUsesOfWith leads to replacing both operands
4719 // with the same value. TODO: Reorganize this.
4720 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4721 LF.UserInst->setOperand(0, FullV);
4723 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4726 DeadInsts.emplace_back(LF.OperandValToReplace);
4729 /// Rewrite all the fixup locations with new values, following the chosen
4731 void LSRInstance::ImplementSolution(
4732 const SmallVectorImpl<const Formula *> &Solution) {
4733 // Keep track of instructions we may have made dead, so that
4734 // we can remove them after we are done working.
4735 SmallVector<WeakVH, 16> DeadInsts;
4737 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
4740 Rewriter.setDebugType(DEBUG_TYPE);
4742 Rewriter.disableCanonicalMode();
4743 Rewriter.enableLSRMode();
4744 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4746 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4747 for (const IVChain &Chain : IVChainVec) {
4748 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
4749 Rewriter.setChainedPhi(PN);
4752 // Expand the new value definitions and update the users.
4753 for (const LSRFixup &Fixup : Fixups) {
4754 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts);
4759 for (const IVChain &Chain : IVChainVec) {
4760 GenerateIVChain(Chain, Rewriter, DeadInsts);
4763 // Clean up after ourselves. This must be done before deleting any
4767 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4770 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
4771 DominatorTree &DT, LoopInfo &LI,
4772 const TargetTransformInfo &TTI)
4773 : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L), Changed(false),
4774 IVIncInsertPos(nullptr) {
4775 // If LoopSimplify form is not available, stay out of trouble.
4776 if (!L->isLoopSimplifyForm())
4779 // If there's no interesting work to be done, bail early.
4780 if (IU.empty()) return;
4782 // If there's too much analysis to be done, bail early. We won't be able to
4783 // model the problem anyway.
4784 unsigned NumUsers = 0;
4785 for (const IVStrideUse &U : IU) {
4786 if (++NumUsers > MaxIVUsers) {
4788 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n");
4791 // Bail out if we have a PHI on an EHPad that gets a value from a
4792 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
4793 // no good place to stick any instructions.
4794 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
4795 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
4796 if (isa<FuncletPadInst>(FirstNonPHI) ||
4797 isa<CatchSwitchInst>(FirstNonPHI))
4798 for (BasicBlock *PredBB : PN->blocks())
4799 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
4805 // All dominating loops must have preheaders, or SCEVExpander may not be able
4806 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4808 // IVUsers analysis should only create users that are dominated by simple loop
4809 // headers. Since this loop should dominate all of its users, its user list
4810 // should be empty if this loop itself is not within a simple loop nest.
4811 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4812 Rung; Rung = Rung->getIDom()) {
4813 BasicBlock *BB = Rung->getBlock();
4814 const Loop *DomLoop = LI.getLoopFor(BB);
4815 if (DomLoop && DomLoop->getHeader() == BB) {
4816 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4821 DEBUG(dbgs() << "\nLSR on loop ";
4822 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4825 // First, perform some low-level loop optimizations.
4827 OptimizeLoopTermCond();
4829 // If loop preparation eliminates all interesting IV users, bail.
4830 if (IU.empty()) return;
4832 // Skip nested loops until we can model them better with formulae.
4834 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4838 // Start collecting data and preparing for the solver.
4840 CollectInterestingTypesAndFactors();
4841 CollectFixupsAndInitialFormulae();
4842 CollectLoopInvariantFixupsAndFormulae();
4844 assert(!Uses.empty() && "IVUsers reported at least one use");
4845 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4846 print_uses(dbgs()));
4848 // Now use the reuse data to generate a bunch of interesting ways
4849 // to formulate the values needed for the uses.
4850 GenerateAllReuseFormulae();
4852 FilterOutUndesirableDedicatedRegisters();
4853 NarrowSearchSpaceUsingHeuristics();
4855 SmallVector<const Formula *, 8> Solution;
4858 // Release memory that is no longer needed.
4863 if (Solution.empty())
4867 // Formulae should be legal.
4868 for (const LSRUse &LU : Uses) {
4869 for (const Formula &F : LU.Formulae)
4870 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4871 F) && "Illegal formula generated!");
4875 // Now that we've decided what we want, make it so.
4876 ImplementSolution(Solution);
4879 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4880 if (Factors.empty() && Types.empty()) return;
4882 OS << "LSR has identified the following interesting factors and types: ";
4885 for (int64_t Factor : Factors) {
4886 if (!First) OS << ", ";
4888 OS << '*' << Factor;
4891 for (Type *Ty : Types) {
4892 if (!First) OS << ", ";
4894 OS << '(' << *Ty << ')';
4899 void LSRInstance::print_fixups(raw_ostream &OS) const {
4900 OS << "LSR is examining the following fixup sites:\n";
4901 for (const LSRFixup &LF : Fixups) {
4908 void LSRInstance::print_uses(raw_ostream &OS) const {
4909 OS << "LSR is examining the following uses:\n";
4910 for (const LSRUse &LU : Uses) {
4914 for (const Formula &F : LU.Formulae) {
4922 void LSRInstance::print(raw_ostream &OS) const {
4923 print_factors_and_types(OS);
4929 void LSRInstance::dump() const {
4930 print(errs()); errs() << '\n';
4935 class LoopStrengthReduce : public LoopPass {
4937 static char ID; // Pass ID, replacement for typeid
4938 LoopStrengthReduce();
4941 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
4942 void getAnalysisUsage(AnalysisUsage &AU) const override;
4947 char LoopStrengthReduce::ID = 0;
4948 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4949 "Loop Strength Reduction", false, false)
4950 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
4951 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4952 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
4953 INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)
4954 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
4955 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4956 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4957 "Loop Strength Reduction", false, false)
4960 Pass *llvm::createLoopStrengthReducePass() {
4961 return new LoopStrengthReduce();
4964 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4965 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4968 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4969 // We split critical edges, so we change the CFG. However, we do update
4970 // many analyses if they are around.
4971 AU.addPreservedID(LoopSimplifyID);
4973 AU.addRequired<LoopInfoWrapperPass>();
4974 AU.addPreserved<LoopInfoWrapperPass>();
4975 AU.addRequiredID(LoopSimplifyID);
4976 AU.addRequired<DominatorTreeWrapperPass>();
4977 AU.addPreserved<DominatorTreeWrapperPass>();
4978 AU.addRequired<ScalarEvolutionWrapperPass>();
4979 AU.addPreserved<ScalarEvolutionWrapperPass>();
4980 // Requiring LoopSimplify a second time here prevents IVUsers from running
4981 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4982 AU.addRequiredID(LoopSimplifyID);
4983 AU.addRequired<IVUsersWrapperPass>();
4984 AU.addPreserved<IVUsersWrapperPass>();
4985 AU.addRequired<TargetTransformInfoWrapperPass>();
4988 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4992 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
4993 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
4994 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
4995 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
4996 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
4997 *L->getHeader()->getParent());
4998 bool Changed = false;
5000 // Run the main LSR transformation.
5001 Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5003 // Remove any extra phis created by processing inner loops.
5004 Changed |= DeleteDeadPHIs(L->getHeader());
5005 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5006 SmallVector<WeakVH, 16> DeadInsts;
5007 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5008 SCEVExpander Rewriter(getAnalysis<ScalarEvolutionWrapperPass>().getSE(), DL,
5011 Rewriter.setDebugType(DEBUG_TYPE);
5013 unsigned numFolded = Rewriter.replaceCongruentIVs(
5014 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
5015 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5016 *L->getHeader()->getParent()));
5019 DeleteTriviallyDeadInstructions(DeadInsts);
5020 DeleteDeadPHIs(L->getHeader());