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/LoopStrengthReduce.h"
57 #include "llvm/ADT/APInt.h"
58 #include "llvm/ADT/DenseMap.h"
59 #include "llvm/ADT/DenseSet.h"
60 #include "llvm/ADT/Hashing.h"
61 #include "llvm/ADT/PointerIntPair.h"
62 #include "llvm/ADT/STLExtras.h"
63 #include "llvm/ADT/SetVector.h"
64 #include "llvm/ADT/SmallBitVector.h"
65 #include "llvm/ADT/SmallPtrSet.h"
66 #include "llvm/ADT/SmallSet.h"
67 #include "llvm/ADT/SmallVector.h"
68 #include "llvm/ADT/iterator_range.h"
69 #include "llvm/Analysis/IVUsers.h"
70 #include "llvm/Analysis/LoopAnalysisManager.h"
71 #include "llvm/Analysis/LoopInfo.h"
72 #include "llvm/Analysis/LoopPass.h"
73 #include "llvm/Analysis/ScalarEvolution.h"
74 #include "llvm/Analysis/ScalarEvolutionExpander.h"
75 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
76 #include "llvm/Analysis/ScalarEvolutionNormalization.h"
77 #include "llvm/Analysis/TargetTransformInfo.h"
78 #include "llvm/IR/BasicBlock.h"
79 #include "llvm/IR/Constant.h"
80 #include "llvm/IR/Constants.h"
81 #include "llvm/IR/DerivedTypes.h"
82 #include "llvm/IR/Dominators.h"
83 #include "llvm/IR/GlobalValue.h"
84 #include "llvm/IR/IRBuilder.h"
85 #include "llvm/IR/InstrTypes.h"
86 #include "llvm/IR/Instruction.h"
87 #include "llvm/IR/Instructions.h"
88 #include "llvm/IR/IntrinsicInst.h"
89 #include "llvm/IR/Intrinsics.h"
90 #include "llvm/IR/Module.h"
91 #include "llvm/IR/OperandTraits.h"
92 #include "llvm/IR/Operator.h"
93 #include "llvm/IR/PassManager.h"
94 #include "llvm/IR/Type.h"
95 #include "llvm/IR/Use.h"
96 #include "llvm/IR/User.h"
97 #include "llvm/IR/Value.h"
98 #include "llvm/IR/ValueHandle.h"
99 #include "llvm/Pass.h"
100 #include "llvm/Support/Casting.h"
101 #include "llvm/Support/CommandLine.h"
102 #include "llvm/Support/Compiler.h"
103 #include "llvm/Support/Debug.h"
104 #include "llvm/Support/ErrorHandling.h"
105 #include "llvm/Support/MathExtras.h"
106 #include "llvm/Support/raw_ostream.h"
107 #include "llvm/Transforms/Scalar.h"
108 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
109 #include "llvm/Transforms/Utils/Local.h"
120 using namespace llvm;
122 #define DEBUG_TYPE "loop-reduce"
124 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
125 /// bail out. This threshold is far beyond the number of users that LSR can
126 /// conceivably solve, so it should not affect generated code, but catches the
127 /// worst cases before LSR burns too much compile time and stack space.
128 static const unsigned MaxIVUsers = 200;
130 // Temporary flag to cleanup congruent phis after LSR phi expansion.
131 // It's currently disabled until we can determine whether it's truly useful or
132 // not. The flag should be removed after the v3.0 release.
133 // This is now needed for ivchains.
134 static cl::opt<bool> EnablePhiElim(
135 "enable-lsr-phielim", cl::Hidden, cl::init(true),
136 cl::desc("Enable LSR phi elimination"));
138 // The flag adds instruction count to solutions cost comparision.
139 static cl::opt<bool> InsnsCost(
140 "lsr-insns-cost", cl::Hidden, cl::init(true),
141 cl::desc("Add instruction count to a LSR cost model"));
143 // Flag to choose how to narrow complex lsr solution
144 static cl::opt<bool> LSRExpNarrow(
145 "lsr-exp-narrow", cl::Hidden, cl::init(false),
146 cl::desc("Narrow LSR complex solution using"
147 " expectation of registers number"));
149 // Flag to narrow search space by filtering non-optimal formulae with
150 // the same ScaledReg and Scale.
151 static cl::opt<bool> FilterSameScaledReg(
152 "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
153 cl::desc("Narrow LSR search space by filtering non-optimal formulae"
154 " with the same ScaledReg and Scale"));
157 // Stress test IV chain generation.
158 static cl::opt<bool> StressIVChain(
159 "stress-ivchain", cl::Hidden, cl::init(false),
160 cl::desc("Stress test LSR IV chains"));
162 static bool StressIVChain = false;
168 /// Used in situations where the accessed memory type is unknown.
169 static const unsigned UnknownAddressSpace =
170 std::numeric_limits<unsigned>::max();
172 Type *MemTy = nullptr;
173 unsigned AddrSpace = UnknownAddressSpace;
175 MemAccessTy() = default;
176 MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
178 bool operator==(MemAccessTy Other) const {
179 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
182 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
184 static MemAccessTy getUnknown(LLVMContext &Ctx,
185 unsigned AS = UnknownAddressSpace) {
186 return MemAccessTy(Type::getVoidTy(Ctx), AS);
190 /// This class holds data which is used to order reuse candidates.
193 /// This represents the set of LSRUse indices which reference
194 /// a particular register.
195 SmallBitVector UsedByIndices;
197 void print(raw_ostream &OS) const;
201 } // end anonymous namespace
203 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
204 void RegSortData::print(raw_ostream &OS) const {
205 OS << "[NumUses=" << UsedByIndices.count() << ']';
208 LLVM_DUMP_METHOD void RegSortData::dump() const {
209 print(errs()); errs() << '\n';
215 /// Map register candidates to information about how they are used.
216 class RegUseTracker {
217 using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
219 RegUsesTy RegUsesMap;
220 SmallVector<const SCEV *, 16> RegSequence;
223 void countRegister(const SCEV *Reg, size_t LUIdx);
224 void dropRegister(const SCEV *Reg, size_t LUIdx);
225 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
227 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
229 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
233 using iterator = SmallVectorImpl<const SCEV *>::iterator;
234 using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
236 iterator begin() { return RegSequence.begin(); }
237 iterator end() { return RegSequence.end(); }
238 const_iterator begin() const { return RegSequence.begin(); }
239 const_iterator end() const { return RegSequence.end(); }
242 } // end anonymous namespace
245 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
246 std::pair<RegUsesTy::iterator, bool> Pair =
247 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
248 RegSortData &RSD = Pair.first->second;
250 RegSequence.push_back(Reg);
251 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
252 RSD.UsedByIndices.set(LUIdx);
256 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
257 RegUsesTy::iterator It = RegUsesMap.find(Reg);
258 assert(It != RegUsesMap.end());
259 RegSortData &RSD = It->second;
260 assert(RSD.UsedByIndices.size() > LUIdx);
261 RSD.UsedByIndices.reset(LUIdx);
265 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
266 assert(LUIdx <= LastLUIdx);
268 // Update RegUses. The data structure is not optimized for this purpose;
269 // we must iterate through it and update each of the bit vectors.
270 for (auto &Pair : RegUsesMap) {
271 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
272 if (LUIdx < UsedByIndices.size())
273 UsedByIndices[LUIdx] =
274 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
275 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
280 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
281 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
282 if (I == RegUsesMap.end())
284 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
285 int i = UsedByIndices.find_first();
286 if (i == -1) return false;
287 if ((size_t)i != LUIdx) return true;
288 return UsedByIndices.find_next(i) != -1;
291 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
292 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
293 assert(I != RegUsesMap.end() && "Unknown register!");
294 return I->second.UsedByIndices;
297 void RegUseTracker::clear() {
304 /// This class holds information that describes a formula for computing
305 /// satisfying a use. It may include broken-out immediates and scaled registers.
307 /// Global base address used for complex addressing.
308 GlobalValue *BaseGV = nullptr;
310 /// Base offset for complex addressing.
311 int64_t BaseOffset = 0;
313 /// Whether any complex addressing has a base register.
314 bool HasBaseReg = false;
316 /// The scale of any complex addressing.
319 /// The list of "base" registers for this use. When this is non-empty. The
320 /// canonical representation of a formula is
321 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
322 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
323 /// 3. The reg containing recurrent expr related with currect loop in the
324 /// formula should be put in the ScaledReg.
325 /// #1 enforces that the scaled register is always used when at least two
326 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
327 /// #2 enforces that 1 * reg is reg.
328 /// #3 ensures invariant regs with respect to current loop can be combined
329 /// together in LSR codegen.
330 /// This invariant can be temporarly broken while building a formula.
331 /// However, every formula inserted into the LSRInstance must be in canonical
333 SmallVector<const SCEV *, 4> BaseRegs;
335 /// The 'scaled' register for this use. This should be non-null when Scale is
337 const SCEV *ScaledReg = nullptr;
339 /// An additional constant offset which added near the use. This requires a
340 /// temporary register, but the offset itself can live in an add immediate
341 /// field rather than a register.
342 int64_t UnfoldedOffset = 0;
346 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
348 bool isCanonical(const Loop &L) const;
350 void canonicalize(const Loop &L);
354 bool hasZeroEnd() const;
356 size_t getNumRegs() const;
357 Type *getType() const;
359 void deleteBaseReg(const SCEV *&S);
361 bool referencesReg(const SCEV *S) const;
362 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
363 const RegUseTracker &RegUses) const;
365 void print(raw_ostream &OS) const;
369 } // end anonymous namespace
371 /// Recursion helper for initialMatch.
372 static void DoInitialMatch(const SCEV *S, Loop *L,
373 SmallVectorImpl<const SCEV *> &Good,
374 SmallVectorImpl<const SCEV *> &Bad,
375 ScalarEvolution &SE) {
376 // Collect expressions which properly dominate the loop header.
377 if (SE.properlyDominates(S, L->getHeader())) {
382 // Look at add operands.
383 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
384 for (const SCEV *S : Add->operands())
385 DoInitialMatch(S, L, Good, Bad, SE);
389 // Look at addrec operands.
390 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
391 if (!AR->getStart()->isZero() && AR->isAffine()) {
392 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
393 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
394 AR->getStepRecurrence(SE),
395 // FIXME: AR->getNoWrapFlags()
396 AR->getLoop(), SCEV::FlagAnyWrap),
401 // Handle a multiplication by -1 (negation) if it didn't fold.
402 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
403 if (Mul->getOperand(0)->isAllOnesValue()) {
404 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
405 const SCEV *NewMul = SE.getMulExpr(Ops);
407 SmallVector<const SCEV *, 4> MyGood;
408 SmallVector<const SCEV *, 4> MyBad;
409 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
410 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
411 SE.getEffectiveSCEVType(NewMul->getType())));
412 for (const SCEV *S : MyGood)
413 Good.push_back(SE.getMulExpr(NegOne, S));
414 for (const SCEV *S : MyBad)
415 Bad.push_back(SE.getMulExpr(NegOne, S));
419 // Ok, we can't do anything interesting. Just stuff the whole thing into a
420 // register and hope for the best.
424 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
425 /// all loop-invariant and loop-computable values in a single base register.
426 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
427 SmallVector<const SCEV *, 4> Good;
428 SmallVector<const SCEV *, 4> Bad;
429 DoInitialMatch(S, L, Good, Bad, SE);
431 const SCEV *Sum = SE.getAddExpr(Good);
433 BaseRegs.push_back(Sum);
437 const SCEV *Sum = SE.getAddExpr(Bad);
439 BaseRegs.push_back(Sum);
445 /// \brief Check whether or not this formula statisfies the canonical
447 /// \see Formula::BaseRegs.
448 bool Formula::isCanonical(const Loop &L) const {
450 return BaseRegs.size() <= 1;
455 if (Scale == 1 && BaseRegs.empty())
458 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
459 if (SAR && SAR->getLoop() == &L)
462 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
463 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
464 // loop, we want to swap the reg in BaseRegs with ScaledReg.
466 find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
467 return isa<const SCEVAddRecExpr>(S) &&
468 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
470 return I == BaseRegs.end();
473 /// \brief Helper method to morph a formula into its canonical representation.
474 /// \see Formula::BaseRegs.
475 /// Every formula having more than one base register, must use the ScaledReg
476 /// field. Otherwise, we would have to do special cases everywhere in LSR
477 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
478 /// On the other hand, 1*reg should be canonicalized into reg.
479 void Formula::canonicalize(const Loop &L) {
482 // So far we did not need this case. This is easy to implement but it is
483 // useless to maintain dead code. Beside it could hurt compile time.
484 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
486 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
488 ScaledReg = BaseRegs.back();
493 // If ScaledReg is an invariant with respect to L, find the reg from
494 // BaseRegs containing the recurrent expr related with Loop L. Swap the
495 // reg with ScaledReg.
496 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
497 if (!SAR || SAR->getLoop() != &L) {
498 auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
500 return isa<const SCEVAddRecExpr>(S) &&
501 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
503 if (I != BaseRegs.end())
504 std::swap(ScaledReg, *I);
508 /// \brief Get rid of the scale in the formula.
509 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
510 /// \return true if it was possible to get rid of the scale, false otherwise.
511 /// \note After this operation the formula may not be in the canonical form.
512 bool Formula::unscale() {
516 BaseRegs.push_back(ScaledReg);
521 bool Formula::hasZeroEnd() const {
522 if (UnfoldedOffset || BaseOffset)
524 if (BaseRegs.size() != 1 || ScaledReg)
529 /// Return the total number of register operands used by this formula. This does
530 /// not include register uses implied by non-constant addrec strides.
531 size_t Formula::getNumRegs() const {
532 return !!ScaledReg + BaseRegs.size();
535 /// Return the type of this formula, if it has one, or null otherwise. This type
536 /// is meaningless except for the bit size.
537 Type *Formula::getType() const {
538 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
539 ScaledReg ? ScaledReg->getType() :
540 BaseGV ? BaseGV->getType() :
544 /// Delete the given base reg from the BaseRegs list.
545 void Formula::deleteBaseReg(const SCEV *&S) {
546 if (&S != &BaseRegs.back())
547 std::swap(S, BaseRegs.back());
551 /// Test if this formula references the given register.
552 bool Formula::referencesReg(const SCEV *S) const {
553 return S == ScaledReg || is_contained(BaseRegs, S);
556 /// Test whether this formula uses registers which are used by uses other than
557 /// the use with the given index.
558 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
559 const RegUseTracker &RegUses) const {
561 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
563 for (const SCEV *BaseReg : BaseRegs)
564 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
569 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
570 void Formula::print(raw_ostream &OS) const {
573 if (!First) OS << " + "; else First = false;
574 BaseGV->printAsOperand(OS, /*PrintType=*/false);
576 if (BaseOffset != 0) {
577 if (!First) OS << " + "; else First = false;
580 for (const SCEV *BaseReg : BaseRegs) {
581 if (!First) OS << " + "; else First = false;
582 OS << "reg(" << *BaseReg << ')';
584 if (HasBaseReg && BaseRegs.empty()) {
585 if (!First) OS << " + "; else First = false;
586 OS << "**error: HasBaseReg**";
587 } else if (!HasBaseReg && !BaseRegs.empty()) {
588 if (!First) OS << " + "; else First = false;
589 OS << "**error: !HasBaseReg**";
592 if (!First) OS << " + "; else First = false;
593 OS << Scale << "*reg(";
600 if (UnfoldedOffset != 0) {
601 if (!First) OS << " + ";
602 OS << "imm(" << UnfoldedOffset << ')';
606 LLVM_DUMP_METHOD void Formula::dump() const {
607 print(errs()); errs() << '\n';
611 /// Return true if the given addrec can be sign-extended without changing its
613 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
615 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
616 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
619 /// Return true if the given add can be sign-extended without changing its
621 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
623 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
624 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
627 /// Return true if the given mul can be sign-extended without changing its
629 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
631 IntegerType::get(SE.getContext(),
632 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
633 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
636 /// Return an expression for LHS /s RHS, if it can be determined and if the
637 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
638 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
639 /// the multiplication may overflow, which is useful when the result will be
640 /// used in a context where the most significant bits are ignored.
641 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
643 bool IgnoreSignificantBits = false) {
644 // Handle the trivial case, which works for any SCEV type.
646 return SE.getConstant(LHS->getType(), 1);
648 // Handle a few RHS special cases.
649 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
651 const APInt &RA = RC->getAPInt();
652 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
654 if (RA.isAllOnesValue())
655 return SE.getMulExpr(LHS, RC);
656 // Handle x /s 1 as x.
661 // Check for a division of a constant by a constant.
662 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
665 const APInt &LA = C->getAPInt();
666 const APInt &RA = RC->getAPInt();
667 if (LA.srem(RA) != 0)
669 return SE.getConstant(LA.sdiv(RA));
672 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
673 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
674 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
675 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
676 IgnoreSignificantBits);
677 if (!Step) return nullptr;
678 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
679 IgnoreSignificantBits);
680 if (!Start) return nullptr;
681 // FlagNW is independent of the start value, step direction, and is
682 // preserved with smaller magnitude steps.
683 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
684 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
689 // Distribute the sdiv over add operands, if the add doesn't overflow.
690 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
691 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
692 SmallVector<const SCEV *, 8> Ops;
693 for (const SCEV *S : Add->operands()) {
694 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
695 if (!Op) return nullptr;
698 return SE.getAddExpr(Ops);
703 // Check for a multiply operand that we can pull RHS out of.
704 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
705 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
706 SmallVector<const SCEV *, 4> Ops;
708 for (const SCEV *S : Mul->operands()) {
710 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
711 IgnoreSignificantBits)) {
717 return Found ? SE.getMulExpr(Ops) : nullptr;
722 // Otherwise we don't know.
726 /// If S involves the addition of a constant integer value, return that integer
727 /// value, and mutate S to point to a new SCEV with that value excluded.
728 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
729 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
730 if (C->getAPInt().getMinSignedBits() <= 64) {
731 S = SE.getConstant(C->getType(), 0);
732 return C->getValue()->getSExtValue();
734 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
735 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
736 int64_t Result = ExtractImmediate(NewOps.front(), SE);
738 S = SE.getAddExpr(NewOps);
740 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
741 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
742 int64_t Result = ExtractImmediate(NewOps.front(), SE);
744 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
745 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
752 /// If S involves the addition of a GlobalValue address, return that symbol, and
753 /// mutate S to point to a new SCEV with that value excluded.
754 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
755 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
756 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
757 S = SE.getConstant(GV->getType(), 0);
760 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
761 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
762 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
764 S = SE.getAddExpr(NewOps);
766 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
767 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
768 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
770 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
771 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
778 /// Returns true if the specified instruction is using the specified value as an
780 static bool isAddressUse(const TargetTransformInfo &TTI,
781 Instruction *Inst, Value *OperandVal) {
782 bool isAddress = isa<LoadInst>(Inst);
783 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
784 if (SI->getPointerOperand() == OperandVal)
786 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
787 // Addressing modes can also be folded into prefetches and a variety
789 switch (II->getIntrinsicID()) {
790 case Intrinsic::memset:
791 case Intrinsic::prefetch:
792 if (II->getArgOperand(0) == OperandVal)
795 case Intrinsic::memmove:
796 case Intrinsic::memcpy:
797 if (II->getArgOperand(0) == OperandVal ||
798 II->getArgOperand(1) == OperandVal)
802 MemIntrinsicInfo IntrInfo;
803 if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
804 if (IntrInfo.PtrVal == OperandVal)
809 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
810 if (RMW->getPointerOperand() == OperandVal)
812 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
813 if (CmpX->getPointerOperand() == OperandVal)
819 /// Return the type of the memory being accessed.
820 static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
822 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
823 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
824 AccessTy.MemTy = SI->getOperand(0)->getType();
825 AccessTy.AddrSpace = SI->getPointerAddressSpace();
826 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
827 AccessTy.AddrSpace = LI->getPointerAddressSpace();
828 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
829 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
830 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
831 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
832 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
833 switch (II->getIntrinsicID()) {
834 case Intrinsic::prefetch:
835 AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
838 MemIntrinsicInfo IntrInfo;
839 if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
841 = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
849 // All pointers have the same requirements, so canonicalize them to an
850 // arbitrary pointer type to minimize variation.
851 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
852 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
853 PTy->getAddressSpace());
858 /// Return true if this AddRec is already a phi in its loop.
859 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
860 for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
861 if (SE.isSCEVable(PN.getType()) &&
862 (SE.getEffectiveSCEVType(PN.getType()) ==
863 SE.getEffectiveSCEVType(AR->getType())) &&
864 SE.getSCEV(&PN) == AR)
870 /// Check if expanding this expression is likely to incur significant cost. This
871 /// is tricky because SCEV doesn't track which expressions are actually computed
872 /// by the current IR.
874 /// We currently allow expansion of IV increments that involve adds,
875 /// multiplication by constants, and AddRecs from existing phis.
877 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
878 /// obvious multiple of the UDivExpr.
879 static bool isHighCostExpansion(const SCEV *S,
880 SmallPtrSetImpl<const SCEV*> &Processed,
881 ScalarEvolution &SE) {
882 // Zero/One operand expressions
883 switch (S->getSCEVType()) {
888 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
891 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
894 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
898 if (!Processed.insert(S).second)
901 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
902 for (const SCEV *S : Add->operands()) {
903 if (isHighCostExpansion(S, Processed, SE))
909 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
910 if (Mul->getNumOperands() == 2) {
911 // Multiplication by a constant is ok
912 if (isa<SCEVConstant>(Mul->getOperand(0)))
913 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
915 // If we have the value of one operand, check if an existing
916 // multiplication already generates this expression.
917 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
918 Value *UVal = U->getValue();
919 for (User *UR : UVal->users()) {
920 // If U is a constant, it may be used by a ConstantExpr.
921 Instruction *UI = dyn_cast<Instruction>(UR);
922 if (UI && UI->getOpcode() == Instruction::Mul &&
923 SE.isSCEVable(UI->getType())) {
924 return SE.getSCEV(UI) == Mul;
931 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
932 if (isExistingPhi(AR, SE))
936 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
940 /// If any of the instructions is the specified set are trivially dead, delete
941 /// them and see if this makes any of their operands subsequently dead.
943 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
944 bool Changed = false;
946 while (!DeadInsts.empty()) {
947 Value *V = DeadInsts.pop_back_val();
948 Instruction *I = dyn_cast_or_null<Instruction>(V);
950 if (!I || !isInstructionTriviallyDead(I))
953 for (Use &O : I->operands())
954 if (Instruction *U = dyn_cast<Instruction>(O)) {
957 DeadInsts.emplace_back(U);
960 I->eraseFromParent();
971 } // end anonymous namespace
973 /// \brief Check if the addressing mode defined by \p F is completely
974 /// folded in \p LU at isel time.
975 /// This includes address-mode folding and special icmp tricks.
976 /// This function returns true if \p LU can accommodate what \p F
977 /// defines and up to 1 base + 1 scaled + offset.
978 /// In other words, if \p F has several base registers, this function may
979 /// still return true. Therefore, users still need to account for
980 /// additional base registers and/or unfolded offsets to derive an
981 /// accurate cost model.
982 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
983 const LSRUse &LU, const Formula &F);
985 // Get the cost of the scaling factor used in F for LU.
986 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
987 const LSRUse &LU, const Formula &F,
992 /// This class is used to measure and compare candidate formulae.
994 TargetTransformInfo::LSRCost C;
1008 bool isLess(Cost &Other, const TargetTransformInfo &TTI);
1013 // Once any of the metrics loses, they must all remain losers.
1015 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1016 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1017 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1018 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1023 assert(isValid() && "invalid cost");
1024 return C.NumRegs == ~0u;
1027 void RateFormula(const TargetTransformInfo &TTI,
1029 SmallPtrSetImpl<const SCEV *> &Regs,
1030 const DenseSet<const SCEV *> &VisitedRegs,
1032 ScalarEvolution &SE, DominatorTree &DT,
1034 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1036 void print(raw_ostream &OS) const;
1040 void RateRegister(const SCEV *Reg,
1041 SmallPtrSetImpl<const SCEV *> &Regs,
1043 ScalarEvolution &SE, DominatorTree &DT);
1044 void RatePrimaryRegister(const SCEV *Reg,
1045 SmallPtrSetImpl<const SCEV *> &Regs,
1047 ScalarEvolution &SE, DominatorTree &DT,
1048 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1051 /// An operand value in an instruction which is to be replaced with some
1052 /// equivalent, possibly strength-reduced, replacement.
1054 /// The instruction which will be updated.
1055 Instruction *UserInst = nullptr;
1057 /// The operand of the instruction which will be replaced. The operand may be
1058 /// used more than once; every instance will be replaced.
1059 Value *OperandValToReplace = nullptr;
1061 /// If this user is to use the post-incremented value of an induction
1062 /// variable, this set is non-empty and holds the loops associated with the
1063 /// induction variable.
1064 PostIncLoopSet PostIncLoops;
1066 /// A constant offset to be added to the LSRUse expression. This allows
1067 /// multiple fixups to share the same LSRUse with different offsets, for
1068 /// example in an unrolled loop.
1071 LSRFixup() = default;
1073 bool isUseFullyOutsideLoop(const Loop *L) const;
1075 void print(raw_ostream &OS) const;
1079 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1080 /// SmallVectors of const SCEV*.
1081 struct UniquifierDenseMapInfo {
1082 static SmallVector<const SCEV *, 4> getEmptyKey() {
1083 SmallVector<const SCEV *, 4> V;
1084 V.push_back(reinterpret_cast<const SCEV *>(-1));
1088 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1089 SmallVector<const SCEV *, 4> V;
1090 V.push_back(reinterpret_cast<const SCEV *>(-2));
1094 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1095 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1098 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1099 const SmallVector<const SCEV *, 4> &RHS) {
1104 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1105 /// as uses invented by LSR itself. It includes information about what kinds of
1106 /// things can be folded into the user, information about the user itself, and
1107 /// information about how the use may be satisfied. TODO: Represent multiple
1108 /// users of the same expression in common?
1110 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1113 /// An enum for a kind of use, indicating what types of scaled and immediate
1114 /// operands it might support.
1116 Basic, ///< A normal use, with no folding.
1117 Special, ///< A special case of basic, allowing -1 scales.
1118 Address, ///< An address use; folding according to TargetLowering
1119 ICmpZero ///< An equality icmp with both operands folded into one.
1120 // TODO: Add a generic icmp too?
1123 using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1126 MemAccessTy AccessTy;
1128 /// The list of operands which are to be replaced.
1129 SmallVector<LSRFixup, 8> Fixups;
1131 /// Keep track of the min and max offsets of the fixups.
1132 int64_t MinOffset = std::numeric_limits<int64_t>::max();
1133 int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1135 /// This records whether all of the fixups using this LSRUse are outside of
1136 /// the loop, in which case some special-case heuristics may be used.
1137 bool AllFixupsOutsideLoop = true;
1139 /// RigidFormula is set to true to guarantee that this use will be associated
1140 /// with a single formula--the one that initially matched. Some SCEV
1141 /// expressions cannot be expanded. This allows LSR to consider the registers
1142 /// used by those expressions without the need to expand them later after
1143 /// changing the formula.
1144 bool RigidFormula = false;
1146 /// This records the widest use type for any fixup using this
1147 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1148 /// fixup widths to be equivalent, because the narrower one may be relying on
1149 /// the implicit truncation to truncate away bogus bits.
1150 Type *WidestFixupType = nullptr;
1152 /// A list of ways to build a value that can satisfy this user. After the
1153 /// list is populated, one of these is selected heuristically and used to
1154 /// formulate a replacement for OperandValToReplace in UserInst.
1155 SmallVector<Formula, 12> Formulae;
1157 /// The set of register candidates used by all formulae in this LSRUse.
1158 SmallPtrSet<const SCEV *, 4> Regs;
1160 LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1162 LSRFixup &getNewFixup() {
1163 Fixups.push_back(LSRFixup());
1164 return Fixups.back();
1167 void pushFixup(LSRFixup &f) {
1168 Fixups.push_back(f);
1169 if (f.Offset > MaxOffset)
1170 MaxOffset = f.Offset;
1171 if (f.Offset < MinOffset)
1172 MinOffset = f.Offset;
1175 bool HasFormulaWithSameRegs(const Formula &F) const;
1176 float getNotSelectedProbability(const SCEV *Reg) const;
1177 bool InsertFormula(const Formula &F, const Loop &L);
1178 void DeleteFormula(Formula &F);
1179 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1181 void print(raw_ostream &OS) const;
1185 } // end anonymous namespace
1187 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1188 LSRUse::KindType Kind, MemAccessTy AccessTy,
1189 GlobalValue *BaseGV, int64_t BaseOffset,
1190 bool HasBaseReg, int64_t Scale,
1191 Instruction *Fixup = nullptr);
1193 /// Tally up interesting quantities from the given register.
1194 void Cost::RateRegister(const SCEV *Reg,
1195 SmallPtrSetImpl<const SCEV *> &Regs,
1197 ScalarEvolution &SE, DominatorTree &DT) {
1198 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1199 // If this is an addrec for another loop, it should be an invariant
1200 // with respect to L since L is the innermost loop (at least
1201 // for now LSR only handles innermost loops).
1202 if (AR->getLoop() != L) {
1203 // If the AddRec exists, consider it's register free and leave it alone.
1204 if (isExistingPhi(AR, SE))
1207 // It is bad to allow LSR for current loop to add induction variables
1208 // for its sibling loops.
1209 if (!AR->getLoop()->contains(L)) {
1214 // Otherwise, it will be an invariant with respect to Loop L.
1218 C.AddRecCost += 1; /// TODO: This should be a function of the stride.
1220 // Add the step value register, if it needs one.
1221 // TODO: The non-affine case isn't precisely modeled here.
1222 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1223 if (!Regs.count(AR->getOperand(1))) {
1224 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
1232 // Rough heuristic; favor registers which don't require extra setup
1233 // instructions in the preheader.
1234 if (!isa<SCEVUnknown>(Reg) &&
1235 !isa<SCEVConstant>(Reg) &&
1236 !(isa<SCEVAddRecExpr>(Reg) &&
1237 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
1238 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
1241 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1242 SE.hasComputableLoopEvolution(Reg, L);
1245 /// Record this register in the set. If we haven't seen it before, rate
1246 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1247 /// one of those regs an instant loser.
1248 void Cost::RatePrimaryRegister(const SCEV *Reg,
1249 SmallPtrSetImpl<const SCEV *> &Regs,
1251 ScalarEvolution &SE, DominatorTree &DT,
1252 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1253 if (LoserRegs && LoserRegs->count(Reg)) {
1257 if (Regs.insert(Reg).second) {
1258 RateRegister(Reg, Regs, L, SE, DT);
1259 if (LoserRegs && isLoser())
1260 LoserRegs->insert(Reg);
1264 void Cost::RateFormula(const TargetTransformInfo &TTI,
1266 SmallPtrSetImpl<const SCEV *> &Regs,
1267 const DenseSet<const SCEV *> &VisitedRegs,
1269 ScalarEvolution &SE, DominatorTree &DT,
1271 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1272 assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1273 // Tally up the registers.
1274 unsigned PrevAddRecCost = C.AddRecCost;
1275 unsigned PrevNumRegs = C.NumRegs;
1276 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1277 if (const SCEV *ScaledReg = F.ScaledReg) {
1278 if (VisitedRegs.count(ScaledReg)) {
1282 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1286 for (const SCEV *BaseReg : F.BaseRegs) {
1287 if (VisitedRegs.count(BaseReg)) {
1291 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1296 // Determine how many (unfolded) adds we'll need inside the loop.
1297 size_t NumBaseParts = F.getNumRegs();
1298 if (NumBaseParts > 1)
1299 // Do not count the base and a possible second register if the target
1300 // allows to fold 2 registers.
1302 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1303 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1305 // Accumulate non-free scaling amounts.
1306 C.ScaleCost += getScalingFactorCost(TTI, LU, F, *L);
1308 // Tally up the non-zero immediates.
1309 for (const LSRFixup &Fixup : LU.Fixups) {
1310 int64_t O = Fixup.Offset;
1311 int64_t Offset = (uint64_t)O + F.BaseOffset;
1313 C.ImmCost += 64; // Handle symbolic values conservatively.
1314 // TODO: This should probably be the pointer size.
1315 else if (Offset != 0)
1316 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1318 // Check with target if this offset with this instruction is
1319 // specifically not supported.
1320 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1321 !isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1322 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1326 // If we don't count instruction cost exit here.
1328 assert(isValid() && "invalid cost");
1332 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1333 // additional instruction (at least fill).
1334 unsigned TTIRegNum = TTI.getNumberOfRegisters(false) - 1;
1335 if (C.NumRegs > TTIRegNum) {
1336 // Cost already exceeded TTIRegNum, then only newly added register can add
1337 // new instructions.
1338 if (PrevNumRegs > TTIRegNum)
1339 C.Insns += (C.NumRegs - PrevNumRegs);
1341 C.Insns += (C.NumRegs - TTIRegNum);
1344 // If ICmpZero formula ends with not 0, it could not be replaced by
1345 // just add or sub. We'll need to compare final result of AddRec.
1346 // That means we'll need an additional instruction.
1347 // For -10 + {0, +, 1}:
1353 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd())
1355 // Each new AddRec adds 1 instruction to calculation.
1356 C.Insns += (C.AddRecCost - PrevAddRecCost);
1358 // BaseAdds adds instructions for unfolded registers.
1359 if (LU.Kind != LSRUse::ICmpZero)
1360 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1361 assert(isValid() && "invalid cost");
1364 /// Set this cost to a losing value.
1366 C.Insns = std::numeric_limits<unsigned>::max();
1367 C.NumRegs = std::numeric_limits<unsigned>::max();
1368 C.AddRecCost = std::numeric_limits<unsigned>::max();
1369 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1370 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1371 C.ImmCost = std::numeric_limits<unsigned>::max();
1372 C.SetupCost = std::numeric_limits<unsigned>::max();
1373 C.ScaleCost = std::numeric_limits<unsigned>::max();
1376 /// Choose the lower cost.
1377 bool Cost::isLess(Cost &Other, const TargetTransformInfo &TTI) {
1378 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1379 C.Insns != Other.C.Insns)
1380 return C.Insns < Other.C.Insns;
1381 return TTI.isLSRCostLess(C, Other.C);
1384 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1385 void Cost::print(raw_ostream &OS) const {
1387 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1388 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1389 if (C.AddRecCost != 0)
1390 OS << ", with addrec cost " << C.AddRecCost;
1391 if (C.NumIVMuls != 0)
1392 OS << ", plus " << C.NumIVMuls << " IV mul"
1393 << (C.NumIVMuls == 1 ? "" : "s");
1394 if (C.NumBaseAdds != 0)
1395 OS << ", plus " << C.NumBaseAdds << " base add"
1396 << (C.NumBaseAdds == 1 ? "" : "s");
1397 if (C.ScaleCost != 0)
1398 OS << ", plus " << C.ScaleCost << " scale cost";
1400 OS << ", plus " << C.ImmCost << " imm cost";
1401 if (C.SetupCost != 0)
1402 OS << ", plus " << C.SetupCost << " setup cost";
1405 LLVM_DUMP_METHOD void Cost::dump() const {
1406 print(errs()); errs() << '\n';
1410 /// Test whether this fixup always uses its value outside of the given loop.
1411 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1412 // PHI nodes use their value in their incoming blocks.
1413 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1414 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1415 if (PN->getIncomingValue(i) == OperandValToReplace &&
1416 L->contains(PN->getIncomingBlock(i)))
1421 return !L->contains(UserInst);
1424 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1425 void LSRFixup::print(raw_ostream &OS) const {
1427 // Store is common and interesting enough to be worth special-casing.
1428 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1430 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1431 } else if (UserInst->getType()->isVoidTy())
1432 OS << UserInst->getOpcodeName();
1434 UserInst->printAsOperand(OS, /*PrintType=*/false);
1436 OS << ", OperandValToReplace=";
1437 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1439 for (const Loop *PIL : PostIncLoops) {
1440 OS << ", PostIncLoop=";
1441 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1445 OS << ", Offset=" << Offset;
1448 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1449 print(errs()); errs() << '\n';
1453 /// Test whether this use as a formula which has the same registers as the given
1455 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1456 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1457 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1458 // Unstable sort by host order ok, because this is only used for uniquifying.
1459 std::sort(Key.begin(), Key.end());
1460 return Uniquifier.count(Key);
1463 /// The function returns a probability of selecting formula without Reg.
1464 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1466 for (const Formula &F : Formulae)
1467 if (F.referencesReg(Reg))
1469 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1472 /// If the given formula has not yet been inserted, add it to the list, and
1473 /// return true. Return false otherwise. The formula must be in canonical form.
1474 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1475 assert(F.isCanonical(L) && "Invalid canonical representation");
1477 if (!Formulae.empty() && RigidFormula)
1480 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1481 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1482 // Unstable sort by host order ok, because this is only used for uniquifying.
1483 std::sort(Key.begin(), Key.end());
1485 if (!Uniquifier.insert(Key).second)
1488 // Using a register to hold the value of 0 is not profitable.
1489 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1490 "Zero allocated in a scaled register!");
1492 for (const SCEV *BaseReg : F.BaseRegs)
1493 assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1496 // Add the formula to the list.
1497 Formulae.push_back(F);
1499 // Record registers now being used by this use.
1500 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1502 Regs.insert(F.ScaledReg);
1507 /// Remove the given formula from this use's list.
1508 void LSRUse::DeleteFormula(Formula &F) {
1509 if (&F != &Formulae.back())
1510 std::swap(F, Formulae.back());
1511 Formulae.pop_back();
1514 /// Recompute the Regs field, and update RegUses.
1515 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1516 // Now that we've filtered out some formulae, recompute the Regs set.
1517 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1519 for (const Formula &F : Formulae) {
1520 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1521 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1524 // Update the RegTracker.
1525 for (const SCEV *S : OldRegs)
1527 RegUses.dropRegister(S, LUIdx);
1530 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1531 void LSRUse::print(raw_ostream &OS) const {
1532 OS << "LSR Use: Kind=";
1534 case Basic: OS << "Basic"; break;
1535 case Special: OS << "Special"; break;
1536 case ICmpZero: OS << "ICmpZero"; break;
1538 OS << "Address of ";
1539 if (AccessTy.MemTy->isPointerTy())
1540 OS << "pointer"; // the full pointer type could be really verbose
1542 OS << *AccessTy.MemTy;
1545 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1548 OS << ", Offsets={";
1549 bool NeedComma = false;
1550 for (const LSRFixup &Fixup : Fixups) {
1551 if (NeedComma) OS << ',';
1557 if (AllFixupsOutsideLoop)
1558 OS << ", all-fixups-outside-loop";
1560 if (WidestFixupType)
1561 OS << ", widest fixup type: " << *WidestFixupType;
1564 LLVM_DUMP_METHOD void LSRUse::dump() const {
1565 print(errs()); errs() << '\n';
1569 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1570 LSRUse::KindType Kind, MemAccessTy AccessTy,
1571 GlobalValue *BaseGV, int64_t BaseOffset,
1572 bool HasBaseReg, int64_t Scale,
1573 Instruction *Fixup/*= nullptr*/) {
1575 case LSRUse::Address:
1576 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1577 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1579 case LSRUse::ICmpZero:
1580 // There's not even a target hook for querying whether it would be legal to
1581 // fold a GV into an ICmp.
1585 // ICmp only has two operands; don't allow more than two non-trivial parts.
1586 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1589 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1590 // putting the scaled register in the other operand of the icmp.
1591 if (Scale != 0 && Scale != -1)
1594 // If we have low-level target information, ask the target if it can fold an
1595 // integer immediate on an icmp.
1596 if (BaseOffset != 0) {
1598 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1599 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1600 // Offs is the ICmp immediate.
1602 // The cast does the right thing with
1603 // std::numeric_limits<int64_t>::min().
1604 BaseOffset = -(uint64_t)BaseOffset;
1605 return TTI.isLegalICmpImmediate(BaseOffset);
1608 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1612 // Only handle single-register values.
1613 return !BaseGV && Scale == 0 && BaseOffset == 0;
1615 case LSRUse::Special:
1616 // Special case Basic to handle -1 scales.
1617 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1620 llvm_unreachable("Invalid LSRUse Kind!");
1623 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1624 int64_t MinOffset, int64_t MaxOffset,
1625 LSRUse::KindType Kind, MemAccessTy AccessTy,
1626 GlobalValue *BaseGV, int64_t BaseOffset,
1627 bool HasBaseReg, int64_t Scale) {
1628 // Check for overflow.
1629 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1632 MinOffset = (uint64_t)BaseOffset + MinOffset;
1633 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1636 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1638 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1639 HasBaseReg, Scale) &&
1640 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1644 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1645 int64_t MinOffset, int64_t MaxOffset,
1646 LSRUse::KindType Kind, MemAccessTy AccessTy,
1647 const Formula &F, const Loop &L) {
1648 // For the purpose of isAMCompletelyFolded either having a canonical formula
1649 // or a scale not equal to zero is correct.
1650 // Problems may arise from non canonical formulae having a scale == 0.
1651 // Strictly speaking it would best to just rely on canonical formulae.
1652 // However, when we generate the scaled formulae, we first check that the
1653 // scaling factor is profitable before computing the actual ScaledReg for
1654 // compile time sake.
1655 assert((F.isCanonical(L) || F.Scale != 0));
1656 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1657 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1660 /// Test whether we know how to expand the current formula.
1661 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1662 int64_t MaxOffset, LSRUse::KindType Kind,
1663 MemAccessTy AccessTy, GlobalValue *BaseGV,
1664 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1665 // We know how to expand completely foldable formulae.
1666 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1667 BaseOffset, HasBaseReg, Scale) ||
1668 // Or formulae that use a base register produced by a sum of base
1671 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1672 BaseGV, BaseOffset, true, 0));
1675 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1676 int64_t MaxOffset, LSRUse::KindType Kind,
1677 MemAccessTy AccessTy, const Formula &F) {
1678 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1679 F.BaseOffset, F.HasBaseReg, F.Scale);
1682 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1683 const LSRUse &LU, const Formula &F) {
1684 // Target may want to look at the user instructions.
1685 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1686 for (const LSRFixup &Fixup : LU.Fixups)
1687 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1688 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1689 F.Scale, Fixup.UserInst))
1694 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1695 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1699 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1700 const LSRUse &LU, const Formula &F,
1705 // If the use is not completely folded in that instruction, we will have to
1706 // pay an extra cost only for scale != 1.
1707 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1709 return F.Scale != 1;
1712 case LSRUse::Address: {
1713 // Check the scaling factor cost with both the min and max offsets.
1714 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1715 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1716 F.Scale, LU.AccessTy.AddrSpace);
1717 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1718 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1719 F.Scale, LU.AccessTy.AddrSpace);
1721 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1722 "Legal addressing mode has an illegal cost!");
1723 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1725 case LSRUse::ICmpZero:
1727 case LSRUse::Special:
1728 // The use is completely folded, i.e., everything is folded into the
1733 llvm_unreachable("Invalid LSRUse Kind!");
1736 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1737 LSRUse::KindType Kind, MemAccessTy AccessTy,
1738 GlobalValue *BaseGV, int64_t BaseOffset,
1740 // Fast-path: zero is always foldable.
1741 if (BaseOffset == 0 && !BaseGV) return true;
1743 // Conservatively, create an address with an immediate and a
1744 // base and a scale.
1745 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1747 // Canonicalize a scale of 1 to a base register if the formula doesn't
1748 // already have a base register.
1749 if (!HasBaseReg && Scale == 1) {
1754 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1758 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1759 ScalarEvolution &SE, int64_t MinOffset,
1760 int64_t MaxOffset, LSRUse::KindType Kind,
1761 MemAccessTy AccessTy, const SCEV *S,
1763 // Fast-path: zero is always foldable.
1764 if (S->isZero()) return true;
1766 // Conservatively, create an address with an immediate and a
1767 // base and a scale.
1768 int64_t BaseOffset = ExtractImmediate(S, SE);
1769 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1771 // If there's anything else involved, it's not foldable.
1772 if (!S->isZero()) return false;
1774 // Fast-path: zero is always foldable.
1775 if (BaseOffset == 0 && !BaseGV) return true;
1777 // Conservatively, create an address with an immediate and a
1778 // base and a scale.
1779 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1781 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1782 BaseOffset, HasBaseReg, Scale);
1787 /// An individual increment in a Chain of IV increments. Relate an IV user to
1788 /// an expression that computes the IV it uses from the IV used by the previous
1789 /// link in the Chain.
1791 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1792 /// original IVOperand. The head of the chain's IVOperand is only valid during
1793 /// chain collection, before LSR replaces IV users. During chain generation,
1794 /// IncExpr can be used to find the new IVOperand that computes the same
1797 Instruction *UserInst;
1799 const SCEV *IncExpr;
1801 IVInc(Instruction *U, Value *O, const SCEV *E)
1802 : UserInst(U), IVOperand(O), IncExpr(E) {}
1805 // The list of IV increments in program order. We typically add the head of a
1806 // chain without finding subsequent links.
1808 SmallVector<IVInc, 1> Incs;
1809 const SCEV *ExprBase = nullptr;
1811 IVChain() = default;
1812 IVChain(const IVInc &Head, const SCEV *Base)
1813 : Incs(1, Head), ExprBase(Base) {}
1815 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1817 // Return the first increment in the chain.
1818 const_iterator begin() const {
1819 assert(!Incs.empty());
1820 return std::next(Incs.begin());
1822 const_iterator end() const {
1826 // Returns true if this chain contains any increments.
1827 bool hasIncs() const { return Incs.size() >= 2; }
1829 // Add an IVInc to the end of this chain.
1830 void add(const IVInc &X) { Incs.push_back(X); }
1832 // Returns the last UserInst in the chain.
1833 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1835 // Returns true if IncExpr can be profitably added to this chain.
1836 bool isProfitableIncrement(const SCEV *OperExpr,
1837 const SCEV *IncExpr,
1841 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1842 /// between FarUsers that definitely cross IV increments and NearUsers that may
1843 /// be used between IV increments.
1845 SmallPtrSet<Instruction*, 4> FarUsers;
1846 SmallPtrSet<Instruction*, 4> NearUsers;
1849 /// This class holds state for the main loop strength reduction logic.
1852 ScalarEvolution &SE;
1855 const TargetTransformInfo &TTI;
1857 bool Changed = false;
1859 /// This is the insert position that the current loop's induction variable
1860 /// increment should be placed. In simple loops, this is the latch block's
1861 /// terminator. But in more complicated cases, this is a position which will
1862 /// dominate all the in-loop post-increment users.
1863 Instruction *IVIncInsertPos = nullptr;
1865 /// Interesting factors between use strides.
1867 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1868 /// default, a SmallDenseSet, because we need to use the full range of
1869 /// int64_ts, and there's currently no good way of doing that with
1871 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1873 /// Interesting use types, to facilitate truncation reuse.
1874 SmallSetVector<Type *, 4> Types;
1876 /// The list of interesting uses.
1877 SmallVector<LSRUse, 16> Uses;
1879 /// Track which uses use which register candidates.
1880 RegUseTracker RegUses;
1882 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1883 // have more than a few IV increment chains in a loop. Missing a Chain falls
1884 // back to normal LSR behavior for those uses.
1885 static const unsigned MaxChains = 8;
1887 /// IV users can form a chain of IV increments.
1888 SmallVector<IVChain, MaxChains> IVChainVec;
1890 /// IV users that belong to profitable IVChains.
1891 SmallPtrSet<Use*, MaxChains> IVIncSet;
1893 void OptimizeShadowIV();
1894 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1895 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1896 void OptimizeLoopTermCond();
1898 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1899 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1900 void FinalizeChain(IVChain &Chain);
1901 void CollectChains();
1902 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1903 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1905 void CollectInterestingTypesAndFactors();
1906 void CollectFixupsAndInitialFormulae();
1908 // Support for sharing of LSRUses between LSRFixups.
1909 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
1912 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1913 LSRUse::KindType Kind, MemAccessTy AccessTy);
1915 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1916 MemAccessTy AccessTy);
1918 void DeleteUse(LSRUse &LU, size_t LUIdx);
1920 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1922 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1923 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1924 void CountRegisters(const Formula &F, size_t LUIdx);
1925 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1927 void CollectLoopInvariantFixupsAndFormulae();
1929 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1930 unsigned Depth = 0);
1932 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1933 const Formula &Base, unsigned Depth,
1934 size_t Idx, bool IsScaledReg = false);
1935 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1936 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1937 const Formula &Base, size_t Idx,
1938 bool IsScaledReg = false);
1939 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1940 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1941 const Formula &Base,
1942 const SmallVectorImpl<int64_t> &Worklist,
1943 size_t Idx, bool IsScaledReg = false);
1944 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1945 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1946 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1947 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1948 void GenerateCrossUseConstantOffsets();
1949 void GenerateAllReuseFormulae();
1951 void FilterOutUndesirableDedicatedRegisters();
1953 size_t EstimateSearchSpaceComplexity() const;
1954 void NarrowSearchSpaceByDetectingSupersets();
1955 void NarrowSearchSpaceByCollapsingUnrolledCode();
1956 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1957 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
1958 void NarrowSearchSpaceByDeletingCostlyFormulas();
1959 void NarrowSearchSpaceByPickingWinnerRegs();
1960 void NarrowSearchSpaceUsingHeuristics();
1962 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1964 SmallVectorImpl<const Formula *> &Workspace,
1965 const Cost &CurCost,
1966 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1967 DenseSet<const SCEV *> &VisitedRegs) const;
1968 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1970 BasicBlock::iterator
1971 HoistInsertPosition(BasicBlock::iterator IP,
1972 const SmallVectorImpl<Instruction *> &Inputs) const;
1973 BasicBlock::iterator
1974 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1977 SCEVExpander &Rewriter) const;
1979 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1980 BasicBlock::iterator IP, SCEVExpander &Rewriter,
1981 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1982 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
1983 const Formula &F, SCEVExpander &Rewriter,
1984 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1985 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1986 SCEVExpander &Rewriter,
1987 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1988 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
1991 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
1992 LoopInfo &LI, const TargetTransformInfo &TTI);
1994 bool getChanged() const { return Changed; }
1996 void print_factors_and_types(raw_ostream &OS) const;
1997 void print_fixups(raw_ostream &OS) const;
1998 void print_uses(raw_ostream &OS) const;
1999 void print(raw_ostream &OS) const;
2003 } // end anonymous namespace
2005 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
2006 /// the cast operation.
2007 void LSRInstance::OptimizeShadowIV() {
2008 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2009 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2012 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2013 UI != E; /* empty */) {
2014 IVUsers::const_iterator CandidateUI = UI;
2016 Instruction *ShadowUse = CandidateUI->getUser();
2017 Type *DestTy = nullptr;
2018 bool IsSigned = false;
2020 /* If shadow use is a int->float cast then insert a second IV
2021 to eliminate this cast.
2023 for (unsigned i = 0; i < n; ++i)
2029 for (unsigned i = 0; i < n; ++i, ++d)
2032 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2034 DestTy = UCast->getDestTy();
2036 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2038 DestTy = SCast->getDestTy();
2040 if (!DestTy) continue;
2042 // If target does not support DestTy natively then do not apply
2043 // this transformation.
2044 if (!TTI.isTypeLegal(DestTy)) continue;
2046 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2048 if (PH->getNumIncomingValues() != 2) continue;
2050 // If the calculation in integers overflows, the result in FP type will
2051 // differ. So we only can do this transformation if we are guaranteed to not
2052 // deal with overflowing values
2053 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2055 if (IsSigned && !AR->hasNoSignedWrap()) continue;
2056 if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2058 Type *SrcTy = PH->getType();
2059 int Mantissa = DestTy->getFPMantissaWidth();
2060 if (Mantissa == -1) continue;
2061 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2064 unsigned Entry, Latch;
2065 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2073 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2074 if (!Init) continue;
2075 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2076 (double)Init->getSExtValue() :
2077 (double)Init->getZExtValue());
2079 BinaryOperator *Incr =
2080 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2081 if (!Incr) continue;
2082 if (Incr->getOpcode() != Instruction::Add
2083 && Incr->getOpcode() != Instruction::Sub)
2086 /* Initialize new IV, double d = 0.0 in above example. */
2087 ConstantInt *C = nullptr;
2088 if (Incr->getOperand(0) == PH)
2089 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2090 else if (Incr->getOperand(1) == PH)
2091 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2097 // Ignore negative constants, as the code below doesn't handle them
2098 // correctly. TODO: Remove this restriction.
2099 if (!C->getValue().isStrictlyPositive()) continue;
2101 /* Add new PHINode. */
2102 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2104 /* create new increment. '++d' in above example. */
2105 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2106 BinaryOperator *NewIncr =
2107 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2108 Instruction::FAdd : Instruction::FSub,
2109 NewPH, CFP, "IV.S.next.", Incr);
2111 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2112 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2114 /* Remove cast operation */
2115 ShadowUse->replaceAllUsesWith(NewPH);
2116 ShadowUse->eraseFromParent();
2122 /// If Cond has an operand that is an expression of an IV, set the IV user and
2123 /// stride information and return true, otherwise return false.
2124 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2125 for (IVStrideUse &U : IU)
2126 if (U.getUser() == Cond) {
2127 // NOTE: we could handle setcc instructions with multiple uses here, but
2128 // InstCombine does it as well for simple uses, it's not clear that it
2129 // occurs enough in real life to handle.
2136 /// Rewrite the loop's terminating condition if it uses a max computation.
2138 /// This is a narrow solution to a specific, but acute, problem. For loops
2144 /// } while (++i < n);
2146 /// the trip count isn't just 'n', because 'n' might not be positive. And
2147 /// unfortunately this can come up even for loops where the user didn't use
2148 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2149 /// will commonly be lowered like this:
2155 /// } while (++i < n);
2158 /// and then it's possible for subsequent optimization to obscure the if
2159 /// test in such a way that indvars can't find it.
2161 /// When indvars can't find the if test in loops like this, it creates a
2162 /// max expression, which allows it to give the loop a canonical
2163 /// induction variable:
2166 /// max = n < 1 ? 1 : n;
2169 /// } while (++i != max);
2171 /// Canonical induction variables are necessary because the loop passes
2172 /// are designed around them. The most obvious example of this is the
2173 /// LoopInfo analysis, which doesn't remember trip count values. It
2174 /// expects to be able to rediscover the trip count each time it is
2175 /// needed, and it does this using a simple analysis that only succeeds if
2176 /// the loop has a canonical induction variable.
2178 /// However, when it comes time to generate code, the maximum operation
2179 /// can be quite costly, especially if it's inside of an outer loop.
2181 /// This function solves this problem by detecting this type of loop and
2182 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2183 /// the instructions for the maximum computation.
2184 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2185 // Check that the loop matches the pattern we're looking for.
2186 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2187 Cond->getPredicate() != CmpInst::ICMP_NE)
2190 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2191 if (!Sel || !Sel->hasOneUse()) return Cond;
2193 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2194 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2196 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2198 // Add one to the backedge-taken count to get the trip count.
2199 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2200 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2202 // Check for a max calculation that matches the pattern. There's no check
2203 // for ICMP_ULE here because the comparison would be with zero, which
2204 // isn't interesting.
2205 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2206 const SCEVNAryExpr *Max = nullptr;
2207 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2208 Pred = ICmpInst::ICMP_SLE;
2210 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2211 Pred = ICmpInst::ICMP_SLT;
2213 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2214 Pred = ICmpInst::ICMP_ULT;
2221 // To handle a max with more than two operands, this optimization would
2222 // require additional checking and setup.
2223 if (Max->getNumOperands() != 2)
2226 const SCEV *MaxLHS = Max->getOperand(0);
2227 const SCEV *MaxRHS = Max->getOperand(1);
2229 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2230 // for a comparison with 1. For <= and >=, a comparison with zero.
2232 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2235 // Check the relevant induction variable for conformance to
2237 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2238 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2239 if (!AR || !AR->isAffine() ||
2240 AR->getStart() != One ||
2241 AR->getStepRecurrence(SE) != One)
2244 assert(AR->getLoop() == L &&
2245 "Loop condition operand is an addrec in a different loop!");
2247 // Check the right operand of the select, and remember it, as it will
2248 // be used in the new comparison instruction.
2249 Value *NewRHS = nullptr;
2250 if (ICmpInst::isTrueWhenEqual(Pred)) {
2251 // Look for n+1, and grab n.
2252 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2253 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2254 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2255 NewRHS = BO->getOperand(0);
2256 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2257 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2258 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2259 NewRHS = BO->getOperand(0);
2262 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2263 NewRHS = Sel->getOperand(1);
2264 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2265 NewRHS = Sel->getOperand(2);
2266 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2267 NewRHS = SU->getValue();
2269 // Max doesn't match expected pattern.
2272 // Determine the new comparison opcode. It may be signed or unsigned,
2273 // and the original comparison may be either equality or inequality.
2274 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2275 Pred = CmpInst::getInversePredicate(Pred);
2277 // Ok, everything looks ok to change the condition into an SLT or SGE and
2278 // delete the max calculation.
2280 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2282 // Delete the max calculation instructions.
2283 Cond->replaceAllUsesWith(NewCond);
2284 CondUse->setUser(NewCond);
2285 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2286 Cond->eraseFromParent();
2287 Sel->eraseFromParent();
2288 if (Cmp->use_empty())
2289 Cmp->eraseFromParent();
2293 /// Change loop terminating condition to use the postinc iv when possible.
2295 LSRInstance::OptimizeLoopTermCond() {
2296 SmallPtrSet<Instruction *, 4> PostIncs;
2298 // We need a different set of heuristics for rotated and non-rotated loops.
2299 // If a loop is rotated then the latch is also the backedge, so inserting
2300 // post-inc expressions just before the latch is ideal. To reduce live ranges
2301 // it also makes sense to rewrite terminating conditions to use post-inc
2304 // If the loop is not rotated then the latch is not a backedge; the latch
2305 // check is done in the loop head. Adding post-inc expressions before the
2306 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2307 // in the loop body. In this case we do *not* want to use post-inc expressions
2308 // in the latch check, and we want to insert post-inc expressions before
2310 BasicBlock *LatchBlock = L->getLoopLatch();
2311 SmallVector<BasicBlock*, 8> ExitingBlocks;
2312 L->getExitingBlocks(ExitingBlocks);
2313 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2314 return LatchBlock != BB;
2316 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2317 IVIncInsertPos = LatchBlock->getTerminator();
2321 // Otherwise treat this as a rotated loop.
2322 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2323 // Get the terminating condition for the loop if possible. If we
2324 // can, we want to change it to use a post-incremented version of its
2325 // induction variable, to allow coalescing the live ranges for the IV into
2326 // one register value.
2328 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2331 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2332 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2335 // Search IVUsesByStride to find Cond's IVUse if there is one.
2336 IVStrideUse *CondUse = nullptr;
2337 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2338 if (!FindIVUserForCond(Cond, CondUse))
2341 // If the trip count is computed in terms of a max (due to ScalarEvolution
2342 // being unable to find a sufficient guard, for example), change the loop
2343 // comparison to use SLT or ULT instead of NE.
2344 // One consequence of doing this now is that it disrupts the count-down
2345 // optimization. That's not always a bad thing though, because in such
2346 // cases it may still be worthwhile to avoid a max.
2347 Cond = OptimizeMax(Cond, CondUse);
2349 // If this exiting block dominates the latch block, it may also use
2350 // the post-inc value if it won't be shared with other uses.
2351 // Check for dominance.
2352 if (!DT.dominates(ExitingBlock, LatchBlock))
2355 // Conservatively avoid trying to use the post-inc value in non-latch
2356 // exits if there may be pre-inc users in intervening blocks.
2357 if (LatchBlock != ExitingBlock)
2358 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2359 // Test if the use is reachable from the exiting block. This dominator
2360 // query is a conservative approximation of reachability.
2361 if (&*UI != CondUse &&
2362 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2363 // Conservatively assume there may be reuse if the quotient of their
2364 // strides could be a legal scale.
2365 const SCEV *A = IU.getStride(*CondUse, L);
2366 const SCEV *B = IU.getStride(*UI, L);
2367 if (!A || !B) continue;
2368 if (SE.getTypeSizeInBits(A->getType()) !=
2369 SE.getTypeSizeInBits(B->getType())) {
2370 if (SE.getTypeSizeInBits(A->getType()) >
2371 SE.getTypeSizeInBits(B->getType()))
2372 B = SE.getSignExtendExpr(B, A->getType());
2374 A = SE.getSignExtendExpr(A, B->getType());
2376 if (const SCEVConstant *D =
2377 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2378 const ConstantInt *C = D->getValue();
2379 // Stride of one or negative one can have reuse with non-addresses.
2380 if (C->isOne() || C->isMinusOne())
2381 goto decline_post_inc;
2382 // Avoid weird situations.
2383 if (C->getValue().getMinSignedBits() >= 64 ||
2384 C->getValue().isMinSignedValue())
2385 goto decline_post_inc;
2386 // Check for possible scaled-address reuse.
2387 MemAccessTy AccessTy = getAccessType(TTI, UI->getUser());
2388 int64_t Scale = C->getSExtValue();
2389 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2391 /*HasBaseReg=*/false, Scale,
2392 AccessTy.AddrSpace))
2393 goto decline_post_inc;
2395 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2397 /*HasBaseReg=*/false, Scale,
2398 AccessTy.AddrSpace))
2399 goto decline_post_inc;
2403 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2406 // It's possible for the setcc instruction to be anywhere in the loop, and
2407 // possible for it to have multiple users. If it is not immediately before
2408 // the exiting block branch, move it.
2409 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2410 if (Cond->hasOneUse()) {
2411 Cond->moveBefore(TermBr);
2413 // Clone the terminating condition and insert into the loopend.
2414 ICmpInst *OldCond = Cond;
2415 Cond = cast<ICmpInst>(Cond->clone());
2416 Cond->setName(L->getHeader()->getName() + ".termcond");
2417 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2419 // Clone the IVUse, as the old use still exists!
2420 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2421 TermBr->replaceUsesOfWith(OldCond, Cond);
2425 // If we get to here, we know that we can transform the setcc instruction to
2426 // use the post-incremented version of the IV, allowing us to coalesce the
2427 // live ranges for the IV correctly.
2428 CondUse->transformToPostInc(L);
2431 PostIncs.insert(Cond);
2435 // Determine an insertion point for the loop induction variable increment. It
2436 // must dominate all the post-inc comparisons we just set up, and it must
2437 // dominate the loop latch edge.
2438 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2439 for (Instruction *Inst : PostIncs) {
2441 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2443 if (BB == Inst->getParent())
2444 IVIncInsertPos = Inst;
2445 else if (BB != IVIncInsertPos->getParent())
2446 IVIncInsertPos = BB->getTerminator();
2450 /// Determine if the given use can accommodate a fixup at the given offset and
2451 /// other details. If so, update the use and return true.
2452 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2453 bool HasBaseReg, LSRUse::KindType Kind,
2454 MemAccessTy AccessTy) {
2455 int64_t NewMinOffset = LU.MinOffset;
2456 int64_t NewMaxOffset = LU.MaxOffset;
2457 MemAccessTy NewAccessTy = AccessTy;
2459 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2460 // something conservative, however this can pessimize in the case that one of
2461 // the uses will have all its uses outside the loop, for example.
2462 if (LU.Kind != Kind)
2465 // Check for a mismatched access type, and fall back conservatively as needed.
2466 // TODO: Be less conservative when the type is similar and can use the same
2467 // addressing modes.
2468 if (Kind == LSRUse::Address) {
2469 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2470 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2471 AccessTy.AddrSpace);
2475 // Conservatively assume HasBaseReg is true for now.
2476 if (NewOffset < LU.MinOffset) {
2477 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2478 LU.MaxOffset - NewOffset, HasBaseReg))
2480 NewMinOffset = NewOffset;
2481 } else if (NewOffset > LU.MaxOffset) {
2482 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2483 NewOffset - LU.MinOffset, HasBaseReg))
2485 NewMaxOffset = NewOffset;
2489 LU.MinOffset = NewMinOffset;
2490 LU.MaxOffset = NewMaxOffset;
2491 LU.AccessTy = NewAccessTy;
2495 /// Return an LSRUse index and an offset value for a fixup which needs the given
2496 /// expression, with the given kind and optional access type. Either reuse an
2497 /// existing use or create a new one, as needed.
2498 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2499 LSRUse::KindType Kind,
2500 MemAccessTy AccessTy) {
2501 const SCEV *Copy = Expr;
2502 int64_t Offset = ExtractImmediate(Expr, SE);
2504 // Basic uses can't accept any offset, for example.
2505 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2506 Offset, /*HasBaseReg=*/ true)) {
2511 std::pair<UseMapTy::iterator, bool> P =
2512 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2514 // A use already existed with this base.
2515 size_t LUIdx = P.first->second;
2516 LSRUse &LU = Uses[LUIdx];
2517 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2519 return std::make_pair(LUIdx, Offset);
2522 // Create a new use.
2523 size_t LUIdx = Uses.size();
2524 P.first->second = LUIdx;
2525 Uses.push_back(LSRUse(Kind, AccessTy));
2526 LSRUse &LU = Uses[LUIdx];
2528 LU.MinOffset = Offset;
2529 LU.MaxOffset = Offset;
2530 return std::make_pair(LUIdx, Offset);
2533 /// Delete the given use from the Uses list.
2534 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2535 if (&LU != &Uses.back())
2536 std::swap(LU, Uses.back());
2540 RegUses.swapAndDropUse(LUIdx, Uses.size());
2543 /// Look for a use distinct from OrigLU which is has a formula that has the same
2544 /// registers as the given formula.
2546 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2547 const LSRUse &OrigLU) {
2548 // Search all uses for the formula. This could be more clever.
2549 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2550 LSRUse &LU = Uses[LUIdx];
2551 // Check whether this use is close enough to OrigLU, to see whether it's
2552 // worthwhile looking through its formulae.
2553 // Ignore ICmpZero uses because they may contain formulae generated by
2554 // GenerateICmpZeroScales, in which case adding fixup offsets may
2556 if (&LU != &OrigLU &&
2557 LU.Kind != LSRUse::ICmpZero &&
2558 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2559 LU.WidestFixupType == OrigLU.WidestFixupType &&
2560 LU.HasFormulaWithSameRegs(OrigF)) {
2561 // Scan through this use's formulae.
2562 for (const Formula &F : LU.Formulae) {
2563 // Check to see if this formula has the same registers and symbols
2565 if (F.BaseRegs == OrigF.BaseRegs &&
2566 F.ScaledReg == OrigF.ScaledReg &&
2567 F.BaseGV == OrigF.BaseGV &&
2568 F.Scale == OrigF.Scale &&
2569 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2570 if (F.BaseOffset == 0)
2572 // This is the formula where all the registers and symbols matched;
2573 // there aren't going to be any others. Since we declined it, we
2574 // can skip the rest of the formulae and proceed to the next LSRUse.
2581 // Nothing looked good.
2585 void LSRInstance::CollectInterestingTypesAndFactors() {
2586 SmallSetVector<const SCEV *, 4> Strides;
2588 // Collect interesting types and strides.
2589 SmallVector<const SCEV *, 4> Worklist;
2590 for (const IVStrideUse &U : IU) {
2591 const SCEV *Expr = IU.getExpr(U);
2593 // Collect interesting types.
2594 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2596 // Add strides for mentioned loops.
2597 Worklist.push_back(Expr);
2599 const SCEV *S = Worklist.pop_back_val();
2600 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2601 if (AR->getLoop() == L)
2602 Strides.insert(AR->getStepRecurrence(SE));
2603 Worklist.push_back(AR->getStart());
2604 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2605 Worklist.append(Add->op_begin(), Add->op_end());
2607 } while (!Worklist.empty());
2610 // Compute interesting factors from the set of interesting strides.
2611 for (SmallSetVector<const SCEV *, 4>::const_iterator
2612 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2613 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2614 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2615 const SCEV *OldStride = *I;
2616 const SCEV *NewStride = *NewStrideIter;
2618 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2619 SE.getTypeSizeInBits(NewStride->getType())) {
2620 if (SE.getTypeSizeInBits(OldStride->getType()) >
2621 SE.getTypeSizeInBits(NewStride->getType()))
2622 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2624 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2626 if (const SCEVConstant *Factor =
2627 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2629 if (Factor->getAPInt().getMinSignedBits() <= 64)
2630 Factors.insert(Factor->getAPInt().getSExtValue());
2631 } else if (const SCEVConstant *Factor =
2632 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2635 if (Factor->getAPInt().getMinSignedBits() <= 64)
2636 Factors.insert(Factor->getAPInt().getSExtValue());
2640 // If all uses use the same type, don't bother looking for truncation-based
2642 if (Types.size() == 1)
2645 DEBUG(print_factors_and_types(dbgs()));
2648 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2649 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2650 /// IVStrideUses, we could partially skip this.
2651 static User::op_iterator
2652 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2653 Loop *L, ScalarEvolution &SE) {
2654 for(; OI != OE; ++OI) {
2655 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2656 if (!SE.isSCEVable(Oper->getType()))
2659 if (const SCEVAddRecExpr *AR =
2660 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2661 if (AR->getLoop() == L)
2669 /// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2670 /// a convenient helper.
2671 static Value *getWideOperand(Value *Oper) {
2672 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2673 return Trunc->getOperand(0);
2677 /// Return true if we allow an IV chain to include both types.
2678 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2679 Type *LType = LVal->getType();
2680 Type *RType = RVal->getType();
2681 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2682 // Different address spaces means (possibly)
2683 // different types of the pointer implementation,
2684 // e.g. i16 vs i32 so disallow that.
2685 (LType->getPointerAddressSpace() ==
2686 RType->getPointerAddressSpace()));
2689 /// Return an approximation of this SCEV expression's "base", or NULL for any
2690 /// constant. Returning the expression itself is conservative. Returning a
2691 /// deeper subexpression is more precise and valid as long as it isn't less
2692 /// complex than another subexpression. For expressions involving multiple
2693 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2694 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2697 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2698 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2699 static const SCEV *getExprBase(const SCEV *S) {
2700 switch (S->getSCEVType()) {
2701 default: // uncluding scUnknown.
2706 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2708 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2710 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2712 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2713 // there's nothing more complex.
2714 // FIXME: not sure if we want to recognize negation.
2715 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2716 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2717 E(Add->op_begin()); I != E; ++I) {
2718 const SCEV *SubExpr = *I;
2719 if (SubExpr->getSCEVType() == scAddExpr)
2720 return getExprBase(SubExpr);
2722 if (SubExpr->getSCEVType() != scMulExpr)
2725 return S; // all operands are scaled, be conservative.
2728 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2732 /// Return true if the chain increment is profitable to expand into a loop
2733 /// invariant value, which may require its own register. A profitable chain
2734 /// increment will be an offset relative to the same base. We allow such offsets
2735 /// to potentially be used as chain increment as long as it's not obviously
2736 /// expensive to expand using real instructions.
2737 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2738 const SCEV *IncExpr,
2739 ScalarEvolution &SE) {
2740 // Aggressively form chains when -stress-ivchain.
2744 // Do not replace a constant offset from IV head with a nonconstant IV
2746 if (!isa<SCEVConstant>(IncExpr)) {
2747 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2748 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2752 SmallPtrSet<const SCEV*, 8> Processed;
2753 return !isHighCostExpansion(IncExpr, Processed, SE);
2756 /// Return true if the number of registers needed for the chain is estimated to
2757 /// be less than the number required for the individual IV users. First prohibit
2758 /// any IV users that keep the IV live across increments (the Users set should
2759 /// be empty). Next count the number and type of increments in the chain.
2761 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2762 /// effectively use postinc addressing modes. Only consider it profitable it the
2763 /// increments can be computed in fewer registers when chained.
2765 /// TODO: Consider IVInc free if it's already used in another chains.
2767 isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2768 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2772 if (!Chain.hasIncs())
2775 if (!Users.empty()) {
2776 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2777 for (Instruction *Inst : Users) {
2778 dbgs() << " " << *Inst << "\n";
2782 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2784 // The chain itself may require a register, so intialize cost to 1.
2787 // A complete chain likely eliminates the need for keeping the original IV in
2788 // a register. LSR does not currently know how to form a complete chain unless
2789 // the header phi already exists.
2790 if (isa<PHINode>(Chain.tailUserInst())
2791 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2794 const SCEV *LastIncExpr = nullptr;
2795 unsigned NumConstIncrements = 0;
2796 unsigned NumVarIncrements = 0;
2797 unsigned NumReusedIncrements = 0;
2798 for (const IVInc &Inc : Chain) {
2799 if (Inc.IncExpr->isZero())
2802 // Incrementing by zero or some constant is neutral. We assume constants can
2803 // be folded into an addressing mode or an add's immediate operand.
2804 if (isa<SCEVConstant>(Inc.IncExpr)) {
2805 ++NumConstIncrements;
2809 if (Inc.IncExpr == LastIncExpr)
2810 ++NumReusedIncrements;
2814 LastIncExpr = Inc.IncExpr;
2816 // An IV chain with a single increment is handled by LSR's postinc
2817 // uses. However, a chain with multiple increments requires keeping the IV's
2818 // value live longer than it needs to be if chained.
2819 if (NumConstIncrements > 1)
2822 // Materializing increment expressions in the preheader that didn't exist in
2823 // the original code may cost a register. For example, sign-extended array
2824 // indices can produce ridiculous increments like this:
2825 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2826 cost += NumVarIncrements;
2828 // Reusing variable increments likely saves a register to hold the multiple of
2830 cost -= NumReusedIncrements;
2832 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2838 /// Add this IV user to an existing chain or make it the head of a new chain.
2839 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2840 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2841 // When IVs are used as types of varying widths, they are generally converted
2842 // to a wider type with some uses remaining narrow under a (free) trunc.
2843 Value *const NextIV = getWideOperand(IVOper);
2844 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2845 const SCEV *const OperExprBase = getExprBase(OperExpr);
2847 // Visit all existing chains. Check if its IVOper can be computed as a
2848 // profitable loop invariant increment from the last link in the Chain.
2849 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2850 const SCEV *LastIncExpr = nullptr;
2851 for (; ChainIdx < NChains; ++ChainIdx) {
2852 IVChain &Chain = IVChainVec[ChainIdx];
2854 // Prune the solution space aggressively by checking that both IV operands
2855 // are expressions that operate on the same unscaled SCEVUnknown. This
2856 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2857 // first avoids creating extra SCEV expressions.
2858 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2861 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2862 if (!isCompatibleIVType(PrevIV, NextIV))
2865 // A phi node terminates a chain.
2866 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2869 // The increment must be loop-invariant so it can be kept in a register.
2870 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2871 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2872 if (!SE.isLoopInvariant(IncExpr, L))
2875 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2876 LastIncExpr = IncExpr;
2880 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2881 // bother for phi nodes, because they must be last in the chain.
2882 if (ChainIdx == NChains) {
2883 if (isa<PHINode>(UserInst))
2885 if (NChains >= MaxChains && !StressIVChain) {
2886 DEBUG(dbgs() << "IV Chain Limit\n");
2889 LastIncExpr = OperExpr;
2890 // IVUsers may have skipped over sign/zero extensions. We don't currently
2891 // attempt to form chains involving extensions unless they can be hoisted
2892 // into this loop's AddRec.
2893 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2896 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2898 ChainUsersVec.resize(NChains);
2899 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2900 << ") IV=" << *LastIncExpr << "\n");
2902 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2903 << ") IV+" << *LastIncExpr << "\n");
2904 // Add this IV user to the end of the chain.
2905 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2907 IVChain &Chain = IVChainVec[ChainIdx];
2909 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2910 // This chain's NearUsers become FarUsers.
2911 if (!LastIncExpr->isZero()) {
2912 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2917 // All other uses of IVOperand become near uses of the chain.
2918 // We currently ignore intermediate values within SCEV expressions, assuming
2919 // they will eventually be used be the current chain, or can be computed
2920 // from one of the chain increments. To be more precise we could
2921 // transitively follow its user and only add leaf IV users to the set.
2922 for (User *U : IVOper->users()) {
2923 Instruction *OtherUse = dyn_cast<Instruction>(U);
2926 // Uses in the chain will no longer be uses if the chain is formed.
2927 // Include the head of the chain in this iteration (not Chain.begin()).
2928 IVChain::const_iterator IncIter = Chain.Incs.begin();
2929 IVChain::const_iterator IncEnd = Chain.Incs.end();
2930 for( ; IncIter != IncEnd; ++IncIter) {
2931 if (IncIter->UserInst == OtherUse)
2934 if (IncIter != IncEnd)
2937 if (SE.isSCEVable(OtherUse->getType())
2938 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2939 && IU.isIVUserOrOperand(OtherUse)) {
2942 NearUsers.insert(OtherUse);
2945 // Since this user is part of the chain, it's no longer considered a use
2947 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2950 /// Populate the vector of Chains.
2952 /// This decreases ILP at the architecture level. Targets with ample registers,
2953 /// multiple memory ports, and no register renaming probably don't want
2954 /// this. However, such targets should probably disable LSR altogether.
2956 /// The job of LSR is to make a reasonable choice of induction variables across
2957 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2958 /// ILP *within the loop* if the target wants it.
2960 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2961 /// will not reorder memory operations, it will recognize this as a chain, but
2962 /// will generate redundant IV increments. Ideally this would be corrected later
2963 /// by a smart scheduler:
2969 /// TODO: Walk the entire domtree within this loop, not just the path to the
2970 /// loop latch. This will discover chains on side paths, but requires
2971 /// maintaining multiple copies of the Chains state.
2972 void LSRInstance::CollectChains() {
2973 DEBUG(dbgs() << "Collecting IV Chains.\n");
2974 SmallVector<ChainUsers, 8> ChainUsersVec;
2976 SmallVector<BasicBlock *,8> LatchPath;
2977 BasicBlock *LoopHeader = L->getHeader();
2978 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2979 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2980 LatchPath.push_back(Rung->getBlock());
2982 LatchPath.push_back(LoopHeader);
2984 // Walk the instruction stream from the loop header to the loop latch.
2985 for (BasicBlock *BB : reverse(LatchPath)) {
2986 for (Instruction &I : *BB) {
2987 // Skip instructions that weren't seen by IVUsers analysis.
2988 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
2991 // Ignore users that are part of a SCEV expression. This way we only
2992 // consider leaf IV Users. This effectively rediscovers a portion of
2993 // IVUsers analysis but in program order this time.
2994 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
2997 // Remove this instruction from any NearUsers set it may be in.
2998 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2999 ChainIdx < NChains; ++ChainIdx) {
3000 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3002 // Search for operands that can be chained.
3003 SmallPtrSet<Instruction*, 4> UniqueOperands;
3004 User::op_iterator IVOpEnd = I.op_end();
3005 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3006 while (IVOpIter != IVOpEnd) {
3007 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3008 if (UniqueOperands.insert(IVOpInst).second)
3009 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3010 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3012 } // Continue walking down the instructions.
3013 } // Continue walking down the domtree.
3014 // Visit phi backedges to determine if the chain can generate the IV postinc.
3015 for (PHINode &PN : L->getHeader()->phis()) {
3016 if (!SE.isSCEVable(PN.getType()))
3020 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3022 ChainInstruction(&PN, IncV, ChainUsersVec);
3024 // Remove any unprofitable chains.
3025 unsigned ChainIdx = 0;
3026 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3027 UsersIdx < NChains; ++UsersIdx) {
3028 if (!isProfitableChain(IVChainVec[UsersIdx],
3029 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3031 // Preserve the chain at UsesIdx.
3032 if (ChainIdx != UsersIdx)
3033 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3034 FinalizeChain(IVChainVec[ChainIdx]);
3037 IVChainVec.resize(ChainIdx);
3040 void LSRInstance::FinalizeChain(IVChain &Chain) {
3041 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3042 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3044 for (const IVInc &Inc : Chain) {
3045 DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
3046 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3047 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3048 IVIncSet.insert(UseI);
3052 /// Return true if the IVInc can be folded into an addressing mode.
3053 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3054 Value *Operand, const TargetTransformInfo &TTI) {
3055 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3056 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3059 if (IncConst->getAPInt().getMinSignedBits() > 64)
3062 MemAccessTy AccessTy = getAccessType(TTI, UserInst);
3063 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3064 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3065 IncOffset, /*HaseBaseReg=*/false))
3071 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3072 /// user's operand from the previous IV user's operand.
3073 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3074 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3075 // Find the new IVOperand for the head of the chain. It may have been replaced
3077 const IVInc &Head = Chain.Incs[0];
3078 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3079 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3080 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3082 Value *IVSrc = nullptr;
3083 while (IVOpIter != IVOpEnd) {
3084 IVSrc = getWideOperand(*IVOpIter);
3086 // If this operand computes the expression that the chain needs, we may use
3087 // it. (Check this after setting IVSrc which is used below.)
3089 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3090 // narrow for the chain, so we can no longer use it. We do allow using a
3091 // wider phi, assuming the LSR checked for free truncation. In that case we
3092 // should already have a truncate on this operand such that
3093 // getSCEV(IVSrc) == IncExpr.
3094 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3095 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3098 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3100 if (IVOpIter == IVOpEnd) {
3101 // Gracefully give up on this chain.
3102 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3106 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3107 Type *IVTy = IVSrc->getType();
3108 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3109 const SCEV *LeftOverExpr = nullptr;
3110 for (const IVInc &Inc : Chain) {
3111 Instruction *InsertPt = Inc.UserInst;
3112 if (isa<PHINode>(InsertPt))
3113 InsertPt = L->getLoopLatch()->getTerminator();
3115 // IVOper will replace the current IV User's operand. IVSrc is the IV
3116 // value currently held in a register.
3117 Value *IVOper = IVSrc;
3118 if (!Inc.IncExpr->isZero()) {
3119 // IncExpr was the result of subtraction of two narrow values, so must
3121 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3122 LeftOverExpr = LeftOverExpr ?
3123 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3125 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3126 // Expand the IV increment.
3127 Rewriter.clearPostInc();
3128 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3129 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3130 SE.getUnknown(IncV));
3131 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3133 // If an IV increment can't be folded, use it as the next IV value.
3134 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3135 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3137 LeftOverExpr = nullptr;
3140 Type *OperTy = Inc.IVOperand->getType();
3141 if (IVTy != OperTy) {
3142 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3143 "cannot extend a chained IV");
3144 IRBuilder<> Builder(InsertPt);
3145 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3147 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3148 DeadInsts.emplace_back(Inc.IVOperand);
3150 // If LSR created a new, wider phi, we may also replace its postinc. We only
3151 // do this if we also found a wide value for the head of the chain.
3152 if (isa<PHINode>(Chain.tailUserInst())) {
3153 for (PHINode &Phi : L->getHeader()->phis()) {
3154 if (!isCompatibleIVType(&Phi, IVSrc))
3156 Instruction *PostIncV = dyn_cast<Instruction>(
3157 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3158 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3160 Value *IVOper = IVSrc;
3161 Type *PostIncTy = PostIncV->getType();
3162 if (IVTy != PostIncTy) {
3163 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3164 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3165 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3166 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3168 Phi.replaceUsesOfWith(PostIncV, IVOper);
3169 DeadInsts.emplace_back(PostIncV);
3174 void LSRInstance::CollectFixupsAndInitialFormulae() {
3175 for (const IVStrideUse &U : IU) {
3176 Instruction *UserInst = U.getUser();
3177 // Skip IV users that are part of profitable IV Chains.
3178 User::op_iterator UseI =
3179 find(UserInst->operands(), U.getOperandValToReplace());
3180 assert(UseI != UserInst->op_end() && "cannot find IV operand");
3181 if (IVIncSet.count(UseI)) {
3182 DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3186 LSRUse::KindType Kind = LSRUse::Basic;
3187 MemAccessTy AccessTy;
3188 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3189 Kind = LSRUse::Address;
3190 AccessTy = getAccessType(TTI, UserInst);
3193 const SCEV *S = IU.getExpr(U);
3194 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3196 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3197 // (N - i == 0), and this allows (N - i) to be the expression that we work
3198 // with rather than just N or i, so we can consider the register
3199 // requirements for both N and i at the same time. Limiting this code to
3200 // equality icmps is not a problem because all interesting loops use
3201 // equality icmps, thanks to IndVarSimplify.
3202 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3203 if (CI->isEquality()) {
3204 // Swap the operands if needed to put the OperandValToReplace on the
3205 // left, for consistency.
3206 Value *NV = CI->getOperand(1);
3207 if (NV == U.getOperandValToReplace()) {
3208 CI->setOperand(1, CI->getOperand(0));
3209 CI->setOperand(0, NV);
3210 NV = CI->getOperand(1);
3214 // x == y --> x - y == 0
3215 const SCEV *N = SE.getSCEV(NV);
3216 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3217 // S is normalized, so normalize N before folding it into S
3218 // to keep the result normalized.
3219 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3220 Kind = LSRUse::ICmpZero;
3221 S = SE.getMinusSCEV(N, S);
3224 // -1 and the negations of all interesting strides (except the negation
3225 // of -1) are now also interesting.
3226 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3227 if (Factors[i] != -1)
3228 Factors.insert(-(uint64_t)Factors[i]);
3232 // Get or create an LSRUse.
3233 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3234 size_t LUIdx = P.first;
3235 int64_t Offset = P.second;
3236 LSRUse &LU = Uses[LUIdx];
3238 // Record the fixup.
3239 LSRFixup &LF = LU.getNewFixup();
3240 LF.UserInst = UserInst;
3241 LF.OperandValToReplace = U.getOperandValToReplace();
3242 LF.PostIncLoops = TmpPostIncLoops;
3244 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3246 if (!LU.WidestFixupType ||
3247 SE.getTypeSizeInBits(LU.WidestFixupType) <
3248 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3249 LU.WidestFixupType = LF.OperandValToReplace->getType();
3251 // If this is the first use of this LSRUse, give it a formula.
3252 if (LU.Formulae.empty()) {
3253 InsertInitialFormula(S, LU, LUIdx);
3254 CountRegisters(LU.Formulae.back(), LUIdx);
3258 DEBUG(print_fixups(dbgs()));
3261 /// Insert a formula for the given expression into the given use, separating out
3262 /// loop-variant portions from loop-invariant and loop-computable portions.
3264 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3265 // Mark uses whose expressions cannot be expanded.
3266 if (!isSafeToExpand(S, SE))
3267 LU.RigidFormula = true;
3270 F.initialMatch(S, L, SE);
3271 bool Inserted = InsertFormula(LU, LUIdx, F);
3272 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3275 /// Insert a simple single-register formula for the given expression into the
3278 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3279 LSRUse &LU, size_t LUIdx) {
3281 F.BaseRegs.push_back(S);
3282 F.HasBaseReg = true;
3283 bool Inserted = InsertFormula(LU, LUIdx, F);
3284 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3287 /// Note which registers are used by the given formula, updating RegUses.
3288 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3290 RegUses.countRegister(F.ScaledReg, LUIdx);
3291 for (const SCEV *BaseReg : F.BaseRegs)
3292 RegUses.countRegister(BaseReg, LUIdx);
3295 /// If the given formula has not yet been inserted, add it to the list, and
3296 /// return true. Return false otherwise.
3297 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3298 // Do not insert formula that we will not be able to expand.
3299 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3300 "Formula is illegal");
3302 if (!LU.InsertFormula(F, *L))
3305 CountRegisters(F, LUIdx);
3309 /// Check for other uses of loop-invariant values which we're tracking. These
3310 /// other uses will pin these values in registers, making them less profitable
3311 /// for elimination.
3312 /// TODO: This currently misses non-constant addrec step registers.
3313 /// TODO: Should this give more weight to users inside the loop?
3315 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3316 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3317 SmallPtrSet<const SCEV *, 32> Visited;
3319 while (!Worklist.empty()) {
3320 const SCEV *S = Worklist.pop_back_val();
3322 // Don't process the same SCEV twice
3323 if (!Visited.insert(S).second)
3326 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3327 Worklist.append(N->op_begin(), N->op_end());
3328 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3329 Worklist.push_back(C->getOperand());
3330 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3331 Worklist.push_back(D->getLHS());
3332 Worklist.push_back(D->getRHS());
3333 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3334 const Value *V = US->getValue();
3335 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3336 // Look for instructions defined outside the loop.
3337 if (L->contains(Inst)) continue;
3338 } else if (isa<UndefValue>(V))
3339 // Undef doesn't have a live range, so it doesn't matter.
3341 for (const Use &U : V->uses()) {
3342 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3343 // Ignore non-instructions.
3346 // Ignore instructions in other functions (as can happen with
3348 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3350 // Ignore instructions not dominated by the loop.
3351 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3352 UserInst->getParent() :
3353 cast<PHINode>(UserInst)->getIncomingBlock(
3354 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3355 if (!DT.dominates(L->getHeader(), UseBB))
3357 // Don't bother if the instruction is in a BB which ends in an EHPad.
3358 if (UseBB->getTerminator()->isEHPad())
3360 // Don't bother rewriting PHIs in catchswitch blocks.
3361 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3363 // Ignore uses which are part of other SCEV expressions, to avoid
3364 // analyzing them multiple times.
3365 if (SE.isSCEVable(UserInst->getType())) {
3366 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3367 // If the user is a no-op, look through to its uses.
3368 if (!isa<SCEVUnknown>(UserS))
3372 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3376 // Ignore icmp instructions which are already being analyzed.
3377 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3378 unsigned OtherIdx = !U.getOperandNo();
3379 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3380 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3384 std::pair<size_t, int64_t> P = getUse(
3385 S, LSRUse::Basic, MemAccessTy());
3386 size_t LUIdx = P.first;
3387 int64_t Offset = P.second;
3388 LSRUse &LU = Uses[LUIdx];
3389 LSRFixup &LF = LU.getNewFixup();
3390 LF.UserInst = const_cast<Instruction *>(UserInst);
3391 LF.OperandValToReplace = U;
3393 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3394 if (!LU.WidestFixupType ||
3395 SE.getTypeSizeInBits(LU.WidestFixupType) <
3396 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3397 LU.WidestFixupType = LF.OperandValToReplace->getType();
3398 InsertSupplementalFormula(US, LU, LUIdx);
3399 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3406 /// Split S into subexpressions which can be pulled out into separate
3407 /// registers. If C is non-null, multiply each subexpression by C.
3409 /// Return remainder expression after factoring the subexpressions captured by
3410 /// Ops. If Ops is complete, return NULL.
3411 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3412 SmallVectorImpl<const SCEV *> &Ops,
3414 ScalarEvolution &SE,
3415 unsigned Depth = 0) {
3416 // Arbitrarily cap recursion to protect compile time.
3420 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3421 // Break out add operands.
3422 for (const SCEV *S : Add->operands()) {
3423 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3425 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3428 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3429 // Split a non-zero base out of an addrec.
3430 if (AR->getStart()->isZero() || !AR->isAffine())
3433 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3434 C, Ops, L, SE, Depth+1);
3435 // Split the non-zero AddRec unless it is part of a nested recurrence that
3436 // does not pertain to this loop.
3437 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3438 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3439 Remainder = nullptr;
3441 if (Remainder != AR->getStart()) {
3443 Remainder = SE.getConstant(AR->getType(), 0);
3444 return SE.getAddRecExpr(Remainder,
3445 AR->getStepRecurrence(SE),
3447 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3450 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3451 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3452 if (Mul->getNumOperands() != 2)
3454 if (const SCEVConstant *Op0 =
3455 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3456 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3457 const SCEV *Remainder =
3458 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3460 Ops.push_back(SE.getMulExpr(C, Remainder));
3467 /// \brief Helper function for LSRInstance::GenerateReassociations.
3468 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3469 const Formula &Base,
3470 unsigned Depth, size_t Idx,
3472 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3473 SmallVector<const SCEV *, 8> AddOps;
3474 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3476 AddOps.push_back(Remainder);
3478 if (AddOps.size() == 1)
3481 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3484 // Loop-variant "unknown" values are uninteresting; we won't be able to
3485 // do anything meaningful with them.
3486 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3489 // Don't pull a constant into a register if the constant could be folded
3490 // into an immediate field.
3491 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3492 LU.AccessTy, *J, Base.getNumRegs() > 1))
3495 // Collect all operands except *J.
3496 SmallVector<const SCEV *, 8> InnerAddOps(
3497 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3498 InnerAddOps.append(std::next(J),
3499 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3501 // Don't leave just a constant behind in a register if the constant could
3502 // be folded into an immediate field.
3503 if (InnerAddOps.size() == 1 &&
3504 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3505 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3508 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3509 if (InnerSum->isZero())
3513 // Add the remaining pieces of the add back into the new formula.
3514 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3515 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3516 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3517 InnerSumSC->getValue()->getZExtValue())) {
3519 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3521 F.ScaledReg = nullptr;
3523 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3524 } else if (IsScaledReg)
3525 F.ScaledReg = InnerSum;
3527 F.BaseRegs[Idx] = InnerSum;
3529 // Add J as its own register, or an unfolded immediate.
3530 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3531 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3532 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3533 SC->getValue()->getZExtValue()))
3535 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3537 F.BaseRegs.push_back(*J);
3538 // We may have changed the number of register in base regs, adjust the
3539 // formula accordingly.
3542 if (InsertFormula(LU, LUIdx, F))
3543 // If that formula hadn't been seen before, recurse to find more like
3545 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3549 /// Split out subexpressions from adds and the bases of addrecs.
3550 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3551 Formula Base, unsigned Depth) {
3552 assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3553 // Arbitrarily cap recursion to protect compile time.
3557 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3558 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3560 if (Base.Scale == 1)
3561 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3562 /* Idx */ -1, /* IsScaledReg */ true);
3565 /// Generate a formula consisting of all of the loop-dominating registers added
3566 /// into a single register.
3567 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3569 // This method is only interesting on a plurality of registers.
3570 if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3573 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3574 // processing the formula.
3578 SmallVector<const SCEV *, 4> Ops;
3579 for (const SCEV *BaseReg : Base.BaseRegs) {
3580 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3581 !SE.hasComputableLoopEvolution(BaseReg, L))
3582 Ops.push_back(BaseReg);
3584 F.BaseRegs.push_back(BaseReg);
3586 if (Ops.size() > 1) {
3587 const SCEV *Sum = SE.getAddExpr(Ops);
3588 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3589 // opportunity to fold something. For now, just ignore such cases
3590 // rather than proceed with zero in a register.
3591 if (!Sum->isZero()) {
3592 F.BaseRegs.push_back(Sum);
3594 (void)InsertFormula(LU, LUIdx, F);
3599 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3600 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3601 const Formula &Base, size_t Idx,
3603 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3604 GlobalValue *GV = ExtractSymbol(G, SE);
3605 if (G->isZero() || !GV)
3609 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3614 F.BaseRegs[Idx] = G;
3615 (void)InsertFormula(LU, LUIdx, F);
3618 /// Generate reuse formulae using symbolic offsets.
3619 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3621 // We can't add a symbolic offset if the address already contains one.
3622 if (Base.BaseGV) return;
3624 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3625 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3626 if (Base.Scale == 1)
3627 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3628 /* IsScaledReg */ true);
3631 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3632 void LSRInstance::GenerateConstantOffsetsImpl(
3633 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3634 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3635 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3636 for (int64_t Offset : Worklist) {
3638 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3639 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3641 // Add the offset to the base register.
3642 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3643 // If it cancelled out, drop the base register, otherwise update it.
3644 if (NewG->isZero()) {
3647 F.ScaledReg = nullptr;
3649 F.deleteBaseReg(F.BaseRegs[Idx]);
3651 } else if (IsScaledReg)
3654 F.BaseRegs[Idx] = NewG;
3656 (void)InsertFormula(LU, LUIdx, F);
3660 int64_t Imm = ExtractImmediate(G, SE);
3661 if (G->isZero() || Imm == 0)
3664 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3665 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3670 F.BaseRegs[Idx] = G;
3671 (void)InsertFormula(LU, LUIdx, F);
3674 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3675 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3677 // TODO: For now, just add the min and max offset, because it usually isn't
3678 // worthwhile looking at everything inbetween.
3679 SmallVector<int64_t, 2> Worklist;
3680 Worklist.push_back(LU.MinOffset);
3681 if (LU.MaxOffset != LU.MinOffset)
3682 Worklist.push_back(LU.MaxOffset);
3684 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3685 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3686 if (Base.Scale == 1)
3687 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3688 /* IsScaledReg */ true);
3691 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3692 /// == y -> x*c == y*c.
3693 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3695 if (LU.Kind != LSRUse::ICmpZero) return;
3697 // Determine the integer type for the base formula.
3698 Type *IntTy = Base.getType();
3700 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3702 // Don't do this if there is more than one offset.
3703 if (LU.MinOffset != LU.MaxOffset) return;
3705 // Check if transformation is valid. It is illegal to multiply pointer.
3706 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3708 for (const SCEV *BaseReg : Base.BaseRegs)
3709 if (BaseReg->getType()->isPointerTy())
3711 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3713 // Check each interesting stride.
3714 for (int64_t Factor : Factors) {
3715 // Check that the multiplication doesn't overflow.
3716 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3718 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3719 if (NewBaseOffset / Factor != Base.BaseOffset)
3721 // If the offset will be truncated at this use, check that it is in bounds.
3722 if (!IntTy->isPointerTy() &&
3723 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3726 // Check that multiplying with the use offset doesn't overflow.
3727 int64_t Offset = LU.MinOffset;
3728 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3730 Offset = (uint64_t)Offset * Factor;
3731 if (Offset / Factor != LU.MinOffset)
3733 // If the offset will be truncated at this use, check that it is in bounds.
3734 if (!IntTy->isPointerTy() &&
3735 !ConstantInt::isValueValidForType(IntTy, Offset))
3739 F.BaseOffset = NewBaseOffset;
3741 // Check that this scale is legal.
3742 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3745 // Compensate for the use having MinOffset built into it.
3746 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3748 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3750 // Check that multiplying with each base register doesn't overflow.
3751 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3752 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3753 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3757 // Check that multiplying with the scaled register doesn't overflow.
3759 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3760 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3764 // Check that multiplying with the unfolded offset doesn't overflow.
3765 if (F.UnfoldedOffset != 0) {
3766 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3769 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3770 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3772 // If the offset will be truncated, check that it is in bounds.
3773 if (!IntTy->isPointerTy() &&
3774 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3778 // If we make it here and it's legal, add it.
3779 (void)InsertFormula(LU, LUIdx, F);
3784 /// Generate stride factor reuse formulae by making use of scaled-offset address
3785 /// modes, for example.
3786 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3787 // Determine the integer type for the base formula.
3788 Type *IntTy = Base.getType();
3791 // If this Formula already has a scaled register, we can't add another one.
3792 // Try to unscale the formula to generate a better scale.
3793 if (Base.Scale != 0 && !Base.unscale())
3796 assert(Base.Scale == 0 && "unscale did not did its job!");
3798 // Check each interesting stride.
3799 for (int64_t Factor : Factors) {
3800 Base.Scale = Factor;
3801 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3802 // Check whether this scale is going to be legal.
3803 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3805 // As a special-case, handle special out-of-loop Basic users specially.
3806 // TODO: Reconsider this special case.
3807 if (LU.Kind == LSRUse::Basic &&
3808 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3809 LU.AccessTy, Base) &&
3810 LU.AllFixupsOutsideLoop)
3811 LU.Kind = LSRUse::Special;
3815 // For an ICmpZero, negating a solitary base register won't lead to
3817 if (LU.Kind == LSRUse::ICmpZero &&
3818 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3820 // For each addrec base reg, if its loop is current loop, apply the scale.
3821 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3822 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3823 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3824 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3825 if (FactorS->isZero())
3827 // Divide out the factor, ignoring high bits, since we'll be
3828 // scaling the value back up in the end.
3829 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3830 // TODO: This could be optimized to avoid all the copying.
3832 F.ScaledReg = Quotient;
3833 F.deleteBaseReg(F.BaseRegs[i]);
3834 // The canonical representation of 1*reg is reg, which is already in
3835 // Base. In that case, do not try to insert the formula, it will be
3837 if (F.Scale == 1 && (F.BaseRegs.empty() ||
3838 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3840 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3841 // non canonical Formula with ScaledReg's loop not being L.
3842 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
3844 (void)InsertFormula(LU, LUIdx, F);
3851 /// Generate reuse formulae from different IV types.
3852 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3853 // Don't bother truncating symbolic values.
3854 if (Base.BaseGV) return;
3856 // Determine the integer type for the base formula.
3857 Type *DstTy = Base.getType();
3859 DstTy = SE.getEffectiveSCEVType(DstTy);
3861 for (Type *SrcTy : Types) {
3862 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3865 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3866 for (const SCEV *&BaseReg : F.BaseRegs)
3867 BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3869 // TODO: This assumes we've done basic processing on all uses and
3870 // have an idea what the register usage is.
3871 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3875 (void)InsertFormula(LU, LUIdx, F);
3882 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3883 /// modifications so that the search phase doesn't have to worry about the data
3884 /// structures moving underneath it.
3888 const SCEV *OrigReg;
3890 WorkItem(size_t LI, int64_t I, const SCEV *R)
3891 : LUIdx(LI), Imm(I), OrigReg(R) {}
3893 void print(raw_ostream &OS) const;
3897 } // end anonymous namespace
3899 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3900 void WorkItem::print(raw_ostream &OS) const {
3901 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3902 << " , add offset " << Imm;
3905 LLVM_DUMP_METHOD void WorkItem::dump() const {
3906 print(errs()); errs() << '\n';
3910 /// Look for registers which are a constant distance apart and try to form reuse
3911 /// opportunities between them.
3912 void LSRInstance::GenerateCrossUseConstantOffsets() {
3913 // Group the registers by their value without any added constant offset.
3914 using ImmMapTy = std::map<int64_t, const SCEV *>;
3916 DenseMap<const SCEV *, ImmMapTy> Map;
3917 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3918 SmallVector<const SCEV *, 8> Sequence;
3919 for (const SCEV *Use : RegUses) {
3920 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
3921 int64_t Imm = ExtractImmediate(Reg, SE);
3922 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
3924 Sequence.push_back(Reg);
3925 Pair.first->second.insert(std::make_pair(Imm, Use));
3926 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
3929 // Now examine each set of registers with the same base value. Build up
3930 // a list of work to do and do the work in a separate step so that we're
3931 // not adding formulae and register counts while we're searching.
3932 SmallVector<WorkItem, 32> WorkItems;
3933 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3934 for (const SCEV *Reg : Sequence) {
3935 const ImmMapTy &Imms = Map.find(Reg)->second;
3937 // It's not worthwhile looking for reuse if there's only one offset.
3938 if (Imms.size() == 1)
3941 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3942 for (const auto &Entry : Imms)
3943 dbgs() << ' ' << Entry.first;
3946 // Examine each offset.
3947 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3949 const SCEV *OrigReg = J->second;
3951 int64_t JImm = J->first;
3952 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3954 if (!isa<SCEVConstant>(OrigReg) &&
3955 UsedByIndicesMap[Reg].count() == 1) {
3956 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3960 // Conservatively examine offsets between this orig reg a few selected
3962 ImmMapTy::const_iterator OtherImms[] = {
3963 Imms.begin(), std::prev(Imms.end()),
3964 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3967 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3968 ImmMapTy::const_iterator M = OtherImms[i];
3969 if (M == J || M == JE) continue;
3971 // Compute the difference between the two.
3972 int64_t Imm = (uint64_t)JImm - M->first;
3973 for (unsigned LUIdx : UsedByIndices.set_bits())
3974 // Make a memo of this use, offset, and register tuple.
3975 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3976 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3983 UsedByIndicesMap.clear();
3984 UniqueItems.clear();
3986 // Now iterate through the worklist and add new formulae.
3987 for (const WorkItem &WI : WorkItems) {
3988 size_t LUIdx = WI.LUIdx;
3989 LSRUse &LU = Uses[LUIdx];
3990 int64_t Imm = WI.Imm;
3991 const SCEV *OrigReg = WI.OrigReg;
3993 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3994 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3995 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3997 // TODO: Use a more targeted data structure.
3998 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3999 Formula F = LU.Formulae[L];
4000 // FIXME: The code for the scaled and unscaled registers looks
4001 // very similar but slightly different. Investigate if they
4002 // could be merged. That way, we would not have to unscale the
4005 // Use the immediate in the scaled register.
4006 if (F.ScaledReg == OrigReg) {
4007 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4008 // Don't create 50 + reg(-50).
4009 if (F.referencesReg(SE.getSCEV(
4010 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4013 NewF.BaseOffset = Offset;
4014 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4017 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4019 // If the new scale is a constant in a register, and adding the constant
4020 // value to the immediate would produce a value closer to zero than the
4021 // immediate itself, then the formula isn't worthwhile.
4022 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4023 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4024 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4025 .ule(std::abs(NewF.BaseOffset)))
4029 NewF.canonicalize(*this->L);
4030 (void)InsertFormula(LU, LUIdx, NewF);
4032 // Use the immediate in a base register.
4033 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4034 const SCEV *BaseReg = F.BaseRegs[N];
4035 if (BaseReg != OrigReg)
4038 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4039 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4040 LU.Kind, LU.AccessTy, NewF)) {
4041 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4044 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4046 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4048 // If the new formula has a constant in a register, and adding the
4049 // constant value to the immediate would produce a value closer to
4050 // zero than the immediate itself, then the formula isn't worthwhile.
4051 for (const SCEV *NewReg : NewF.BaseRegs)
4052 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4053 if ((C->getAPInt() + NewF.BaseOffset)
4055 .slt(std::abs(NewF.BaseOffset)) &&
4056 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4057 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4061 NewF.canonicalize(*this->L);
4062 (void)InsertFormula(LU, LUIdx, NewF);
4071 /// Generate formulae for each use.
4073 LSRInstance::GenerateAllReuseFormulae() {
4074 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4075 // queries are more precise.
4076 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4077 LSRUse &LU = Uses[LUIdx];
4078 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4079 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4080 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4081 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4083 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4084 LSRUse &LU = Uses[LUIdx];
4085 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4086 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4087 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4088 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4089 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4090 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4091 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4092 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4094 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4095 LSRUse &LU = Uses[LUIdx];
4096 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4097 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4100 GenerateCrossUseConstantOffsets();
4102 DEBUG(dbgs() << "\n"
4103 "After generating reuse formulae:\n";
4104 print_uses(dbgs()));
4107 /// If there are multiple formulae with the same set of registers used
4108 /// by other uses, pick the best one and delete the others.
4109 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4110 DenseSet<const SCEV *> VisitedRegs;
4111 SmallPtrSet<const SCEV *, 16> Regs;
4112 SmallPtrSet<const SCEV *, 16> LoserRegs;
4114 bool ChangedFormulae = false;
4117 // Collect the best formula for each unique set of shared registers. This
4118 // is reset for each use.
4119 using BestFormulaeTy =
4120 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4122 BestFormulaeTy BestFormulae;
4124 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4125 LSRUse &LU = Uses[LUIdx];
4126 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4129 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4130 FIdx != NumForms; ++FIdx) {
4131 Formula &F = LU.Formulae[FIdx];
4133 // Some formulas are instant losers. For example, they may depend on
4134 // nonexistent AddRecs from other loops. These need to be filtered
4135 // immediately, otherwise heuristics could choose them over others leading
4136 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4137 // avoids the need to recompute this information across formulae using the
4138 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4139 // the corresponding bad register from the Regs set.
4142 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, SE, DT, LU, &LoserRegs);
4143 if (CostF.isLoser()) {
4144 // During initial formula generation, undesirable formulae are generated
4145 // by uses within other loops that have some non-trivial address mode or
4146 // use the postinc form of the IV. LSR needs to provide these formulae
4147 // as the basis of rediscovering the desired formula that uses an AddRec
4148 // corresponding to the existing phi. Once all formulae have been
4149 // generated, these initial losers may be pruned.
4150 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4154 SmallVector<const SCEV *, 4> Key;
4155 for (const SCEV *Reg : F.BaseRegs) {
4156 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4160 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4161 Key.push_back(F.ScaledReg);
4162 // Unstable sort by host order ok, because this is only used for
4164 std::sort(Key.begin(), Key.end());
4166 std::pair<BestFormulaeTy::const_iterator, bool> P =
4167 BestFormulae.insert(std::make_pair(Key, FIdx));
4171 Formula &Best = LU.Formulae[P.first->second];
4175 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, SE, DT, LU);
4176 if (CostF.isLess(CostBest, TTI))
4178 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4180 " in favor of formula "; Best.print(dbgs());
4184 ChangedFormulae = true;
4186 LU.DeleteFormula(F);
4192 // Now that we've filtered out some formulae, recompute the Regs set.
4194 LU.RecomputeRegs(LUIdx, RegUses);
4196 // Reset this to prepare for the next use.
4197 BestFormulae.clear();
4200 DEBUG(if (ChangedFormulae) {
4202 "After filtering out undesirable candidates:\n";
4207 // This is a rough guess that seems to work fairly well.
4208 static const size_t ComplexityLimit = std::numeric_limits<uint16_t>::max();
4210 /// Estimate the worst-case number of solutions the solver might have to
4211 /// consider. It almost never considers this many solutions because it prune the
4212 /// search space, but the pruning isn't always sufficient.
4213 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4215 for (const LSRUse &LU : Uses) {
4216 size_t FSize = LU.Formulae.size();
4217 if (FSize >= ComplexityLimit) {
4218 Power = ComplexityLimit;
4222 if (Power >= ComplexityLimit)
4228 /// When one formula uses a superset of the registers of another formula, it
4229 /// won't help reduce register pressure (though it may not necessarily hurt
4230 /// register pressure); remove it to simplify the system.
4231 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4232 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4233 DEBUG(dbgs() << "The search space is too complex.\n");
4235 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4236 "which use a superset of registers used by other "
4239 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4240 LSRUse &LU = Uses[LUIdx];
4242 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4243 Formula &F = LU.Formulae[i];
4244 // Look for a formula with a constant or GV in a register. If the use
4245 // also has a formula with that same value in an immediate field,
4246 // delete the one that uses a register.
4247 for (SmallVectorImpl<const SCEV *>::const_iterator
4248 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4249 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4251 NewF.BaseOffset += C->getValue()->getSExtValue();
4252 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4253 (I - F.BaseRegs.begin()));
4254 if (LU.HasFormulaWithSameRegs(NewF)) {
4255 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4256 LU.DeleteFormula(F);
4262 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4263 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4267 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4268 (I - F.BaseRegs.begin()));
4269 if (LU.HasFormulaWithSameRegs(NewF)) {
4270 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4272 LU.DeleteFormula(F);
4283 LU.RecomputeRegs(LUIdx, RegUses);
4286 DEBUG(dbgs() << "After pre-selection:\n";
4287 print_uses(dbgs()));
4291 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4292 /// allocate a single register for them.
4293 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4294 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4297 DEBUG(dbgs() << "The search space is too complex.\n"
4298 "Narrowing the search space by assuming that uses separated "
4299 "by a constant offset will use the same registers.\n");
4301 // This is especially useful for unrolled loops.
4303 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4304 LSRUse &LU = Uses[LUIdx];
4305 for (const Formula &F : LU.Formulae) {
4306 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4309 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4313 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4314 LU.Kind, LU.AccessTy))
4317 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4319 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4321 // Transfer the fixups of LU to LUThatHas.
4322 for (LSRFixup &Fixup : LU.Fixups) {
4323 Fixup.Offset += F.BaseOffset;
4324 LUThatHas->pushFixup(Fixup);
4325 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4328 // Delete formulae from the new use which are no longer legal.
4330 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4331 Formula &F = LUThatHas->Formulae[i];
4332 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4333 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4334 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4336 LUThatHas->DeleteFormula(F);
4344 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4346 // Delete the old use.
4347 DeleteUse(LU, LUIdx);
4354 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4357 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4358 /// we've done more filtering, as it may be able to find more formulae to
4360 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4361 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4362 DEBUG(dbgs() << "The search space is too complex.\n");
4364 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4365 "undesirable dedicated registers.\n");
4367 FilterOutUndesirableDedicatedRegisters();
4369 DEBUG(dbgs() << "After pre-selection:\n";
4370 print_uses(dbgs()));
4374 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4375 /// Pick the best one and delete the others.
4376 /// This narrowing heuristic is to keep as many formulae with different
4377 /// Scale and ScaledReg pair as possible while narrowing the search space.
4378 /// The benefit is that it is more likely to find out a better solution
4379 /// from a formulae set with more Scale and ScaledReg variations than
4380 /// a formulae set with the same Scale and ScaledReg. The picking winner
4381 /// reg heurstic will often keep the formulae with the same Scale and
4382 /// ScaledReg and filter others, and we want to avoid that if possible.
4383 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4384 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4387 DEBUG(dbgs() << "The search space is too complex.\n"
4388 "Narrowing the search space by choosing the best Formula "
4389 "from the Formulae with the same Scale and ScaledReg.\n");
4391 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4392 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4394 BestFormulaeTy BestFormulae;
4396 bool ChangedFormulae = false;
4398 DenseSet<const SCEV *> VisitedRegs;
4399 SmallPtrSet<const SCEV *, 16> Regs;
4401 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4402 LSRUse &LU = Uses[LUIdx];
4403 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4405 // Return true if Formula FA is better than Formula FB.
4406 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4407 // First we will try to choose the Formula with fewer new registers.
4408 // For a register used by current Formula, the more the register is
4409 // shared among LSRUses, the less we increase the register number
4410 // counter of the formula.
4411 size_t FARegNum = 0;
4412 for (const SCEV *Reg : FA.BaseRegs) {
4413 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4414 FARegNum += (NumUses - UsedByIndices.count() + 1);
4416 size_t FBRegNum = 0;
4417 for (const SCEV *Reg : FB.BaseRegs) {
4418 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4419 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4421 if (FARegNum != FBRegNum)
4422 return FARegNum < FBRegNum;
4424 // If the new register numbers are the same, choose the Formula with
4426 Cost CostFA, CostFB;
4428 CostFA.RateFormula(TTI, FA, Regs, VisitedRegs, L, SE, DT, LU);
4430 CostFB.RateFormula(TTI, FB, Regs, VisitedRegs, L, SE, DT, LU);
4431 return CostFA.isLess(CostFB, TTI);
4435 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4437 Formula &F = LU.Formulae[FIdx];
4440 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4444 Formula &Best = LU.Formulae[P.first->second];
4445 if (IsBetterThan(F, Best))
4447 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4449 " in favor of formula ";
4450 Best.print(dbgs()); dbgs() << '\n');
4452 ChangedFormulae = true;
4454 LU.DeleteFormula(F);
4460 LU.RecomputeRegs(LUIdx, RegUses);
4462 // Reset this to prepare for the next use.
4463 BestFormulae.clear();
4466 DEBUG(if (ChangedFormulae) {
4468 "After filtering out undesirable candidates:\n";
4473 /// The function delete formulas with high registers number expectation.
4474 /// Assuming we don't know the value of each formula (already delete
4475 /// all inefficient), generate probability of not selecting for each
4479 /// reg(a) + reg({0,+,1})
4480 /// reg(a) + reg({-1,+,1}) + 1
4483 /// reg(b) + reg({0,+,1})
4484 /// reg(b) + reg({-1,+,1}) + 1
4487 /// reg(c) + reg(b) + reg({0,+,1})
4488 /// reg(c) + reg({b,+,1})
4490 /// Probability of not selecting
4492 /// reg(a) (1/3) * 1 * 1
4493 /// reg(b) 1 * (1/3) * (1/2)
4494 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4495 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4496 /// reg({a,+,1}) (2/3) * 1 * 1
4497 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4498 /// reg(c) 1 * 1 * 0
4500 /// Now count registers number mathematical expectation for each formula:
4501 /// Note that for each use we exclude probability if not selecting for the use.
4502 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4503 /// probabilty 1/3 of not selecting for Use1).
4505 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4506 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4509 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4510 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4511 /// reg({b,+,1}) 2/3
4513 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4514 /// reg(c) + reg({b,+,1}) 1 + 2/3
4515 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4516 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4518 // Ok, we have too many of formulae on our hands to conveniently handle.
4519 // Use a rough heuristic to thin out the list.
4521 // Set of Regs wich will be 100% used in final solution.
4522 // Used in each formula of a solution (in example above this is reg(c)).
4523 // We can skip them in calculations.
4524 SmallPtrSet<const SCEV *, 4> UniqRegs;
4525 DEBUG(dbgs() << "The search space is too complex.\n");
4527 // Map each register to probability of not selecting
4528 DenseMap <const SCEV *, float> RegNumMap;
4529 for (const SCEV *Reg : RegUses) {
4530 if (UniqRegs.count(Reg))
4533 for (const LSRUse &LU : Uses) {
4534 if (!LU.Regs.count(Reg))
4536 float P = LU.getNotSelectedProbability(Reg);
4540 UniqRegs.insert(Reg);
4542 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4545 DEBUG(dbgs() << "Narrowing the search space by deleting costly formulas\n");
4547 // Delete formulas where registers number expectation is high.
4548 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4549 LSRUse &LU = Uses[LUIdx];
4550 // If nothing to delete - continue.
4551 if (LU.Formulae.size() < 2)
4553 // This is temporary solution to test performance. Float should be
4554 // replaced with round independent type (based on integers) to avoid
4555 // different results for different target builds.
4556 float FMinRegNum = LU.Formulae[0].getNumRegs();
4557 float FMinARegNum = LU.Formulae[0].getNumRegs();
4559 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4560 Formula &F = LU.Formulae[i];
4563 for (const SCEV *BaseReg : F.BaseRegs) {
4564 if (UniqRegs.count(BaseReg))
4566 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4567 if (isa<SCEVAddRecExpr>(BaseReg))
4569 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4571 if (const SCEV *ScaledReg = F.ScaledReg) {
4572 if (!UniqRegs.count(ScaledReg)) {
4574 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4575 if (isa<SCEVAddRecExpr>(ScaledReg))
4577 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4580 if (FMinRegNum > FRegNum ||
4581 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4582 FMinRegNum = FRegNum;
4583 FMinARegNum = FARegNum;
4587 DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4588 dbgs() << " with min reg num " << FMinRegNum << '\n');
4590 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4591 while (LU.Formulae.size() != 1) {
4592 DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4594 LU.Formulae.pop_back();
4596 LU.RecomputeRegs(LUIdx, RegUses);
4597 assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4598 Formula &F = LU.Formulae[0];
4599 DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4600 // When we choose the formula, the regs become unique.
4601 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4603 UniqRegs.insert(F.ScaledReg);
4605 DEBUG(dbgs() << "After pre-selection:\n";
4606 print_uses(dbgs()));
4609 /// Pick a register which seems likely to be profitable, and then in any use
4610 /// which has any reference to that register, delete all formulae which do not
4611 /// reference that register.
4612 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4613 // With all other options exhausted, loop until the system is simple
4614 // enough to handle.
4615 SmallPtrSet<const SCEV *, 4> Taken;
4616 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4617 // Ok, we have too many of formulae on our hands to conveniently handle.
4618 // Use a rough heuristic to thin out the list.
4619 DEBUG(dbgs() << "The search space is too complex.\n");
4621 // Pick the register which is used by the most LSRUses, which is likely
4622 // to be a good reuse register candidate.
4623 const SCEV *Best = nullptr;
4624 unsigned BestNum = 0;
4625 for (const SCEV *Reg : RegUses) {
4626 if (Taken.count(Reg))
4630 BestNum = RegUses.getUsedByIndices(Reg).count();
4632 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4633 if (Count > BestNum) {
4640 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4641 << " will yield profitable reuse.\n");
4644 // In any use with formulae which references this register, delete formulae
4645 // which don't reference it.
4646 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4647 LSRUse &LU = Uses[LUIdx];
4648 if (!LU.Regs.count(Best)) continue;
4651 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4652 Formula &F = LU.Formulae[i];
4653 if (!F.referencesReg(Best)) {
4654 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4655 LU.DeleteFormula(F);
4659 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4665 LU.RecomputeRegs(LUIdx, RegUses);
4668 DEBUG(dbgs() << "After pre-selection:\n";
4669 print_uses(dbgs()));
4673 /// If there are an extraordinary number of formulae to choose from, use some
4674 /// rough heuristics to prune down the number of formulae. This keeps the main
4675 /// solver from taking an extraordinary amount of time in some worst-case
4677 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4678 NarrowSearchSpaceByDetectingSupersets();
4679 NarrowSearchSpaceByCollapsingUnrolledCode();
4680 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4681 if (FilterSameScaledReg)
4682 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4684 NarrowSearchSpaceByDeletingCostlyFormulas();
4686 NarrowSearchSpaceByPickingWinnerRegs();
4689 /// This is the recursive solver.
4690 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4692 SmallVectorImpl<const Formula *> &Workspace,
4693 const Cost &CurCost,
4694 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4695 DenseSet<const SCEV *> &VisitedRegs) const {
4698 // - use more aggressive filtering
4699 // - sort the formula so that the most profitable solutions are found first
4700 // - sort the uses too
4702 // - don't compute a cost, and then compare. compare while computing a cost
4704 // - track register sets with SmallBitVector
4706 const LSRUse &LU = Uses[Workspace.size()];
4708 // If this use references any register that's already a part of the
4709 // in-progress solution, consider it a requirement that a formula must
4710 // reference that register in order to be considered. This prunes out
4711 // unprofitable searching.
4712 SmallSetVector<const SCEV *, 4> ReqRegs;
4713 for (const SCEV *S : CurRegs)
4714 if (LU.Regs.count(S))
4717 SmallPtrSet<const SCEV *, 16> NewRegs;
4719 for (const Formula &F : LU.Formulae) {
4720 // Ignore formulae which may not be ideal in terms of register reuse of
4721 // ReqRegs. The formula should use all required registers before
4722 // introducing new ones.
4723 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4724 for (const SCEV *Reg : ReqRegs) {
4725 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4726 is_contained(F.BaseRegs, Reg)) {
4728 if (NumReqRegsToFind == 0)
4732 if (NumReqRegsToFind != 0) {
4733 // If none of the formulae satisfied the required registers, then we could
4734 // clear ReqRegs and try again. Currently, we simply give up in this case.
4738 // Evaluate the cost of the current formula. If it's already worse than
4739 // the current best, prune the search at that point.
4742 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, SE, DT, LU);
4743 if (NewCost.isLess(SolutionCost, TTI)) {
4744 Workspace.push_back(&F);
4745 if (Workspace.size() != Uses.size()) {
4746 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4747 NewRegs, VisitedRegs);
4748 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4749 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4751 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4752 dbgs() << ".\n Regs:";
4753 for (const SCEV *S : NewRegs)
4754 dbgs() << ' ' << *S;
4757 SolutionCost = NewCost;
4758 Solution = Workspace;
4760 Workspace.pop_back();
4765 /// Choose one formula from each use. Return the results in the given Solution
4767 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4768 SmallVector<const Formula *, 8> Workspace;
4770 SolutionCost.Lose();
4772 SmallPtrSet<const SCEV *, 16> CurRegs;
4773 DenseSet<const SCEV *> VisitedRegs;
4774 Workspace.reserve(Uses.size());
4776 // SolveRecurse does all the work.
4777 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4778 CurRegs, VisitedRegs);
4779 if (Solution.empty()) {
4780 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4784 // Ok, we've now made all our decisions.
4785 DEBUG(dbgs() << "\n"
4786 "The chosen solution requires "; SolutionCost.print(dbgs());
4788 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4790 Uses[i].print(dbgs());
4793 Solution[i]->print(dbgs());
4797 assert(Solution.size() == Uses.size() && "Malformed solution!");
4800 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4801 /// we can go while still being dominated by the input positions. This helps
4802 /// canonicalize the insert position, which encourages sharing.
4803 BasicBlock::iterator
4804 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4805 const SmallVectorImpl<Instruction *> &Inputs)
4807 Instruction *Tentative = &*IP;
4809 bool AllDominate = true;
4810 Instruction *BetterPos = nullptr;
4811 // Don't bother attempting to insert before a catchswitch, their basic block
4812 // cannot have other non-PHI instructions.
4813 if (isa<CatchSwitchInst>(Tentative))
4816 for (Instruction *Inst : Inputs) {
4817 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4818 AllDominate = false;
4821 // Attempt to find an insert position in the middle of the block,
4822 // instead of at the end, so that it can be used for other expansions.
4823 if (Tentative->getParent() == Inst->getParent() &&
4824 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4825 BetterPos = &*std::next(BasicBlock::iterator(Inst));
4830 IP = BetterPos->getIterator();
4832 IP = Tentative->getIterator();
4834 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4835 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4838 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4839 if (!Rung) return IP;
4840 Rung = Rung->getIDom();
4841 if (!Rung) return IP;
4842 IDom = Rung->getBlock();
4844 // Don't climb into a loop though.
4845 const Loop *IDomLoop = LI.getLoopFor(IDom);
4846 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4847 if (IDomDepth <= IPLoopDepth &&
4848 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4852 Tentative = IDom->getTerminator();
4858 /// Determine an input position which will be dominated by the operands and
4859 /// which will dominate the result.
4860 BasicBlock::iterator
4861 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4864 SCEVExpander &Rewriter) const {
4865 // Collect some instructions which must be dominated by the
4866 // expanding replacement. These must be dominated by any operands that
4867 // will be required in the expansion.
4868 SmallVector<Instruction *, 4> Inputs;
4869 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4870 Inputs.push_back(I);
4871 if (LU.Kind == LSRUse::ICmpZero)
4872 if (Instruction *I =
4873 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4874 Inputs.push_back(I);
4875 if (LF.PostIncLoops.count(L)) {
4876 if (LF.isUseFullyOutsideLoop(L))
4877 Inputs.push_back(L->getLoopLatch()->getTerminator());
4879 Inputs.push_back(IVIncInsertPos);
4881 // The expansion must also be dominated by the increment positions of any
4882 // loops it for which it is using post-inc mode.
4883 for (const Loop *PIL : LF.PostIncLoops) {
4884 if (PIL == L) continue;
4886 // Be dominated by the loop exit.
4887 SmallVector<BasicBlock *, 4> ExitingBlocks;
4888 PIL->getExitingBlocks(ExitingBlocks);
4889 if (!ExitingBlocks.empty()) {
4890 BasicBlock *BB = ExitingBlocks[0];
4891 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4892 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4893 Inputs.push_back(BB->getTerminator());
4897 assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
4898 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4899 "Insertion point must be a normal instruction");
4901 // Then, climb up the immediate dominator tree as far as we can go while
4902 // still being dominated by the input positions.
4903 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4905 // Don't insert instructions before PHI nodes.
4906 while (isa<PHINode>(IP)) ++IP;
4908 // Ignore landingpad instructions.
4909 while (IP->isEHPad()) ++IP;
4911 // Ignore debug intrinsics.
4912 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4914 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4915 // IP consistent across expansions and allows the previously inserted
4916 // instructions to be reused by subsequent expansion.
4917 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
4923 /// Emit instructions for the leading candidate expression for this LSRUse (this
4924 /// is called "expanding").
4925 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
4926 const Formula &F, BasicBlock::iterator IP,
4927 SCEVExpander &Rewriter,
4928 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
4929 if (LU.RigidFormula)
4930 return LF.OperandValToReplace;
4932 // Determine an input position which will be dominated by the operands and
4933 // which will dominate the result.
4934 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4935 Rewriter.setInsertPoint(&*IP);
4937 // Inform the Rewriter if we have a post-increment use, so that it can
4938 // perform an advantageous expansion.
4939 Rewriter.setPostInc(LF.PostIncLoops);
4941 // This is the type that the user actually needs.
4942 Type *OpTy = LF.OperandValToReplace->getType();
4943 // This will be the type that we'll initially expand to.
4944 Type *Ty = F.getType();
4946 // No type known; just expand directly to the ultimate type.
4948 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4949 // Expand directly to the ultimate type if it's the right size.
4951 // This is the type to do integer arithmetic in.
4952 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4954 // Build up a list of operands to add together to form the full base.
4955 SmallVector<const SCEV *, 8> Ops;
4957 // Expand the BaseRegs portion.
4958 for (const SCEV *Reg : F.BaseRegs) {
4959 assert(!Reg->isZero() && "Zero allocated in a base register!");
4961 // If we're expanding for a post-inc user, make the post-inc adjustment.
4962 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
4963 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
4966 // Expand the ScaledReg portion.
4967 Value *ICmpScaledV = nullptr;
4969 const SCEV *ScaledS = F.ScaledReg;
4971 // If we're expanding for a post-inc user, make the post-inc adjustment.
4972 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4973 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
4975 if (LU.Kind == LSRUse::ICmpZero) {
4976 // Expand ScaleReg as if it was part of the base regs.
4979 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
4981 // An interesting way of "folding" with an icmp is to use a negated
4982 // scale, which we'll implement by inserting it into the other operand
4984 assert(F.Scale == -1 &&
4985 "The only scale supported by ICmpZero uses is -1!");
4986 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
4989 // Otherwise just expand the scaled register and an explicit scale,
4990 // which is expected to be matched as part of the address.
4992 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4993 // Unless the addressing mode will not be folded.
4994 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4995 isAMCompletelyFolded(TTI, LU, F)) {
4996 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4998 Ops.push_back(SE.getUnknown(FullV));
5000 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5003 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5004 Ops.push_back(ScaledS);
5008 // Expand the GV portion.
5010 // Flush the operand list to suppress SCEVExpander hoisting.
5012 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5014 Ops.push_back(SE.getUnknown(FullV));
5016 Ops.push_back(SE.getUnknown(F.BaseGV));
5019 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5020 // unfolded offsets. LSR assumes they both live next to their uses.
5022 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5024 Ops.push_back(SE.getUnknown(FullV));
5027 // Expand the immediate portion.
5028 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5030 if (LU.Kind == LSRUse::ICmpZero) {
5031 // The other interesting way of "folding" with an ICmpZero is to use a
5032 // negated immediate.
5034 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5036 Ops.push_back(SE.getUnknown(ICmpScaledV));
5037 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5040 // Just add the immediate values. These again are expected to be matched
5041 // as part of the address.
5042 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5046 // Expand the unfolded offset portion.
5047 int64_t UnfoldedOffset = F.UnfoldedOffset;
5048 if (UnfoldedOffset != 0) {
5049 // Just add the immediate values.
5050 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5054 // Emit instructions summing all the operands.
5055 const SCEV *FullS = Ops.empty() ?
5056 SE.getConstant(IntTy, 0) :
5058 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5060 // We're done expanding now, so reset the rewriter.
5061 Rewriter.clearPostInc();
5063 // An ICmpZero Formula represents an ICmp which we're handling as a
5064 // comparison against zero. Now that we've expanded an expression for that
5065 // form, update the ICmp's other operand.
5066 if (LU.Kind == LSRUse::ICmpZero) {
5067 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5068 DeadInsts.emplace_back(CI->getOperand(1));
5069 assert(!F.BaseGV && "ICmp does not support folding a global value and "
5070 "a scale at the same time!");
5071 if (F.Scale == -1) {
5072 if (ICmpScaledV->getType() != OpTy) {
5074 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5076 ICmpScaledV, OpTy, "tmp", CI);
5079 CI->setOperand(1, ICmpScaledV);
5081 // A scale of 1 means that the scale has been expanded as part of the
5083 assert((F.Scale == 0 || F.Scale == 1) &&
5084 "ICmp does not support folding a global value and "
5085 "a scale at the same time!");
5086 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5088 if (C->getType() != OpTy)
5089 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5093 CI->setOperand(1, C);
5100 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5101 /// effectively happens in their predecessor blocks, so the expression may need
5102 /// to be expanded in multiple places.
5103 void LSRInstance::RewriteForPHI(
5104 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5105 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5106 DenseMap<BasicBlock *, Value *> Inserted;
5107 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5108 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5109 BasicBlock *BB = PN->getIncomingBlock(i);
5111 // If this is a critical edge, split the edge so that we do not insert
5112 // the code on all predecessor/successor paths. We do this unless this
5113 // is the canonical backedge for this loop, which complicates post-inc
5115 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5116 !isa<IndirectBrInst>(BB->getTerminator()) &&
5117 !isa<CatchSwitchInst>(BB->getTerminator())) {
5118 BasicBlock *Parent = PN->getParent();
5119 Loop *PNLoop = LI.getLoopFor(Parent);
5120 if (!PNLoop || Parent != PNLoop->getHeader()) {
5121 // Split the critical edge.
5122 BasicBlock *NewBB = nullptr;
5123 if (!Parent->isLandingPad()) {
5124 NewBB = SplitCriticalEdge(BB, Parent,
5125 CriticalEdgeSplittingOptions(&DT, &LI)
5126 .setMergeIdenticalEdges()
5127 .setDontDeleteUselessPHIs());
5129 SmallVector<BasicBlock*, 2> NewBBs;
5130 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5133 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5134 // phi predecessors are identical. The simple thing to do is skip
5135 // splitting in this case rather than complicate the API.
5137 // If PN is outside of the loop and BB is in the loop, we want to
5138 // move the block to be immediately before the PHI block, not
5139 // immediately after BB.
5140 if (L->contains(BB) && !L->contains(PN))
5141 NewBB->moveBefore(PN->getParent());
5143 // Splitting the edge can reduce the number of PHI entries we have.
5144 e = PN->getNumIncomingValues();
5146 i = PN->getBasicBlockIndex(BB);
5151 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5152 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5154 PN->setIncomingValue(i, Pair.first->second);
5156 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5157 Rewriter, DeadInsts);
5159 // If this is reuse-by-noop-cast, insert the noop cast.
5160 Type *OpTy = LF.OperandValToReplace->getType();
5161 if (FullV->getType() != OpTy)
5163 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5165 FullV, LF.OperandValToReplace->getType(),
5166 "tmp", BB->getTerminator());
5168 PN->setIncomingValue(i, FullV);
5169 Pair.first->second = FullV;
5174 /// Emit instructions for the leading candidate expression for this LSRUse (this
5175 /// is called "expanding"), and update the UserInst to reference the newly
5177 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5178 const Formula &F, SCEVExpander &Rewriter,
5179 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5180 // First, find an insertion point that dominates UserInst. For PHI nodes,
5181 // find the nearest block which dominates all the relevant uses.
5182 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5183 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5186 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5188 // If this is reuse-by-noop-cast, insert the noop cast.
5189 Type *OpTy = LF.OperandValToReplace->getType();
5190 if (FullV->getType() != OpTy) {
5192 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5193 FullV, OpTy, "tmp", LF.UserInst);
5197 // Update the user. ICmpZero is handled specially here (for now) because
5198 // Expand may have updated one of the operands of the icmp already, and
5199 // its new value may happen to be equal to LF.OperandValToReplace, in
5200 // which case doing replaceUsesOfWith leads to replacing both operands
5201 // with the same value. TODO: Reorganize this.
5202 if (LU.Kind == LSRUse::ICmpZero)
5203 LF.UserInst->setOperand(0, FullV);
5205 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5208 DeadInsts.emplace_back(LF.OperandValToReplace);
5211 /// Rewrite all the fixup locations with new values, following the chosen
5213 void LSRInstance::ImplementSolution(
5214 const SmallVectorImpl<const Formula *> &Solution) {
5215 // Keep track of instructions we may have made dead, so that
5216 // we can remove them after we are done working.
5217 SmallVector<WeakTrackingVH, 16> DeadInsts;
5219 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
5222 Rewriter.setDebugType(DEBUG_TYPE);
5224 Rewriter.disableCanonicalMode();
5225 Rewriter.enableLSRMode();
5226 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5228 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5229 for (const IVChain &Chain : IVChainVec) {
5230 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5231 Rewriter.setChainedPhi(PN);
5234 // Expand the new value definitions and update the users.
5235 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5236 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5237 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5241 for (const IVChain &Chain : IVChainVec) {
5242 GenerateIVChain(Chain, Rewriter, DeadInsts);
5245 // Clean up after ourselves. This must be done before deleting any
5249 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5252 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5253 DominatorTree &DT, LoopInfo &LI,
5254 const TargetTransformInfo &TTI)
5255 : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L) {
5256 // If LoopSimplify form is not available, stay out of trouble.
5257 if (!L->isLoopSimplifyForm())
5260 // If there's no interesting work to be done, bail early.
5261 if (IU.empty()) return;
5263 // If there's too much analysis to be done, bail early. We won't be able to
5264 // model the problem anyway.
5265 unsigned NumUsers = 0;
5266 for (const IVStrideUse &U : IU) {
5267 if (++NumUsers > MaxIVUsers) {
5269 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n");
5272 // Bail out if we have a PHI on an EHPad that gets a value from a
5273 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5274 // no good place to stick any instructions.
5275 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5276 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5277 if (isa<FuncletPadInst>(FirstNonPHI) ||
5278 isa<CatchSwitchInst>(FirstNonPHI))
5279 for (BasicBlock *PredBB : PN->blocks())
5280 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5286 // All dominating loops must have preheaders, or SCEVExpander may not be able
5287 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5289 // IVUsers analysis should only create users that are dominated by simple loop
5290 // headers. Since this loop should dominate all of its users, its user list
5291 // should be empty if this loop itself is not within a simple loop nest.
5292 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5293 Rung; Rung = Rung->getIDom()) {
5294 BasicBlock *BB = Rung->getBlock();
5295 const Loop *DomLoop = LI.getLoopFor(BB);
5296 if (DomLoop && DomLoop->getHeader() == BB) {
5297 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
5302 DEBUG(dbgs() << "\nLSR on loop ";
5303 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
5306 // First, perform some low-level loop optimizations.
5308 OptimizeLoopTermCond();
5310 // If loop preparation eliminates all interesting IV users, bail.
5311 if (IU.empty()) return;
5313 // Skip nested loops until we can model them better with formulae.
5315 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
5319 // Start collecting data and preparing for the solver.
5321 CollectInterestingTypesAndFactors();
5322 CollectFixupsAndInitialFormulae();
5323 CollectLoopInvariantFixupsAndFormulae();
5325 assert(!Uses.empty() && "IVUsers reported at least one use");
5326 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
5327 print_uses(dbgs()));
5329 // Now use the reuse data to generate a bunch of interesting ways
5330 // to formulate the values needed for the uses.
5331 GenerateAllReuseFormulae();
5333 FilterOutUndesirableDedicatedRegisters();
5334 NarrowSearchSpaceUsingHeuristics();
5336 SmallVector<const Formula *, 8> Solution;
5339 // Release memory that is no longer needed.
5344 if (Solution.empty())
5348 // Formulae should be legal.
5349 for (const LSRUse &LU : Uses) {
5350 for (const Formula &F : LU.Formulae)
5351 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
5352 F) && "Illegal formula generated!");
5356 // Now that we've decided what we want, make it so.
5357 ImplementSolution(Solution);
5360 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5361 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5362 if (Factors.empty() && Types.empty()) return;
5364 OS << "LSR has identified the following interesting factors and types: ";
5367 for (int64_t Factor : Factors) {
5368 if (!First) OS << ", ";
5370 OS << '*' << Factor;
5373 for (Type *Ty : Types) {
5374 if (!First) OS << ", ";
5376 OS << '(' << *Ty << ')';
5381 void LSRInstance::print_fixups(raw_ostream &OS) const {
5382 OS << "LSR is examining the following fixup sites:\n";
5383 for (const LSRUse &LU : Uses)
5384 for (const LSRFixup &LF : LU.Fixups) {
5391 void LSRInstance::print_uses(raw_ostream &OS) const {
5392 OS << "LSR is examining the following uses:\n";
5393 for (const LSRUse &LU : Uses) {
5397 for (const Formula &F : LU.Formulae) {
5405 void LSRInstance::print(raw_ostream &OS) const {
5406 print_factors_and_types(OS);
5411 LLVM_DUMP_METHOD void LSRInstance::dump() const {
5412 print(errs()); errs() << '\n';
5418 class LoopStrengthReduce : public LoopPass {
5420 static char ID; // Pass ID, replacement for typeid
5422 LoopStrengthReduce();
5425 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5426 void getAnalysisUsage(AnalysisUsage &AU) const override;
5429 } // end anonymous namespace
5431 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5432 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5435 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5436 // We split critical edges, so we change the CFG. However, we do update
5437 // many analyses if they are around.
5438 AU.addPreservedID(LoopSimplifyID);
5440 AU.addRequired<LoopInfoWrapperPass>();
5441 AU.addPreserved<LoopInfoWrapperPass>();
5442 AU.addRequiredID(LoopSimplifyID);
5443 AU.addRequired<DominatorTreeWrapperPass>();
5444 AU.addPreserved<DominatorTreeWrapperPass>();
5445 AU.addRequired<ScalarEvolutionWrapperPass>();
5446 AU.addPreserved<ScalarEvolutionWrapperPass>();
5447 // Requiring LoopSimplify a second time here prevents IVUsers from running
5448 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5449 AU.addRequiredID(LoopSimplifyID);
5450 AU.addRequired<IVUsersWrapperPass>();
5451 AU.addPreserved<IVUsersWrapperPass>();
5452 AU.addRequired<TargetTransformInfoWrapperPass>();
5455 static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5456 DominatorTree &DT, LoopInfo &LI,
5457 const TargetTransformInfo &TTI) {
5458 bool Changed = false;
5460 // Run the main LSR transformation.
5461 Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5463 // Remove any extra phis created by processing inner loops.
5464 Changed |= DeleteDeadPHIs(L->getHeader());
5465 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5466 SmallVector<WeakTrackingVH, 16> DeadInsts;
5467 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5468 SCEVExpander Rewriter(SE, DL, "lsr");
5470 Rewriter.setDebugType(DEBUG_TYPE);
5472 unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5475 DeleteTriviallyDeadInstructions(DeadInsts);
5476 DeleteDeadPHIs(L->getHeader());
5482 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5486 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5487 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5488 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5489 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5490 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5491 *L->getHeader()->getParent());
5492 return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5495 PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
5496 LoopStandardAnalysisResults &AR,
5498 if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5499 AR.DT, AR.LI, AR.TTI))
5500 return PreservedAnalyses::all();
5502 return getLoopPassPreservedAnalyses();
5505 char LoopStrengthReduce::ID = 0;
5507 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5508 "Loop Strength Reduction", false, false)
5509 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5510 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5511 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5512 INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)
5513 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5514 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5515 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5516 "Loop Strength Reduction", false, false)
5518 Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }