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() << ']';
148 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
149 void RegSortData::dump() const {
150 print(errs()); errs() << '\n';
156 /// Map register candidates to information about how they are used.
157 class RegUseTracker {
158 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
160 RegUsesTy RegUsesMap;
161 SmallVector<const SCEV *, 16> RegSequence;
164 void countRegister(const SCEV *Reg, size_t LUIdx);
165 void dropRegister(const SCEV *Reg, size_t LUIdx);
166 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
168 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
170 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
174 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
175 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
176 iterator begin() { return RegSequence.begin(); }
177 iterator end() { return RegSequence.end(); }
178 const_iterator begin() const { return RegSequence.begin(); }
179 const_iterator end() const { return RegSequence.end(); }
185 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
186 std::pair<RegUsesTy::iterator, bool> Pair =
187 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
188 RegSortData &RSD = Pair.first->second;
190 RegSequence.push_back(Reg);
191 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
192 RSD.UsedByIndices.set(LUIdx);
196 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
197 RegUsesTy::iterator It = RegUsesMap.find(Reg);
198 assert(It != RegUsesMap.end());
199 RegSortData &RSD = It->second;
200 assert(RSD.UsedByIndices.size() > LUIdx);
201 RSD.UsedByIndices.reset(LUIdx);
205 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
206 assert(LUIdx <= LastLUIdx);
208 // Update RegUses. The data structure is not optimized for this purpose;
209 // we must iterate through it and update each of the bit vectors.
210 for (auto &Pair : RegUsesMap) {
211 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
212 if (LUIdx < UsedByIndices.size())
213 UsedByIndices[LUIdx] =
214 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
215 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
220 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
221 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
222 if (I == RegUsesMap.end())
224 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
225 int i = UsedByIndices.find_first();
226 if (i == -1) return false;
227 if ((size_t)i != LUIdx) return true;
228 return UsedByIndices.find_next(i) != -1;
231 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
232 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
233 assert(I != RegUsesMap.end() && "Unknown register!");
234 return I->second.UsedByIndices;
237 void RegUseTracker::clear() {
244 /// This class holds information that describes a formula for computing
245 /// satisfying a use. It may include broken-out immediates and scaled registers.
247 /// Global base address used for complex addressing.
250 /// Base offset for complex addressing.
253 /// Whether any complex addressing has a base register.
256 /// The scale of any complex addressing.
259 /// The list of "base" registers for this use. When this is non-empty. The
260 /// canonical representation of a formula is
261 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
262 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
263 /// #1 enforces that the scaled register is always used when at least two
264 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
265 /// #2 enforces that 1 * reg is reg.
266 /// This invariant can be temporarly broken while building a formula.
267 /// However, every formula inserted into the LSRInstance must be in canonical
269 SmallVector<const SCEV *, 4> BaseRegs;
271 /// The 'scaled' register for this use. This should be non-null when Scale is
273 const SCEV *ScaledReg;
275 /// An additional constant offset which added near the use. This requires a
276 /// temporary register, but the offset itself can live in an add immediate
277 /// field rather than a register.
278 int64_t UnfoldedOffset;
281 : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
282 ScaledReg(nullptr), UnfoldedOffset(0) {}
284 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
286 bool isCanonical() const;
292 size_t getNumRegs() const;
293 Type *getType() const;
295 void deleteBaseReg(const SCEV *&S);
297 bool referencesReg(const SCEV *S) const;
298 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
299 const RegUseTracker &RegUses) const;
301 void print(raw_ostream &OS) const;
307 /// Recursion helper for initialMatch.
308 static void DoInitialMatch(const SCEV *S, Loop *L,
309 SmallVectorImpl<const SCEV *> &Good,
310 SmallVectorImpl<const SCEV *> &Bad,
311 ScalarEvolution &SE) {
312 // Collect expressions which properly dominate the loop header.
313 if (SE.properlyDominates(S, L->getHeader())) {
318 // Look at add operands.
319 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
320 for (const SCEV *S : Add->operands())
321 DoInitialMatch(S, L, Good, Bad, SE);
325 // Look at addrec operands.
326 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
327 if (!AR->getStart()->isZero()) {
328 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
329 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
330 AR->getStepRecurrence(SE),
331 // FIXME: AR->getNoWrapFlags()
332 AR->getLoop(), SCEV::FlagAnyWrap),
337 // Handle a multiplication by -1 (negation) if it didn't fold.
338 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
339 if (Mul->getOperand(0)->isAllOnesValue()) {
340 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
341 const SCEV *NewMul = SE.getMulExpr(Ops);
343 SmallVector<const SCEV *, 4> MyGood;
344 SmallVector<const SCEV *, 4> MyBad;
345 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
346 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
347 SE.getEffectiveSCEVType(NewMul->getType())));
348 for (const SCEV *S : MyGood)
349 Good.push_back(SE.getMulExpr(NegOne, S));
350 for (const SCEV *S : MyBad)
351 Bad.push_back(SE.getMulExpr(NegOne, S));
355 // Ok, we can't do anything interesting. Just stuff the whole thing into a
356 // register and hope for the best.
360 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
361 /// all loop-invariant and loop-computable values in a single base register.
362 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
363 SmallVector<const SCEV *, 4> Good;
364 SmallVector<const SCEV *, 4> Bad;
365 DoInitialMatch(S, L, Good, Bad, SE);
367 const SCEV *Sum = SE.getAddExpr(Good);
369 BaseRegs.push_back(Sum);
373 const SCEV *Sum = SE.getAddExpr(Bad);
375 BaseRegs.push_back(Sum);
381 /// \brief Check whether or not this formula statisfies the canonical
383 /// \see Formula::BaseRegs.
384 bool Formula::isCanonical() const {
386 return Scale != 1 || !BaseRegs.empty();
387 return BaseRegs.size() <= 1;
390 /// \brief Helper method to morph a formula into its canonical representation.
391 /// \see Formula::BaseRegs.
392 /// Every formula having more than one base register, must use the ScaledReg
393 /// field. Otherwise, we would have to do special cases everywhere in LSR
394 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
395 /// On the other hand, 1*reg should be canonicalized into reg.
396 void Formula::canonicalize() {
399 // So far we did not need this case. This is easy to implement but it is
400 // useless to maintain dead code. Beside it could hurt compile time.
401 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
402 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
403 ScaledReg = BaseRegs.back();
406 size_t BaseRegsSize = BaseRegs.size();
408 // If ScaledReg is an invariant, try to find a variant expression.
409 while (Try < BaseRegsSize && !isa<SCEVAddRecExpr>(ScaledReg))
410 std::swap(ScaledReg, BaseRegs[Try++]);
413 /// \brief Get rid of the scale in the formula.
414 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
415 /// \return true if it was possible to get rid of the scale, false otherwise.
416 /// \note After this operation the formula may not be in the canonical form.
417 bool Formula::unscale() {
421 BaseRegs.push_back(ScaledReg);
426 /// Return the total number of register operands used by this formula. This does
427 /// not include register uses implied by non-constant addrec strides.
428 size_t Formula::getNumRegs() const {
429 return !!ScaledReg + BaseRegs.size();
432 /// Return the type of this formula, if it has one, or null otherwise. This type
433 /// is meaningless except for the bit size.
434 Type *Formula::getType() const {
435 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
436 ScaledReg ? ScaledReg->getType() :
437 BaseGV ? BaseGV->getType() :
441 /// Delete the given base reg from the BaseRegs list.
442 void Formula::deleteBaseReg(const SCEV *&S) {
443 if (&S != &BaseRegs.back())
444 std::swap(S, BaseRegs.back());
448 /// Test if this formula references the given register.
449 bool Formula::referencesReg(const SCEV *S) const {
450 return S == ScaledReg ||
451 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
454 /// Test whether this formula uses registers which are used by uses other than
455 /// the use with the given index.
456 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
457 const RegUseTracker &RegUses) const {
459 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
461 for (const SCEV *BaseReg : BaseRegs)
462 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
467 void Formula::print(raw_ostream &OS) const {
470 if (!First) OS << " + "; else First = false;
471 BaseGV->printAsOperand(OS, /*PrintType=*/false);
473 if (BaseOffset != 0) {
474 if (!First) OS << " + "; else First = false;
477 for (const SCEV *BaseReg : BaseRegs) {
478 if (!First) OS << " + "; else First = false;
479 OS << "reg(" << *BaseReg << ')';
481 if (HasBaseReg && BaseRegs.empty()) {
482 if (!First) OS << " + "; else First = false;
483 OS << "**error: HasBaseReg**";
484 } else if (!HasBaseReg && !BaseRegs.empty()) {
485 if (!First) OS << " + "; else First = false;
486 OS << "**error: !HasBaseReg**";
489 if (!First) OS << " + "; else First = false;
490 OS << Scale << "*reg(";
497 if (UnfoldedOffset != 0) {
498 if (!First) OS << " + ";
499 OS << "imm(" << UnfoldedOffset << ')';
503 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
504 void Formula::dump() const {
505 print(errs()); errs() << '\n';
509 /// Return true if the given addrec can be sign-extended without changing its
511 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
513 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
514 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
517 /// Return true if the given add can be sign-extended without changing its
519 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
521 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
522 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
525 /// Return true if the given mul can be sign-extended without changing its
527 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
529 IntegerType::get(SE.getContext(),
530 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
531 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
534 /// Return an expression for LHS /s RHS, if it can be determined and if the
535 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
536 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
537 /// the multiplication may overflow, which is useful when the result will be
538 /// used in a context where the most significant bits are ignored.
539 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
541 bool IgnoreSignificantBits = false) {
542 // Handle the trivial case, which works for any SCEV type.
544 return SE.getConstant(LHS->getType(), 1);
546 // Handle a few RHS special cases.
547 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
549 const APInt &RA = RC->getValue()->getValue();
550 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
552 if (RA.isAllOnesValue())
553 return SE.getMulExpr(LHS, RC);
554 // Handle x /s 1 as x.
559 // Check for a division of a constant by a constant.
560 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
563 const APInt &LA = C->getValue()->getValue();
564 const APInt &RA = RC->getValue()->getValue();
565 if (LA.srem(RA) != 0)
567 return SE.getConstant(LA.sdiv(RA));
570 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
571 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
572 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
573 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
574 IgnoreSignificantBits);
575 if (!Step) return nullptr;
576 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
577 IgnoreSignificantBits);
578 if (!Start) return nullptr;
579 // FlagNW is independent of the start value, step direction, and is
580 // preserved with smaller magnitude steps.
581 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
582 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
587 // Distribute the sdiv over add operands, if the add doesn't overflow.
588 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
589 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
590 SmallVector<const SCEV *, 8> Ops;
591 for (const SCEV *S : Add->operands()) {
592 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
593 if (!Op) return nullptr;
596 return SE.getAddExpr(Ops);
601 // Check for a multiply operand that we can pull RHS out of.
602 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
603 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
604 SmallVector<const SCEV *, 4> Ops;
606 for (const SCEV *S : Mul->operands()) {
608 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
609 IgnoreSignificantBits)) {
615 return Found ? SE.getMulExpr(Ops) : nullptr;
620 // Otherwise we don't know.
624 /// If S involves the addition of a constant integer value, return that integer
625 /// value, and mutate S to point to a new SCEV with that value excluded.
626 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
627 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
628 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
629 S = SE.getConstant(C->getType(), 0);
630 return C->getValue()->getSExtValue();
632 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
633 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
634 int64_t Result = ExtractImmediate(NewOps.front(), SE);
636 S = SE.getAddExpr(NewOps);
638 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
639 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
640 int64_t Result = ExtractImmediate(NewOps.front(), SE);
642 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
643 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
650 /// If S involves the addition of a GlobalValue address, return that symbol, and
651 /// mutate S to point to a new SCEV with that value excluded.
652 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
653 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
654 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
655 S = SE.getConstant(GV->getType(), 0);
658 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
659 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
660 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
662 S = SE.getAddExpr(NewOps);
664 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
665 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
666 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
668 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
669 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
676 /// Returns true if the specified instruction is using the specified value as an
678 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
679 bool isAddress = isa<LoadInst>(Inst);
680 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
681 if (SI->getOperand(1) == OperandVal)
683 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
684 // Addressing modes can also be folded into prefetches and a variety
686 switch (II->getIntrinsicID()) {
688 case Intrinsic::prefetch:
689 case Intrinsic::x86_sse_storeu_ps:
690 case Intrinsic::x86_sse2_storeu_pd:
691 case Intrinsic::x86_sse2_storeu_dq:
692 case Intrinsic::x86_sse2_storel_dq:
693 if (II->getArgOperand(0) == OperandVal)
701 /// Return the type of the memory being accessed.
702 static MemAccessTy getAccessType(const Instruction *Inst) {
703 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
704 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
705 AccessTy.MemTy = SI->getOperand(0)->getType();
706 AccessTy.AddrSpace = SI->getPointerAddressSpace();
707 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
708 AccessTy.AddrSpace = LI->getPointerAddressSpace();
709 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
710 // Addressing modes can also be folded into prefetches and a variety
712 switch (II->getIntrinsicID()) {
714 case Intrinsic::x86_sse_storeu_ps:
715 case Intrinsic::x86_sse2_storeu_pd:
716 case Intrinsic::x86_sse2_storeu_dq:
717 case Intrinsic::x86_sse2_storel_dq:
718 AccessTy.MemTy = II->getArgOperand(0)->getType();
723 // All pointers have the same requirements, so canonicalize them to an
724 // arbitrary pointer type to minimize variation.
725 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
726 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
727 PTy->getAddressSpace());
732 /// Return true if this AddRec is already a phi in its loop.
733 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
734 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
735 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
736 if (SE.isSCEVable(PN->getType()) &&
737 (SE.getEffectiveSCEVType(PN->getType()) ==
738 SE.getEffectiveSCEVType(AR->getType())) &&
739 SE.getSCEV(PN) == AR)
745 /// Check if expanding this expression is likely to incur significant cost. This
746 /// is tricky because SCEV doesn't track which expressions are actually computed
747 /// by the current IR.
749 /// We currently allow expansion of IV increments that involve adds,
750 /// multiplication by constants, and AddRecs from existing phis.
752 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
753 /// obvious multiple of the UDivExpr.
754 static bool isHighCostExpansion(const SCEV *S,
755 SmallPtrSetImpl<const SCEV*> &Processed,
756 ScalarEvolution &SE) {
757 // Zero/One operand expressions
758 switch (S->getSCEVType()) {
763 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
766 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
769 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
773 if (!Processed.insert(S).second)
776 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
777 for (const SCEV *S : Add->operands()) {
778 if (isHighCostExpansion(S, Processed, SE))
784 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
785 if (Mul->getNumOperands() == 2) {
786 // Multiplication by a constant is ok
787 if (isa<SCEVConstant>(Mul->getOperand(0)))
788 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
790 // If we have the value of one operand, check if an existing
791 // multiplication already generates this expression.
792 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
793 Value *UVal = U->getValue();
794 for (User *UR : UVal->users()) {
795 // If U is a constant, it may be used by a ConstantExpr.
796 Instruction *UI = dyn_cast<Instruction>(UR);
797 if (UI && UI->getOpcode() == Instruction::Mul &&
798 SE.isSCEVable(UI->getType())) {
799 return SE.getSCEV(UI) == Mul;
806 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
807 if (isExistingPhi(AR, SE))
811 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
815 /// If any of the instructions is the specified set are trivially dead, delete
816 /// them and see if this makes any of their operands subsequently dead.
818 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
819 bool Changed = false;
821 while (!DeadInsts.empty()) {
822 Value *V = DeadInsts.pop_back_val();
823 Instruction *I = dyn_cast_or_null<Instruction>(V);
825 if (!I || !isInstructionTriviallyDead(I))
828 for (Use &O : I->operands())
829 if (Instruction *U = dyn_cast<Instruction>(O)) {
832 DeadInsts.emplace_back(U);
835 I->eraseFromParent();
846 /// \brief Check if the addressing mode defined by \p F is completely
847 /// folded in \p LU at isel time.
848 /// This includes address-mode folding and special icmp tricks.
849 /// This function returns true if \p LU can accommodate what \p F
850 /// defines and up to 1 base + 1 scaled + offset.
851 /// In other words, if \p F has several base registers, this function may
852 /// still return true. Therefore, users still need to account for
853 /// additional base registers and/or unfolded offsets to derive an
854 /// accurate cost model.
855 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
856 const LSRUse &LU, const Formula &F);
857 // Get the cost of the scaling factor used in F for LU.
858 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
859 const LSRUse &LU, const Formula &F);
863 /// This class is used to measure and compare candidate formulae.
865 /// TODO: Some of these could be merged. Also, a lexical ordering
866 /// isn't always optimal.
870 unsigned NumBaseAdds;
877 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
878 SetupCost(0), ScaleCost(0) {}
880 bool operator<(const Cost &Other) const;
885 // Once any of the metrics loses, they must all remain losers.
887 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
888 | ImmCost | SetupCost | ScaleCost) != ~0u)
889 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
890 & ImmCost & SetupCost & ScaleCost) == ~0u);
895 assert(isValid() && "invalid cost");
896 return NumRegs == ~0u;
899 void RateFormula(const TargetTransformInfo &TTI,
901 SmallPtrSetImpl<const SCEV *> &Regs,
902 const DenseSet<const SCEV *> &VisitedRegs,
904 const SmallVectorImpl<int64_t> &Offsets,
905 ScalarEvolution &SE, DominatorTree &DT,
907 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
909 void print(raw_ostream &OS) const;
913 void RateRegister(const SCEV *Reg,
914 SmallPtrSetImpl<const SCEV *> &Regs,
916 ScalarEvolution &SE, DominatorTree &DT);
917 void RatePrimaryRegister(const SCEV *Reg,
918 SmallPtrSetImpl<const SCEV *> &Regs,
920 ScalarEvolution &SE, DominatorTree &DT,
921 SmallPtrSetImpl<const SCEV *> *LoserRegs);
926 /// Tally up interesting quantities from the given register.
927 void Cost::RateRegister(const SCEV *Reg,
928 SmallPtrSetImpl<const SCEV *> &Regs,
930 ScalarEvolution &SE, DominatorTree &DT) {
931 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
932 // If this is an addrec for another loop, don't second-guess its addrec phi
933 // nodes. LSR isn't currently smart enough to reason about more than one
934 // loop at a time. LSR has already run on inner loops, will not run on outer
935 // loops, and cannot be expected to change sibling loops.
936 if (AR->getLoop() != L) {
937 // If the AddRec exists, consider it's register free and leave it alone.
938 if (isExistingPhi(AR, SE))
941 // Otherwise, do not consider this formula at all.
945 AddRecCost += 1; /// TODO: This should be a function of the stride.
947 // Add the step value register, if it needs one.
948 // TODO: The non-affine case isn't precisely modeled here.
949 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
950 if (!Regs.count(AR->getOperand(1))) {
951 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
959 // Rough heuristic; favor registers which don't require extra setup
960 // instructions in the preheader.
961 if (!isa<SCEVUnknown>(Reg) &&
962 !isa<SCEVConstant>(Reg) &&
963 !(isa<SCEVAddRecExpr>(Reg) &&
964 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
965 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
968 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
969 SE.hasComputableLoopEvolution(Reg, L);
972 /// Record this register in the set. If we haven't seen it before, rate
973 /// it. Optional LoserRegs provides a way to declare any formula that refers to
974 /// one of those regs an instant loser.
975 void Cost::RatePrimaryRegister(const SCEV *Reg,
976 SmallPtrSetImpl<const SCEV *> &Regs,
978 ScalarEvolution &SE, DominatorTree &DT,
979 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
980 if (LoserRegs && LoserRegs->count(Reg)) {
984 if (Regs.insert(Reg).second) {
985 RateRegister(Reg, Regs, L, SE, DT);
986 if (LoserRegs && isLoser())
987 LoserRegs->insert(Reg);
991 void Cost::RateFormula(const TargetTransformInfo &TTI,
993 SmallPtrSetImpl<const SCEV *> &Regs,
994 const DenseSet<const SCEV *> &VisitedRegs,
996 const SmallVectorImpl<int64_t> &Offsets,
997 ScalarEvolution &SE, DominatorTree &DT,
999 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1000 assert(F.isCanonical() && "Cost is accurate only for canonical formula");
1001 // Tally up the registers.
1002 if (const SCEV *ScaledReg = F.ScaledReg) {
1003 if (VisitedRegs.count(ScaledReg)) {
1007 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1011 for (const SCEV *BaseReg : F.BaseRegs) {
1012 if (VisitedRegs.count(BaseReg)) {
1016 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1021 // Determine how many (unfolded) adds we'll need inside the loop.
1022 size_t NumBaseParts = F.getNumRegs();
1023 if (NumBaseParts > 1)
1024 // Do not count the base and a possible second register if the target
1025 // allows to fold 2 registers.
1027 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1028 NumBaseAdds += (F.UnfoldedOffset != 0);
1030 // Accumulate non-free scaling amounts.
1031 ScaleCost += getScalingFactorCost(TTI, LU, F);
1033 // Tally up the non-zero immediates.
1034 for (int64_t O : Offsets) {
1035 int64_t Offset = (uint64_t)O + F.BaseOffset;
1037 ImmCost += 64; // Handle symbolic values conservatively.
1038 // TODO: This should probably be the pointer size.
1039 else if (Offset != 0)
1040 ImmCost += APInt(64, Offset, true).getMinSignedBits();
1042 assert(isValid() && "invalid cost");
1045 /// Set this cost to a losing value.
1056 /// Choose the lower cost.
1057 bool Cost::operator<(const Cost &Other) const {
1058 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
1059 ImmCost, SetupCost) <
1060 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
1061 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
1065 void Cost::print(raw_ostream &OS) const {
1066 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
1067 if (AddRecCost != 0)
1068 OS << ", with addrec cost " << AddRecCost;
1070 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
1071 if (NumBaseAdds != 0)
1072 OS << ", plus " << NumBaseAdds << " base add"
1073 << (NumBaseAdds == 1 ? "" : "s");
1075 OS << ", plus " << ScaleCost << " scale cost";
1077 OS << ", plus " << ImmCost << " imm cost";
1079 OS << ", plus " << SetupCost << " setup cost";
1082 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1083 void Cost::dump() const {
1084 print(errs()); errs() << '\n';
1090 /// An operand value in an instruction which is to be replaced with some
1091 /// equivalent, possibly strength-reduced, replacement.
1093 /// The instruction which will be updated.
1094 Instruction *UserInst;
1096 /// The operand of the instruction which will be replaced. The operand may be
1097 /// used more than once; every instance will be replaced.
1098 Value *OperandValToReplace;
1100 /// If this user is to use the post-incremented value of an induction
1101 /// variable, this variable is non-null and holds the loop associated with the
1102 /// induction variable.
1103 PostIncLoopSet PostIncLoops;
1105 /// The index of the LSRUse describing the expression which this fixup needs,
1106 /// minus an offset (below).
1109 /// A constant offset to be added to the LSRUse expression. This allows
1110 /// multiple fixups to share the same LSRUse with different offsets, for
1111 /// example in an unrolled loop.
1114 bool isUseFullyOutsideLoop(const Loop *L) const;
1118 void print(raw_ostream &OS) const;
1124 LSRFixup::LSRFixup()
1125 : UserInst(nullptr), OperandValToReplace(nullptr), LUIdx(~size_t(0)),
1128 /// Test whether this fixup always uses its value outside of the given loop.
1129 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1130 // PHI nodes use their value in their incoming blocks.
1131 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1132 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1133 if (PN->getIncomingValue(i) == OperandValToReplace &&
1134 L->contains(PN->getIncomingBlock(i)))
1139 return !L->contains(UserInst);
1142 void LSRFixup::print(raw_ostream &OS) const {
1144 // Store is common and interesting enough to be worth special-casing.
1145 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1147 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1148 } else if (UserInst->getType()->isVoidTy())
1149 OS << UserInst->getOpcodeName();
1151 UserInst->printAsOperand(OS, /*PrintType=*/false);
1153 OS << ", OperandValToReplace=";
1154 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1156 for (const Loop *PIL : PostIncLoops) {
1157 OS << ", PostIncLoop=";
1158 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1161 if (LUIdx != ~size_t(0))
1162 OS << ", LUIdx=" << LUIdx;
1165 OS << ", Offset=" << Offset;
1168 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1169 void LSRFixup::dump() const {
1170 print(errs()); errs() << '\n';
1176 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1177 /// SmallVectors of const SCEV*.
1178 struct UniquifierDenseMapInfo {
1179 static SmallVector<const SCEV *, 4> getEmptyKey() {
1180 SmallVector<const SCEV *, 4> V;
1181 V.push_back(reinterpret_cast<const SCEV *>(-1));
1185 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1186 SmallVector<const SCEV *, 4> V;
1187 V.push_back(reinterpret_cast<const SCEV *>(-2));
1191 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1192 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1195 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1196 const SmallVector<const SCEV *, 4> &RHS) {
1201 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1202 /// as uses invented by LSR itself. It includes information about what kinds of
1203 /// things can be folded into the user, information about the user itself, and
1204 /// information about how the use may be satisfied. TODO: Represent multiple
1205 /// users of the same expression in common?
1207 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1210 /// An enum for a kind of use, indicating what types of scaled and immediate
1211 /// operands it might support.
1213 Basic, ///< A normal use, with no folding.
1214 Special, ///< A special case of basic, allowing -1 scales.
1215 Address, ///< An address use; folding according to TargetLowering
1216 ICmpZero ///< An equality icmp with both operands folded into one.
1217 // TODO: Add a generic icmp too?
1220 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1223 MemAccessTy AccessTy;
1225 SmallVector<int64_t, 8> Offsets;
1229 /// This records whether all of the fixups using this LSRUse are outside of
1230 /// the loop, in which case some special-case heuristics may be used.
1231 bool AllFixupsOutsideLoop;
1233 /// RigidFormula is set to true to guarantee that this use will be associated
1234 /// with a single formula--the one that initially matched. Some SCEV
1235 /// expressions cannot be expanded. This allows LSR to consider the registers
1236 /// used by those expressions without the need to expand them later after
1237 /// changing the formula.
1240 /// This records the widest use type for any fixup using this
1241 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1242 /// fixup widths to be equivalent, because the narrower one may be relying on
1243 /// the implicit truncation to truncate away bogus bits.
1244 Type *WidestFixupType;
1246 /// A list of ways to build a value that can satisfy this user. After the
1247 /// list is populated, one of these is selected heuristically and used to
1248 /// formulate a replacement for OperandValToReplace in UserInst.
1249 SmallVector<Formula, 12> Formulae;
1251 /// The set of register candidates used by all formulae in this LSRUse.
1252 SmallPtrSet<const SCEV *, 4> Regs;
1254 LSRUse(KindType K, MemAccessTy AT)
1255 : Kind(K), AccessTy(AT), MinOffset(INT64_MAX), MaxOffset(INT64_MIN),
1256 AllFixupsOutsideLoop(true), RigidFormula(false),
1257 WidestFixupType(nullptr) {}
1259 bool HasFormulaWithSameRegs(const Formula &F) const;
1260 bool InsertFormula(const Formula &F);
1261 void DeleteFormula(Formula &F);
1262 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1264 void print(raw_ostream &OS) const;
1270 /// Test whether this use as a formula which has the same registers as the given
1272 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1273 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1274 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1275 // Unstable sort by host order ok, because this is only used for uniquifying.
1276 std::sort(Key.begin(), Key.end());
1277 return Uniquifier.count(Key);
1280 /// If the given formula has not yet been inserted, add it to the list, and
1281 /// return true. Return false otherwise. The formula must be in canonical form.
1282 bool LSRUse::InsertFormula(const Formula &F) {
1283 assert(F.isCanonical() && "Invalid canonical representation");
1285 if (!Formulae.empty() && RigidFormula)
1288 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1289 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1290 // Unstable sort by host order ok, because this is only used for uniquifying.
1291 std::sort(Key.begin(), Key.end());
1293 if (!Uniquifier.insert(Key).second)
1296 // Using a register to hold the value of 0 is not profitable.
1297 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1298 "Zero allocated in a scaled register!");
1300 for (const SCEV *BaseReg : F.BaseRegs)
1301 assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1304 // Add the formula to the list.
1305 Formulae.push_back(F);
1307 // Record registers now being used by this use.
1308 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1310 Regs.insert(F.ScaledReg);
1315 /// Remove the given formula from this use's list.
1316 void LSRUse::DeleteFormula(Formula &F) {
1317 if (&F != &Formulae.back())
1318 std::swap(F, Formulae.back());
1319 Formulae.pop_back();
1322 /// Recompute the Regs field, and update RegUses.
1323 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1324 // Now that we've filtered out some formulae, recompute the Regs set.
1325 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1327 for (const Formula &F : Formulae) {
1328 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1329 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1332 // Update the RegTracker.
1333 for (const SCEV *S : OldRegs)
1335 RegUses.dropRegister(S, LUIdx);
1338 void LSRUse::print(raw_ostream &OS) const {
1339 OS << "LSR Use: Kind=";
1341 case Basic: OS << "Basic"; break;
1342 case Special: OS << "Special"; break;
1343 case ICmpZero: OS << "ICmpZero"; break;
1345 OS << "Address of ";
1346 if (AccessTy.MemTy->isPointerTy())
1347 OS << "pointer"; // the full pointer type could be really verbose
1349 OS << *AccessTy.MemTy;
1352 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1355 OS << ", Offsets={";
1356 bool NeedComma = false;
1357 for (int64_t O : Offsets) {
1358 if (NeedComma) OS << ',';
1364 if (AllFixupsOutsideLoop)
1365 OS << ", all-fixups-outside-loop";
1367 if (WidestFixupType)
1368 OS << ", widest fixup type: " << *WidestFixupType;
1371 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1372 void LSRUse::dump() const {
1373 print(errs()); errs() << '\n';
1377 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1378 LSRUse::KindType Kind, MemAccessTy AccessTy,
1379 GlobalValue *BaseGV, int64_t BaseOffset,
1380 bool HasBaseReg, int64_t Scale) {
1382 case LSRUse::Address:
1383 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1384 HasBaseReg, Scale, AccessTy.AddrSpace);
1386 case LSRUse::ICmpZero:
1387 // There's not even a target hook for querying whether it would be legal to
1388 // fold a GV into an ICmp.
1392 // ICmp only has two operands; don't allow more than two non-trivial parts.
1393 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1396 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1397 // putting the scaled register in the other operand of the icmp.
1398 if (Scale != 0 && Scale != -1)
1401 // If we have low-level target information, ask the target if it can fold an
1402 // integer immediate on an icmp.
1403 if (BaseOffset != 0) {
1405 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1406 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1407 // Offs is the ICmp immediate.
1409 // The cast does the right thing with INT64_MIN.
1410 BaseOffset = -(uint64_t)BaseOffset;
1411 return TTI.isLegalICmpImmediate(BaseOffset);
1414 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1418 // Only handle single-register values.
1419 return !BaseGV && Scale == 0 && BaseOffset == 0;
1421 case LSRUse::Special:
1422 // Special case Basic to handle -1 scales.
1423 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1426 llvm_unreachable("Invalid LSRUse Kind!");
1429 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1430 int64_t MinOffset, int64_t MaxOffset,
1431 LSRUse::KindType Kind, MemAccessTy AccessTy,
1432 GlobalValue *BaseGV, int64_t BaseOffset,
1433 bool HasBaseReg, int64_t Scale) {
1434 // Check for overflow.
1435 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1438 MinOffset = (uint64_t)BaseOffset + MinOffset;
1439 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1442 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1444 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1445 HasBaseReg, Scale) &&
1446 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1450 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1451 int64_t MinOffset, int64_t MaxOffset,
1452 LSRUse::KindType Kind, MemAccessTy AccessTy,
1454 // For the purpose of isAMCompletelyFolded either having a canonical formula
1455 // or a scale not equal to zero is correct.
1456 // Problems may arise from non canonical formulae having a scale == 0.
1457 // Strictly speaking it would best to just rely on canonical formulae.
1458 // However, when we generate the scaled formulae, we first check that the
1459 // scaling factor is profitable before computing the actual ScaledReg for
1460 // compile time sake.
1461 assert((F.isCanonical() || F.Scale != 0));
1462 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1463 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1466 /// Test whether we know how to expand the current formula.
1467 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1468 int64_t MaxOffset, LSRUse::KindType Kind,
1469 MemAccessTy AccessTy, GlobalValue *BaseGV,
1470 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1471 // We know how to expand completely foldable formulae.
1472 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1473 BaseOffset, HasBaseReg, Scale) ||
1474 // Or formulae that use a base register produced by a sum of base
1477 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1478 BaseGV, BaseOffset, true, 0));
1481 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1482 int64_t MaxOffset, LSRUse::KindType Kind,
1483 MemAccessTy AccessTy, const Formula &F) {
1484 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1485 F.BaseOffset, F.HasBaseReg, F.Scale);
1488 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1489 const LSRUse &LU, const Formula &F) {
1490 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1491 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1495 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1496 const LSRUse &LU, const Formula &F) {
1500 // If the use is not completely folded in that instruction, we will have to
1501 // pay an extra cost only for scale != 1.
1502 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1504 return F.Scale != 1;
1507 case LSRUse::Address: {
1508 // Check the scaling factor cost with both the min and max offsets.
1509 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1510 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1511 F.Scale, LU.AccessTy.AddrSpace);
1512 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1513 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1514 F.Scale, LU.AccessTy.AddrSpace);
1516 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1517 "Legal addressing mode has an illegal cost!");
1518 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1520 case LSRUse::ICmpZero:
1522 case LSRUse::Special:
1523 // The use is completely folded, i.e., everything is folded into the
1528 llvm_unreachable("Invalid LSRUse Kind!");
1531 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1532 LSRUse::KindType Kind, MemAccessTy AccessTy,
1533 GlobalValue *BaseGV, int64_t BaseOffset,
1535 // Fast-path: zero is always foldable.
1536 if (BaseOffset == 0 && !BaseGV) return true;
1538 // Conservatively, create an address with an immediate and a
1539 // base and a scale.
1540 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1542 // Canonicalize a scale of 1 to a base register if the formula doesn't
1543 // already have a base register.
1544 if (!HasBaseReg && Scale == 1) {
1549 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1553 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1554 ScalarEvolution &SE, int64_t MinOffset,
1555 int64_t MaxOffset, LSRUse::KindType Kind,
1556 MemAccessTy AccessTy, const SCEV *S,
1558 // Fast-path: zero is always foldable.
1559 if (S->isZero()) return true;
1561 // Conservatively, create an address with an immediate and a
1562 // base and a scale.
1563 int64_t BaseOffset = ExtractImmediate(S, SE);
1564 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1566 // If there's anything else involved, it's not foldable.
1567 if (!S->isZero()) return false;
1569 // Fast-path: zero is always foldable.
1570 if (BaseOffset == 0 && !BaseGV) return true;
1572 // Conservatively, create an address with an immediate and a
1573 // base and a scale.
1574 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1576 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1577 BaseOffset, HasBaseReg, Scale);
1582 /// An individual increment in a Chain of IV increments. Relate an IV user to
1583 /// an expression that computes the IV it uses from the IV used by the previous
1584 /// link in the Chain.
1586 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1587 /// original IVOperand. The head of the chain's IVOperand is only valid during
1588 /// chain collection, before LSR replaces IV users. During chain generation,
1589 /// IncExpr can be used to find the new IVOperand that computes the same
1592 Instruction *UserInst;
1594 const SCEV *IncExpr;
1596 IVInc(Instruction *U, Value *O, const SCEV *E):
1597 UserInst(U), IVOperand(O), IncExpr(E) {}
1600 // The list of IV increments in program order. We typically add the head of a
1601 // chain without finding subsequent links.
1603 SmallVector<IVInc,1> Incs;
1604 const SCEV *ExprBase;
1606 IVChain() : ExprBase(nullptr) {}
1608 IVChain(const IVInc &Head, const SCEV *Base)
1609 : Incs(1, Head), ExprBase(Base) {}
1611 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1613 // Return the first increment in the chain.
1614 const_iterator begin() const {
1615 assert(!Incs.empty());
1616 return std::next(Incs.begin());
1618 const_iterator end() const {
1622 // Returns true if this chain contains any increments.
1623 bool hasIncs() const { return Incs.size() >= 2; }
1625 // Add an IVInc to the end of this chain.
1626 void add(const IVInc &X) { Incs.push_back(X); }
1628 // Returns the last UserInst in the chain.
1629 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1631 // Returns true if IncExpr can be profitably added to this chain.
1632 bool isProfitableIncrement(const SCEV *OperExpr,
1633 const SCEV *IncExpr,
1637 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1638 /// between FarUsers that definitely cross IV increments and NearUsers that may
1639 /// be used between IV increments.
1641 SmallPtrSet<Instruction*, 4> FarUsers;
1642 SmallPtrSet<Instruction*, 4> NearUsers;
1645 /// This class holds state for the main loop strength reduction logic.
1648 ScalarEvolution &SE;
1651 const TargetTransformInfo &TTI;
1655 /// This is the insert position that the current loop's induction variable
1656 /// increment should be placed. In simple loops, this is the latch block's
1657 /// terminator. But in more complicated cases, this is a position which will
1658 /// dominate all the in-loop post-increment users.
1659 Instruction *IVIncInsertPos;
1661 /// Interesting factors between use strides.
1662 SmallSetVector<int64_t, 8> Factors;
1664 /// Interesting use types, to facilitate truncation reuse.
1665 SmallSetVector<Type *, 4> Types;
1667 /// The list of operands which are to be replaced.
1668 SmallVector<LSRFixup, 16> Fixups;
1670 /// The list of interesting uses.
1671 SmallVector<LSRUse, 16> Uses;
1673 /// Track which uses use which register candidates.
1674 RegUseTracker RegUses;
1676 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1677 // have more than a few IV increment chains in a loop. Missing a Chain falls
1678 // back to normal LSR behavior for those uses.
1679 static const unsigned MaxChains = 8;
1681 /// IV users can form a chain of IV increments.
1682 SmallVector<IVChain, MaxChains> IVChainVec;
1684 /// IV users that belong to profitable IVChains.
1685 SmallPtrSet<Use*, MaxChains> IVIncSet;
1687 void OptimizeShadowIV();
1688 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1689 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1690 void OptimizeLoopTermCond();
1692 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1693 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1694 void FinalizeChain(IVChain &Chain);
1695 void CollectChains();
1696 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1697 SmallVectorImpl<WeakVH> &DeadInsts);
1699 void CollectInterestingTypesAndFactors();
1700 void CollectFixupsAndInitialFormulae();
1702 LSRFixup &getNewFixup() {
1703 Fixups.push_back(LSRFixup());
1704 return Fixups.back();
1707 // Support for sharing of LSRUses between LSRFixups.
1708 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1711 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1712 LSRUse::KindType Kind, MemAccessTy AccessTy);
1714 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1715 MemAccessTy AccessTy);
1717 void DeleteUse(LSRUse &LU, size_t LUIdx);
1719 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1721 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1722 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1723 void CountRegisters(const Formula &F, size_t LUIdx);
1724 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1726 void CollectLoopInvariantFixupsAndFormulae();
1728 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1729 unsigned Depth = 0);
1731 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1732 const Formula &Base, unsigned Depth,
1733 size_t Idx, bool IsScaledReg = false);
1734 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1735 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1736 const Formula &Base, size_t Idx,
1737 bool IsScaledReg = false);
1738 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1739 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1740 const Formula &Base,
1741 const SmallVectorImpl<int64_t> &Worklist,
1742 size_t Idx, bool IsScaledReg = false);
1743 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1744 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1745 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1746 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1747 void GenerateCrossUseConstantOffsets();
1748 void GenerateAllReuseFormulae();
1750 void FilterOutUndesirableDedicatedRegisters();
1752 size_t EstimateSearchSpaceComplexity() const;
1753 void NarrowSearchSpaceByDetectingSupersets();
1754 void NarrowSearchSpaceByCollapsingUnrolledCode();
1755 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1756 void NarrowSearchSpaceByPickingWinnerRegs();
1757 void NarrowSearchSpaceUsingHeuristics();
1759 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1761 SmallVectorImpl<const Formula *> &Workspace,
1762 const Cost &CurCost,
1763 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1764 DenseSet<const SCEV *> &VisitedRegs) const;
1765 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1767 BasicBlock::iterator
1768 HoistInsertPosition(BasicBlock::iterator IP,
1769 const SmallVectorImpl<Instruction *> &Inputs) const;
1770 BasicBlock::iterator
1771 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1774 SCEVExpander &Rewriter) const;
1776 Value *Expand(const LSRFixup &LF,
1778 BasicBlock::iterator IP,
1779 SCEVExpander &Rewriter,
1780 SmallVectorImpl<WeakVH> &DeadInsts) const;
1781 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1783 SCEVExpander &Rewriter,
1784 SmallVectorImpl<WeakVH> &DeadInsts,
1786 void Rewrite(const LSRFixup &LF,
1788 SCEVExpander &Rewriter,
1789 SmallVectorImpl<WeakVH> &DeadInsts,
1791 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1795 LSRInstance(Loop *L, Pass *P);
1797 bool getChanged() const { return Changed; }
1799 void print_factors_and_types(raw_ostream &OS) const;
1800 void print_fixups(raw_ostream &OS) const;
1801 void print_uses(raw_ostream &OS) const;
1802 void print(raw_ostream &OS) const;
1808 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
1809 /// the cast operation.
1810 void LSRInstance::OptimizeShadowIV() {
1811 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1812 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1815 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1816 UI != E; /* empty */) {
1817 IVUsers::const_iterator CandidateUI = UI;
1819 Instruction *ShadowUse = CandidateUI->getUser();
1820 Type *DestTy = nullptr;
1821 bool IsSigned = false;
1823 /* If shadow use is a int->float cast then insert a second IV
1824 to eliminate this cast.
1826 for (unsigned i = 0; i < n; ++i)
1832 for (unsigned i = 0; i < n; ++i, ++d)
1835 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1837 DestTy = UCast->getDestTy();
1839 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1841 DestTy = SCast->getDestTy();
1843 if (!DestTy) continue;
1845 // If target does not support DestTy natively then do not apply
1846 // this transformation.
1847 if (!TTI.isTypeLegal(DestTy)) continue;
1849 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1851 if (PH->getNumIncomingValues() != 2) continue;
1853 Type *SrcTy = PH->getType();
1854 int Mantissa = DestTy->getFPMantissaWidth();
1855 if (Mantissa == -1) continue;
1856 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1859 unsigned Entry, Latch;
1860 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1868 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1869 if (!Init) continue;
1870 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1871 (double)Init->getSExtValue() :
1872 (double)Init->getZExtValue());
1874 BinaryOperator *Incr =
1875 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1876 if (!Incr) continue;
1877 if (Incr->getOpcode() != Instruction::Add
1878 && Incr->getOpcode() != Instruction::Sub)
1881 /* Initialize new IV, double d = 0.0 in above example. */
1882 ConstantInt *C = nullptr;
1883 if (Incr->getOperand(0) == PH)
1884 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1885 else if (Incr->getOperand(1) == PH)
1886 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1892 // Ignore negative constants, as the code below doesn't handle them
1893 // correctly. TODO: Remove this restriction.
1894 if (!C->getValue().isStrictlyPositive()) continue;
1896 /* Add new PHINode. */
1897 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1899 /* create new increment. '++d' in above example. */
1900 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1901 BinaryOperator *NewIncr =
1902 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1903 Instruction::FAdd : Instruction::FSub,
1904 NewPH, CFP, "IV.S.next.", Incr);
1906 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1907 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1909 /* Remove cast operation */
1910 ShadowUse->replaceAllUsesWith(NewPH);
1911 ShadowUse->eraseFromParent();
1917 /// If Cond has an operand that is an expression of an IV, set the IV user and
1918 /// stride information and return true, otherwise return false.
1919 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1920 for (IVStrideUse &U : IU)
1921 if (U.getUser() == Cond) {
1922 // NOTE: we could handle setcc instructions with multiple uses here, but
1923 // InstCombine does it as well for simple uses, it's not clear that it
1924 // occurs enough in real life to handle.
1931 /// Rewrite the loop's terminating condition if it uses a max computation.
1933 /// This is a narrow solution to a specific, but acute, problem. For loops
1939 /// } while (++i < n);
1941 /// the trip count isn't just 'n', because 'n' might not be positive. And
1942 /// unfortunately this can come up even for loops where the user didn't use
1943 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1944 /// will commonly be lowered like this:
1950 /// } while (++i < n);
1953 /// and then it's possible for subsequent optimization to obscure the if
1954 /// test in such a way that indvars can't find it.
1956 /// When indvars can't find the if test in loops like this, it creates a
1957 /// max expression, which allows it to give the loop a canonical
1958 /// induction variable:
1961 /// max = n < 1 ? 1 : n;
1964 /// } while (++i != max);
1966 /// Canonical induction variables are necessary because the loop passes
1967 /// are designed around them. The most obvious example of this is the
1968 /// LoopInfo analysis, which doesn't remember trip count values. It
1969 /// expects to be able to rediscover the trip count each time it is
1970 /// needed, and it does this using a simple analysis that only succeeds if
1971 /// the loop has a canonical induction variable.
1973 /// However, when it comes time to generate code, the maximum operation
1974 /// can be quite costly, especially if it's inside of an outer loop.
1976 /// This function solves this problem by detecting this type of loop and
1977 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1978 /// the instructions for the maximum computation.
1980 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1981 // Check that the loop matches the pattern we're looking for.
1982 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1983 Cond->getPredicate() != CmpInst::ICMP_NE)
1986 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1987 if (!Sel || !Sel->hasOneUse()) return Cond;
1989 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1990 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1992 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1994 // Add one to the backedge-taken count to get the trip count.
1995 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1996 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1998 // Check for a max calculation that matches the pattern. There's no check
1999 // for ICMP_ULE here because the comparison would be with zero, which
2000 // isn't interesting.
2001 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2002 const SCEVNAryExpr *Max = nullptr;
2003 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2004 Pred = ICmpInst::ICMP_SLE;
2006 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2007 Pred = ICmpInst::ICMP_SLT;
2009 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2010 Pred = ICmpInst::ICMP_ULT;
2017 // To handle a max with more than two operands, this optimization would
2018 // require additional checking and setup.
2019 if (Max->getNumOperands() != 2)
2022 const SCEV *MaxLHS = Max->getOperand(0);
2023 const SCEV *MaxRHS = Max->getOperand(1);
2025 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2026 // for a comparison with 1. For <= and >=, a comparison with zero.
2028 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2031 // Check the relevant induction variable for conformance to
2033 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2034 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2035 if (!AR || !AR->isAffine() ||
2036 AR->getStart() != One ||
2037 AR->getStepRecurrence(SE) != One)
2040 assert(AR->getLoop() == L &&
2041 "Loop condition operand is an addrec in a different loop!");
2043 // Check the right operand of the select, and remember it, as it will
2044 // be used in the new comparison instruction.
2045 Value *NewRHS = nullptr;
2046 if (ICmpInst::isTrueWhenEqual(Pred)) {
2047 // Look for n+1, and grab n.
2048 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2049 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2050 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2051 NewRHS = BO->getOperand(0);
2052 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2053 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2054 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2055 NewRHS = BO->getOperand(0);
2058 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2059 NewRHS = Sel->getOperand(1);
2060 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2061 NewRHS = Sel->getOperand(2);
2062 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2063 NewRHS = SU->getValue();
2065 // Max doesn't match expected pattern.
2068 // Determine the new comparison opcode. It may be signed or unsigned,
2069 // and the original comparison may be either equality or inequality.
2070 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2071 Pred = CmpInst::getInversePredicate(Pred);
2073 // Ok, everything looks ok to change the condition into an SLT or SGE and
2074 // delete the max calculation.
2076 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2078 // Delete the max calculation instructions.
2079 Cond->replaceAllUsesWith(NewCond);
2080 CondUse->setUser(NewCond);
2081 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2082 Cond->eraseFromParent();
2083 Sel->eraseFromParent();
2084 if (Cmp->use_empty())
2085 Cmp->eraseFromParent();
2089 /// Change loop terminating condition to use the postinc iv when possible.
2091 LSRInstance::OptimizeLoopTermCond() {
2092 SmallPtrSet<Instruction *, 4> PostIncs;
2094 BasicBlock *LatchBlock = L->getLoopLatch();
2095 SmallVector<BasicBlock*, 8> ExitingBlocks;
2096 L->getExitingBlocks(ExitingBlocks);
2098 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2100 // Get the terminating condition for the loop if possible. If we
2101 // can, we want to change it to use a post-incremented version of its
2102 // induction variable, to allow coalescing the live ranges for the IV into
2103 // one register value.
2105 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2108 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2109 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2112 // Search IVUsesByStride to find Cond's IVUse if there is one.
2113 IVStrideUse *CondUse = nullptr;
2114 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2115 if (!FindIVUserForCond(Cond, CondUse))
2118 // If the trip count is computed in terms of a max (due to ScalarEvolution
2119 // being unable to find a sufficient guard, for example), change the loop
2120 // comparison to use SLT or ULT instead of NE.
2121 // One consequence of doing this now is that it disrupts the count-down
2122 // optimization. That's not always a bad thing though, because in such
2123 // cases it may still be worthwhile to avoid a max.
2124 Cond = OptimizeMax(Cond, CondUse);
2126 // If this exiting block dominates the latch block, it may also use
2127 // the post-inc value if it won't be shared with other uses.
2128 // Check for dominance.
2129 if (!DT.dominates(ExitingBlock, LatchBlock))
2132 // Conservatively avoid trying to use the post-inc value in non-latch
2133 // exits if there may be pre-inc users in intervening blocks.
2134 if (LatchBlock != ExitingBlock)
2135 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2136 // Test if the use is reachable from the exiting block. This dominator
2137 // query is a conservative approximation of reachability.
2138 if (&*UI != CondUse &&
2139 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2140 // Conservatively assume there may be reuse if the quotient of their
2141 // strides could be a legal scale.
2142 const SCEV *A = IU.getStride(*CondUse, L);
2143 const SCEV *B = IU.getStride(*UI, L);
2144 if (!A || !B) continue;
2145 if (SE.getTypeSizeInBits(A->getType()) !=
2146 SE.getTypeSizeInBits(B->getType())) {
2147 if (SE.getTypeSizeInBits(A->getType()) >
2148 SE.getTypeSizeInBits(B->getType()))
2149 B = SE.getSignExtendExpr(B, A->getType());
2151 A = SE.getSignExtendExpr(A, B->getType());
2153 if (const SCEVConstant *D =
2154 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2155 const ConstantInt *C = D->getValue();
2156 // Stride of one or negative one can have reuse with non-addresses.
2157 if (C->isOne() || C->isAllOnesValue())
2158 goto decline_post_inc;
2159 // Avoid weird situations.
2160 if (C->getValue().getMinSignedBits() >= 64 ||
2161 C->getValue().isMinSignedValue())
2162 goto decline_post_inc;
2163 // Check for possible scaled-address reuse.
2164 MemAccessTy AccessTy = getAccessType(UI->getUser());
2165 int64_t Scale = C->getSExtValue();
2166 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2168 /*HasBaseReg=*/false, Scale,
2169 AccessTy.AddrSpace))
2170 goto decline_post_inc;
2172 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2174 /*HasBaseReg=*/false, Scale,
2175 AccessTy.AddrSpace))
2176 goto decline_post_inc;
2180 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2183 // It's possible for the setcc instruction to be anywhere in the loop, and
2184 // possible for it to have multiple users. If it is not immediately before
2185 // the exiting block branch, move it.
2186 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2187 if (Cond->hasOneUse()) {
2188 Cond->moveBefore(TermBr);
2190 // Clone the terminating condition and insert into the loopend.
2191 ICmpInst *OldCond = Cond;
2192 Cond = cast<ICmpInst>(Cond->clone());
2193 Cond->setName(L->getHeader()->getName() + ".termcond");
2194 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2196 // Clone the IVUse, as the old use still exists!
2197 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2198 TermBr->replaceUsesOfWith(OldCond, Cond);
2202 // If we get to here, we know that we can transform the setcc instruction to
2203 // use the post-incremented version of the IV, allowing us to coalesce the
2204 // live ranges for the IV correctly.
2205 CondUse->transformToPostInc(L);
2208 PostIncs.insert(Cond);
2212 // Determine an insertion point for the loop induction variable increment. It
2213 // must dominate all the post-inc comparisons we just set up, and it must
2214 // dominate the loop latch edge.
2215 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2216 for (Instruction *Inst : PostIncs) {
2218 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2220 if (BB == Inst->getParent())
2221 IVIncInsertPos = Inst;
2222 else if (BB != IVIncInsertPos->getParent())
2223 IVIncInsertPos = BB->getTerminator();
2227 /// Determine if the given use can accommodate a fixup at the given offset and
2228 /// other details. If so, update the use and return true.
2229 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2230 bool HasBaseReg, LSRUse::KindType Kind,
2231 MemAccessTy AccessTy) {
2232 int64_t NewMinOffset = LU.MinOffset;
2233 int64_t NewMaxOffset = LU.MaxOffset;
2234 MemAccessTy NewAccessTy = AccessTy;
2236 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2237 // something conservative, however this can pessimize in the case that one of
2238 // the uses will have all its uses outside the loop, for example.
2239 if (LU.Kind != Kind)
2242 // Check for a mismatched access type, and fall back conservatively as needed.
2243 // TODO: Be less conservative when the type is similar and can use the same
2244 // addressing modes.
2245 if (Kind == LSRUse::Address) {
2246 if (AccessTy != LU.AccessTy)
2247 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext());
2250 // Conservatively assume HasBaseReg is true for now.
2251 if (NewOffset < LU.MinOffset) {
2252 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2253 LU.MaxOffset - NewOffset, HasBaseReg))
2255 NewMinOffset = NewOffset;
2256 } else if (NewOffset > LU.MaxOffset) {
2257 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2258 NewOffset - LU.MinOffset, HasBaseReg))
2260 NewMaxOffset = NewOffset;
2264 LU.MinOffset = NewMinOffset;
2265 LU.MaxOffset = NewMaxOffset;
2266 LU.AccessTy = NewAccessTy;
2267 if (NewOffset != LU.Offsets.back())
2268 LU.Offsets.push_back(NewOffset);
2272 /// Return an LSRUse index and an offset value for a fixup which needs the given
2273 /// expression, with the given kind and optional access type. Either reuse an
2274 /// existing use or create a new one, as needed.
2275 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2276 LSRUse::KindType Kind,
2277 MemAccessTy AccessTy) {
2278 const SCEV *Copy = Expr;
2279 int64_t Offset = ExtractImmediate(Expr, SE);
2281 // Basic uses can't accept any offset, for example.
2282 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2283 Offset, /*HasBaseReg=*/ true)) {
2288 std::pair<UseMapTy::iterator, bool> P =
2289 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2291 // A use already existed with this base.
2292 size_t LUIdx = P.first->second;
2293 LSRUse &LU = Uses[LUIdx];
2294 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2296 return std::make_pair(LUIdx, Offset);
2299 // Create a new use.
2300 size_t LUIdx = Uses.size();
2301 P.first->second = LUIdx;
2302 Uses.push_back(LSRUse(Kind, AccessTy));
2303 LSRUse &LU = Uses[LUIdx];
2305 // We don't need to track redundant offsets, but we don't need to go out
2306 // of our way here to avoid them.
2307 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2308 LU.Offsets.push_back(Offset);
2310 LU.MinOffset = Offset;
2311 LU.MaxOffset = Offset;
2312 return std::make_pair(LUIdx, Offset);
2315 /// Delete the given use from the Uses list.
2316 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2317 if (&LU != &Uses.back())
2318 std::swap(LU, Uses.back());
2322 RegUses.swapAndDropUse(LUIdx, Uses.size());
2325 /// Look for a use distinct from OrigLU which is has a formula that has the same
2326 /// registers as the given formula.
2328 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2329 const LSRUse &OrigLU) {
2330 // Search all uses for the formula. This could be more clever.
2331 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2332 LSRUse &LU = Uses[LUIdx];
2333 // Check whether this use is close enough to OrigLU, to see whether it's
2334 // worthwhile looking through its formulae.
2335 // Ignore ICmpZero uses because they may contain formulae generated by
2336 // GenerateICmpZeroScales, in which case adding fixup offsets may
2338 if (&LU != &OrigLU &&
2339 LU.Kind != LSRUse::ICmpZero &&
2340 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2341 LU.WidestFixupType == OrigLU.WidestFixupType &&
2342 LU.HasFormulaWithSameRegs(OrigF)) {
2343 // Scan through this use's formulae.
2344 for (const Formula &F : LU.Formulae) {
2345 // Check to see if this formula has the same registers and symbols
2347 if (F.BaseRegs == OrigF.BaseRegs &&
2348 F.ScaledReg == OrigF.ScaledReg &&
2349 F.BaseGV == OrigF.BaseGV &&
2350 F.Scale == OrigF.Scale &&
2351 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2352 if (F.BaseOffset == 0)
2354 // This is the formula where all the registers and symbols matched;
2355 // there aren't going to be any others. Since we declined it, we
2356 // can skip the rest of the formulae and proceed to the next LSRUse.
2363 // Nothing looked good.
2367 void LSRInstance::CollectInterestingTypesAndFactors() {
2368 SmallSetVector<const SCEV *, 4> Strides;
2370 // Collect interesting types and strides.
2371 SmallVector<const SCEV *, 4> Worklist;
2372 for (const IVStrideUse &U : IU) {
2373 const SCEV *Expr = IU.getExpr(U);
2375 // Collect interesting types.
2376 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2378 // Add strides for mentioned loops.
2379 Worklist.push_back(Expr);
2381 const SCEV *S = Worklist.pop_back_val();
2382 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2383 if (AR->getLoop() == L)
2384 Strides.insert(AR->getStepRecurrence(SE));
2385 Worklist.push_back(AR->getStart());
2386 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2387 Worklist.append(Add->op_begin(), Add->op_end());
2389 } while (!Worklist.empty());
2392 // Compute interesting factors from the set of interesting strides.
2393 for (SmallSetVector<const SCEV *, 4>::const_iterator
2394 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2395 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2396 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2397 const SCEV *OldStride = *I;
2398 const SCEV *NewStride = *NewStrideIter;
2400 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2401 SE.getTypeSizeInBits(NewStride->getType())) {
2402 if (SE.getTypeSizeInBits(OldStride->getType()) >
2403 SE.getTypeSizeInBits(NewStride->getType()))
2404 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2406 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2408 if (const SCEVConstant *Factor =
2409 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2411 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2412 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2413 } else if (const SCEVConstant *Factor =
2414 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2417 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2418 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2422 // If all uses use the same type, don't bother looking for truncation-based
2424 if (Types.size() == 1)
2427 DEBUG(print_factors_and_types(dbgs()));
2430 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2431 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2432 /// IVStrideUses, we could partially skip this.
2433 static User::op_iterator
2434 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2435 Loop *L, ScalarEvolution &SE) {
2436 for(; OI != OE; ++OI) {
2437 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2438 if (!SE.isSCEVable(Oper->getType()))
2441 if (const SCEVAddRecExpr *AR =
2442 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2443 if (AR->getLoop() == L)
2451 /// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2452 /// a convenient helper.
2453 static Value *getWideOperand(Value *Oper) {
2454 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2455 return Trunc->getOperand(0);
2459 /// Return true if we allow an IV chain to include both types.
2460 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2461 Type *LType = LVal->getType();
2462 Type *RType = RVal->getType();
2463 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2466 /// Return an approximation of this SCEV expression's "base", or NULL for any
2467 /// constant. Returning the expression itself is conservative. Returning a
2468 /// deeper subexpression is more precise and valid as long as it isn't less
2469 /// complex than another subexpression. For expressions involving multiple
2470 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2471 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2474 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2475 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2476 static const SCEV *getExprBase(const SCEV *S) {
2477 switch (S->getSCEVType()) {
2478 default: // uncluding scUnknown.
2483 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2485 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2487 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2489 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2490 // there's nothing more complex.
2491 // FIXME: not sure if we want to recognize negation.
2492 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2493 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2494 E(Add->op_begin()); I != E; ++I) {
2495 const SCEV *SubExpr = *I;
2496 if (SubExpr->getSCEVType() == scAddExpr)
2497 return getExprBase(SubExpr);
2499 if (SubExpr->getSCEVType() != scMulExpr)
2502 return S; // all operands are scaled, be conservative.
2505 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2509 /// Return true if the chain increment is profitable to expand into a loop
2510 /// invariant value, which may require its own register. A profitable chain
2511 /// increment will be an offset relative to the same base. We allow such offsets
2512 /// to potentially be used as chain increment as long as it's not obviously
2513 /// expensive to expand using real instructions.
2514 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2515 const SCEV *IncExpr,
2516 ScalarEvolution &SE) {
2517 // Aggressively form chains when -stress-ivchain.
2521 // Do not replace a constant offset from IV head with a nonconstant IV
2523 if (!isa<SCEVConstant>(IncExpr)) {
2524 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2525 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2529 SmallPtrSet<const SCEV*, 8> Processed;
2530 return !isHighCostExpansion(IncExpr, Processed, SE);
2533 /// Return true if the number of registers needed for the chain is estimated to
2534 /// be less than the number required for the individual IV users. First prohibit
2535 /// any IV users that keep the IV live across increments (the Users set should
2536 /// be empty). Next count the number and type of increments in the chain.
2538 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2539 /// effectively use postinc addressing modes. Only consider it profitable it the
2540 /// increments can be computed in fewer registers when chained.
2542 /// TODO: Consider IVInc free if it's already used in another chains.
2544 isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2545 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2549 if (!Chain.hasIncs())
2552 if (!Users.empty()) {
2553 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2554 for (Instruction *Inst : Users) {
2555 dbgs() << " " << *Inst << "\n";
2559 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2561 // The chain itself may require a register, so intialize cost to 1.
2564 // A complete chain likely eliminates the need for keeping the original IV in
2565 // a register. LSR does not currently know how to form a complete chain unless
2566 // the header phi already exists.
2567 if (isa<PHINode>(Chain.tailUserInst())
2568 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2571 const SCEV *LastIncExpr = nullptr;
2572 unsigned NumConstIncrements = 0;
2573 unsigned NumVarIncrements = 0;
2574 unsigned NumReusedIncrements = 0;
2575 for (const IVInc &Inc : Chain) {
2576 if (Inc.IncExpr->isZero())
2579 // Incrementing by zero or some constant is neutral. We assume constants can
2580 // be folded into an addressing mode or an add's immediate operand.
2581 if (isa<SCEVConstant>(Inc.IncExpr)) {
2582 ++NumConstIncrements;
2586 if (Inc.IncExpr == LastIncExpr)
2587 ++NumReusedIncrements;
2591 LastIncExpr = Inc.IncExpr;
2593 // An IV chain with a single increment is handled by LSR's postinc
2594 // uses. However, a chain with multiple increments requires keeping the IV's
2595 // value live longer than it needs to be if chained.
2596 if (NumConstIncrements > 1)
2599 // Materializing increment expressions in the preheader that didn't exist in
2600 // the original code may cost a register. For example, sign-extended array
2601 // indices can produce ridiculous increments like this:
2602 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2603 cost += NumVarIncrements;
2605 // Reusing variable increments likely saves a register to hold the multiple of
2607 cost -= NumReusedIncrements;
2609 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2615 /// Add this IV user to an existing chain or make it the head of a new chain.
2616 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2617 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2618 // When IVs are used as types of varying widths, they are generally converted
2619 // to a wider type with some uses remaining narrow under a (free) trunc.
2620 Value *const NextIV = getWideOperand(IVOper);
2621 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2622 const SCEV *const OperExprBase = getExprBase(OperExpr);
2624 // Visit all existing chains. Check if its IVOper can be computed as a
2625 // profitable loop invariant increment from the last link in the Chain.
2626 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2627 const SCEV *LastIncExpr = nullptr;
2628 for (; ChainIdx < NChains; ++ChainIdx) {
2629 IVChain &Chain = IVChainVec[ChainIdx];
2631 // Prune the solution space aggressively by checking that both IV operands
2632 // are expressions that operate on the same unscaled SCEVUnknown. This
2633 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2634 // first avoids creating extra SCEV expressions.
2635 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2638 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2639 if (!isCompatibleIVType(PrevIV, NextIV))
2642 // A phi node terminates a chain.
2643 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2646 // The increment must be loop-invariant so it can be kept in a register.
2647 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2648 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2649 if (!SE.isLoopInvariant(IncExpr, L))
2652 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2653 LastIncExpr = IncExpr;
2657 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2658 // bother for phi nodes, because they must be last in the chain.
2659 if (ChainIdx == NChains) {
2660 if (isa<PHINode>(UserInst))
2662 if (NChains >= MaxChains && !StressIVChain) {
2663 DEBUG(dbgs() << "IV Chain Limit\n");
2666 LastIncExpr = OperExpr;
2667 // IVUsers may have skipped over sign/zero extensions. We don't currently
2668 // attempt to form chains involving extensions unless they can be hoisted
2669 // into this loop's AddRec.
2670 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2673 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2675 ChainUsersVec.resize(NChains);
2676 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2677 << ") IV=" << *LastIncExpr << "\n");
2679 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2680 << ") IV+" << *LastIncExpr << "\n");
2681 // Add this IV user to the end of the chain.
2682 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2684 IVChain &Chain = IVChainVec[ChainIdx];
2686 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2687 // This chain's NearUsers become FarUsers.
2688 if (!LastIncExpr->isZero()) {
2689 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2694 // All other uses of IVOperand become near uses of the chain.
2695 // We currently ignore intermediate values within SCEV expressions, assuming
2696 // they will eventually be used be the current chain, or can be computed
2697 // from one of the chain increments. To be more precise we could
2698 // transitively follow its user and only add leaf IV users to the set.
2699 for (User *U : IVOper->users()) {
2700 Instruction *OtherUse = dyn_cast<Instruction>(U);
2703 // Uses in the chain will no longer be uses if the chain is formed.
2704 // Include the head of the chain in this iteration (not Chain.begin()).
2705 IVChain::const_iterator IncIter = Chain.Incs.begin();
2706 IVChain::const_iterator IncEnd = Chain.Incs.end();
2707 for( ; IncIter != IncEnd; ++IncIter) {
2708 if (IncIter->UserInst == OtherUse)
2711 if (IncIter != IncEnd)
2714 if (SE.isSCEVable(OtherUse->getType())
2715 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2716 && IU.isIVUserOrOperand(OtherUse)) {
2719 NearUsers.insert(OtherUse);
2722 // Since this user is part of the chain, it's no longer considered a use
2724 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2727 /// Populate the vector of Chains.
2729 /// This decreases ILP at the architecture level. Targets with ample registers,
2730 /// multiple memory ports, and no register renaming probably don't want
2731 /// this. However, such targets should probably disable LSR altogether.
2733 /// The job of LSR is to make a reasonable choice of induction variables across
2734 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2735 /// ILP *within the loop* if the target wants it.
2737 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2738 /// will not reorder memory operations, it will recognize this as a chain, but
2739 /// will generate redundant IV increments. Ideally this would be corrected later
2740 /// by a smart scheduler:
2746 /// TODO: Walk the entire domtree within this loop, not just the path to the
2747 /// loop latch. This will discover chains on side paths, but requires
2748 /// maintaining multiple copies of the Chains state.
2749 void LSRInstance::CollectChains() {
2750 DEBUG(dbgs() << "Collecting IV Chains.\n");
2751 SmallVector<ChainUsers, 8> ChainUsersVec;
2753 SmallVector<BasicBlock *,8> LatchPath;
2754 BasicBlock *LoopHeader = L->getHeader();
2755 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2756 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2757 LatchPath.push_back(Rung->getBlock());
2759 LatchPath.push_back(LoopHeader);
2761 // Walk the instruction stream from the loop header to the loop latch.
2762 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2763 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2764 BBIter != BBEnd; ++BBIter) {
2765 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2767 // Skip instructions that weren't seen by IVUsers analysis.
2768 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&*I))
2771 // Ignore users that are part of a SCEV expression. This way we only
2772 // consider leaf IV Users. This effectively rediscovers a portion of
2773 // IVUsers analysis but in program order this time.
2774 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(&*I)))
2777 // Remove this instruction from any NearUsers set it may be in.
2778 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2779 ChainIdx < NChains; ++ChainIdx) {
2780 ChainUsersVec[ChainIdx].NearUsers.erase(&*I);
2782 // Search for operands that can be chained.
2783 SmallPtrSet<Instruction*, 4> UniqueOperands;
2784 User::op_iterator IVOpEnd = I->op_end();
2785 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2786 while (IVOpIter != IVOpEnd) {
2787 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2788 if (UniqueOperands.insert(IVOpInst).second)
2789 ChainInstruction(&*I, IVOpInst, ChainUsersVec);
2790 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2792 } // Continue walking down the instructions.
2793 } // Continue walking down the domtree.
2794 // Visit phi backedges to determine if the chain can generate the IV postinc.
2795 for (BasicBlock::iterator I = L->getHeader()->begin();
2796 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2797 if (!SE.isSCEVable(PN->getType()))
2801 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2803 ChainInstruction(PN, IncV, ChainUsersVec);
2805 // Remove any unprofitable chains.
2806 unsigned ChainIdx = 0;
2807 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2808 UsersIdx < NChains; ++UsersIdx) {
2809 if (!isProfitableChain(IVChainVec[UsersIdx],
2810 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2812 // Preserve the chain at UsesIdx.
2813 if (ChainIdx != UsersIdx)
2814 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2815 FinalizeChain(IVChainVec[ChainIdx]);
2818 IVChainVec.resize(ChainIdx);
2821 void LSRInstance::FinalizeChain(IVChain &Chain) {
2822 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2823 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2825 for (const IVInc &Inc : Chain) {
2826 DEBUG(dbgs() << " Inc: " << Inc.UserInst << "\n");
2827 auto UseI = std::find(Inc.UserInst->op_begin(), Inc.UserInst->op_end(),
2829 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
2830 IVIncSet.insert(UseI);
2834 /// Return true if the IVInc can be folded into an addressing mode.
2835 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2836 Value *Operand, const TargetTransformInfo &TTI) {
2837 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2838 if (!IncConst || !isAddressUse(UserInst, Operand))
2841 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2844 MemAccessTy AccessTy = getAccessType(UserInst);
2845 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2846 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
2847 IncOffset, /*HaseBaseReg=*/false))
2853 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
2854 /// user's operand from the previous IV user's operand.
2855 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2856 SmallVectorImpl<WeakVH> &DeadInsts) {
2857 // Find the new IVOperand for the head of the chain. It may have been replaced
2859 const IVInc &Head = Chain.Incs[0];
2860 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2861 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2862 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2864 Value *IVSrc = nullptr;
2865 while (IVOpIter != IVOpEnd) {
2866 IVSrc = getWideOperand(*IVOpIter);
2868 // If this operand computes the expression that the chain needs, we may use
2869 // it. (Check this after setting IVSrc which is used below.)
2871 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2872 // narrow for the chain, so we can no longer use it. We do allow using a
2873 // wider phi, assuming the LSR checked for free truncation. In that case we
2874 // should already have a truncate on this operand such that
2875 // getSCEV(IVSrc) == IncExpr.
2876 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2877 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2880 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2882 if (IVOpIter == IVOpEnd) {
2883 // Gracefully give up on this chain.
2884 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2888 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2889 Type *IVTy = IVSrc->getType();
2890 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2891 const SCEV *LeftOverExpr = nullptr;
2892 for (const IVInc &Inc : Chain) {
2893 Instruction *InsertPt = Inc.UserInst;
2894 if (isa<PHINode>(InsertPt))
2895 InsertPt = L->getLoopLatch()->getTerminator();
2897 // IVOper will replace the current IV User's operand. IVSrc is the IV
2898 // value currently held in a register.
2899 Value *IVOper = IVSrc;
2900 if (!Inc.IncExpr->isZero()) {
2901 // IncExpr was the result of subtraction of two narrow values, so must
2903 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
2904 LeftOverExpr = LeftOverExpr ?
2905 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2907 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2908 // Expand the IV increment.
2909 Rewriter.clearPostInc();
2910 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2911 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2912 SE.getUnknown(IncV));
2913 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2915 // If an IV increment can't be folded, use it as the next IV value.
2916 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
2917 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2919 LeftOverExpr = nullptr;
2922 Type *OperTy = Inc.IVOperand->getType();
2923 if (IVTy != OperTy) {
2924 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2925 "cannot extend a chained IV");
2926 IRBuilder<> Builder(InsertPt);
2927 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2929 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
2930 DeadInsts.emplace_back(Inc.IVOperand);
2932 // If LSR created a new, wider phi, we may also replace its postinc. We only
2933 // do this if we also found a wide value for the head of the chain.
2934 if (isa<PHINode>(Chain.tailUserInst())) {
2935 for (BasicBlock::iterator I = L->getHeader()->begin();
2936 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2937 if (!isCompatibleIVType(Phi, IVSrc))
2939 Instruction *PostIncV = dyn_cast<Instruction>(
2940 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2941 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2943 Value *IVOper = IVSrc;
2944 Type *PostIncTy = PostIncV->getType();
2945 if (IVTy != PostIncTy) {
2946 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2947 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2948 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2949 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2951 Phi->replaceUsesOfWith(PostIncV, IVOper);
2952 DeadInsts.emplace_back(PostIncV);
2957 void LSRInstance::CollectFixupsAndInitialFormulae() {
2958 for (const IVStrideUse &U : IU) {
2959 Instruction *UserInst = U.getUser();
2960 // Skip IV users that are part of profitable IV Chains.
2961 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2962 U.getOperandValToReplace());
2963 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2964 if (IVIncSet.count(UseI))
2968 LSRFixup &LF = getNewFixup();
2969 LF.UserInst = UserInst;
2970 LF.OperandValToReplace = U.getOperandValToReplace();
2971 LF.PostIncLoops = U.getPostIncLoops();
2973 LSRUse::KindType Kind = LSRUse::Basic;
2974 MemAccessTy AccessTy;
2975 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2976 Kind = LSRUse::Address;
2977 AccessTy = getAccessType(LF.UserInst);
2980 const SCEV *S = IU.getExpr(U);
2982 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2983 // (N - i == 0), and this allows (N - i) to be the expression that we work
2984 // with rather than just N or i, so we can consider the register
2985 // requirements for both N and i at the same time. Limiting this code to
2986 // equality icmps is not a problem because all interesting loops use
2987 // equality icmps, thanks to IndVarSimplify.
2988 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2989 if (CI->isEquality()) {
2990 // Swap the operands if needed to put the OperandValToReplace on the
2991 // left, for consistency.
2992 Value *NV = CI->getOperand(1);
2993 if (NV == LF.OperandValToReplace) {
2994 CI->setOperand(1, CI->getOperand(0));
2995 CI->setOperand(0, NV);
2996 NV = CI->getOperand(1);
3000 // x == y --> x - y == 0
3001 const SCEV *N = SE.getSCEV(NV);
3002 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3003 // S is normalized, so normalize N before folding it into S
3004 // to keep the result normalized.
3005 N = TransformForPostIncUse(Normalize, N, CI, nullptr,
3006 LF.PostIncLoops, SE, DT);
3007 Kind = LSRUse::ICmpZero;
3008 S = SE.getMinusSCEV(N, S);
3011 // -1 and the negations of all interesting strides (except the negation
3012 // of -1) are now also interesting.
3013 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3014 if (Factors[i] != -1)
3015 Factors.insert(-(uint64_t)Factors[i]);
3019 // Set up the initial formula for this use.
3020 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3022 LF.Offset = P.second;
3023 LSRUse &LU = Uses[LF.LUIdx];
3024 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3025 if (!LU.WidestFixupType ||
3026 SE.getTypeSizeInBits(LU.WidestFixupType) <
3027 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3028 LU.WidestFixupType = LF.OperandValToReplace->getType();
3030 // If this is the first use of this LSRUse, give it a formula.
3031 if (LU.Formulae.empty()) {
3032 InsertInitialFormula(S, LU, LF.LUIdx);
3033 CountRegisters(LU.Formulae.back(), LF.LUIdx);
3037 DEBUG(print_fixups(dbgs()));
3040 /// Insert a formula for the given expression into the given use, separating out
3041 /// loop-variant portions from loop-invariant and loop-computable portions.
3043 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3044 // Mark uses whose expressions cannot be expanded.
3045 if (!isSafeToExpand(S, SE))
3046 LU.RigidFormula = true;
3049 F.initialMatch(S, L, SE);
3050 bool Inserted = InsertFormula(LU, LUIdx, F);
3051 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3054 /// Insert a simple single-register formula for the given expression into the
3057 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3058 LSRUse &LU, size_t LUIdx) {
3060 F.BaseRegs.push_back(S);
3061 F.HasBaseReg = true;
3062 bool Inserted = InsertFormula(LU, LUIdx, F);
3063 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3066 /// Note which registers are used by the given formula, updating RegUses.
3067 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3069 RegUses.countRegister(F.ScaledReg, LUIdx);
3070 for (const SCEV *BaseReg : F.BaseRegs)
3071 RegUses.countRegister(BaseReg, LUIdx);
3074 /// If the given formula has not yet been inserted, add it to the list, and
3075 /// return true. Return false otherwise.
3076 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3077 // Do not insert formula that we will not be able to expand.
3078 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3079 "Formula is illegal");
3080 if (!LU.InsertFormula(F))
3083 CountRegisters(F, LUIdx);
3087 /// Check for other uses of loop-invariant values which we're tracking. These
3088 /// other uses will pin these values in registers, making them less profitable
3089 /// for elimination.
3090 /// TODO: This currently misses non-constant addrec step registers.
3091 /// TODO: Should this give more weight to users inside the loop?
3093 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3094 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3095 SmallPtrSet<const SCEV *, 32> Visited;
3097 while (!Worklist.empty()) {
3098 const SCEV *S = Worklist.pop_back_val();
3100 // Don't process the same SCEV twice
3101 if (!Visited.insert(S).second)
3104 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3105 Worklist.append(N->op_begin(), N->op_end());
3106 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3107 Worklist.push_back(C->getOperand());
3108 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3109 Worklist.push_back(D->getLHS());
3110 Worklist.push_back(D->getRHS());
3111 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3112 const Value *V = US->getValue();
3113 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3114 // Look for instructions defined outside the loop.
3115 if (L->contains(Inst)) continue;
3116 } else if (isa<UndefValue>(V))
3117 // Undef doesn't have a live range, so it doesn't matter.
3119 for (const Use &U : V->uses()) {
3120 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3121 // Ignore non-instructions.
3124 // Ignore instructions in other functions (as can happen with
3126 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3128 // Ignore instructions not dominated by the loop.
3129 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3130 UserInst->getParent() :
3131 cast<PHINode>(UserInst)->getIncomingBlock(
3132 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3133 if (!DT.dominates(L->getHeader(), UseBB))
3135 // Ignore uses which are part of other SCEV expressions, to avoid
3136 // analyzing them multiple times.
3137 if (SE.isSCEVable(UserInst->getType())) {
3138 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3139 // If the user is a no-op, look through to its uses.
3140 if (!isa<SCEVUnknown>(UserS))
3144 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3148 // Ignore icmp instructions which are already being analyzed.
3149 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3150 unsigned OtherIdx = !U.getOperandNo();
3151 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3152 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3156 LSRFixup &LF = getNewFixup();
3157 LF.UserInst = const_cast<Instruction *>(UserInst);
3158 LF.OperandValToReplace = U;
3159 std::pair<size_t, int64_t> P = getUse(
3160 S, LSRUse::Basic, MemAccessTy());
3162 LF.Offset = P.second;
3163 LSRUse &LU = Uses[LF.LUIdx];
3164 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3165 if (!LU.WidestFixupType ||
3166 SE.getTypeSizeInBits(LU.WidestFixupType) <
3167 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3168 LU.WidestFixupType = LF.OperandValToReplace->getType();
3169 InsertSupplementalFormula(US, LU, LF.LUIdx);
3170 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3177 /// Split S into subexpressions which can be pulled out into separate
3178 /// registers. If C is non-null, multiply each subexpression by C.
3180 /// Return remainder expression after factoring the subexpressions captured by
3181 /// Ops. If Ops is complete, return NULL.
3182 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3183 SmallVectorImpl<const SCEV *> &Ops,
3185 ScalarEvolution &SE,
3186 unsigned Depth = 0) {
3187 // Arbitrarily cap recursion to protect compile time.
3191 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3192 // Break out add operands.
3193 for (const SCEV *S : Add->operands()) {
3194 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3196 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3199 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3200 // Split a non-zero base out of an addrec.
3201 if (AR->getStart()->isZero())
3204 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3205 C, Ops, L, SE, Depth+1);
3206 // Split the non-zero AddRec unless it is part of a nested recurrence that
3207 // does not pertain to this loop.
3208 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3209 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3210 Remainder = nullptr;
3212 if (Remainder != AR->getStart()) {
3214 Remainder = SE.getConstant(AR->getType(), 0);
3215 return SE.getAddRecExpr(Remainder,
3216 AR->getStepRecurrence(SE),
3218 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3221 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3222 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3223 if (Mul->getNumOperands() != 2)
3225 if (const SCEVConstant *Op0 =
3226 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3227 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3228 const SCEV *Remainder =
3229 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3231 Ops.push_back(SE.getMulExpr(C, Remainder));
3238 /// \brief Helper function for LSRInstance::GenerateReassociations.
3239 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3240 const Formula &Base,
3241 unsigned Depth, size_t Idx,
3243 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3244 SmallVector<const SCEV *, 8> AddOps;
3245 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3247 AddOps.push_back(Remainder);
3249 if (AddOps.size() == 1)
3252 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3256 // Loop-variant "unknown" values are uninteresting; we won't be able to
3257 // do anything meaningful with them.
3258 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3261 // Don't pull a constant into a register if the constant could be folded
3262 // into an immediate field.
3263 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3264 LU.AccessTy, *J, Base.getNumRegs() > 1))
3267 // Collect all operands except *J.
3268 SmallVector<const SCEV *, 8> InnerAddOps(
3269 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3270 InnerAddOps.append(std::next(J),
3271 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3273 // Don't leave just a constant behind in a register if the constant could
3274 // be folded into an immediate field.
3275 if (InnerAddOps.size() == 1 &&
3276 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3277 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3280 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3281 if (InnerSum->isZero())
3285 // Add the remaining pieces of the add back into the new formula.
3286 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3287 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3288 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3289 InnerSumSC->getValue()->getZExtValue())) {
3291 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3293 F.ScaledReg = nullptr;
3295 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3296 } else if (IsScaledReg)
3297 F.ScaledReg = InnerSum;
3299 F.BaseRegs[Idx] = InnerSum;
3301 // Add J as its own register, or an unfolded immediate.
3302 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3303 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3304 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3305 SC->getValue()->getZExtValue()))
3307 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3309 F.BaseRegs.push_back(*J);
3310 // We may have changed the number of register in base regs, adjust the
3311 // formula accordingly.
3314 if (InsertFormula(LU, LUIdx, F))
3315 // If that formula hadn't been seen before, recurse to find more like
3317 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3321 /// Split out subexpressions from adds and the bases of addrecs.
3322 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3323 Formula Base, unsigned Depth) {
3324 assert(Base.isCanonical() && "Input must be in the canonical form");
3325 // Arbitrarily cap recursion to protect compile time.
3329 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3330 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3332 if (Base.Scale == 1)
3333 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3334 /* Idx */ -1, /* IsScaledReg */ true);
3337 /// Generate a formula consisting of all of the loop-dominating registers added
3338 /// into a single register.
3339 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3341 // This method is only interesting on a plurality of registers.
3342 if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3345 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3346 // processing the formula.
3350 SmallVector<const SCEV *, 4> Ops;
3351 for (const SCEV *BaseReg : Base.BaseRegs) {
3352 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3353 !SE.hasComputableLoopEvolution(BaseReg, L))
3354 Ops.push_back(BaseReg);
3356 F.BaseRegs.push_back(BaseReg);
3358 if (Ops.size() > 1) {
3359 const SCEV *Sum = SE.getAddExpr(Ops);
3360 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3361 // opportunity to fold something. For now, just ignore such cases
3362 // rather than proceed with zero in a register.
3363 if (!Sum->isZero()) {
3364 F.BaseRegs.push_back(Sum);
3366 (void)InsertFormula(LU, LUIdx, F);
3371 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3372 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3373 const Formula &Base, size_t Idx,
3375 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3376 GlobalValue *GV = ExtractSymbol(G, SE);
3377 if (G->isZero() || !GV)
3381 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3386 F.BaseRegs[Idx] = G;
3387 (void)InsertFormula(LU, LUIdx, F);
3390 /// Generate reuse formulae using symbolic offsets.
3391 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3393 // We can't add a symbolic offset if the address already contains one.
3394 if (Base.BaseGV) return;
3396 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3397 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3398 if (Base.Scale == 1)
3399 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3400 /* IsScaledReg */ true);
3403 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3404 void LSRInstance::GenerateConstantOffsetsImpl(
3405 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3406 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3407 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3408 for (int64_t Offset : Worklist) {
3410 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3411 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3413 // Add the offset to the base register.
3414 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3415 // If it cancelled out, drop the base register, otherwise update it.
3416 if (NewG->isZero()) {
3419 F.ScaledReg = nullptr;
3421 F.deleteBaseReg(F.BaseRegs[Idx]);
3423 } else if (IsScaledReg)
3426 F.BaseRegs[Idx] = NewG;
3428 (void)InsertFormula(LU, LUIdx, F);
3432 int64_t Imm = ExtractImmediate(G, SE);
3433 if (G->isZero() || Imm == 0)
3436 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3437 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3442 F.BaseRegs[Idx] = G;
3443 (void)InsertFormula(LU, LUIdx, F);
3446 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3447 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3449 // TODO: For now, just add the min and max offset, because it usually isn't
3450 // worthwhile looking at everything inbetween.
3451 SmallVector<int64_t, 2> Worklist;
3452 Worklist.push_back(LU.MinOffset);
3453 if (LU.MaxOffset != LU.MinOffset)
3454 Worklist.push_back(LU.MaxOffset);
3456 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3457 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3458 if (Base.Scale == 1)
3459 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3460 /* IsScaledReg */ true);
3463 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3464 /// == y -> x*c == y*c.
3465 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3467 if (LU.Kind != LSRUse::ICmpZero) return;
3469 // Determine the integer type for the base formula.
3470 Type *IntTy = Base.getType();
3472 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3474 // Don't do this if there is more than one offset.
3475 if (LU.MinOffset != LU.MaxOffset) return;
3477 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3479 // Check each interesting stride.
3480 for (int64_t Factor : Factors) {
3481 // Check that the multiplication doesn't overflow.
3482 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3484 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3485 if (NewBaseOffset / Factor != Base.BaseOffset)
3487 // If the offset will be truncated at this use, check that it is in bounds.
3488 if (!IntTy->isPointerTy() &&
3489 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3492 // Check that multiplying with the use offset doesn't overflow.
3493 int64_t Offset = LU.MinOffset;
3494 if (Offset == INT64_MIN && Factor == -1)
3496 Offset = (uint64_t)Offset * Factor;
3497 if (Offset / Factor != LU.MinOffset)
3499 // If the offset will be truncated at this use, check that it is in bounds.
3500 if (!IntTy->isPointerTy() &&
3501 !ConstantInt::isValueValidForType(IntTy, Offset))
3505 F.BaseOffset = NewBaseOffset;
3507 // Check that this scale is legal.
3508 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3511 // Compensate for the use having MinOffset built into it.
3512 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3514 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3516 // Check that multiplying with each base register doesn't overflow.
3517 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3518 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3519 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3523 // Check that multiplying with the scaled register doesn't overflow.
3525 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3526 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3530 // Check that multiplying with the unfolded offset doesn't overflow.
3531 if (F.UnfoldedOffset != 0) {
3532 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3534 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3535 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3537 // If the offset will be truncated, check that it is in bounds.
3538 if (!IntTy->isPointerTy() &&
3539 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3543 // If we make it here and it's legal, add it.
3544 (void)InsertFormula(LU, LUIdx, F);
3549 /// Generate stride factor reuse formulae by making use of scaled-offset address
3550 /// modes, for example.
3551 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3552 // Determine the integer type for the base formula.
3553 Type *IntTy = Base.getType();
3556 // If this Formula already has a scaled register, we can't add another one.
3557 // Try to unscale the formula to generate a better scale.
3558 if (Base.Scale != 0 && !Base.unscale())
3561 assert(Base.Scale == 0 && "unscale did not did its job!");
3563 // Check each interesting stride.
3564 for (int64_t Factor : Factors) {
3565 Base.Scale = Factor;
3566 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3567 // Check whether this scale is going to be legal.
3568 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3570 // As a special-case, handle special out-of-loop Basic users specially.
3571 // TODO: Reconsider this special case.
3572 if (LU.Kind == LSRUse::Basic &&
3573 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3574 LU.AccessTy, Base) &&
3575 LU.AllFixupsOutsideLoop)
3576 LU.Kind = LSRUse::Special;
3580 // For an ICmpZero, negating a solitary base register won't lead to
3582 if (LU.Kind == LSRUse::ICmpZero &&
3583 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3585 // For each addrec base reg, apply the scale, if possible.
3586 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3587 if (const SCEVAddRecExpr *AR =
3588 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3589 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3590 if (FactorS->isZero())
3592 // Divide out the factor, ignoring high bits, since we'll be
3593 // scaling the value back up in the end.
3594 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3595 // TODO: This could be optimized to avoid all the copying.
3597 F.ScaledReg = Quotient;
3598 F.deleteBaseReg(F.BaseRegs[i]);
3599 // The canonical representation of 1*reg is reg, which is already in
3600 // Base. In that case, do not try to insert the formula, it will be
3602 if (F.Scale == 1 && F.BaseRegs.empty())
3604 (void)InsertFormula(LU, LUIdx, F);
3610 /// Generate reuse formulae from different IV types.
3611 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3612 // Don't bother truncating symbolic values.
3613 if (Base.BaseGV) return;
3615 // Determine the integer type for the base formula.
3616 Type *DstTy = Base.getType();
3618 DstTy = SE.getEffectiveSCEVType(DstTy);
3620 for (Type *SrcTy : Types) {
3621 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3624 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3625 for (const SCEV *&BaseReg : F.BaseRegs)
3626 BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3628 // TODO: This assumes we've done basic processing on all uses and
3629 // have an idea what the register usage is.
3630 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3633 (void)InsertFormula(LU, LUIdx, F);
3640 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3641 /// modifications so that the search phase doesn't have to worry about the data
3642 /// structures moving underneath it.
3646 const SCEV *OrigReg;
3648 WorkItem(size_t LI, int64_t I, const SCEV *R)
3649 : LUIdx(LI), Imm(I), OrigReg(R) {}
3651 void print(raw_ostream &OS) const;
3657 void WorkItem::print(raw_ostream &OS) const {
3658 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3659 << " , add offset " << Imm;
3662 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3663 void WorkItem::dump() const {
3664 print(errs()); errs() << '\n';
3668 /// Look for registers which are a constant distance apart and try to form reuse
3669 /// opportunities between them.
3670 void LSRInstance::GenerateCrossUseConstantOffsets() {
3671 // Group the registers by their value without any added constant offset.
3672 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3673 DenseMap<const SCEV *, ImmMapTy> Map;
3674 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3675 SmallVector<const SCEV *, 8> Sequence;
3676 for (const SCEV *Use : RegUses) {
3677 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
3678 int64_t Imm = ExtractImmediate(Reg, SE);
3679 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
3681 Sequence.push_back(Reg);
3682 Pair.first->second.insert(std::make_pair(Imm, Use));
3683 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
3686 // Now examine each set of registers with the same base value. Build up
3687 // a list of work to do and do the work in a separate step so that we're
3688 // not adding formulae and register counts while we're searching.
3689 SmallVector<WorkItem, 32> WorkItems;
3690 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3691 for (const SCEV *Reg : Sequence) {
3692 const ImmMapTy &Imms = Map.find(Reg)->second;
3694 // It's not worthwhile looking for reuse if there's only one offset.
3695 if (Imms.size() == 1)
3698 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3699 for (const auto &Entry : Imms)
3700 dbgs() << ' ' << Entry.first;
3703 // Examine each offset.
3704 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3706 const SCEV *OrigReg = J->second;
3708 int64_t JImm = J->first;
3709 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3711 if (!isa<SCEVConstant>(OrigReg) &&
3712 UsedByIndicesMap[Reg].count() == 1) {
3713 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3717 // Conservatively examine offsets between this orig reg a few selected
3719 ImmMapTy::const_iterator OtherImms[] = {
3720 Imms.begin(), std::prev(Imms.end()),
3721 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3724 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3725 ImmMapTy::const_iterator M = OtherImms[i];
3726 if (M == J || M == JE) continue;
3728 // Compute the difference between the two.
3729 int64_t Imm = (uint64_t)JImm - M->first;
3730 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3731 LUIdx = UsedByIndices.find_next(LUIdx))
3732 // Make a memo of this use, offset, and register tuple.
3733 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3734 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3741 UsedByIndicesMap.clear();
3742 UniqueItems.clear();
3744 // Now iterate through the worklist and add new formulae.
3745 for (const WorkItem &WI : WorkItems) {
3746 size_t LUIdx = WI.LUIdx;
3747 LSRUse &LU = Uses[LUIdx];
3748 int64_t Imm = WI.Imm;
3749 const SCEV *OrigReg = WI.OrigReg;
3751 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3752 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3753 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3755 // TODO: Use a more targeted data structure.
3756 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3757 Formula F = LU.Formulae[L];
3758 // FIXME: The code for the scaled and unscaled registers looks
3759 // very similar but slightly different. Investigate if they
3760 // could be merged. That way, we would not have to unscale the
3763 // Use the immediate in the scaled register.
3764 if (F.ScaledReg == OrigReg) {
3765 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3766 // Don't create 50 + reg(-50).
3767 if (F.referencesReg(SE.getSCEV(
3768 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3771 NewF.BaseOffset = Offset;
3772 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3775 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3777 // If the new scale is a constant in a register, and adding the constant
3778 // value to the immediate would produce a value closer to zero than the
3779 // immediate itself, then the formula isn't worthwhile.
3780 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3781 if (C->getValue()->isNegative() !=
3782 (NewF.BaseOffset < 0) &&
3783 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3784 .ule(std::abs(NewF.BaseOffset)))
3788 NewF.canonicalize();
3789 (void)InsertFormula(LU, LUIdx, NewF);
3791 // Use the immediate in a base register.
3792 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3793 const SCEV *BaseReg = F.BaseRegs[N];
3794 if (BaseReg != OrigReg)
3797 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3798 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3799 LU.Kind, LU.AccessTy, NewF)) {
3800 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3803 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3805 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3807 // If the new formula has a constant in a register, and adding the
3808 // constant value to the immediate would produce a value closer to
3809 // zero than the immediate itself, then the formula isn't worthwhile.
3810 for (const SCEV *NewReg : NewF.BaseRegs)
3811 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
3812 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3813 std::abs(NewF.BaseOffset)) &&
3814 (C->getValue()->getValue() +
3815 NewF.BaseOffset).countTrailingZeros() >=
3816 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3820 NewF.canonicalize();
3821 (void)InsertFormula(LU, LUIdx, NewF);
3830 /// Generate formulae for each use.
3832 LSRInstance::GenerateAllReuseFormulae() {
3833 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3834 // queries are more precise.
3835 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3836 LSRUse &LU = Uses[LUIdx];
3837 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3838 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3839 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3840 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3842 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3843 LSRUse &LU = Uses[LUIdx];
3844 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3845 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3846 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3847 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3848 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3849 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3850 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3851 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3853 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3854 LSRUse &LU = Uses[LUIdx];
3855 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3856 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3859 GenerateCrossUseConstantOffsets();
3861 DEBUG(dbgs() << "\n"
3862 "After generating reuse formulae:\n";
3863 print_uses(dbgs()));
3866 /// If there are multiple formulae with the same set of registers used
3867 /// by other uses, pick the best one and delete the others.
3868 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3869 DenseSet<const SCEV *> VisitedRegs;
3870 SmallPtrSet<const SCEV *, 16> Regs;
3871 SmallPtrSet<const SCEV *, 16> LoserRegs;
3873 bool ChangedFormulae = false;
3876 // Collect the best formula for each unique set of shared registers. This
3877 // is reset for each use.
3878 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3880 BestFormulaeTy BestFormulae;
3882 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3883 LSRUse &LU = Uses[LUIdx];
3884 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3887 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3888 FIdx != NumForms; ++FIdx) {
3889 Formula &F = LU.Formulae[FIdx];
3891 // Some formulas are instant losers. For example, they may depend on
3892 // nonexistent AddRecs from other loops. These need to be filtered
3893 // immediately, otherwise heuristics could choose them over others leading
3894 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3895 // avoids the need to recompute this information across formulae using the
3896 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3897 // the corresponding bad register from the Regs set.
3900 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3902 if (CostF.isLoser()) {
3903 // During initial formula generation, undesirable formulae are generated
3904 // by uses within other loops that have some non-trivial address mode or
3905 // use the postinc form of the IV. LSR needs to provide these formulae
3906 // as the basis of rediscovering the desired formula that uses an AddRec
3907 // corresponding to the existing phi. Once all formulae have been
3908 // generated, these initial losers may be pruned.
3909 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3913 SmallVector<const SCEV *, 4> Key;
3914 for (const SCEV *Reg : F.BaseRegs) {
3915 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3919 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3920 Key.push_back(F.ScaledReg);
3921 // Unstable sort by host order ok, because this is only used for
3923 std::sort(Key.begin(), Key.end());
3925 std::pair<BestFormulaeTy::const_iterator, bool> P =
3926 BestFormulae.insert(std::make_pair(Key, FIdx));
3930 Formula &Best = LU.Formulae[P.first->second];
3934 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3936 if (CostF < CostBest)
3938 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3940 " in favor of formula "; Best.print(dbgs());
3944 ChangedFormulae = true;
3946 LU.DeleteFormula(F);
3952 // Now that we've filtered out some formulae, recompute the Regs set.
3954 LU.RecomputeRegs(LUIdx, RegUses);
3956 // Reset this to prepare for the next use.
3957 BestFormulae.clear();
3960 DEBUG(if (ChangedFormulae) {
3962 "After filtering out undesirable candidates:\n";
3967 // This is a rough guess that seems to work fairly well.
3968 static const size_t ComplexityLimit = UINT16_MAX;
3970 /// Estimate the worst-case number of solutions the solver might have to
3971 /// consider. It almost never considers this many solutions because it prune the
3972 /// search space, but the pruning isn't always sufficient.
3973 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3975 for (const LSRUse &LU : Uses) {
3976 size_t FSize = LU.Formulae.size();
3977 if (FSize >= ComplexityLimit) {
3978 Power = ComplexityLimit;
3982 if (Power >= ComplexityLimit)
3988 /// When one formula uses a superset of the registers of another formula, it
3989 /// won't help reduce register pressure (though it may not necessarily hurt
3990 /// register pressure); remove it to simplify the system.
3991 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3992 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3993 DEBUG(dbgs() << "The search space is too complex.\n");
3995 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3996 "which use a superset of registers used by other "
3999 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4000 LSRUse &LU = Uses[LUIdx];
4002 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4003 Formula &F = LU.Formulae[i];
4004 // Look for a formula with a constant or GV in a register. If the use
4005 // also has a formula with that same value in an immediate field,
4006 // delete the one that uses a register.
4007 for (SmallVectorImpl<const SCEV *>::const_iterator
4008 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4009 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4011 NewF.BaseOffset += C->getValue()->getSExtValue();
4012 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4013 (I - F.BaseRegs.begin()));
4014 if (LU.HasFormulaWithSameRegs(NewF)) {
4015 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4016 LU.DeleteFormula(F);
4022 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4023 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4027 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4028 (I - F.BaseRegs.begin()));
4029 if (LU.HasFormulaWithSameRegs(NewF)) {
4030 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4032 LU.DeleteFormula(F);
4043 LU.RecomputeRegs(LUIdx, RegUses);
4046 DEBUG(dbgs() << "After pre-selection:\n";
4047 print_uses(dbgs()));
4051 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4052 /// allocate a single register for them.
4053 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4054 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4057 DEBUG(dbgs() << "The search space is too complex.\n"
4058 "Narrowing the search space by assuming that uses separated "
4059 "by a constant offset will use the same registers.\n");
4061 // This is especially useful for unrolled loops.
4063 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4064 LSRUse &LU = Uses[LUIdx];
4065 for (const Formula &F : LU.Formulae) {
4066 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4069 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4073 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4074 LU.Kind, LU.AccessTy))
4077 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4079 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4081 // Update the relocs to reference the new use.
4082 for (LSRFixup &Fixup : Fixups) {
4083 if (Fixup.LUIdx == LUIdx) {
4084 Fixup.LUIdx = LUThatHas - &Uses.front();
4085 Fixup.Offset += F.BaseOffset;
4086 // Add the new offset to LUThatHas' offset list.
4087 if (LUThatHas->Offsets.back() != Fixup.Offset) {
4088 LUThatHas->Offsets.push_back(Fixup.Offset);
4089 if (Fixup.Offset > LUThatHas->MaxOffset)
4090 LUThatHas->MaxOffset = Fixup.Offset;
4091 if (Fixup.Offset < LUThatHas->MinOffset)
4092 LUThatHas->MinOffset = Fixup.Offset;
4094 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4096 if (Fixup.LUIdx == NumUses-1)
4097 Fixup.LUIdx = LUIdx;
4100 // Delete formulae from the new use which are no longer legal.
4102 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4103 Formula &F = LUThatHas->Formulae[i];
4104 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4105 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4106 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4108 LUThatHas->DeleteFormula(F);
4116 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4118 // Delete the old use.
4119 DeleteUse(LU, LUIdx);
4126 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4129 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4130 /// we've done more filtering, as it may be able to find more formulae to
4132 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4133 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4134 DEBUG(dbgs() << "The search space is too complex.\n");
4136 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4137 "undesirable dedicated registers.\n");
4139 FilterOutUndesirableDedicatedRegisters();
4141 DEBUG(dbgs() << "After pre-selection:\n";
4142 print_uses(dbgs()));
4146 /// Pick a register which seems likely to be profitable, and then in any use
4147 /// which has any reference to that register, delete all formulae which do not
4148 /// reference that register.
4149 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4150 // With all other options exhausted, loop until the system is simple
4151 // enough to handle.
4152 SmallPtrSet<const SCEV *, 4> Taken;
4153 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4154 // Ok, we have too many of formulae on our hands to conveniently handle.
4155 // Use a rough heuristic to thin out the list.
4156 DEBUG(dbgs() << "The search space is too complex.\n");
4158 // Pick the register which is used by the most LSRUses, which is likely
4159 // to be a good reuse register candidate.
4160 const SCEV *Best = nullptr;
4161 unsigned BestNum = 0;
4162 for (const SCEV *Reg : RegUses) {
4163 if (Taken.count(Reg))
4168 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4169 if (Count > BestNum) {
4176 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4177 << " will yield profitable reuse.\n");
4180 // In any use with formulae which references this register, delete formulae
4181 // which don't reference it.
4182 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4183 LSRUse &LU = Uses[LUIdx];
4184 if (!LU.Regs.count(Best)) continue;
4187 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4188 Formula &F = LU.Formulae[i];
4189 if (!F.referencesReg(Best)) {
4190 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4191 LU.DeleteFormula(F);
4195 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4201 LU.RecomputeRegs(LUIdx, RegUses);
4204 DEBUG(dbgs() << "After pre-selection:\n";
4205 print_uses(dbgs()));
4209 /// If there are an extraordinary number of formulae to choose from, use some
4210 /// rough heuristics to prune down the number of formulae. This keeps the main
4211 /// solver from taking an extraordinary amount of time in some worst-case
4213 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4214 NarrowSearchSpaceByDetectingSupersets();
4215 NarrowSearchSpaceByCollapsingUnrolledCode();
4216 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4217 NarrowSearchSpaceByPickingWinnerRegs();
4220 /// This is the recursive solver.
4221 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4223 SmallVectorImpl<const Formula *> &Workspace,
4224 const Cost &CurCost,
4225 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4226 DenseSet<const SCEV *> &VisitedRegs) const {
4229 // - use more aggressive filtering
4230 // - sort the formula so that the most profitable solutions are found first
4231 // - sort the uses too
4233 // - don't compute a cost, and then compare. compare while computing a cost
4235 // - track register sets with SmallBitVector
4237 const LSRUse &LU = Uses[Workspace.size()];
4239 // If this use references any register that's already a part of the
4240 // in-progress solution, consider it a requirement that a formula must
4241 // reference that register in order to be considered. This prunes out
4242 // unprofitable searching.
4243 SmallSetVector<const SCEV *, 4> ReqRegs;
4244 for (const SCEV *S : CurRegs)
4245 if (LU.Regs.count(S))
4248 SmallPtrSet<const SCEV *, 16> NewRegs;
4250 for (const Formula &F : LU.Formulae) {
4251 // Ignore formulae which may not be ideal in terms of register reuse of
4252 // ReqRegs. The formula should use all required registers before
4253 // introducing new ones.
4254 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4255 for (const SCEV *Reg : ReqRegs) {
4256 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4257 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) !=
4260 if (NumReqRegsToFind == 0)
4264 if (NumReqRegsToFind != 0) {
4265 // If none of the formulae satisfied the required registers, then we could
4266 // clear ReqRegs and try again. Currently, we simply give up in this case.
4270 // Evaluate the cost of the current formula. If it's already worse than
4271 // the current best, prune the search at that point.
4274 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4276 if (NewCost < SolutionCost) {
4277 Workspace.push_back(&F);
4278 if (Workspace.size() != Uses.size()) {
4279 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4280 NewRegs, VisitedRegs);
4281 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4282 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4284 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4285 dbgs() << ".\n Regs:";
4286 for (const SCEV *S : NewRegs)
4287 dbgs() << ' ' << *S;
4290 SolutionCost = NewCost;
4291 Solution = Workspace;
4293 Workspace.pop_back();
4298 /// Choose one formula from each use. Return the results in the given Solution
4300 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4301 SmallVector<const Formula *, 8> Workspace;
4303 SolutionCost.Lose();
4305 SmallPtrSet<const SCEV *, 16> CurRegs;
4306 DenseSet<const SCEV *> VisitedRegs;
4307 Workspace.reserve(Uses.size());
4309 // SolveRecurse does all the work.
4310 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4311 CurRegs, VisitedRegs);
4312 if (Solution.empty()) {
4313 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4317 // Ok, we've now made all our decisions.
4318 DEBUG(dbgs() << "\n"
4319 "The chosen solution requires "; SolutionCost.print(dbgs());
4321 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4323 Uses[i].print(dbgs());
4326 Solution[i]->print(dbgs());
4330 assert(Solution.size() == Uses.size() && "Malformed solution!");
4333 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4334 /// we can go while still being dominated by the input positions. This helps
4335 /// canonicalize the insert position, which encourages sharing.
4336 BasicBlock::iterator
4337 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4338 const SmallVectorImpl<Instruction *> &Inputs)
4341 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4342 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4345 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4346 if (!Rung) return IP;
4347 Rung = Rung->getIDom();
4348 if (!Rung) return IP;
4349 IDom = Rung->getBlock();
4351 // Don't climb into a loop though.
4352 const Loop *IDomLoop = LI.getLoopFor(IDom);
4353 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4354 if (IDomDepth <= IPLoopDepth &&
4355 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4359 bool AllDominate = true;
4360 Instruction *BetterPos = nullptr;
4361 Instruction *Tentative = IDom->getTerminator();
4362 for (Instruction *Inst : Inputs) {
4363 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4364 AllDominate = false;
4367 // Attempt to find an insert position in the middle of the block,
4368 // instead of at the end, so that it can be used for other expansions.
4369 if (IDom == Inst->getParent() &&
4370 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4371 BetterPos = &*std::next(BasicBlock::iterator(Inst));
4376 IP = BetterPos->getIterator();
4378 IP = Tentative->getIterator();
4384 /// Determine an input position which will be dominated by the operands and
4385 /// which will dominate the result.
4386 BasicBlock::iterator
4387 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4390 SCEVExpander &Rewriter) const {
4391 // Collect some instructions which must be dominated by the
4392 // expanding replacement. These must be dominated by any operands that
4393 // will be required in the expansion.
4394 SmallVector<Instruction *, 4> Inputs;
4395 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4396 Inputs.push_back(I);
4397 if (LU.Kind == LSRUse::ICmpZero)
4398 if (Instruction *I =
4399 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4400 Inputs.push_back(I);
4401 if (LF.PostIncLoops.count(L)) {
4402 if (LF.isUseFullyOutsideLoop(L))
4403 Inputs.push_back(L->getLoopLatch()->getTerminator());
4405 Inputs.push_back(IVIncInsertPos);
4407 // The expansion must also be dominated by the increment positions of any
4408 // loops it for which it is using post-inc mode.
4409 for (const Loop *PIL : LF.PostIncLoops) {
4410 if (PIL == L) continue;
4412 // Be dominated by the loop exit.
4413 SmallVector<BasicBlock *, 4> ExitingBlocks;
4414 PIL->getExitingBlocks(ExitingBlocks);
4415 if (!ExitingBlocks.empty()) {
4416 BasicBlock *BB = ExitingBlocks[0];
4417 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4418 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4419 Inputs.push_back(BB->getTerminator());
4423 assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
4424 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4425 "Insertion point must be a normal instruction");
4427 // Then, climb up the immediate dominator tree as far as we can go while
4428 // still being dominated by the input positions.
4429 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4431 // Don't insert instructions before PHI nodes.
4432 while (isa<PHINode>(IP)) ++IP;
4434 // Ignore landingpad instructions.
4435 while (IP->isEHPad()) ++IP;
4437 // Ignore debug intrinsics.
4438 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4440 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4441 // IP consistent across expansions and allows the previously inserted
4442 // instructions to be reused by subsequent expansion.
4443 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
4449 /// Emit instructions for the leading candidate expression for this LSRUse (this
4450 /// is called "expanding").
4451 Value *LSRInstance::Expand(const LSRFixup &LF,
4453 BasicBlock::iterator IP,
4454 SCEVExpander &Rewriter,
4455 SmallVectorImpl<WeakVH> &DeadInsts) const {
4456 const LSRUse &LU = Uses[LF.LUIdx];
4457 if (LU.RigidFormula)
4458 return LF.OperandValToReplace;
4460 // Determine an input position which will be dominated by the operands and
4461 // which will dominate the result.
4462 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4464 // Inform the Rewriter if we have a post-increment use, so that it can
4465 // perform an advantageous expansion.
4466 Rewriter.setPostInc(LF.PostIncLoops);
4468 // This is the type that the user actually needs.
4469 Type *OpTy = LF.OperandValToReplace->getType();
4470 // This will be the type that we'll initially expand to.
4471 Type *Ty = F.getType();
4473 // No type known; just expand directly to the ultimate type.
4475 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4476 // Expand directly to the ultimate type if it's the right size.
4478 // This is the type to do integer arithmetic in.
4479 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4481 // Build up a list of operands to add together to form the full base.
4482 SmallVector<const SCEV *, 8> Ops;
4484 // Expand the BaseRegs portion.
4485 for (const SCEV *Reg : F.BaseRegs) {
4486 assert(!Reg->isZero() && "Zero allocated in a base register!");
4488 // If we're expanding for a post-inc user, make the post-inc adjustment.
4489 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4490 Reg = TransformForPostIncUse(Denormalize, Reg,
4491 LF.UserInst, LF.OperandValToReplace,
4494 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr, &*IP)));
4497 // Expand the ScaledReg portion.
4498 Value *ICmpScaledV = nullptr;
4500 const SCEV *ScaledS = F.ScaledReg;
4502 // If we're expanding for a post-inc user, make the post-inc adjustment.
4503 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4504 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4505 LF.UserInst, LF.OperandValToReplace,
4508 if (LU.Kind == LSRUse::ICmpZero) {
4509 // Expand ScaleReg as if it was part of the base regs.
4512 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, &*IP)));
4514 // An interesting way of "folding" with an icmp is to use a negated
4515 // scale, which we'll implement by inserting it into the other operand
4517 assert(F.Scale == -1 &&
4518 "The only scale supported by ICmpZero uses is -1!");
4519 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr, &*IP);
4522 // Otherwise just expand the scaled register and an explicit scale,
4523 // which is expected to be matched as part of the address.
4525 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4526 // Unless the addressing mode will not be folded.
4527 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4528 isAMCompletelyFolded(TTI, LU, F)) {
4529 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, &*IP);
4531 Ops.push_back(SE.getUnknown(FullV));
4533 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, &*IP));
4536 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4537 Ops.push_back(ScaledS);
4541 // Expand the GV portion.
4543 // Flush the operand list to suppress SCEVExpander hoisting.
4545 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, &*IP);
4547 Ops.push_back(SE.getUnknown(FullV));
4549 Ops.push_back(SE.getUnknown(F.BaseGV));
4552 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4553 // unfolded offsets. LSR assumes they both live next to their uses.
4555 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, &*IP);
4557 Ops.push_back(SE.getUnknown(FullV));
4560 // Expand the immediate portion.
4561 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4563 if (LU.Kind == LSRUse::ICmpZero) {
4564 // The other interesting way of "folding" with an ICmpZero is to use a
4565 // negated immediate.
4567 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4569 Ops.push_back(SE.getUnknown(ICmpScaledV));
4570 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4573 // Just add the immediate values. These again are expected to be matched
4574 // as part of the address.
4575 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4579 // Expand the unfolded offset portion.
4580 int64_t UnfoldedOffset = F.UnfoldedOffset;
4581 if (UnfoldedOffset != 0) {
4582 // Just add the immediate values.
4583 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4587 // Emit instructions summing all the operands.
4588 const SCEV *FullS = Ops.empty() ?
4589 SE.getConstant(IntTy, 0) :
4591 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, &*IP);
4593 // We're done expanding now, so reset the rewriter.
4594 Rewriter.clearPostInc();
4596 // An ICmpZero Formula represents an ICmp which we're handling as a
4597 // comparison against zero. Now that we've expanded an expression for that
4598 // form, update the ICmp's other operand.
4599 if (LU.Kind == LSRUse::ICmpZero) {
4600 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4601 DeadInsts.emplace_back(CI->getOperand(1));
4602 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4603 "a scale at the same time!");
4604 if (F.Scale == -1) {
4605 if (ICmpScaledV->getType() != OpTy) {
4607 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4609 ICmpScaledV, OpTy, "tmp", CI);
4612 CI->setOperand(1, ICmpScaledV);
4614 // A scale of 1 means that the scale has been expanded as part of the
4616 assert((F.Scale == 0 || F.Scale == 1) &&
4617 "ICmp does not support folding a global value and "
4618 "a scale at the same time!");
4619 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4621 if (C->getType() != OpTy)
4622 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4626 CI->setOperand(1, C);
4633 /// Helper for Rewrite. PHI nodes are special because the use of their operands
4634 /// effectively happens in their predecessor blocks, so the expression may need
4635 /// to be expanded in multiple places.
4636 void LSRInstance::RewriteForPHI(PHINode *PN,
4639 SCEVExpander &Rewriter,
4640 SmallVectorImpl<WeakVH> &DeadInsts,
4642 DenseMap<BasicBlock *, Value *> Inserted;
4643 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4644 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4645 BasicBlock *BB = PN->getIncomingBlock(i);
4647 // If this is a critical edge, split the edge so that we do not insert
4648 // the code on all predecessor/successor paths. We do this unless this
4649 // is the canonical backedge for this loop, which complicates post-inc
4651 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4652 !isa<IndirectBrInst>(BB->getTerminator())) {
4653 BasicBlock *Parent = PN->getParent();
4654 Loop *PNLoop = LI.getLoopFor(Parent);
4655 if (!PNLoop || Parent != PNLoop->getHeader()) {
4656 // Split the critical edge.
4657 BasicBlock *NewBB = nullptr;
4658 if (!Parent->isLandingPad()) {
4659 NewBB = SplitCriticalEdge(BB, Parent,
4660 CriticalEdgeSplittingOptions(&DT, &LI)
4661 .setMergeIdenticalEdges()
4662 .setDontDeleteUselessPHIs());
4664 SmallVector<BasicBlock*, 2> NewBBs;
4665 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
4668 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4669 // phi predecessors are identical. The simple thing to do is skip
4670 // splitting in this case rather than complicate the API.
4672 // If PN is outside of the loop and BB is in the loop, we want to
4673 // move the block to be immediately before the PHI block, not
4674 // immediately after BB.
4675 if (L->contains(BB) && !L->contains(PN))
4676 NewBB->moveBefore(PN->getParent());
4678 // Splitting the edge can reduce the number of PHI entries we have.
4679 e = PN->getNumIncomingValues();
4681 i = PN->getBasicBlockIndex(BB);
4686 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4687 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
4689 PN->setIncomingValue(i, Pair.first->second);
4691 Value *FullV = Expand(LF, F, BB->getTerminator()->getIterator(),
4692 Rewriter, DeadInsts);
4694 // If this is reuse-by-noop-cast, insert the noop cast.
4695 Type *OpTy = LF.OperandValToReplace->getType();
4696 if (FullV->getType() != OpTy)
4698 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4700 FullV, LF.OperandValToReplace->getType(),
4701 "tmp", BB->getTerminator());
4703 PN->setIncomingValue(i, FullV);
4704 Pair.first->second = FullV;
4709 /// Emit instructions for the leading candidate expression for this LSRUse (this
4710 /// is called "expanding"), and update the UserInst to reference the newly
4712 void LSRInstance::Rewrite(const LSRFixup &LF,
4714 SCEVExpander &Rewriter,
4715 SmallVectorImpl<WeakVH> &DeadInsts,
4717 // First, find an insertion point that dominates UserInst. For PHI nodes,
4718 // find the nearest block which dominates all the relevant uses.
4719 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4720 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4723 Expand(LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
4725 // If this is reuse-by-noop-cast, insert the noop cast.
4726 Type *OpTy = LF.OperandValToReplace->getType();
4727 if (FullV->getType() != OpTy) {
4729 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4730 FullV, OpTy, "tmp", LF.UserInst);
4734 // Update the user. ICmpZero is handled specially here (for now) because
4735 // Expand may have updated one of the operands of the icmp already, and
4736 // its new value may happen to be equal to LF.OperandValToReplace, in
4737 // which case doing replaceUsesOfWith leads to replacing both operands
4738 // with the same value. TODO: Reorganize this.
4739 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4740 LF.UserInst->setOperand(0, FullV);
4742 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4745 DeadInsts.emplace_back(LF.OperandValToReplace);
4748 /// Rewrite all the fixup locations with new values, following the chosen
4751 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4753 // Keep track of instructions we may have made dead, so that
4754 // we can remove them after we are done working.
4755 SmallVector<WeakVH, 16> DeadInsts;
4757 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
4760 Rewriter.setDebugType(DEBUG_TYPE);
4762 Rewriter.disableCanonicalMode();
4763 Rewriter.enableLSRMode();
4764 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4766 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4767 for (const IVChain &Chain : IVChainVec) {
4768 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
4769 Rewriter.setChainedPhi(PN);
4772 // Expand the new value definitions and update the users.
4773 for (const LSRFixup &Fixup : Fixups) {
4774 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4779 for (const IVChain &Chain : IVChainVec) {
4780 GenerateIVChain(Chain, Rewriter, DeadInsts);
4783 // Clean up after ourselves. This must be done before deleting any
4787 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4790 LSRInstance::LSRInstance(Loop *L, Pass *P)
4791 : IU(P->getAnalysis<IVUsers>()),
4792 SE(P->getAnalysis<ScalarEvolutionWrapperPass>().getSE()),
4793 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4794 LI(P->getAnalysis<LoopInfoWrapperPass>().getLoopInfo()),
4795 TTI(P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
4796 *L->getHeader()->getParent())),
4797 L(L), Changed(false), IVIncInsertPos(nullptr) {
4798 // If LoopSimplify form is not available, stay out of trouble.
4799 if (!L->isLoopSimplifyForm())
4802 // If there's no interesting work to be done, bail early.
4803 if (IU.empty()) return;
4805 // If there's too much analysis to be done, bail early. We won't be able to
4806 // model the problem anyway.
4807 unsigned NumUsers = 0;
4808 for (const IVStrideUse &U : IU) {
4809 if (++NumUsers > MaxIVUsers) {
4811 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n");
4817 // All dominating loops must have preheaders, or SCEVExpander may not be able
4818 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4820 // IVUsers analysis should only create users that are dominated by simple loop
4821 // headers. Since this loop should dominate all of its users, its user list
4822 // should be empty if this loop itself is not within a simple loop nest.
4823 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4824 Rung; Rung = Rung->getIDom()) {
4825 BasicBlock *BB = Rung->getBlock();
4826 const Loop *DomLoop = LI.getLoopFor(BB);
4827 if (DomLoop && DomLoop->getHeader() == BB) {
4828 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4833 DEBUG(dbgs() << "\nLSR on loop ";
4834 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4837 // First, perform some low-level loop optimizations.
4839 OptimizeLoopTermCond();
4841 // If loop preparation eliminates all interesting IV users, bail.
4842 if (IU.empty()) return;
4844 // Skip nested loops until we can model them better with formulae.
4846 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4850 // Start collecting data and preparing for the solver.
4852 CollectInterestingTypesAndFactors();
4853 CollectFixupsAndInitialFormulae();
4854 CollectLoopInvariantFixupsAndFormulae();
4856 assert(!Uses.empty() && "IVUsers reported at least one use");
4857 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4858 print_uses(dbgs()));
4860 // Now use the reuse data to generate a bunch of interesting ways
4861 // to formulate the values needed for the uses.
4862 GenerateAllReuseFormulae();
4864 FilterOutUndesirableDedicatedRegisters();
4865 NarrowSearchSpaceUsingHeuristics();
4867 SmallVector<const Formula *, 8> Solution;
4870 // Release memory that is no longer needed.
4875 if (Solution.empty())
4879 // Formulae should be legal.
4880 for (const LSRUse &LU : Uses) {
4881 for (const Formula &F : LU.Formulae)
4882 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4883 F) && "Illegal formula generated!");
4887 // Now that we've decided what we want, make it so.
4888 ImplementSolution(Solution, P);
4891 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4892 if (Factors.empty() && Types.empty()) return;
4894 OS << "LSR has identified the following interesting factors and types: ";
4897 for (int64_t Factor : Factors) {
4898 if (!First) OS << ", ";
4900 OS << '*' << Factor;
4903 for (Type *Ty : Types) {
4904 if (!First) OS << ", ";
4906 OS << '(' << *Ty << ')';
4911 void LSRInstance::print_fixups(raw_ostream &OS) const {
4912 OS << "LSR is examining the following fixup sites:\n";
4913 for (const LSRFixup &LF : Fixups) {
4920 void LSRInstance::print_uses(raw_ostream &OS) const {
4921 OS << "LSR is examining the following uses:\n";
4922 for (const LSRUse &LU : Uses) {
4926 for (const Formula &F : LU.Formulae) {
4934 void LSRInstance::print(raw_ostream &OS) const {
4935 print_factors_and_types(OS);
4940 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4941 void LSRInstance::dump() const {
4942 print(errs()); errs() << '\n';
4948 class LoopStrengthReduce : public LoopPass {
4950 static char ID; // Pass ID, replacement for typeid
4951 LoopStrengthReduce();
4954 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
4955 void getAnalysisUsage(AnalysisUsage &AU) const override;
4960 char LoopStrengthReduce::ID = 0;
4961 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4962 "Loop Strength Reduction", false, false)
4963 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
4964 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4965 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
4966 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4967 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
4968 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4969 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4970 "Loop Strength Reduction", false, false)
4973 Pass *llvm::createLoopStrengthReducePass() {
4974 return new LoopStrengthReduce();
4977 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4978 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4981 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4982 // We split critical edges, so we change the CFG. However, we do update
4983 // many analyses if they are around.
4984 AU.addPreservedID(LoopSimplifyID);
4986 AU.addRequired<LoopInfoWrapperPass>();
4987 AU.addPreserved<LoopInfoWrapperPass>();
4988 AU.addRequiredID(LoopSimplifyID);
4989 AU.addRequired<DominatorTreeWrapperPass>();
4990 AU.addPreserved<DominatorTreeWrapperPass>();
4991 AU.addRequired<ScalarEvolutionWrapperPass>();
4992 AU.addPreserved<ScalarEvolutionWrapperPass>();
4993 // Requiring LoopSimplify a second time here prevents IVUsers from running
4994 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4995 AU.addRequiredID(LoopSimplifyID);
4996 AU.addRequired<IVUsers>();
4997 AU.addPreserved<IVUsers>();
4998 AU.addRequired<TargetTransformInfoWrapperPass>();
5001 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5002 if (skipOptnoneFunction(L))
5005 bool Changed = false;
5007 // Run the main LSR transformation.
5008 Changed |= LSRInstance(L, this).getChanged();
5010 // Remove any extra phis created by processing inner loops.
5011 Changed |= DeleteDeadPHIs(L->getHeader());
5012 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5013 SmallVector<WeakVH, 16> DeadInsts;
5014 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5015 SCEVExpander Rewriter(getAnalysis<ScalarEvolutionWrapperPass>().getSE(), DL,
5018 Rewriter.setDebugType(DEBUG_TYPE);
5020 unsigned numFolded = Rewriter.replaceCongruentIVs(
5021 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
5022 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5023 *L->getHeader()->getParent()));
5026 DeleteTriviallyDeadInstructions(DeadInsts);
5027 DeleteDeadPHIs(L->getHeader());