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 case Intrinsic::x86_sse_storeu_ps:
688 case Intrinsic::x86_sse2_storeu_pd:
689 case Intrinsic::x86_sse2_storeu_dq:
690 case Intrinsic::x86_sse2_storel_dq:
691 if (II->getArgOperand(0) == OperandVal)
699 /// Return the type of the memory being accessed.
700 static MemAccessTy getAccessType(const Instruction *Inst) {
701 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
702 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
703 AccessTy.MemTy = SI->getOperand(0)->getType();
704 AccessTy.AddrSpace = SI->getPointerAddressSpace();
705 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
706 AccessTy.AddrSpace = LI->getPointerAddressSpace();
707 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
708 // Addressing modes can also be folded into prefetches and a variety
710 switch (II->getIntrinsicID()) {
712 case Intrinsic::x86_sse_storeu_ps:
713 case Intrinsic::x86_sse2_storeu_pd:
714 case Intrinsic::x86_sse2_storeu_dq:
715 case Intrinsic::x86_sse2_storel_dq:
716 AccessTy.MemTy = II->getArgOperand(0)->getType();
721 // All pointers have the same requirements, so canonicalize them to an
722 // arbitrary pointer type to minimize variation.
723 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
724 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
725 PTy->getAddressSpace());
730 /// Return true if this AddRec is already a phi in its loop.
731 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
732 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
733 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
734 if (SE.isSCEVable(PN->getType()) &&
735 (SE.getEffectiveSCEVType(PN->getType()) ==
736 SE.getEffectiveSCEVType(AR->getType())) &&
737 SE.getSCEV(PN) == AR)
743 /// Check if expanding this expression is likely to incur significant cost. This
744 /// is tricky because SCEV doesn't track which expressions are actually computed
745 /// by the current IR.
747 /// We currently allow expansion of IV increments that involve adds,
748 /// multiplication by constants, and AddRecs from existing phis.
750 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
751 /// obvious multiple of the UDivExpr.
752 static bool isHighCostExpansion(const SCEV *S,
753 SmallPtrSetImpl<const SCEV*> &Processed,
754 ScalarEvolution &SE) {
755 // Zero/One operand expressions
756 switch (S->getSCEVType()) {
761 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
764 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
767 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
771 if (!Processed.insert(S).second)
774 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
775 for (const SCEV *S : Add->operands()) {
776 if (isHighCostExpansion(S, Processed, SE))
782 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
783 if (Mul->getNumOperands() == 2) {
784 // Multiplication by a constant is ok
785 if (isa<SCEVConstant>(Mul->getOperand(0)))
786 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
788 // If we have the value of one operand, check if an existing
789 // multiplication already generates this expression.
790 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
791 Value *UVal = U->getValue();
792 for (User *UR : UVal->users()) {
793 // If U is a constant, it may be used by a ConstantExpr.
794 Instruction *UI = dyn_cast<Instruction>(UR);
795 if (UI && UI->getOpcode() == Instruction::Mul &&
796 SE.isSCEVable(UI->getType())) {
797 return SE.getSCEV(UI) == Mul;
804 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
805 if (isExistingPhi(AR, SE))
809 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
813 /// If any of the instructions is the specified set are trivially dead, delete
814 /// them and see if this makes any of their operands subsequently dead.
816 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
817 bool Changed = false;
819 while (!DeadInsts.empty()) {
820 Value *V = DeadInsts.pop_back_val();
821 Instruction *I = dyn_cast_or_null<Instruction>(V);
823 if (!I || !isInstructionTriviallyDead(I))
826 for (Use &O : I->operands())
827 if (Instruction *U = dyn_cast<Instruction>(O)) {
830 DeadInsts.emplace_back(U);
833 I->eraseFromParent();
844 /// \brief Check if the addressing mode defined by \p F is completely
845 /// folded in \p LU at isel time.
846 /// This includes address-mode folding and special icmp tricks.
847 /// This function returns true if \p LU can accommodate what \p F
848 /// defines and up to 1 base + 1 scaled + offset.
849 /// In other words, if \p F has several base registers, this function may
850 /// still return true. Therefore, users still need to account for
851 /// additional base registers and/or unfolded offsets to derive an
852 /// accurate cost model.
853 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
854 const LSRUse &LU, const Formula &F);
855 // Get the cost of the scaling factor used in F for LU.
856 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
857 const LSRUse &LU, const Formula &F);
861 /// This class is used to measure and compare candidate formulae.
863 /// TODO: Some of these could be merged. Also, a lexical ordering
864 /// isn't always optimal.
868 unsigned NumBaseAdds;
875 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
876 SetupCost(0), ScaleCost(0) {}
878 bool operator<(const Cost &Other) const;
883 // Once any of the metrics loses, they must all remain losers.
885 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
886 | ImmCost | SetupCost | ScaleCost) != ~0u)
887 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
888 & ImmCost & SetupCost & ScaleCost) == ~0u);
893 assert(isValid() && "invalid cost");
894 return NumRegs == ~0u;
897 void RateFormula(const TargetTransformInfo &TTI,
899 SmallPtrSetImpl<const SCEV *> &Regs,
900 const DenseSet<const SCEV *> &VisitedRegs,
902 const SmallVectorImpl<int64_t> &Offsets,
903 ScalarEvolution &SE, DominatorTree &DT,
905 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
907 void print(raw_ostream &OS) const;
911 void RateRegister(const SCEV *Reg,
912 SmallPtrSetImpl<const SCEV *> &Regs,
914 ScalarEvolution &SE, DominatorTree &DT);
915 void RatePrimaryRegister(const SCEV *Reg,
916 SmallPtrSetImpl<const SCEV *> &Regs,
918 ScalarEvolution &SE, DominatorTree &DT,
919 SmallPtrSetImpl<const SCEV *> *LoserRegs);
924 /// Tally up interesting quantities from the given register.
925 void Cost::RateRegister(const SCEV *Reg,
926 SmallPtrSetImpl<const SCEV *> &Regs,
928 ScalarEvolution &SE, DominatorTree &DT) {
929 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
930 // If this is an addrec for another loop, don't second-guess its addrec phi
931 // nodes. LSR isn't currently smart enough to reason about more than one
932 // loop at a time. LSR has already run on inner loops, will not run on outer
933 // loops, and cannot be expected to change sibling loops.
934 if (AR->getLoop() != L) {
935 // If the AddRec exists, consider it's register free and leave it alone.
936 if (isExistingPhi(AR, SE))
939 // Otherwise, do not consider this formula at all.
943 AddRecCost += 1; /// TODO: This should be a function of the stride.
945 // Add the step value register, if it needs one.
946 // TODO: The non-affine case isn't precisely modeled here.
947 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
948 if (!Regs.count(AR->getOperand(1))) {
949 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
957 // Rough heuristic; favor registers which don't require extra setup
958 // instructions in the preheader.
959 if (!isa<SCEVUnknown>(Reg) &&
960 !isa<SCEVConstant>(Reg) &&
961 !(isa<SCEVAddRecExpr>(Reg) &&
962 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
963 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
966 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
967 SE.hasComputableLoopEvolution(Reg, L);
970 /// Record this register in the set. If we haven't seen it before, rate
971 /// it. Optional LoserRegs provides a way to declare any formula that refers to
972 /// one of those regs an instant loser.
973 void Cost::RatePrimaryRegister(const SCEV *Reg,
974 SmallPtrSetImpl<const SCEV *> &Regs,
976 ScalarEvolution &SE, DominatorTree &DT,
977 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
978 if (LoserRegs && LoserRegs->count(Reg)) {
982 if (Regs.insert(Reg).second) {
983 RateRegister(Reg, Regs, L, SE, DT);
984 if (LoserRegs && isLoser())
985 LoserRegs->insert(Reg);
989 void Cost::RateFormula(const TargetTransformInfo &TTI,
991 SmallPtrSetImpl<const SCEV *> &Regs,
992 const DenseSet<const SCEV *> &VisitedRegs,
994 const SmallVectorImpl<int64_t> &Offsets,
995 ScalarEvolution &SE, DominatorTree &DT,
997 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
998 assert(F.isCanonical() && "Cost is accurate only for canonical formula");
999 // Tally up the registers.
1000 if (const SCEV *ScaledReg = F.ScaledReg) {
1001 if (VisitedRegs.count(ScaledReg)) {
1005 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1009 for (const SCEV *BaseReg : F.BaseRegs) {
1010 if (VisitedRegs.count(BaseReg)) {
1014 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1019 // Determine how many (unfolded) adds we'll need inside the loop.
1020 size_t NumBaseParts = F.getNumRegs();
1021 if (NumBaseParts > 1)
1022 // Do not count the base and a possible second register if the target
1023 // allows to fold 2 registers.
1025 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1026 NumBaseAdds += (F.UnfoldedOffset != 0);
1028 // Accumulate non-free scaling amounts.
1029 ScaleCost += getScalingFactorCost(TTI, LU, F);
1031 // Tally up the non-zero immediates.
1032 for (int64_t O : Offsets) {
1033 int64_t Offset = (uint64_t)O + F.BaseOffset;
1035 ImmCost += 64; // Handle symbolic values conservatively.
1036 // TODO: This should probably be the pointer size.
1037 else if (Offset != 0)
1038 ImmCost += APInt(64, Offset, true).getMinSignedBits();
1040 assert(isValid() && "invalid cost");
1043 /// Set this cost to a losing value.
1054 /// Choose the lower cost.
1055 bool Cost::operator<(const Cost &Other) const {
1056 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
1057 ImmCost, SetupCost) <
1058 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
1059 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
1063 void Cost::print(raw_ostream &OS) const {
1064 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
1065 if (AddRecCost != 0)
1066 OS << ", with addrec cost " << AddRecCost;
1068 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
1069 if (NumBaseAdds != 0)
1070 OS << ", plus " << NumBaseAdds << " base add"
1071 << (NumBaseAdds == 1 ? "" : "s");
1073 OS << ", plus " << ScaleCost << " scale cost";
1075 OS << ", plus " << ImmCost << " imm cost";
1077 OS << ", plus " << SetupCost << " setup cost";
1081 void Cost::dump() const {
1082 print(errs()); errs() << '\n';
1087 /// An operand value in an instruction which is to be replaced with some
1088 /// equivalent, possibly strength-reduced, replacement.
1090 /// The instruction which will be updated.
1091 Instruction *UserInst;
1093 /// The operand of the instruction which will be replaced. The operand may be
1094 /// used more than once; every instance will be replaced.
1095 Value *OperandValToReplace;
1097 /// If this user is to use the post-incremented value of an induction
1098 /// variable, this variable is non-null and holds the loop associated with the
1099 /// induction variable.
1100 PostIncLoopSet PostIncLoops;
1102 /// The index of the LSRUse describing the expression which this fixup needs,
1103 /// minus an offset (below).
1106 /// A constant offset to be added to the LSRUse expression. This allows
1107 /// multiple fixups to share the same LSRUse with different offsets, for
1108 /// example in an unrolled loop.
1111 bool isUseFullyOutsideLoop(const Loop *L) const;
1115 void print(raw_ostream &OS) const;
1121 LSRFixup::LSRFixup()
1122 : UserInst(nullptr), OperandValToReplace(nullptr), LUIdx(~size_t(0)),
1125 /// Test whether this fixup always uses its value outside of the given loop.
1126 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1127 // PHI nodes use their value in their incoming blocks.
1128 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1129 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1130 if (PN->getIncomingValue(i) == OperandValToReplace &&
1131 L->contains(PN->getIncomingBlock(i)))
1136 return !L->contains(UserInst);
1139 void LSRFixup::print(raw_ostream &OS) const {
1141 // Store is common and interesting enough to be worth special-casing.
1142 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1144 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1145 } else if (UserInst->getType()->isVoidTy())
1146 OS << UserInst->getOpcodeName();
1148 UserInst->printAsOperand(OS, /*PrintType=*/false);
1150 OS << ", OperandValToReplace=";
1151 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1153 for (const Loop *PIL : PostIncLoops) {
1154 OS << ", PostIncLoop=";
1155 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1158 if (LUIdx != ~size_t(0))
1159 OS << ", LUIdx=" << LUIdx;
1162 OS << ", Offset=" << Offset;
1166 void LSRFixup::dump() const {
1167 print(errs()); errs() << '\n';
1172 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1173 /// SmallVectors of const SCEV*.
1174 struct UniquifierDenseMapInfo {
1175 static SmallVector<const SCEV *, 4> getEmptyKey() {
1176 SmallVector<const SCEV *, 4> V;
1177 V.push_back(reinterpret_cast<const SCEV *>(-1));
1181 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1182 SmallVector<const SCEV *, 4> V;
1183 V.push_back(reinterpret_cast<const SCEV *>(-2));
1187 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1188 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1191 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1192 const SmallVector<const SCEV *, 4> &RHS) {
1197 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1198 /// as uses invented by LSR itself. It includes information about what kinds of
1199 /// things can be folded into the user, information about the user itself, and
1200 /// information about how the use may be satisfied. TODO: Represent multiple
1201 /// users of the same expression in common?
1203 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1206 /// An enum for a kind of use, indicating what types of scaled and immediate
1207 /// operands it might support.
1209 Basic, ///< A normal use, with no folding.
1210 Special, ///< A special case of basic, allowing -1 scales.
1211 Address, ///< An address use; folding according to TargetLowering
1212 ICmpZero ///< An equality icmp with both operands folded into one.
1213 // TODO: Add a generic icmp too?
1216 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1219 MemAccessTy AccessTy;
1221 SmallVector<int64_t, 8> Offsets;
1225 /// This records whether all of the fixups using this LSRUse are outside of
1226 /// the loop, in which case some special-case heuristics may be used.
1227 bool AllFixupsOutsideLoop;
1229 /// RigidFormula is set to true to guarantee that this use will be associated
1230 /// with a single formula--the one that initially matched. Some SCEV
1231 /// expressions cannot be expanded. This allows LSR to consider the registers
1232 /// used by those expressions without the need to expand them later after
1233 /// changing the formula.
1236 /// This records the widest use type for any fixup using this
1237 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1238 /// fixup widths to be equivalent, because the narrower one may be relying on
1239 /// the implicit truncation to truncate away bogus bits.
1240 Type *WidestFixupType;
1242 /// A list of ways to build a value that can satisfy this user. After the
1243 /// list is populated, one of these is selected heuristically and used to
1244 /// formulate a replacement for OperandValToReplace in UserInst.
1245 SmallVector<Formula, 12> Formulae;
1247 /// The set of register candidates used by all formulae in this LSRUse.
1248 SmallPtrSet<const SCEV *, 4> Regs;
1250 LSRUse(KindType K, MemAccessTy AT)
1251 : Kind(K), AccessTy(AT), MinOffset(INT64_MAX), MaxOffset(INT64_MIN),
1252 AllFixupsOutsideLoop(true), RigidFormula(false),
1253 WidestFixupType(nullptr) {}
1255 bool HasFormulaWithSameRegs(const Formula &F) const;
1256 bool InsertFormula(const Formula &F);
1257 void DeleteFormula(Formula &F);
1258 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1260 void print(raw_ostream &OS) const;
1266 /// Test whether this use as a formula which has the same registers as the given
1268 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1269 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1270 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1271 // Unstable sort by host order ok, because this is only used for uniquifying.
1272 std::sort(Key.begin(), Key.end());
1273 return Uniquifier.count(Key);
1276 /// If the given formula has not yet been inserted, add it to the list, and
1277 /// return true. Return false otherwise. The formula must be in canonical form.
1278 bool LSRUse::InsertFormula(const Formula &F) {
1279 assert(F.isCanonical() && "Invalid canonical representation");
1281 if (!Formulae.empty() && RigidFormula)
1284 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1285 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1286 // Unstable sort by host order ok, because this is only used for uniquifying.
1287 std::sort(Key.begin(), Key.end());
1289 if (!Uniquifier.insert(Key).second)
1292 // Using a register to hold the value of 0 is not profitable.
1293 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1294 "Zero allocated in a scaled register!");
1296 for (const SCEV *BaseReg : F.BaseRegs)
1297 assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1300 // Add the formula to the list.
1301 Formulae.push_back(F);
1303 // Record registers now being used by this use.
1304 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1306 Regs.insert(F.ScaledReg);
1311 /// Remove the given formula from this use's list.
1312 void LSRUse::DeleteFormula(Formula &F) {
1313 if (&F != &Formulae.back())
1314 std::swap(F, Formulae.back());
1315 Formulae.pop_back();
1318 /// Recompute the Regs field, and update RegUses.
1319 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1320 // Now that we've filtered out some formulae, recompute the Regs set.
1321 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1323 for (const Formula &F : Formulae) {
1324 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1325 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1328 // Update the RegTracker.
1329 for (const SCEV *S : OldRegs)
1331 RegUses.dropRegister(S, LUIdx);
1334 void LSRUse::print(raw_ostream &OS) const {
1335 OS << "LSR Use: Kind=";
1337 case Basic: OS << "Basic"; break;
1338 case Special: OS << "Special"; break;
1339 case ICmpZero: OS << "ICmpZero"; break;
1341 OS << "Address of ";
1342 if (AccessTy.MemTy->isPointerTy())
1343 OS << "pointer"; // the full pointer type could be really verbose
1345 OS << *AccessTy.MemTy;
1348 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1351 OS << ", Offsets={";
1352 bool NeedComma = false;
1353 for (int64_t O : Offsets) {
1354 if (NeedComma) OS << ',';
1360 if (AllFixupsOutsideLoop)
1361 OS << ", all-fixups-outside-loop";
1363 if (WidestFixupType)
1364 OS << ", widest fixup type: " << *WidestFixupType;
1368 void LSRUse::dump() const {
1369 print(errs()); errs() << '\n';
1372 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1373 LSRUse::KindType Kind, MemAccessTy AccessTy,
1374 GlobalValue *BaseGV, int64_t BaseOffset,
1375 bool HasBaseReg, int64_t Scale) {
1377 case LSRUse::Address:
1378 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1379 HasBaseReg, Scale, AccessTy.AddrSpace);
1381 case LSRUse::ICmpZero:
1382 // There's not even a target hook for querying whether it would be legal to
1383 // fold a GV into an ICmp.
1387 // ICmp only has two operands; don't allow more than two non-trivial parts.
1388 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1391 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1392 // putting the scaled register in the other operand of the icmp.
1393 if (Scale != 0 && Scale != -1)
1396 // If we have low-level target information, ask the target if it can fold an
1397 // integer immediate on an icmp.
1398 if (BaseOffset != 0) {
1400 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1401 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1402 // Offs is the ICmp immediate.
1404 // The cast does the right thing with INT64_MIN.
1405 BaseOffset = -(uint64_t)BaseOffset;
1406 return TTI.isLegalICmpImmediate(BaseOffset);
1409 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1413 // Only handle single-register values.
1414 return !BaseGV && Scale == 0 && BaseOffset == 0;
1416 case LSRUse::Special:
1417 // Special case Basic to handle -1 scales.
1418 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1421 llvm_unreachable("Invalid LSRUse Kind!");
1424 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1425 int64_t MinOffset, int64_t MaxOffset,
1426 LSRUse::KindType Kind, MemAccessTy AccessTy,
1427 GlobalValue *BaseGV, int64_t BaseOffset,
1428 bool HasBaseReg, int64_t Scale) {
1429 // Check for overflow.
1430 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1433 MinOffset = (uint64_t)BaseOffset + MinOffset;
1434 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1437 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1439 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1440 HasBaseReg, Scale) &&
1441 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1445 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1446 int64_t MinOffset, int64_t MaxOffset,
1447 LSRUse::KindType Kind, MemAccessTy AccessTy,
1449 // For the purpose of isAMCompletelyFolded either having a canonical formula
1450 // or a scale not equal to zero is correct.
1451 // Problems may arise from non canonical formulae having a scale == 0.
1452 // Strictly speaking it would best to just rely on canonical formulae.
1453 // However, when we generate the scaled formulae, we first check that the
1454 // scaling factor is profitable before computing the actual ScaledReg for
1455 // compile time sake.
1456 assert((F.isCanonical() || F.Scale != 0));
1457 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1458 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1461 /// Test whether we know how to expand the current formula.
1462 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1463 int64_t MaxOffset, LSRUse::KindType Kind,
1464 MemAccessTy AccessTy, GlobalValue *BaseGV,
1465 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1466 // We know how to expand completely foldable formulae.
1467 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1468 BaseOffset, HasBaseReg, Scale) ||
1469 // Or formulae that use a base register produced by a sum of base
1472 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1473 BaseGV, BaseOffset, true, 0));
1476 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1477 int64_t MaxOffset, LSRUse::KindType Kind,
1478 MemAccessTy AccessTy, const Formula &F) {
1479 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1480 F.BaseOffset, F.HasBaseReg, F.Scale);
1483 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1484 const LSRUse &LU, const Formula &F) {
1485 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1486 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1490 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1491 const LSRUse &LU, const Formula &F) {
1495 // If the use is not completely folded in that instruction, we will have to
1496 // pay an extra cost only for scale != 1.
1497 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1499 return F.Scale != 1;
1502 case LSRUse::Address: {
1503 // Check the scaling factor cost with both the min and max offsets.
1504 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1505 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1506 F.Scale, LU.AccessTy.AddrSpace);
1507 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1508 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1509 F.Scale, LU.AccessTy.AddrSpace);
1511 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1512 "Legal addressing mode has an illegal cost!");
1513 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1515 case LSRUse::ICmpZero:
1517 case LSRUse::Special:
1518 // The use is completely folded, i.e., everything is folded into the
1523 llvm_unreachable("Invalid LSRUse Kind!");
1526 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1527 LSRUse::KindType Kind, MemAccessTy AccessTy,
1528 GlobalValue *BaseGV, int64_t BaseOffset,
1530 // Fast-path: zero is always foldable.
1531 if (BaseOffset == 0 && !BaseGV) return true;
1533 // Conservatively, create an address with an immediate and a
1534 // base and a scale.
1535 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1537 // Canonicalize a scale of 1 to a base register if the formula doesn't
1538 // already have a base register.
1539 if (!HasBaseReg && Scale == 1) {
1544 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1548 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1549 ScalarEvolution &SE, int64_t MinOffset,
1550 int64_t MaxOffset, LSRUse::KindType Kind,
1551 MemAccessTy AccessTy, const SCEV *S,
1553 // Fast-path: zero is always foldable.
1554 if (S->isZero()) return true;
1556 // Conservatively, create an address with an immediate and a
1557 // base and a scale.
1558 int64_t BaseOffset = ExtractImmediate(S, SE);
1559 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1561 // If there's anything else involved, it's not foldable.
1562 if (!S->isZero()) return false;
1564 // Fast-path: zero is always foldable.
1565 if (BaseOffset == 0 && !BaseGV) return true;
1567 // Conservatively, create an address with an immediate and a
1568 // base and a scale.
1569 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1571 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1572 BaseOffset, HasBaseReg, Scale);
1577 /// An individual increment in a Chain of IV increments. Relate an IV user to
1578 /// an expression that computes the IV it uses from the IV used by the previous
1579 /// link in the Chain.
1581 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1582 /// original IVOperand. The head of the chain's IVOperand is only valid during
1583 /// chain collection, before LSR replaces IV users. During chain generation,
1584 /// IncExpr can be used to find the new IVOperand that computes the same
1587 Instruction *UserInst;
1589 const SCEV *IncExpr;
1591 IVInc(Instruction *U, Value *O, const SCEV *E):
1592 UserInst(U), IVOperand(O), IncExpr(E) {}
1595 // The list of IV increments in program order. We typically add the head of a
1596 // chain without finding subsequent links.
1598 SmallVector<IVInc,1> Incs;
1599 const SCEV *ExprBase;
1601 IVChain() : ExprBase(nullptr) {}
1603 IVChain(const IVInc &Head, const SCEV *Base)
1604 : Incs(1, Head), ExprBase(Base) {}
1606 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1608 // Return the first increment in the chain.
1609 const_iterator begin() const {
1610 assert(!Incs.empty());
1611 return std::next(Incs.begin());
1613 const_iterator end() const {
1617 // Returns true if this chain contains any increments.
1618 bool hasIncs() const { return Incs.size() >= 2; }
1620 // Add an IVInc to the end of this chain.
1621 void add(const IVInc &X) { Incs.push_back(X); }
1623 // Returns the last UserInst in the chain.
1624 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1626 // Returns true if IncExpr can be profitably added to this chain.
1627 bool isProfitableIncrement(const SCEV *OperExpr,
1628 const SCEV *IncExpr,
1632 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1633 /// between FarUsers that definitely cross IV increments and NearUsers that may
1634 /// be used between IV increments.
1636 SmallPtrSet<Instruction*, 4> FarUsers;
1637 SmallPtrSet<Instruction*, 4> NearUsers;
1640 /// This class holds state for the main loop strength reduction logic.
1643 ScalarEvolution &SE;
1646 const TargetTransformInfo &TTI;
1650 /// This is the insert position that the current loop's induction variable
1651 /// increment should be placed. In simple loops, this is the latch block's
1652 /// terminator. But in more complicated cases, this is a position which will
1653 /// dominate all the in-loop post-increment users.
1654 Instruction *IVIncInsertPos;
1656 /// Interesting factors between use strides.
1657 SmallSetVector<int64_t, 8> Factors;
1659 /// Interesting use types, to facilitate truncation reuse.
1660 SmallSetVector<Type *, 4> Types;
1662 /// The list of operands which are to be replaced.
1663 SmallVector<LSRFixup, 16> Fixups;
1665 /// The list of interesting uses.
1666 SmallVector<LSRUse, 16> Uses;
1668 /// Track which uses use which register candidates.
1669 RegUseTracker RegUses;
1671 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1672 // have more than a few IV increment chains in a loop. Missing a Chain falls
1673 // back to normal LSR behavior for those uses.
1674 static const unsigned MaxChains = 8;
1676 /// IV users can form a chain of IV increments.
1677 SmallVector<IVChain, MaxChains> IVChainVec;
1679 /// IV users that belong to profitable IVChains.
1680 SmallPtrSet<Use*, MaxChains> IVIncSet;
1682 void OptimizeShadowIV();
1683 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1684 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1685 void OptimizeLoopTermCond();
1687 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1688 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1689 void FinalizeChain(IVChain &Chain);
1690 void CollectChains();
1691 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1692 SmallVectorImpl<WeakVH> &DeadInsts);
1694 void CollectInterestingTypesAndFactors();
1695 void CollectFixupsAndInitialFormulae();
1697 LSRFixup &getNewFixup() {
1698 Fixups.push_back(LSRFixup());
1699 return Fixups.back();
1702 // Support for sharing of LSRUses between LSRFixups.
1703 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1706 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1707 LSRUse::KindType Kind, MemAccessTy AccessTy);
1709 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1710 MemAccessTy AccessTy);
1712 void DeleteUse(LSRUse &LU, size_t LUIdx);
1714 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1716 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1717 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1718 void CountRegisters(const Formula &F, size_t LUIdx);
1719 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1721 void CollectLoopInvariantFixupsAndFormulae();
1723 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1724 unsigned Depth = 0);
1726 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1727 const Formula &Base, unsigned Depth,
1728 size_t Idx, bool IsScaledReg = false);
1729 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1730 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1731 const Formula &Base, size_t Idx,
1732 bool IsScaledReg = false);
1733 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1734 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1735 const Formula &Base,
1736 const SmallVectorImpl<int64_t> &Worklist,
1737 size_t Idx, bool IsScaledReg = false);
1738 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1739 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1740 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1741 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1742 void GenerateCrossUseConstantOffsets();
1743 void GenerateAllReuseFormulae();
1745 void FilterOutUndesirableDedicatedRegisters();
1747 size_t EstimateSearchSpaceComplexity() const;
1748 void NarrowSearchSpaceByDetectingSupersets();
1749 void NarrowSearchSpaceByCollapsingUnrolledCode();
1750 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1751 void NarrowSearchSpaceByPickingWinnerRegs();
1752 void NarrowSearchSpaceUsingHeuristics();
1754 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1756 SmallVectorImpl<const Formula *> &Workspace,
1757 const Cost &CurCost,
1758 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1759 DenseSet<const SCEV *> &VisitedRegs) const;
1760 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1762 BasicBlock::iterator
1763 HoistInsertPosition(BasicBlock::iterator IP,
1764 const SmallVectorImpl<Instruction *> &Inputs) const;
1765 BasicBlock::iterator
1766 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1769 SCEVExpander &Rewriter) const;
1771 Value *Expand(const LSRFixup &LF,
1773 BasicBlock::iterator IP,
1774 SCEVExpander &Rewriter,
1775 SmallVectorImpl<WeakVH> &DeadInsts) const;
1776 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1778 SCEVExpander &Rewriter,
1779 SmallVectorImpl<WeakVH> &DeadInsts) const;
1780 void Rewrite(const LSRFixup &LF,
1782 SCEVExpander &Rewriter,
1783 SmallVectorImpl<WeakVH> &DeadInsts) const;
1784 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
1787 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
1788 LoopInfo &LI, const TargetTransformInfo &TTI);
1790 bool getChanged() const { return Changed; }
1792 void print_factors_and_types(raw_ostream &OS) const;
1793 void print_fixups(raw_ostream &OS) const;
1794 void print_uses(raw_ostream &OS) const;
1795 void print(raw_ostream &OS) const;
1801 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
1802 /// the cast operation.
1803 void LSRInstance::OptimizeShadowIV() {
1804 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1805 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1808 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1809 UI != E; /* empty */) {
1810 IVUsers::const_iterator CandidateUI = UI;
1812 Instruction *ShadowUse = CandidateUI->getUser();
1813 Type *DestTy = nullptr;
1814 bool IsSigned = false;
1816 /* If shadow use is a int->float cast then insert a second IV
1817 to eliminate this cast.
1819 for (unsigned i = 0; i < n; ++i)
1825 for (unsigned i = 0; i < n; ++i, ++d)
1828 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1830 DestTy = UCast->getDestTy();
1832 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1834 DestTy = SCast->getDestTy();
1836 if (!DestTy) continue;
1838 // If target does not support DestTy natively then do not apply
1839 // this transformation.
1840 if (!TTI.isTypeLegal(DestTy)) continue;
1842 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1844 if (PH->getNumIncomingValues() != 2) continue;
1846 Type *SrcTy = PH->getType();
1847 int Mantissa = DestTy->getFPMantissaWidth();
1848 if (Mantissa == -1) continue;
1849 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1852 unsigned Entry, Latch;
1853 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1861 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1862 if (!Init) continue;
1863 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1864 (double)Init->getSExtValue() :
1865 (double)Init->getZExtValue());
1867 BinaryOperator *Incr =
1868 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1869 if (!Incr) continue;
1870 if (Incr->getOpcode() != Instruction::Add
1871 && Incr->getOpcode() != Instruction::Sub)
1874 /* Initialize new IV, double d = 0.0 in above example. */
1875 ConstantInt *C = nullptr;
1876 if (Incr->getOperand(0) == PH)
1877 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1878 else if (Incr->getOperand(1) == PH)
1879 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1885 // Ignore negative constants, as the code below doesn't handle them
1886 // correctly. TODO: Remove this restriction.
1887 if (!C->getValue().isStrictlyPositive()) continue;
1889 /* Add new PHINode. */
1890 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1892 /* create new increment. '++d' in above example. */
1893 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1894 BinaryOperator *NewIncr =
1895 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1896 Instruction::FAdd : Instruction::FSub,
1897 NewPH, CFP, "IV.S.next.", Incr);
1899 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1900 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1902 /* Remove cast operation */
1903 ShadowUse->replaceAllUsesWith(NewPH);
1904 ShadowUse->eraseFromParent();
1910 /// If Cond has an operand that is an expression of an IV, set the IV user and
1911 /// stride information and return true, otherwise return false.
1912 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1913 for (IVStrideUse &U : IU)
1914 if (U.getUser() == Cond) {
1915 // NOTE: we could handle setcc instructions with multiple uses here, but
1916 // InstCombine does it as well for simple uses, it's not clear that it
1917 // occurs enough in real life to handle.
1924 /// Rewrite the loop's terminating condition if it uses a max computation.
1926 /// This is a narrow solution to a specific, but acute, problem. For loops
1932 /// } while (++i < n);
1934 /// the trip count isn't just 'n', because 'n' might not be positive. And
1935 /// unfortunately this can come up even for loops where the user didn't use
1936 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1937 /// will commonly be lowered like this:
1943 /// } while (++i < n);
1946 /// and then it's possible for subsequent optimization to obscure the if
1947 /// test in such a way that indvars can't find it.
1949 /// When indvars can't find the if test in loops like this, it creates a
1950 /// max expression, which allows it to give the loop a canonical
1951 /// induction variable:
1954 /// max = n < 1 ? 1 : n;
1957 /// } while (++i != max);
1959 /// Canonical induction variables are necessary because the loop passes
1960 /// are designed around them. The most obvious example of this is the
1961 /// LoopInfo analysis, which doesn't remember trip count values. It
1962 /// expects to be able to rediscover the trip count each time it is
1963 /// needed, and it does this using a simple analysis that only succeeds if
1964 /// the loop has a canonical induction variable.
1966 /// However, when it comes time to generate code, the maximum operation
1967 /// can be quite costly, especially if it's inside of an outer loop.
1969 /// This function solves this problem by detecting this type of loop and
1970 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1971 /// the instructions for the maximum computation.
1973 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1974 // Check that the loop matches the pattern we're looking for.
1975 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1976 Cond->getPredicate() != CmpInst::ICMP_NE)
1979 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1980 if (!Sel || !Sel->hasOneUse()) return Cond;
1982 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1983 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1985 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1987 // Add one to the backedge-taken count to get the trip count.
1988 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1989 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1991 // Check for a max calculation that matches the pattern. There's no check
1992 // for ICMP_ULE here because the comparison would be with zero, which
1993 // isn't interesting.
1994 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1995 const SCEVNAryExpr *Max = nullptr;
1996 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1997 Pred = ICmpInst::ICMP_SLE;
1999 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2000 Pred = ICmpInst::ICMP_SLT;
2002 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2003 Pred = ICmpInst::ICMP_ULT;
2010 // To handle a max with more than two operands, this optimization would
2011 // require additional checking and setup.
2012 if (Max->getNumOperands() != 2)
2015 const SCEV *MaxLHS = Max->getOperand(0);
2016 const SCEV *MaxRHS = Max->getOperand(1);
2018 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2019 // for a comparison with 1. For <= and >=, a comparison with zero.
2021 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2024 // Check the relevant induction variable for conformance to
2026 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2027 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2028 if (!AR || !AR->isAffine() ||
2029 AR->getStart() != One ||
2030 AR->getStepRecurrence(SE) != One)
2033 assert(AR->getLoop() == L &&
2034 "Loop condition operand is an addrec in a different loop!");
2036 // Check the right operand of the select, and remember it, as it will
2037 // be used in the new comparison instruction.
2038 Value *NewRHS = nullptr;
2039 if (ICmpInst::isTrueWhenEqual(Pred)) {
2040 // Look for n+1, and grab n.
2041 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2042 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2043 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2044 NewRHS = BO->getOperand(0);
2045 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2046 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2047 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2048 NewRHS = BO->getOperand(0);
2051 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2052 NewRHS = Sel->getOperand(1);
2053 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2054 NewRHS = Sel->getOperand(2);
2055 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2056 NewRHS = SU->getValue();
2058 // Max doesn't match expected pattern.
2061 // Determine the new comparison opcode. It may be signed or unsigned,
2062 // and the original comparison may be either equality or inequality.
2063 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2064 Pred = CmpInst::getInversePredicate(Pred);
2066 // Ok, everything looks ok to change the condition into an SLT or SGE and
2067 // delete the max calculation.
2069 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2071 // Delete the max calculation instructions.
2072 Cond->replaceAllUsesWith(NewCond);
2073 CondUse->setUser(NewCond);
2074 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2075 Cond->eraseFromParent();
2076 Sel->eraseFromParent();
2077 if (Cmp->use_empty())
2078 Cmp->eraseFromParent();
2082 /// Change loop terminating condition to use the postinc iv when possible.
2084 LSRInstance::OptimizeLoopTermCond() {
2085 SmallPtrSet<Instruction *, 4> PostIncs;
2087 BasicBlock *LatchBlock = L->getLoopLatch();
2088 SmallVector<BasicBlock*, 8> ExitingBlocks;
2089 L->getExitingBlocks(ExitingBlocks);
2091 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2093 // Get the terminating condition for the loop if possible. If we
2094 // can, we want to change it to use a post-incremented version of its
2095 // induction variable, to allow coalescing the live ranges for the IV into
2096 // one register value.
2098 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2101 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2102 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2105 // Search IVUsesByStride to find Cond's IVUse if there is one.
2106 IVStrideUse *CondUse = nullptr;
2107 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2108 if (!FindIVUserForCond(Cond, CondUse))
2111 // If the trip count is computed in terms of a max (due to ScalarEvolution
2112 // being unable to find a sufficient guard, for example), change the loop
2113 // comparison to use SLT or ULT instead of NE.
2114 // One consequence of doing this now is that it disrupts the count-down
2115 // optimization. That's not always a bad thing though, because in such
2116 // cases it may still be worthwhile to avoid a max.
2117 Cond = OptimizeMax(Cond, CondUse);
2119 // If this exiting block dominates the latch block, it may also use
2120 // the post-inc value if it won't be shared with other uses.
2121 // Check for dominance.
2122 if (!DT.dominates(ExitingBlock, LatchBlock))
2125 // Conservatively avoid trying to use the post-inc value in non-latch
2126 // exits if there may be pre-inc users in intervening blocks.
2127 if (LatchBlock != ExitingBlock)
2128 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2129 // Test if the use is reachable from the exiting block. This dominator
2130 // query is a conservative approximation of reachability.
2131 if (&*UI != CondUse &&
2132 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2133 // Conservatively assume there may be reuse if the quotient of their
2134 // strides could be a legal scale.
2135 const SCEV *A = IU.getStride(*CondUse, L);
2136 const SCEV *B = IU.getStride(*UI, L);
2137 if (!A || !B) continue;
2138 if (SE.getTypeSizeInBits(A->getType()) !=
2139 SE.getTypeSizeInBits(B->getType())) {
2140 if (SE.getTypeSizeInBits(A->getType()) >
2141 SE.getTypeSizeInBits(B->getType()))
2142 B = SE.getSignExtendExpr(B, A->getType());
2144 A = SE.getSignExtendExpr(A, B->getType());
2146 if (const SCEVConstant *D =
2147 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2148 const ConstantInt *C = D->getValue();
2149 // Stride of one or negative one can have reuse with non-addresses.
2150 if (C->isOne() || C->isAllOnesValue())
2151 goto decline_post_inc;
2152 // Avoid weird situations.
2153 if (C->getValue().getMinSignedBits() >= 64 ||
2154 C->getValue().isMinSignedValue())
2155 goto decline_post_inc;
2156 // Check for possible scaled-address reuse.
2157 MemAccessTy AccessTy = getAccessType(UI->getUser());
2158 int64_t Scale = C->getSExtValue();
2159 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2161 /*HasBaseReg=*/false, Scale,
2162 AccessTy.AddrSpace))
2163 goto decline_post_inc;
2165 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2167 /*HasBaseReg=*/false, Scale,
2168 AccessTy.AddrSpace))
2169 goto decline_post_inc;
2173 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2176 // It's possible for the setcc instruction to be anywhere in the loop, and
2177 // possible for it to have multiple users. If it is not immediately before
2178 // the exiting block branch, move it.
2179 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2180 if (Cond->hasOneUse()) {
2181 Cond->moveBefore(TermBr);
2183 // Clone the terminating condition and insert into the loopend.
2184 ICmpInst *OldCond = Cond;
2185 Cond = cast<ICmpInst>(Cond->clone());
2186 Cond->setName(L->getHeader()->getName() + ".termcond");
2187 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2189 // Clone the IVUse, as the old use still exists!
2190 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2191 TermBr->replaceUsesOfWith(OldCond, Cond);
2195 // If we get to here, we know that we can transform the setcc instruction to
2196 // use the post-incremented version of the IV, allowing us to coalesce the
2197 // live ranges for the IV correctly.
2198 CondUse->transformToPostInc(L);
2201 PostIncs.insert(Cond);
2205 // Determine an insertion point for the loop induction variable increment. It
2206 // must dominate all the post-inc comparisons we just set up, and it must
2207 // dominate the loop latch edge.
2208 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2209 for (Instruction *Inst : PostIncs) {
2211 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2213 if (BB == Inst->getParent())
2214 IVIncInsertPos = Inst;
2215 else if (BB != IVIncInsertPos->getParent())
2216 IVIncInsertPos = BB->getTerminator();
2220 /// Determine if the given use can accommodate a fixup at the given offset and
2221 /// other details. If so, update the use and return true.
2222 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2223 bool HasBaseReg, LSRUse::KindType Kind,
2224 MemAccessTy AccessTy) {
2225 int64_t NewMinOffset = LU.MinOffset;
2226 int64_t NewMaxOffset = LU.MaxOffset;
2227 MemAccessTy NewAccessTy = AccessTy;
2229 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2230 // something conservative, however this can pessimize in the case that one of
2231 // the uses will have all its uses outside the loop, for example.
2232 if (LU.Kind != Kind)
2235 // Check for a mismatched access type, and fall back conservatively as needed.
2236 // TODO: Be less conservative when the type is similar and can use the same
2237 // addressing modes.
2238 if (Kind == LSRUse::Address) {
2239 if (AccessTy != LU.AccessTy)
2240 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext());
2243 // Conservatively assume HasBaseReg is true for now.
2244 if (NewOffset < LU.MinOffset) {
2245 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2246 LU.MaxOffset - NewOffset, HasBaseReg))
2248 NewMinOffset = NewOffset;
2249 } else if (NewOffset > LU.MaxOffset) {
2250 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2251 NewOffset - LU.MinOffset, HasBaseReg))
2253 NewMaxOffset = NewOffset;
2257 LU.MinOffset = NewMinOffset;
2258 LU.MaxOffset = NewMaxOffset;
2259 LU.AccessTy = NewAccessTy;
2260 if (NewOffset != LU.Offsets.back())
2261 LU.Offsets.push_back(NewOffset);
2265 /// Return an LSRUse index and an offset value for a fixup which needs the given
2266 /// expression, with the given kind and optional access type. Either reuse an
2267 /// existing use or create a new one, as needed.
2268 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2269 LSRUse::KindType Kind,
2270 MemAccessTy AccessTy) {
2271 const SCEV *Copy = Expr;
2272 int64_t Offset = ExtractImmediate(Expr, SE);
2274 // Basic uses can't accept any offset, for example.
2275 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2276 Offset, /*HasBaseReg=*/ true)) {
2281 std::pair<UseMapTy::iterator, bool> P =
2282 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2284 // A use already existed with this base.
2285 size_t LUIdx = P.first->second;
2286 LSRUse &LU = Uses[LUIdx];
2287 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2289 return std::make_pair(LUIdx, Offset);
2292 // Create a new use.
2293 size_t LUIdx = Uses.size();
2294 P.first->second = LUIdx;
2295 Uses.push_back(LSRUse(Kind, AccessTy));
2296 LSRUse &LU = Uses[LUIdx];
2298 // We don't need to track redundant offsets, but we don't need to go out
2299 // of our way here to avoid them.
2300 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2301 LU.Offsets.push_back(Offset);
2303 LU.MinOffset = Offset;
2304 LU.MaxOffset = Offset;
2305 return std::make_pair(LUIdx, Offset);
2308 /// Delete the given use from the Uses list.
2309 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2310 if (&LU != &Uses.back())
2311 std::swap(LU, Uses.back());
2315 RegUses.swapAndDropUse(LUIdx, Uses.size());
2318 /// Look for a use distinct from OrigLU which is has a formula that has the same
2319 /// registers as the given formula.
2321 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2322 const LSRUse &OrigLU) {
2323 // Search all uses for the formula. This could be more clever.
2324 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2325 LSRUse &LU = Uses[LUIdx];
2326 // Check whether this use is close enough to OrigLU, to see whether it's
2327 // worthwhile looking through its formulae.
2328 // Ignore ICmpZero uses because they may contain formulae generated by
2329 // GenerateICmpZeroScales, in which case adding fixup offsets may
2331 if (&LU != &OrigLU &&
2332 LU.Kind != LSRUse::ICmpZero &&
2333 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2334 LU.WidestFixupType == OrigLU.WidestFixupType &&
2335 LU.HasFormulaWithSameRegs(OrigF)) {
2336 // Scan through this use's formulae.
2337 for (const Formula &F : LU.Formulae) {
2338 // Check to see if this formula has the same registers and symbols
2340 if (F.BaseRegs == OrigF.BaseRegs &&
2341 F.ScaledReg == OrigF.ScaledReg &&
2342 F.BaseGV == OrigF.BaseGV &&
2343 F.Scale == OrigF.Scale &&
2344 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2345 if (F.BaseOffset == 0)
2347 // This is the formula where all the registers and symbols matched;
2348 // there aren't going to be any others. Since we declined it, we
2349 // can skip the rest of the formulae and proceed to the next LSRUse.
2356 // Nothing looked good.
2360 void LSRInstance::CollectInterestingTypesAndFactors() {
2361 SmallSetVector<const SCEV *, 4> Strides;
2363 // Collect interesting types and strides.
2364 SmallVector<const SCEV *, 4> Worklist;
2365 for (const IVStrideUse &U : IU) {
2366 const SCEV *Expr = IU.getExpr(U);
2368 // Collect interesting types.
2369 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2371 // Add strides for mentioned loops.
2372 Worklist.push_back(Expr);
2374 const SCEV *S = Worklist.pop_back_val();
2375 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2376 if (AR->getLoop() == L)
2377 Strides.insert(AR->getStepRecurrence(SE));
2378 Worklist.push_back(AR->getStart());
2379 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2380 Worklist.append(Add->op_begin(), Add->op_end());
2382 } while (!Worklist.empty());
2385 // Compute interesting factors from the set of interesting strides.
2386 for (SmallSetVector<const SCEV *, 4>::const_iterator
2387 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2388 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2389 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2390 const SCEV *OldStride = *I;
2391 const SCEV *NewStride = *NewStrideIter;
2393 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2394 SE.getTypeSizeInBits(NewStride->getType())) {
2395 if (SE.getTypeSizeInBits(OldStride->getType()) >
2396 SE.getTypeSizeInBits(NewStride->getType()))
2397 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2399 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2401 if (const SCEVConstant *Factor =
2402 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2404 if (Factor->getAPInt().getMinSignedBits() <= 64)
2405 Factors.insert(Factor->getAPInt().getSExtValue());
2406 } else if (const SCEVConstant *Factor =
2407 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2410 if (Factor->getAPInt().getMinSignedBits() <= 64)
2411 Factors.insert(Factor->getAPInt().getSExtValue());
2415 // If all uses use the same type, don't bother looking for truncation-based
2417 if (Types.size() == 1)
2420 DEBUG(print_factors_and_types(dbgs()));
2423 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2424 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2425 /// IVStrideUses, we could partially skip this.
2426 static User::op_iterator
2427 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2428 Loop *L, ScalarEvolution &SE) {
2429 for(; OI != OE; ++OI) {
2430 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2431 if (!SE.isSCEVable(Oper->getType()))
2434 if (const SCEVAddRecExpr *AR =
2435 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2436 if (AR->getLoop() == L)
2444 /// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2445 /// a convenient helper.
2446 static Value *getWideOperand(Value *Oper) {
2447 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2448 return Trunc->getOperand(0);
2452 /// Return true if we allow an IV chain to include both types.
2453 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2454 Type *LType = LVal->getType();
2455 Type *RType = RVal->getType();
2456 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2459 /// Return an approximation of this SCEV expression's "base", or NULL for any
2460 /// constant. Returning the expression itself is conservative. Returning a
2461 /// deeper subexpression is more precise and valid as long as it isn't less
2462 /// complex than another subexpression. For expressions involving multiple
2463 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2464 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2467 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2468 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2469 static const SCEV *getExprBase(const SCEV *S) {
2470 switch (S->getSCEVType()) {
2471 default: // uncluding scUnknown.
2476 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2478 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2480 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2482 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2483 // there's nothing more complex.
2484 // FIXME: not sure if we want to recognize negation.
2485 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2486 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2487 E(Add->op_begin()); I != E; ++I) {
2488 const SCEV *SubExpr = *I;
2489 if (SubExpr->getSCEVType() == scAddExpr)
2490 return getExprBase(SubExpr);
2492 if (SubExpr->getSCEVType() != scMulExpr)
2495 return S; // all operands are scaled, be conservative.
2498 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2502 /// Return true if the chain increment is profitable to expand into a loop
2503 /// invariant value, which may require its own register. A profitable chain
2504 /// increment will be an offset relative to the same base. We allow such offsets
2505 /// to potentially be used as chain increment as long as it's not obviously
2506 /// expensive to expand using real instructions.
2507 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2508 const SCEV *IncExpr,
2509 ScalarEvolution &SE) {
2510 // Aggressively form chains when -stress-ivchain.
2514 // Do not replace a constant offset from IV head with a nonconstant IV
2516 if (!isa<SCEVConstant>(IncExpr)) {
2517 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2518 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2522 SmallPtrSet<const SCEV*, 8> Processed;
2523 return !isHighCostExpansion(IncExpr, Processed, SE);
2526 /// Return true if the number of registers needed for the chain is estimated to
2527 /// be less than the number required for the individual IV users. First prohibit
2528 /// any IV users that keep the IV live across increments (the Users set should
2529 /// be empty). Next count the number and type of increments in the chain.
2531 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2532 /// effectively use postinc addressing modes. Only consider it profitable it the
2533 /// increments can be computed in fewer registers when chained.
2535 /// TODO: Consider IVInc free if it's already used in another chains.
2537 isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2538 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2542 if (!Chain.hasIncs())
2545 if (!Users.empty()) {
2546 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2547 for (Instruction *Inst : Users) {
2548 dbgs() << " " << *Inst << "\n";
2552 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2554 // The chain itself may require a register, so intialize cost to 1.
2557 // A complete chain likely eliminates the need for keeping the original IV in
2558 // a register. LSR does not currently know how to form a complete chain unless
2559 // the header phi already exists.
2560 if (isa<PHINode>(Chain.tailUserInst())
2561 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2564 const SCEV *LastIncExpr = nullptr;
2565 unsigned NumConstIncrements = 0;
2566 unsigned NumVarIncrements = 0;
2567 unsigned NumReusedIncrements = 0;
2568 for (const IVInc &Inc : Chain) {
2569 if (Inc.IncExpr->isZero())
2572 // Incrementing by zero or some constant is neutral. We assume constants can
2573 // be folded into an addressing mode or an add's immediate operand.
2574 if (isa<SCEVConstant>(Inc.IncExpr)) {
2575 ++NumConstIncrements;
2579 if (Inc.IncExpr == LastIncExpr)
2580 ++NumReusedIncrements;
2584 LastIncExpr = Inc.IncExpr;
2586 // An IV chain with a single increment is handled by LSR's postinc
2587 // uses. However, a chain with multiple increments requires keeping the IV's
2588 // value live longer than it needs to be if chained.
2589 if (NumConstIncrements > 1)
2592 // Materializing increment expressions in the preheader that didn't exist in
2593 // the original code may cost a register. For example, sign-extended array
2594 // indices can produce ridiculous increments like this:
2595 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2596 cost += NumVarIncrements;
2598 // Reusing variable increments likely saves a register to hold the multiple of
2600 cost -= NumReusedIncrements;
2602 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2608 /// Add this IV user to an existing chain or make it the head of a new chain.
2609 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2610 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2611 // When IVs are used as types of varying widths, they are generally converted
2612 // to a wider type with some uses remaining narrow under a (free) trunc.
2613 Value *const NextIV = getWideOperand(IVOper);
2614 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2615 const SCEV *const OperExprBase = getExprBase(OperExpr);
2617 // Visit all existing chains. Check if its IVOper can be computed as a
2618 // profitable loop invariant increment from the last link in the Chain.
2619 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2620 const SCEV *LastIncExpr = nullptr;
2621 for (; ChainIdx < NChains; ++ChainIdx) {
2622 IVChain &Chain = IVChainVec[ChainIdx];
2624 // Prune the solution space aggressively by checking that both IV operands
2625 // are expressions that operate on the same unscaled SCEVUnknown. This
2626 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2627 // first avoids creating extra SCEV expressions.
2628 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2631 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2632 if (!isCompatibleIVType(PrevIV, NextIV))
2635 // A phi node terminates a chain.
2636 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2639 // The increment must be loop-invariant so it can be kept in a register.
2640 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2641 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2642 if (!SE.isLoopInvariant(IncExpr, L))
2645 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2646 LastIncExpr = IncExpr;
2650 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2651 // bother for phi nodes, because they must be last in the chain.
2652 if (ChainIdx == NChains) {
2653 if (isa<PHINode>(UserInst))
2655 if (NChains >= MaxChains && !StressIVChain) {
2656 DEBUG(dbgs() << "IV Chain Limit\n");
2659 LastIncExpr = OperExpr;
2660 // IVUsers may have skipped over sign/zero extensions. We don't currently
2661 // attempt to form chains involving extensions unless they can be hoisted
2662 // into this loop's AddRec.
2663 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2666 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2668 ChainUsersVec.resize(NChains);
2669 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2670 << ") IV=" << *LastIncExpr << "\n");
2672 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2673 << ") IV+" << *LastIncExpr << "\n");
2674 // Add this IV user to the end of the chain.
2675 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2677 IVChain &Chain = IVChainVec[ChainIdx];
2679 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2680 // This chain's NearUsers become FarUsers.
2681 if (!LastIncExpr->isZero()) {
2682 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2687 // All other uses of IVOperand become near uses of the chain.
2688 // We currently ignore intermediate values within SCEV expressions, assuming
2689 // they will eventually be used be the current chain, or can be computed
2690 // from one of the chain increments. To be more precise we could
2691 // transitively follow its user and only add leaf IV users to the set.
2692 for (User *U : IVOper->users()) {
2693 Instruction *OtherUse = dyn_cast<Instruction>(U);
2696 // Uses in the chain will no longer be uses if the chain is formed.
2697 // Include the head of the chain in this iteration (not Chain.begin()).
2698 IVChain::const_iterator IncIter = Chain.Incs.begin();
2699 IVChain::const_iterator IncEnd = Chain.Incs.end();
2700 for( ; IncIter != IncEnd; ++IncIter) {
2701 if (IncIter->UserInst == OtherUse)
2704 if (IncIter != IncEnd)
2707 if (SE.isSCEVable(OtherUse->getType())
2708 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2709 && IU.isIVUserOrOperand(OtherUse)) {
2712 NearUsers.insert(OtherUse);
2715 // Since this user is part of the chain, it's no longer considered a use
2717 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2720 /// Populate the vector of Chains.
2722 /// This decreases ILP at the architecture level. Targets with ample registers,
2723 /// multiple memory ports, and no register renaming probably don't want
2724 /// this. However, such targets should probably disable LSR altogether.
2726 /// The job of LSR is to make a reasonable choice of induction variables across
2727 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2728 /// ILP *within the loop* if the target wants it.
2730 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2731 /// will not reorder memory operations, it will recognize this as a chain, but
2732 /// will generate redundant IV increments. Ideally this would be corrected later
2733 /// by a smart scheduler:
2739 /// TODO: Walk the entire domtree within this loop, not just the path to the
2740 /// loop latch. This will discover chains on side paths, but requires
2741 /// maintaining multiple copies of the Chains state.
2742 void LSRInstance::CollectChains() {
2743 DEBUG(dbgs() << "Collecting IV Chains.\n");
2744 SmallVector<ChainUsers, 8> ChainUsersVec;
2746 SmallVector<BasicBlock *,8> LatchPath;
2747 BasicBlock *LoopHeader = L->getHeader();
2748 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2749 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2750 LatchPath.push_back(Rung->getBlock());
2752 LatchPath.push_back(LoopHeader);
2754 // Walk the instruction stream from the loop header to the loop latch.
2755 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2756 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2757 BBIter != BBEnd; ++BBIter) {
2758 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2760 // Skip instructions that weren't seen by IVUsers analysis.
2761 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&*I))
2764 // Ignore users that are part of a SCEV expression. This way we only
2765 // consider leaf IV Users. This effectively rediscovers a portion of
2766 // IVUsers analysis but in program order this time.
2767 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(&*I)))
2770 // Remove this instruction from any NearUsers set it may be in.
2771 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2772 ChainIdx < NChains; ++ChainIdx) {
2773 ChainUsersVec[ChainIdx].NearUsers.erase(&*I);
2775 // Search for operands that can be chained.
2776 SmallPtrSet<Instruction*, 4> UniqueOperands;
2777 User::op_iterator IVOpEnd = I->op_end();
2778 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2779 while (IVOpIter != IVOpEnd) {
2780 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2781 if (UniqueOperands.insert(IVOpInst).second)
2782 ChainInstruction(&*I, IVOpInst, ChainUsersVec);
2783 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2785 } // Continue walking down the instructions.
2786 } // Continue walking down the domtree.
2787 // Visit phi backedges to determine if the chain can generate the IV postinc.
2788 for (BasicBlock::iterator I = L->getHeader()->begin();
2789 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2790 if (!SE.isSCEVable(PN->getType()))
2794 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2796 ChainInstruction(PN, IncV, ChainUsersVec);
2798 // Remove any unprofitable chains.
2799 unsigned ChainIdx = 0;
2800 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2801 UsersIdx < NChains; ++UsersIdx) {
2802 if (!isProfitableChain(IVChainVec[UsersIdx],
2803 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2805 // Preserve the chain at UsesIdx.
2806 if (ChainIdx != UsersIdx)
2807 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2808 FinalizeChain(IVChainVec[ChainIdx]);
2811 IVChainVec.resize(ChainIdx);
2814 void LSRInstance::FinalizeChain(IVChain &Chain) {
2815 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2816 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2818 for (const IVInc &Inc : Chain) {
2819 DEBUG(dbgs() << " Inc: " << Inc.UserInst << "\n");
2820 auto UseI = std::find(Inc.UserInst->op_begin(), Inc.UserInst->op_end(),
2822 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
2823 IVIncSet.insert(UseI);
2827 /// Return true if the IVInc can be folded into an addressing mode.
2828 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2829 Value *Operand, const TargetTransformInfo &TTI) {
2830 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2831 if (!IncConst || !isAddressUse(UserInst, Operand))
2834 if (IncConst->getAPInt().getMinSignedBits() > 64)
2837 MemAccessTy AccessTy = getAccessType(UserInst);
2838 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2839 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
2840 IncOffset, /*HaseBaseReg=*/false))
2846 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
2847 /// user's operand from the previous IV user's operand.
2848 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2849 SmallVectorImpl<WeakVH> &DeadInsts) {
2850 // Find the new IVOperand for the head of the chain. It may have been replaced
2852 const IVInc &Head = Chain.Incs[0];
2853 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2854 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2855 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2857 Value *IVSrc = nullptr;
2858 while (IVOpIter != IVOpEnd) {
2859 IVSrc = getWideOperand(*IVOpIter);
2861 // If this operand computes the expression that the chain needs, we may use
2862 // it. (Check this after setting IVSrc which is used below.)
2864 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2865 // narrow for the chain, so we can no longer use it. We do allow using a
2866 // wider phi, assuming the LSR checked for free truncation. In that case we
2867 // should already have a truncate on this operand such that
2868 // getSCEV(IVSrc) == IncExpr.
2869 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2870 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2873 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2875 if (IVOpIter == IVOpEnd) {
2876 // Gracefully give up on this chain.
2877 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2881 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2882 Type *IVTy = IVSrc->getType();
2883 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2884 const SCEV *LeftOverExpr = nullptr;
2885 for (const IVInc &Inc : Chain) {
2886 Instruction *InsertPt = Inc.UserInst;
2887 if (isa<PHINode>(InsertPt))
2888 InsertPt = L->getLoopLatch()->getTerminator();
2890 // IVOper will replace the current IV User's operand. IVSrc is the IV
2891 // value currently held in a register.
2892 Value *IVOper = IVSrc;
2893 if (!Inc.IncExpr->isZero()) {
2894 // IncExpr was the result of subtraction of two narrow values, so must
2896 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
2897 LeftOverExpr = LeftOverExpr ?
2898 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2900 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2901 // Expand the IV increment.
2902 Rewriter.clearPostInc();
2903 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2904 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2905 SE.getUnknown(IncV));
2906 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2908 // If an IV increment can't be folded, use it as the next IV value.
2909 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
2910 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2912 LeftOverExpr = nullptr;
2915 Type *OperTy = Inc.IVOperand->getType();
2916 if (IVTy != OperTy) {
2917 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2918 "cannot extend a chained IV");
2919 IRBuilder<> Builder(InsertPt);
2920 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2922 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
2923 DeadInsts.emplace_back(Inc.IVOperand);
2925 // If LSR created a new, wider phi, we may also replace its postinc. We only
2926 // do this if we also found a wide value for the head of the chain.
2927 if (isa<PHINode>(Chain.tailUserInst())) {
2928 for (BasicBlock::iterator I = L->getHeader()->begin();
2929 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2930 if (!isCompatibleIVType(Phi, IVSrc))
2932 Instruction *PostIncV = dyn_cast<Instruction>(
2933 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2934 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2936 Value *IVOper = IVSrc;
2937 Type *PostIncTy = PostIncV->getType();
2938 if (IVTy != PostIncTy) {
2939 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2940 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2941 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2942 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2944 Phi->replaceUsesOfWith(PostIncV, IVOper);
2945 DeadInsts.emplace_back(PostIncV);
2950 void LSRInstance::CollectFixupsAndInitialFormulae() {
2951 for (const IVStrideUse &U : IU) {
2952 Instruction *UserInst = U.getUser();
2953 // Skip IV users that are part of profitable IV Chains.
2954 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2955 U.getOperandValToReplace());
2956 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2957 if (IVIncSet.count(UseI))
2961 LSRFixup &LF = getNewFixup();
2962 LF.UserInst = UserInst;
2963 LF.OperandValToReplace = U.getOperandValToReplace();
2964 LF.PostIncLoops = U.getPostIncLoops();
2966 LSRUse::KindType Kind = LSRUse::Basic;
2967 MemAccessTy AccessTy;
2968 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2969 Kind = LSRUse::Address;
2970 AccessTy = getAccessType(LF.UserInst);
2973 const SCEV *S = IU.getExpr(U);
2975 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2976 // (N - i == 0), and this allows (N - i) to be the expression that we work
2977 // with rather than just N or i, so we can consider the register
2978 // requirements for both N and i at the same time. Limiting this code to
2979 // equality icmps is not a problem because all interesting loops use
2980 // equality icmps, thanks to IndVarSimplify.
2981 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2982 if (CI->isEquality()) {
2983 // Swap the operands if needed to put the OperandValToReplace on the
2984 // left, for consistency.
2985 Value *NV = CI->getOperand(1);
2986 if (NV == LF.OperandValToReplace) {
2987 CI->setOperand(1, CI->getOperand(0));
2988 CI->setOperand(0, NV);
2989 NV = CI->getOperand(1);
2993 // x == y --> x - y == 0
2994 const SCEV *N = SE.getSCEV(NV);
2995 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2996 // S is normalized, so normalize N before folding it into S
2997 // to keep the result normalized.
2998 N = TransformForPostIncUse(Normalize, N, CI, nullptr,
2999 LF.PostIncLoops, SE, DT);
3000 Kind = LSRUse::ICmpZero;
3001 S = SE.getMinusSCEV(N, S);
3004 // -1 and the negations of all interesting strides (except the negation
3005 // of -1) are now also interesting.
3006 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3007 if (Factors[i] != -1)
3008 Factors.insert(-(uint64_t)Factors[i]);
3012 // Set up the initial formula for this use.
3013 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3015 LF.Offset = P.second;
3016 LSRUse &LU = Uses[LF.LUIdx];
3017 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3018 if (!LU.WidestFixupType ||
3019 SE.getTypeSizeInBits(LU.WidestFixupType) <
3020 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3021 LU.WidestFixupType = LF.OperandValToReplace->getType();
3023 // If this is the first use of this LSRUse, give it a formula.
3024 if (LU.Formulae.empty()) {
3025 InsertInitialFormula(S, LU, LF.LUIdx);
3026 CountRegisters(LU.Formulae.back(), LF.LUIdx);
3030 DEBUG(print_fixups(dbgs()));
3033 /// Insert a formula for the given expression into the given use, separating out
3034 /// loop-variant portions from loop-invariant and loop-computable portions.
3036 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3037 // Mark uses whose expressions cannot be expanded.
3038 if (!isSafeToExpand(S, SE))
3039 LU.RigidFormula = true;
3042 F.initialMatch(S, L, SE);
3043 bool Inserted = InsertFormula(LU, LUIdx, F);
3044 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3047 /// Insert a simple single-register formula for the given expression into the
3050 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3051 LSRUse &LU, size_t LUIdx) {
3053 F.BaseRegs.push_back(S);
3054 F.HasBaseReg = true;
3055 bool Inserted = InsertFormula(LU, LUIdx, F);
3056 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3059 /// Note which registers are used by the given formula, updating RegUses.
3060 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3062 RegUses.countRegister(F.ScaledReg, LUIdx);
3063 for (const SCEV *BaseReg : F.BaseRegs)
3064 RegUses.countRegister(BaseReg, LUIdx);
3067 /// If the given formula has not yet been inserted, add it to the list, and
3068 /// return true. Return false otherwise.
3069 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3070 // Do not insert formula that we will not be able to expand.
3071 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3072 "Formula is illegal");
3073 if (!LU.InsertFormula(F))
3076 CountRegisters(F, LUIdx);
3080 /// Check for other uses of loop-invariant values which we're tracking. These
3081 /// other uses will pin these values in registers, making them less profitable
3082 /// for elimination.
3083 /// TODO: This currently misses non-constant addrec step registers.
3084 /// TODO: Should this give more weight to users inside the loop?
3086 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3087 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3088 SmallPtrSet<const SCEV *, 32> Visited;
3090 while (!Worklist.empty()) {
3091 const SCEV *S = Worklist.pop_back_val();
3093 // Don't process the same SCEV twice
3094 if (!Visited.insert(S).second)
3097 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3098 Worklist.append(N->op_begin(), N->op_end());
3099 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3100 Worklist.push_back(C->getOperand());
3101 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3102 Worklist.push_back(D->getLHS());
3103 Worklist.push_back(D->getRHS());
3104 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3105 const Value *V = US->getValue();
3106 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3107 // Look for instructions defined outside the loop.
3108 if (L->contains(Inst)) continue;
3109 } else if (isa<UndefValue>(V))
3110 // Undef doesn't have a live range, so it doesn't matter.
3112 for (const Use &U : V->uses()) {
3113 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3114 // Ignore non-instructions.
3117 // Ignore instructions in other functions (as can happen with
3119 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3121 // Ignore instructions not dominated by the loop.
3122 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3123 UserInst->getParent() :
3124 cast<PHINode>(UserInst)->getIncomingBlock(
3125 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3126 if (!DT.dominates(L->getHeader(), UseBB))
3128 // Don't bother if the instruction is in a BB which ends in an EHPad.
3129 if (UseBB->getTerminator()->isEHPad())
3131 // Ignore uses which are part of other SCEV expressions, to avoid
3132 // analyzing them multiple times.
3133 if (SE.isSCEVable(UserInst->getType())) {
3134 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3135 // If the user is a no-op, look through to its uses.
3136 if (!isa<SCEVUnknown>(UserS))
3140 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3144 // Ignore icmp instructions which are already being analyzed.
3145 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3146 unsigned OtherIdx = !U.getOperandNo();
3147 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3148 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3152 LSRFixup &LF = getNewFixup();
3153 LF.UserInst = const_cast<Instruction *>(UserInst);
3154 LF.OperandValToReplace = U;
3155 std::pair<size_t, int64_t> P = getUse(
3156 S, LSRUse::Basic, MemAccessTy());
3158 LF.Offset = P.second;
3159 LSRUse &LU = Uses[LF.LUIdx];
3160 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3161 if (!LU.WidestFixupType ||
3162 SE.getTypeSizeInBits(LU.WidestFixupType) <
3163 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3164 LU.WidestFixupType = LF.OperandValToReplace->getType();
3165 InsertSupplementalFormula(US, LU, LF.LUIdx);
3166 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3173 /// Split S into subexpressions which can be pulled out into separate
3174 /// registers. If C is non-null, multiply each subexpression by C.
3176 /// Return remainder expression after factoring the subexpressions captured by
3177 /// Ops. If Ops is complete, return NULL.
3178 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3179 SmallVectorImpl<const SCEV *> &Ops,
3181 ScalarEvolution &SE,
3182 unsigned Depth = 0) {
3183 // Arbitrarily cap recursion to protect compile time.
3187 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3188 // Break out add operands.
3189 for (const SCEV *S : Add->operands()) {
3190 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3192 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3195 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3196 // Split a non-zero base out of an addrec.
3197 if (AR->getStart()->isZero())
3200 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3201 C, Ops, L, SE, Depth+1);
3202 // Split the non-zero AddRec unless it is part of a nested recurrence that
3203 // does not pertain to this loop.
3204 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3205 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3206 Remainder = nullptr;
3208 if (Remainder != AR->getStart()) {
3210 Remainder = SE.getConstant(AR->getType(), 0);
3211 return SE.getAddRecExpr(Remainder,
3212 AR->getStepRecurrence(SE),
3214 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3217 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3218 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3219 if (Mul->getNumOperands() != 2)
3221 if (const SCEVConstant *Op0 =
3222 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3223 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3224 const SCEV *Remainder =
3225 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3227 Ops.push_back(SE.getMulExpr(C, Remainder));
3234 /// \brief Helper function for LSRInstance::GenerateReassociations.
3235 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3236 const Formula &Base,
3237 unsigned Depth, size_t Idx,
3239 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3240 SmallVector<const SCEV *, 8> AddOps;
3241 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3243 AddOps.push_back(Remainder);
3245 if (AddOps.size() == 1)
3248 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3252 // Loop-variant "unknown" values are uninteresting; we won't be able to
3253 // do anything meaningful with them.
3254 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3257 // Don't pull a constant into a register if the constant could be folded
3258 // into an immediate field.
3259 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3260 LU.AccessTy, *J, Base.getNumRegs() > 1))
3263 // Collect all operands except *J.
3264 SmallVector<const SCEV *, 8> InnerAddOps(
3265 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3266 InnerAddOps.append(std::next(J),
3267 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3269 // Don't leave just a constant behind in a register if the constant could
3270 // be folded into an immediate field.
3271 if (InnerAddOps.size() == 1 &&
3272 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3273 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3276 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3277 if (InnerSum->isZero())
3281 // Add the remaining pieces of the add back into the new formula.
3282 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3283 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3284 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3285 InnerSumSC->getValue()->getZExtValue())) {
3287 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3289 F.ScaledReg = nullptr;
3291 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3292 } else if (IsScaledReg)
3293 F.ScaledReg = InnerSum;
3295 F.BaseRegs[Idx] = InnerSum;
3297 // Add J as its own register, or an unfolded immediate.
3298 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3299 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3300 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3301 SC->getValue()->getZExtValue()))
3303 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3305 F.BaseRegs.push_back(*J);
3306 // We may have changed the number of register&nbs