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/Analysis/IVUsers.h"
69 #include "llvm/Analysis/LoopInfo.h"
70 #include "llvm/Analysis/LoopPass.h"
71 #include "llvm/Analysis/ScalarEvolution.h"
72 #include "llvm/Analysis/ScalarEvolutionExpander.h"
73 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
74 #include "llvm/Analysis/ScalarEvolutionNormalization.h"
75 #include "llvm/Analysis/TargetTransformInfo.h"
76 #include "llvm/IR/BasicBlock.h"
77 #include "llvm/IR/Constant.h"
78 #include "llvm/IR/Constants.h"
79 #include "llvm/IR/DerivedTypes.h"
80 #include "llvm/IR/Dominators.h"
81 #include "llvm/IR/GlobalValue.h"
82 #include "llvm/IR/IRBuilder.h"
83 #include "llvm/IR/Instruction.h"
84 #include "llvm/IR/Instructions.h"
85 #include "llvm/IR/IntrinsicInst.h"
86 #include "llvm/IR/Module.h"
87 #include "llvm/IR/OperandTraits.h"
88 #include "llvm/IR/Operator.h"
89 #include "llvm/IR/Type.h"
90 #include "llvm/IR/Value.h"
91 #include "llvm/IR/ValueHandle.h"
92 #include "llvm/Pass.h"
93 #include "llvm/Support/Casting.h"
94 #include "llvm/Support/CommandLine.h"
95 #include "llvm/Support/Compiler.h"
96 #include "llvm/Support/Debug.h"
97 #include "llvm/Support/ErrorHandling.h"
98 #include "llvm/Support/MathExtras.h"
99 #include "llvm/Support/raw_ostream.h"
100 #include "llvm/Transforms/Scalar.h"
101 #include "llvm/Transforms/Scalar/LoopPassManager.h"
102 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
103 #include "llvm/Transforms/Utils/Local.h"
114 using namespace llvm;
116 #define DEBUG_TYPE "loop-reduce"
118 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
119 /// bail out. This threshold is far beyond the number of users that LSR can
120 /// conceivably solve, so it should not affect generated code, but catches the
121 /// worst cases before LSR burns too much compile time and stack space.
122 static const unsigned MaxIVUsers = 200;
124 // Temporary flag to cleanup congruent phis after LSR phi expansion.
125 // It's currently disabled until we can determine whether it's truly useful or
126 // not. The flag should be removed after the v3.0 release.
127 // This is now needed for ivchains.
128 static cl::opt<bool> EnablePhiElim(
129 "enable-lsr-phielim", cl::Hidden, cl::init(true),
130 cl::desc("Enable LSR phi elimination"));
132 // The flag adds instruction count to solutions cost comparision.
133 static cl::opt<bool> InsnsCost(
134 "lsr-insns-cost", cl::Hidden, cl::init(false),
135 cl::desc("Add instruction count to a LSR cost model"));
137 // Flag to choose how to narrow complex lsr solution
138 static cl::opt<bool> LSRExpNarrow(
139 "lsr-exp-narrow", cl::Hidden, cl::init(false),
140 cl::desc("Narrow LSR complex solution using"
141 " expectation of registers number"));
143 // Flag to narrow search space by filtering non-optimal formulae with
144 // the same ScaledReg and Scale.
145 static cl::opt<bool> FilterSameScaledReg(
146 "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
147 cl::desc("Narrow LSR search space by filtering non-optimal formulae"
148 " with the same ScaledReg and Scale"));
151 // Stress test IV chain generation.
152 static cl::opt<bool> StressIVChain(
153 "stress-ivchain", cl::Hidden, cl::init(false),
154 cl::desc("Stress test LSR IV chains"));
156 static bool StressIVChain = false;
162 /// Used in situations where the accessed memory type is unknown.
163 static const unsigned UnknownAddressSpace = ~0u;
168 MemAccessTy() : MemTy(nullptr), AddrSpace(UnknownAddressSpace) {}
170 MemAccessTy(Type *Ty, unsigned AS) :
171 MemTy(Ty), AddrSpace(AS) {}
173 bool operator==(MemAccessTy Other) const {
174 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
177 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
179 static MemAccessTy getUnknown(LLVMContext &Ctx,
180 unsigned AS = UnknownAddressSpace) {
181 return MemAccessTy(Type::getVoidTy(Ctx), AS);
185 /// This class holds data which is used to order reuse candidates.
188 /// This represents the set of LSRUse indices which reference
189 /// a particular register.
190 SmallBitVector UsedByIndices;
192 void print(raw_ostream &OS) const;
196 } // end anonymous namespace
198 void RegSortData::print(raw_ostream &OS) const {
199 OS << "[NumUses=" << UsedByIndices.count() << ']';
202 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
203 LLVM_DUMP_METHOD void RegSortData::dump() const {
204 print(errs()); errs() << '\n';
210 /// Map register candidates to information about how they are used.
211 class RegUseTracker {
212 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
214 RegUsesTy RegUsesMap;
215 SmallVector<const SCEV *, 16> RegSequence;
218 void countRegister(const SCEV *Reg, size_t LUIdx);
219 void dropRegister(const SCEV *Reg, size_t LUIdx);
220 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
222 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
224 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
228 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
229 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
230 iterator begin() { return RegSequence.begin(); }
231 iterator end() { return RegSequence.end(); }
232 const_iterator begin() const { return RegSequence.begin(); }
233 const_iterator end() const { return RegSequence.end(); }
236 } // end anonymous namespace
239 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
240 std::pair<RegUsesTy::iterator, bool> Pair =
241 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
242 RegSortData &RSD = Pair.first->second;
244 RegSequence.push_back(Reg);
245 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
246 RSD.UsedByIndices.set(LUIdx);
250 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
251 RegUsesTy::iterator It = RegUsesMap.find(Reg);
252 assert(It != RegUsesMap.end());
253 RegSortData &RSD = It->second;
254 assert(RSD.UsedByIndices.size() > LUIdx);
255 RSD.UsedByIndices.reset(LUIdx);
259 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
260 assert(LUIdx <= LastLUIdx);
262 // Update RegUses. The data structure is not optimized for this purpose;
263 // we must iterate through it and update each of the bit vectors.
264 for (auto &Pair : RegUsesMap) {
265 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
266 if (LUIdx < UsedByIndices.size())
267 UsedByIndices[LUIdx] =
268 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
269 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
274 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
275 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
276 if (I == RegUsesMap.end())
278 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
279 int i = UsedByIndices.find_first();
280 if (i == -1) return false;
281 if ((size_t)i != LUIdx) return true;
282 return UsedByIndices.find_next(i) != -1;
285 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
286 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
287 assert(I != RegUsesMap.end() && "Unknown register!");
288 return I->second.UsedByIndices;
291 void RegUseTracker::clear() {
298 /// This class holds information that describes a formula for computing
299 /// satisfying a use. It may include broken-out immediates and scaled registers.
301 /// Global base address used for complex addressing.
304 /// Base offset for complex addressing.
307 /// Whether any complex addressing has a base register.
310 /// The scale of any complex addressing.
313 /// The list of "base" registers for this use. When this is non-empty. The
314 /// canonical representation of a formula is
315 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
316 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
317 /// 3. The reg containing recurrent expr related with currect loop in the
318 /// formula should be put in the ScaledReg.
319 /// #1 enforces that the scaled register is always used when at least two
320 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
321 /// #2 enforces that 1 * reg is reg.
322 /// #3 ensures invariant regs with respect to current loop can be combined
323 /// together in LSR codegen.
324 /// This invariant can be temporarly broken while building a formula.
325 /// However, every formula inserted into the LSRInstance must be in canonical
327 SmallVector<const SCEV *, 4> BaseRegs;
329 /// The 'scaled' register for this use. This should be non-null when Scale is
331 const SCEV *ScaledReg;
333 /// An additional constant offset which added near the use. This requires a
334 /// temporary register, but the offset itself can live in an add immediate
335 /// field rather than a register.
336 int64_t UnfoldedOffset;
339 : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
340 ScaledReg(nullptr), UnfoldedOffset(0) {}
342 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
344 bool isCanonical(const Loop &L) const;
346 void canonicalize(const Loop &L);
350 bool hasZeroEnd() const;
352 size_t getNumRegs() const;
353 Type *getType() const;
355 void deleteBaseReg(const SCEV *&S);
357 bool referencesReg(const SCEV *S) const;
358 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
359 const RegUseTracker &RegUses) const;
361 void print(raw_ostream &OS) const;
365 } // end anonymous namespace
367 /// Recursion helper for initialMatch.
368 static void DoInitialMatch(const SCEV *S, Loop *L,
369 SmallVectorImpl<const SCEV *> &Good,
370 SmallVectorImpl<const SCEV *> &Bad,
371 ScalarEvolution &SE) {
372 // Collect expressions which properly dominate the loop header.
373 if (SE.properlyDominates(S, L->getHeader())) {
378 // Look at add operands.
379 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
380 for (const SCEV *S : Add->operands())
381 DoInitialMatch(S, L, Good, Bad, SE);
385 // Look at addrec operands.
386 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
387 if (!AR->getStart()->isZero() && AR->isAffine()) {
388 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
389 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
390 AR->getStepRecurrence(SE),
391 // FIXME: AR->getNoWrapFlags()
392 AR->getLoop(), SCEV::FlagAnyWrap),
397 // Handle a multiplication by -1 (negation) if it didn't fold.
398 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
399 if (Mul->getOperand(0)->isAllOnesValue()) {
400 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
401 const SCEV *NewMul = SE.getMulExpr(Ops);
403 SmallVector<const SCEV *, 4> MyGood;
404 SmallVector<const SCEV *, 4> MyBad;
405 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
406 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
407 SE.getEffectiveSCEVType(NewMul->getType())));
408 for (const SCEV *S : MyGood)
409 Good.push_back(SE.getMulExpr(NegOne, S));
410 for (const SCEV *S : MyBad)
411 Bad.push_back(SE.getMulExpr(NegOne, S));
415 // Ok, we can't do anything interesting. Just stuff the whole thing into a
416 // register and hope for the best.
420 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
421 /// all loop-invariant and loop-computable values in a single base register.
422 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
423 SmallVector<const SCEV *, 4> Good;
424 SmallVector<const SCEV *, 4> Bad;
425 DoInitialMatch(S, L, Good, Bad, SE);
427 const SCEV *Sum = SE.getAddExpr(Good);
429 BaseRegs.push_back(Sum);
433 const SCEV *Sum = SE.getAddExpr(Bad);
435 BaseRegs.push_back(Sum);
441 /// \brief Check whether or not this formula statisfies the canonical
443 /// \see Formula::BaseRegs.
444 bool Formula::isCanonical(const Loop &L) const {
446 return BaseRegs.size() <= 1;
451 if (Scale == 1 && BaseRegs.empty())
454 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
455 if (SAR && SAR->getLoop() == &L)
458 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
459 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
460 // loop, we want to swap the reg in BaseRegs with ScaledReg.
462 find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
463 return isa<const SCEVAddRecExpr>(S) &&
464 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
466 return I == BaseRegs.end();
469 /// \brief Helper method to morph a formula into its canonical representation.
470 /// \see Formula::BaseRegs.
471 /// Every formula having more than one base register, must use the ScaledReg
472 /// field. Otherwise, we would have to do special cases everywhere in LSR
473 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
474 /// On the other hand, 1*reg should be canonicalized into reg.
475 void Formula::canonicalize(const Loop &L) {
478 // So far we did not need this case. This is easy to implement but it is
479 // useless to maintain dead code. Beside it could hurt compile time.
480 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
482 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
484 ScaledReg = BaseRegs.back();
489 // If ScaledReg is an invariant with respect to L, find the reg from
490 // BaseRegs containing the recurrent expr related with Loop L. Swap the
491 // reg with ScaledReg.
492 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
493 if (!SAR || SAR->getLoop() != &L) {
494 auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
496 return isa<const SCEVAddRecExpr>(S) &&
497 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
499 if (I != BaseRegs.end())
500 std::swap(ScaledReg, *I);
504 /// \brief Get rid of the scale in the formula.
505 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
506 /// \return true if it was possible to get rid of the scale, false otherwise.
507 /// \note After this operation the formula may not be in the canonical form.
508 bool Formula::unscale() {
512 BaseRegs.push_back(ScaledReg);
517 bool Formula::hasZeroEnd() const {
518 if (UnfoldedOffset || BaseOffset)
520 if (BaseRegs.size() != 1 || ScaledReg)
525 /// Return the total number of register operands used by this formula. This does
526 /// not include register uses implied by non-constant addrec strides.
527 size_t Formula::getNumRegs() const {
528 return !!ScaledReg + BaseRegs.size();
531 /// Return the type of this formula, if it has one, or null otherwise. This type
532 /// is meaningless except for the bit size.
533 Type *Formula::getType() const {
534 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
535 ScaledReg ? ScaledReg->getType() :
536 BaseGV ? BaseGV->getType() :
540 /// Delete the given base reg from the BaseRegs list.
541 void Formula::deleteBaseReg(const SCEV *&S) {
542 if (&S != &BaseRegs.back())
543 std::swap(S, BaseRegs.back());
547 /// Test if this formula references the given register.
548 bool Formula::referencesReg(const SCEV *S) const {
549 return S == ScaledReg || is_contained(BaseRegs, S);
552 /// Test whether this formula uses registers which are used by uses other than
553 /// the use with the given index.
554 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
555 const RegUseTracker &RegUses) const {
557 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
559 for (const SCEV *BaseReg : BaseRegs)
560 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
565 void Formula::print(raw_ostream &OS) const {
568 if (!First) OS << " + "; else First = false;
569 BaseGV->printAsOperand(OS, /*PrintType=*/false);
571 if (BaseOffset != 0) {
572 if (!First) OS << " + "; else First = false;
575 for (const SCEV *BaseReg : BaseRegs) {
576 if (!First) OS << " + "; else First = false;
577 OS << "reg(" << *BaseReg << ')';
579 if (HasBaseReg && BaseRegs.empty()) {
580 if (!First) OS << " + "; else First = false;
581 OS << "**error: HasBaseReg**";
582 } else if (!HasBaseReg && !BaseRegs.empty()) {
583 if (!First) OS << " + "; else First = false;
584 OS << "**error: !HasBaseReg**";
587 if (!First) OS << " + "; else First = false;
588 OS << Scale << "*reg(";
595 if (UnfoldedOffset != 0) {
596 if (!First) OS << " + ";
597 OS << "imm(" << UnfoldedOffset << ')';
601 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
602 LLVM_DUMP_METHOD void Formula::dump() const {
603 print(errs()); errs() << '\n';
607 /// Return true if the given addrec can be sign-extended without changing its
609 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
611 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
612 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
615 /// Return true if the given add can be sign-extended without changing its
617 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
619 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
620 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
623 /// Return true if the given mul can be sign-extended without changing its
625 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
627 IntegerType::get(SE.getContext(),
628 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
629 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
632 /// Return an expression for LHS /s RHS, if it can be determined and if the
633 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
634 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
635 /// the multiplication may overflow, which is useful when the result will be
636 /// used in a context where the most significant bits are ignored.
637 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
639 bool IgnoreSignificantBits = false) {
640 // Handle the trivial case, which works for any SCEV type.
642 return SE.getConstant(LHS->getType(), 1);
644 // Handle a few RHS special cases.
645 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
647 const APInt &RA = RC->getAPInt();
648 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
650 if (RA.isAllOnesValue())
651 return SE.getMulExpr(LHS, RC);
652 // Handle x /s 1 as x.
657 // Check for a division of a constant by a constant.
658 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
661 const APInt &LA = C->getAPInt();
662 const APInt &RA = RC->getAPInt();
663 if (LA.srem(RA) != 0)
665 return SE.getConstant(LA.sdiv(RA));
668 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
669 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
670 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
671 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
672 IgnoreSignificantBits);
673 if (!Step) return nullptr;
674 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
675 IgnoreSignificantBits);
676 if (!Start) return nullptr;
677 // FlagNW is independent of the start value, step direction, and is
678 // preserved with smaller magnitude steps.
679 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
680 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
685 // Distribute the sdiv over add operands, if the add doesn't overflow.
686 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
687 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
688 SmallVector<const SCEV *, 8> Ops;
689 for (const SCEV *S : Add->operands()) {
690 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
691 if (!Op) return nullptr;
694 return SE.getAddExpr(Ops);
699 // Check for a multiply operand that we can pull RHS out of.
700 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
701 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
702 SmallVector<const SCEV *, 4> Ops;
704 for (const SCEV *S : Mul->operands()) {
706 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
707 IgnoreSignificantBits)) {
713 return Found ? SE.getMulExpr(Ops) : nullptr;
718 // Otherwise we don't know.
722 /// If S involves the addition of a constant integer value, return that integer
723 /// value, and mutate S to point to a new SCEV with that value excluded.
724 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
725 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
726 if (C->getAPInt().getMinSignedBits() <= 64) {
727 S = SE.getConstant(C->getType(), 0);
728 return C->getValue()->getSExtValue();
730 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
731 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
732 int64_t Result = ExtractImmediate(NewOps.front(), SE);
734 S = SE.getAddExpr(NewOps);
736 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
737 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
738 int64_t Result = ExtractImmediate(NewOps.front(), SE);
740 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
741 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
748 /// If S involves the addition of a GlobalValue address, return that symbol, and
749 /// mutate S to point to a new SCEV with that value excluded.
750 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
751 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
752 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
753 S = SE.getConstant(GV->getType(), 0);
756 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
757 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
758 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
760 S = SE.getAddExpr(NewOps);
762 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
763 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
764 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
766 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
767 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
774 /// Returns true if the specified instruction is using the specified value as an
776 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
777 bool isAddress = isa<LoadInst>(Inst);
778 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
779 if (SI->getPointerOperand() == OperandVal)
781 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
782 // Addressing modes can also be folded into prefetches and a variety
784 switch (II->getIntrinsicID()) {
786 case Intrinsic::prefetch:
787 if (II->getArgOperand(0) == OperandVal)
791 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
792 if (RMW->getPointerOperand() == OperandVal)
794 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
795 if (CmpX->getPointerOperand() == OperandVal)
801 /// Return the type of the memory being accessed.
802 static MemAccessTy getAccessType(const Instruction *Inst) {
803 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
804 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
805 AccessTy.MemTy = SI->getOperand(0)->getType();
806 AccessTy.AddrSpace = SI->getPointerAddressSpace();
807 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
808 AccessTy.AddrSpace = LI->getPointerAddressSpace();
809 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
810 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
811 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
812 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
815 // All pointers have the same requirements, so canonicalize them to an
816 // arbitrary pointer type to minimize variation.
817 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
818 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
819 PTy->getAddressSpace());
824 /// Return true if this AddRec is already a phi in its loop.
825 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
826 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
827 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
828 if (SE.isSCEVable(PN->getType()) &&
829 (SE.getEffectiveSCEVType(PN->getType()) ==
830 SE.getEffectiveSCEVType(AR->getType())) &&
831 SE.getSCEV(PN) == AR)
837 /// Check if expanding this expression is likely to incur significant cost. This
838 /// is tricky because SCEV doesn't track which expressions are actually computed
839 /// by the current IR.
841 /// We currently allow expansion of IV increments that involve adds,
842 /// multiplication by constants, and AddRecs from existing phis.
844 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
845 /// obvious multiple of the UDivExpr.
846 static bool isHighCostExpansion(const SCEV *S,
847 SmallPtrSetImpl<const SCEV*> &Processed,
848 ScalarEvolution &SE) {
849 // Zero/One operand expressions
850 switch (S->getSCEVType()) {
855 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
858 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
861 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
865 if (!Processed.insert(S).second)
868 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
869 for (const SCEV *S : Add->operands()) {
870 if (isHighCostExpansion(S, Processed, SE))
876 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
877 if (Mul->getNumOperands() == 2) {
878 // Multiplication by a constant is ok
879 if (isa<SCEVConstant>(Mul->getOperand(0)))
880 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
882 // If we have the value of one operand, check if an existing
883 // multiplication already generates this expression.
884 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
885 Value *UVal = U->getValue();
886 for (User *UR : UVal->users()) {
887 // If U is a constant, it may be used by a ConstantExpr.
888 Instruction *UI = dyn_cast<Instruction>(UR);
889 if (UI && UI->getOpcode() == Instruction::Mul &&
890 SE.isSCEVable(UI->getType())) {
891 return SE.getSCEV(UI) == Mul;
898 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
899 if (isExistingPhi(AR, SE))
903 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
907 /// If any of the instructions is the specified set are trivially dead, delete
908 /// them and see if this makes any of their operands subsequently dead.
910 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
911 bool Changed = false;
913 while (!DeadInsts.empty()) {
914 Value *V = DeadInsts.pop_back_val();
915 Instruction *I = dyn_cast_or_null<Instruction>(V);
917 if (!I || !isInstructionTriviallyDead(I))
920 for (Use &O : I->operands())
921 if (Instruction *U = dyn_cast<Instruction>(O)) {
924 DeadInsts.emplace_back(U);
927 I->eraseFromParent();
938 } // end anonymous namespace
940 /// \brief Check if the addressing mode defined by \p F is completely
941 /// folded in \p LU at isel time.
942 /// This includes address-mode folding and special icmp tricks.
943 /// This function returns true if \p LU can accommodate what \p F
944 /// defines and up to 1 base + 1 scaled + offset.
945 /// In other words, if \p F has several base registers, this function may
946 /// still return true. Therefore, users still need to account for
947 /// additional base registers and/or unfolded offsets to derive an
948 /// accurate cost model.
949 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
950 const LSRUse &LU, const Formula &F);
951 // Get the cost of the scaling factor used in F for LU.
952 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
953 const LSRUse &LU, const Formula &F,
958 /// This class is used to measure and compare candidate formulae.
960 TargetTransformInfo::LSRCost C;
974 bool isLess(Cost &Other, const TargetTransformInfo &TTI);
979 // Once any of the metrics loses, they must all remain losers.
981 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
982 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
983 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
984 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
989 assert(isValid() && "invalid cost");
990 return C.NumRegs == ~0u;
993 void RateFormula(const TargetTransformInfo &TTI,
995 SmallPtrSetImpl<const SCEV *> &Regs,
996 const DenseSet<const SCEV *> &VisitedRegs,
998 ScalarEvolution &SE, DominatorTree &DT,
1000 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1002 void print(raw_ostream &OS) const;
1006 void RateRegister(const SCEV *Reg,
1007 SmallPtrSetImpl<const SCEV *> &Regs,
1009 ScalarEvolution &SE, DominatorTree &DT);
1010 void RatePrimaryRegister(const SCEV *Reg,
1011 SmallPtrSetImpl<const SCEV *> &Regs,
1013 ScalarEvolution &SE, DominatorTree &DT,
1014 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1017 /// An operand value in an instruction which is to be replaced with some
1018 /// equivalent, possibly strength-reduced, replacement.
1020 /// The instruction which will be updated.
1021 Instruction *UserInst;
1023 /// The operand of the instruction which will be replaced. The operand may be
1024 /// used more than once; every instance will be replaced.
1025 Value *OperandValToReplace;
1027 /// If this user is to use the post-incremented value of an induction
1028 /// variable, this variable is non-null and holds the loop associated with the
1029 /// induction variable.
1030 PostIncLoopSet PostIncLoops;
1032 /// A constant offset to be added to the LSRUse expression. This allows
1033 /// multiple fixups to share the same LSRUse with different offsets, for
1034 /// example in an unrolled loop.
1037 bool isUseFullyOutsideLoop(const Loop *L) const;
1041 void print(raw_ostream &OS) const;
1045 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1046 /// SmallVectors of const SCEV*.
1047 struct UniquifierDenseMapInfo {
1048 static SmallVector<const SCEV *, 4> getEmptyKey() {
1049 SmallVector<const SCEV *, 4> V;
1050 V.push_back(reinterpret_cast<const SCEV *>(-1));
1054 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1055 SmallVector<const SCEV *, 4> V;
1056 V.push_back(reinterpret_cast<const SCEV *>(-2));
1060 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1061 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1064 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1065 const SmallVector<const SCEV *, 4> &RHS) {
1070 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1071 /// as uses invented by LSR itself. It includes information about what kinds of
1072 /// things can be folded into the user, information about the user itself, and
1073 /// information about how the use may be satisfied. TODO: Represent multiple
1074 /// users of the same expression in common?
1076 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1079 /// An enum for a kind of use, indicating what types of scaled and immediate
1080 /// operands it might support.
1082 Basic, ///< A normal use, with no folding.
1083 Special, ///< A special case of basic, allowing -1 scales.
1084 Address, ///< An address use; folding according to TargetLowering
1085 ICmpZero ///< An equality icmp with both operands folded into one.
1086 // TODO: Add a generic icmp too?
1089 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1092 MemAccessTy AccessTy;
1094 /// The list of operands which are to be replaced.
1095 SmallVector<LSRFixup, 8> Fixups;
1097 /// Keep track of the min and max offsets of the fixups.
1101 /// This records whether all of the fixups using this LSRUse are outside of
1102 /// the loop, in which case some special-case heuristics may be used.
1103 bool AllFixupsOutsideLoop;
1105 /// RigidFormula is set to true to guarantee that this use will be associated
1106 /// with a single formula--the one that initially matched. Some SCEV
1107 /// expressions cannot be expanded. This allows LSR to consider the registers
1108 /// used by those expressions without the need to expand them later after
1109 /// changing the formula.
1112 /// This records the widest use type for any fixup using this
1113 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1114 /// fixup widths to be equivalent, because the narrower one may be relying on
1115 /// the implicit truncation to truncate away bogus bits.
1116 Type *WidestFixupType;
1118 /// A list of ways to build a value that can satisfy this user. After the
1119 /// list is populated, one of these is selected heuristically and used to
1120 /// formulate a replacement for OperandValToReplace in UserInst.
1121 SmallVector<Formula, 12> Formulae;
1123 /// The set of register candidates used by all formulae in this LSRUse.
1124 SmallPtrSet<const SCEV *, 4> Regs;
1126 LSRUse(KindType K, MemAccessTy AT)
1127 : Kind(K), AccessTy(AT), MinOffset(INT64_MAX), MaxOffset(INT64_MIN),
1128 AllFixupsOutsideLoop(true), RigidFormula(false),
1129 WidestFixupType(nullptr) {}
1131 LSRFixup &getNewFixup() {
1132 Fixups.push_back(LSRFixup());
1133 return Fixups.back();
1136 void pushFixup(LSRFixup &f) {
1137 Fixups.push_back(f);
1138 if (f.Offset > MaxOffset)
1139 MaxOffset = f.Offset;
1140 if (f.Offset < MinOffset)
1141 MinOffset = f.Offset;
1144 bool HasFormulaWithSameRegs(const Formula &F) const;
1145 float getNotSelectedProbability(const SCEV *Reg) const;
1146 bool InsertFormula(const Formula &F, const Loop &L);
1147 void DeleteFormula(Formula &F);
1148 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1150 void print(raw_ostream &OS) const;
1154 } // end anonymous namespace
1156 /// Tally up interesting quantities from the given register.
1157 void Cost::RateRegister(const SCEV *Reg,
1158 SmallPtrSetImpl<const SCEV *> &Regs,
1160 ScalarEvolution &SE, DominatorTree &DT) {
1161 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1162 // If this is an addrec for another loop, it should be an invariant
1163 // with respect to L since L is the innermost loop (at least
1164 // for now LSR only handles innermost loops).
1165 if (AR->getLoop() != L) {
1166 // If the AddRec exists, consider it's register free and leave it alone.
1167 if (isExistingPhi(AR, SE))
1170 // It is bad to allow LSR for current loop to add induction variables
1171 // for its sibling loops.
1172 if (!AR->getLoop()->contains(L)) {
1177 // Otherwise, it will be an invariant with respect to Loop L.
1181 C.AddRecCost += 1; /// TODO: This should be a function of the stride.
1183 // Add the step value register, if it needs one.
1184 // TODO: The non-affine case isn't precisely modeled here.
1185 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1186 if (!Regs.count(AR->getOperand(1))) {
1187 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
1195 // Rough heuristic; favor registers which don't require extra setup
1196 // instructions in the preheader.
1197 if (!isa<SCEVUnknown>(Reg) &&
1198 !isa<SCEVConstant>(Reg) &&
1199 !(isa<SCEVAddRecExpr>(Reg) &&
1200 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
1201 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
1204 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1205 SE.hasComputableLoopEvolution(Reg, L);
1208 /// Record this register in the set. If we haven't seen it before, rate
1209 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1210 /// one of those regs an instant loser.
1211 void Cost::RatePrimaryRegister(const SCEV *Reg,
1212 SmallPtrSetImpl<const SCEV *> &Regs,
1214 ScalarEvolution &SE, DominatorTree &DT,
1215 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1216 if (LoserRegs && LoserRegs->count(Reg)) {
1220 if (Regs.insert(Reg).second) {
1221 RateRegister(Reg, Regs, L, SE, DT);
1222 if (LoserRegs && isLoser())
1223 LoserRegs->insert(Reg);
1227 void Cost::RateFormula(const TargetTransformInfo &TTI,
1229 SmallPtrSetImpl<const SCEV *> &Regs,
1230 const DenseSet<const SCEV *> &VisitedRegs,
1232 ScalarEvolution &SE, DominatorTree &DT,
1234 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1235 assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1236 // Tally up the registers.
1237 unsigned PrevAddRecCost = C.AddRecCost;
1238 unsigned PrevNumRegs = C.NumRegs;
1239 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1240 if (const SCEV *ScaledReg = F.ScaledReg) {
1241 if (VisitedRegs.count(ScaledReg)) {
1245 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1249 for (const SCEV *BaseReg : F.BaseRegs) {
1250 if (VisitedRegs.count(BaseReg)) {
1254 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1259 // Determine how many (unfolded) adds we'll need inside the loop.
1260 size_t NumBaseParts = F.getNumRegs();
1261 if (NumBaseParts > 1)
1262 // Do not count the base and a possible second register if the target
1263 // allows to fold 2 registers.
1265 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1266 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1268 // Accumulate non-free scaling amounts.
1269 C.ScaleCost += getScalingFactorCost(TTI, LU, F, *L);
1271 // Tally up the non-zero immediates.
1272 for (const LSRFixup &Fixup : LU.Fixups) {
1273 int64_t O = Fixup.Offset;
1274 int64_t Offset = (uint64_t)O + F.BaseOffset;
1276 C.ImmCost += 64; // Handle symbolic values conservatively.
1277 // TODO: This should probably be the pointer size.
1278 else if (Offset != 0)
1279 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1281 // Check with target if this offset with this instruction is
1282 // specifically not supported.
1283 if ((isa<LoadInst>(Fixup.UserInst) || isa<StoreInst>(Fixup.UserInst)) &&
1284 !TTI.isFoldableMemAccessOffset(Fixup.UserInst, Offset))
1288 // If we don't count instruction cost exit here.
1290 assert(isValid() && "invalid cost");
1294 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1295 // additional instruction (at least fill).
1296 unsigned TTIRegNum = TTI.getNumberOfRegisters(false) - 1;
1297 if (C.NumRegs > TTIRegNum) {
1298 // Cost already exceeded TTIRegNum, then only newly added register can add
1299 // new instructions.
1300 if (PrevNumRegs > TTIRegNum)
1301 C.Insns += (C.NumRegs - PrevNumRegs);
1303 C.Insns += (C.NumRegs - TTIRegNum);
1306 // If ICmpZero formula ends with not 0, it could not be replaced by
1307 // just add or sub. We'll need to compare final result of AddRec.
1308 // That means we'll need an additional instruction.
1309 // For -10 + {0, +, 1}:
1315 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd())
1317 // Each new AddRec adds 1 instruction to calculation.
1318 C.Insns += (C.AddRecCost - PrevAddRecCost);
1320 // BaseAdds adds instructions for unfolded registers.
1321 if (LU.Kind != LSRUse::ICmpZero)
1322 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1323 assert(isValid() && "invalid cost");
1326 /// Set this cost to a losing value.
1332 C.NumBaseAdds = ~0u;
1338 /// Choose the lower cost.
1339 bool Cost::isLess(Cost &Other, const TargetTransformInfo &TTI) {
1340 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1341 C.Insns != Other.C.Insns)
1342 return C.Insns < Other.C.Insns;
1343 return TTI.isLSRCostLess(C, Other.C);
1346 void Cost::print(raw_ostream &OS) const {
1348 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1349 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1350 if (C.AddRecCost != 0)
1351 OS << ", with addrec cost " << C.AddRecCost;
1352 if (C.NumIVMuls != 0)
1353 OS << ", plus " << C.NumIVMuls << " IV mul"
1354 << (C.NumIVMuls == 1 ? "" : "s");
1355 if (C.NumBaseAdds != 0)
1356 OS << ", plus " << C.NumBaseAdds << " base add"
1357 << (C.NumBaseAdds == 1 ? "" : "s");
1358 if (C.ScaleCost != 0)
1359 OS << ", plus " << C.ScaleCost << " scale cost";
1361 OS << ", plus " << C.ImmCost << " imm cost";
1362 if (C.SetupCost != 0)
1363 OS << ", plus " << C.SetupCost << " setup cost";
1366 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1367 LLVM_DUMP_METHOD void Cost::dump() const {
1368 print(errs()); errs() << '\n';
1372 LSRFixup::LSRFixup()
1373 : UserInst(nullptr), OperandValToReplace(nullptr),
1376 /// Test whether this fixup always uses its value outside of the given loop.
1377 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1378 // PHI nodes use their value in their incoming blocks.
1379 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1380 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1381 if (PN->getIncomingValue(i) == OperandValToReplace &&
1382 L->contains(PN->getIncomingBlock(i)))
1387 return !L->contains(UserInst);
1390 void LSRFixup::print(raw_ostream &OS) const {
1392 // Store is common and interesting enough to be worth special-casing.
1393 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1395 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1396 } else if (UserInst->getType()->isVoidTy())
1397 OS << UserInst->getOpcodeName();
1399 UserInst->printAsOperand(OS, /*PrintType=*/false);
1401 OS << ", OperandValToReplace=";
1402 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1404 for (const Loop *PIL : PostIncLoops) {
1405 OS << ", PostIncLoop=";
1406 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1410 OS << ", Offset=" << Offset;
1413 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1414 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1415 print(errs()); errs() << '\n';
1419 /// Test whether this use as a formula which has the same registers as the given
1421 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1422 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1423 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1424 // Unstable sort by host order ok, because this is only used for uniquifying.
1425 std::sort(Key.begin(), Key.end());
1426 return Uniquifier.count(Key);
1429 /// The function returns a probability of selecting formula without Reg.
1430 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1432 for (const Formula &F : Formulae)
1433 if (F.referencesReg(Reg))
1435 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1438 /// If the given formula has not yet been inserted, add it to the list, and
1439 /// return true. Return false otherwise. The formula must be in canonical form.
1440 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1441 assert(F.isCanonical(L) && "Invalid canonical representation");
1443 if (!Formulae.empty() && RigidFormula)
1446 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1447 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1448 // Unstable sort by host order ok, because this is only used for uniquifying.
1449 std::sort(Key.begin(), Key.end());
1451 if (!Uniquifier.insert(Key).second)
1454 // Using a register to hold the value of 0 is not profitable.
1455 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1456 "Zero allocated in a scaled register!");
1458 for (const SCEV *BaseReg : F.BaseRegs)
1459 assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1462 // Add the formula to the list.
1463 Formulae.push_back(F);
1465 // Record registers now being used by this use.
1466 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1468 Regs.insert(F.ScaledReg);
1473 /// Remove the given formula from this use's list.
1474 void LSRUse::DeleteFormula(Formula &F) {
1475 if (&F != &Formulae.back())
1476 std::swap(F, Formulae.back());
1477 Formulae.pop_back();
1480 /// Recompute the Regs field, and update RegUses.
1481 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1482 // Now that we've filtered out some formulae, recompute the Regs set.
1483 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1485 for (const Formula &F : Formulae) {
1486 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1487 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1490 // Update the RegTracker.
1491 for (const SCEV *S : OldRegs)
1493 RegUses.dropRegister(S, LUIdx);
1496 void LSRUse::print(raw_ostream &OS) const {
1497 OS << "LSR Use: Kind=";
1499 case Basic: OS << "Basic"; break;
1500 case Special: OS << "Special"; break;
1501 case ICmpZero: OS << "ICmpZero"; break;
1503 OS << "Address of ";
1504 if (AccessTy.MemTy->isPointerTy())
1505 OS << "pointer"; // the full pointer type could be really verbose
1507 OS << *AccessTy.MemTy;
1510 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1513 OS << ", Offsets={";
1514 bool NeedComma = false;
1515 for (const LSRFixup &Fixup : Fixups) {
1516 if (NeedComma) OS << ',';
1522 if (AllFixupsOutsideLoop)
1523 OS << ", all-fixups-outside-loop";
1525 if (WidestFixupType)
1526 OS << ", widest fixup type: " << *WidestFixupType;
1529 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1530 LLVM_DUMP_METHOD void LSRUse::dump() const {
1531 print(errs()); errs() << '\n';
1535 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1536 LSRUse::KindType Kind, MemAccessTy AccessTy,
1537 GlobalValue *BaseGV, int64_t BaseOffset,
1538 bool HasBaseReg, int64_t Scale) {
1540 case LSRUse::Address:
1541 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1542 HasBaseReg, Scale, AccessTy.AddrSpace);
1544 case LSRUse::ICmpZero:
1545 // There's not even a target hook for querying whether it would be legal to
1546 // fold a GV into an ICmp.
1550 // ICmp only has two operands; don't allow more than two non-trivial parts.
1551 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1554 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1555 // putting the scaled register in the other operand of the icmp.
1556 if (Scale != 0 && Scale != -1)
1559 // If we have low-level target information, ask the target if it can fold an
1560 // integer immediate on an icmp.
1561 if (BaseOffset != 0) {
1563 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1564 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1565 // Offs is the ICmp immediate.
1567 // The cast does the right thing with INT64_MIN.
1568 BaseOffset = -(uint64_t)BaseOffset;
1569 return TTI.isLegalICmpImmediate(BaseOffset);
1572 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1576 // Only handle single-register values.
1577 return !BaseGV && Scale == 0 && BaseOffset == 0;
1579 case LSRUse::Special:
1580 // Special case Basic to handle -1 scales.
1581 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1584 llvm_unreachable("Invalid LSRUse Kind!");
1587 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1588 int64_t MinOffset, int64_t MaxOffset,
1589 LSRUse::KindType Kind, MemAccessTy AccessTy,
1590 GlobalValue *BaseGV, int64_t BaseOffset,
1591 bool HasBaseReg, int64_t Scale) {
1592 // Check for overflow.
1593 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1596 MinOffset = (uint64_t)BaseOffset + MinOffset;
1597 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1600 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1602 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1603 HasBaseReg, Scale) &&
1604 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1608 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1609 int64_t MinOffset, int64_t MaxOffset,
1610 LSRUse::KindType Kind, MemAccessTy AccessTy,
1611 const Formula &F, const Loop &L) {
1612 // For the purpose of isAMCompletelyFolded either having a canonical formula
1613 // or a scale not equal to zero is correct.
1614 // Problems may arise from non canonical formulae having a scale == 0.
1615 // Strictly speaking it would best to just rely on canonical formulae.
1616 // However, when we generate the scaled formulae, we first check that the
1617 // scaling factor is profitable before computing the actual ScaledReg for
1618 // compile time sake.
1619 assert((F.isCanonical(L) || F.Scale != 0));
1620 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1621 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1624 /// Test whether we know how to expand the current formula.
1625 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1626 int64_t MaxOffset, LSRUse::KindType Kind,
1627 MemAccessTy AccessTy, GlobalValue *BaseGV,
1628 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1629 // We know how to expand completely foldable formulae.
1630 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1631 BaseOffset, HasBaseReg, Scale) ||
1632 // Or formulae that use a base register produced by a sum of base
1635 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1636 BaseGV, BaseOffset, true, 0));
1639 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1640 int64_t MaxOffset, LSRUse::KindType Kind,
1641 MemAccessTy AccessTy, const Formula &F) {
1642 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1643 F.BaseOffset, F.HasBaseReg, F.Scale);
1646 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1647 const LSRUse &LU, const Formula &F) {
1648 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1649 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1653 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1654 const LSRUse &LU, const Formula &F,
1659 // If the use is not completely folded in that instruction, we will have to
1660 // pay an extra cost only for scale != 1.
1661 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1663 return F.Scale != 1;
1666 case LSRUse::Address: {
1667 // Check the scaling factor cost with both the min and max offsets.
1668 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1669 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1670 F.Scale, LU.AccessTy.AddrSpace);
1671 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1672 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1673 F.Scale, LU.AccessTy.AddrSpace);
1675 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1676 "Legal addressing mode has an illegal cost!");
1677 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1679 case LSRUse::ICmpZero:
1681 case LSRUse::Special:
1682 // The use is completely folded, i.e., everything is folded into the
1687 llvm_unreachable("Invalid LSRUse Kind!");
1690 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1691 LSRUse::KindType Kind, MemAccessTy AccessTy,
1692 GlobalValue *BaseGV, int64_t BaseOffset,
1694 // Fast-path: zero is always foldable.
1695 if (BaseOffset == 0 && !BaseGV) return true;
1697 // Conservatively, create an address with an immediate and a
1698 // base and a scale.
1699 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1701 // Canonicalize a scale of 1 to a base register if the formula doesn't
1702 // already have a base register.
1703 if (!HasBaseReg && Scale == 1) {
1708 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1712 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1713 ScalarEvolution &SE, int64_t MinOffset,
1714 int64_t MaxOffset, LSRUse::KindType Kind,
1715 MemAccessTy AccessTy, const SCEV *S,
1717 // Fast-path: zero is always foldable.
1718 if (S->isZero()) return true;
1720 // Conservatively, create an address with an immediate and a
1721 // base and a scale.
1722 int64_t BaseOffset = ExtractImmediate(S, SE);
1723 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1725 // If there's anything else involved, it's not foldable.
1726 if (!S->isZero()) return false;
1728 // Fast-path: zero is always foldable.
1729 if (BaseOffset == 0 && !BaseGV) return true;
1731 // Conservatively, create an address with an immediate and a
1732 // base and a scale.
1733 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1735 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1736 BaseOffset, HasBaseReg, Scale);
1741 /// An individual increment in a Chain of IV increments. Relate an IV user to
1742 /// an expression that computes the IV it uses from the IV used by the previous
1743 /// link in the Chain.
1745 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1746 /// original IVOperand. The head of the chain's IVOperand is only valid during
1747 /// chain collection, before LSR replaces IV users. During chain generation,
1748 /// IncExpr can be used to find the new IVOperand that computes the same
1751 Instruction *UserInst;
1753 const SCEV *IncExpr;
1755 IVInc(Instruction *U, Value *O, const SCEV *E):
1756 UserInst(U), IVOperand(O), IncExpr(E) {}
1759 // The list of IV increments in program order. We typically add the head of a
1760 // chain without finding subsequent links.
1762 SmallVector<IVInc,1> Incs;
1763 const SCEV *ExprBase;
1765 IVChain() : ExprBase(nullptr) {}
1767 IVChain(const IVInc &Head, const SCEV *Base)
1768 : Incs(1, Head), ExprBase(Base) {}
1770 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1772 // Return the first increment in the chain.
1773 const_iterator begin() const {
1774 assert(!Incs.empty());
1775 return std::next(Incs.begin());
1777 const_iterator end() const {
1781 // Returns true if this chain contains any increments.
1782 bool hasIncs() const { return Incs.size() >= 2; }
1784 // Add an IVInc to the end of this chain.
1785 void add(const IVInc &X) { Incs.push_back(X); }
1787 // Returns the last UserInst in the chain.
1788 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1790 // Returns true if IncExpr can be profitably added to this chain.
1791 bool isProfitableIncrement(const SCEV *OperExpr,
1792 const SCEV *IncExpr,
1796 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1797 /// between FarUsers that definitely cross IV increments and NearUsers that may
1798 /// be used between IV increments.
1800 SmallPtrSet<Instruction*, 4> FarUsers;
1801 SmallPtrSet<Instruction*, 4> NearUsers;
1804 /// This class holds state for the main loop strength reduction logic.
1807 ScalarEvolution &SE;
1810 const TargetTransformInfo &TTI;
1814 /// This is the insert position that the current loop's induction variable
1815 /// increment should be placed. In simple loops, this is the latch block's
1816 /// terminator. But in more complicated cases, this is a position which will
1817 /// dominate all the in-loop post-increment users.
1818 Instruction *IVIncInsertPos;
1820 /// Interesting factors between use strides.
1822 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1823 /// default, a SmallDenseSet, because we need to use the full range of
1824 /// int64_ts, and there's currently no good way of doing that with
1826 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1828 /// Interesting use types, to facilitate truncation reuse.
1829 SmallSetVector<Type *, 4> Types;
1831 /// The list of interesting uses.
1832 SmallVector<LSRUse, 16> Uses;
1834 /// Track which uses use which register candidates.
1835 RegUseTracker RegUses;
1837 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1838 // have more than a few IV increment chains in a loop. Missing a Chain falls
1839 // back to normal LSR behavior for those uses.
1840 static const unsigned MaxChains = 8;
1842 /// IV users can form a chain of IV increments.
1843 SmallVector<IVChain, MaxChains> IVChainVec;
1845 /// IV users that belong to profitable IVChains.
1846 SmallPtrSet<Use*, MaxChains> IVIncSet;
1848 void OptimizeShadowIV();
1849 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1850 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1851 void OptimizeLoopTermCond();
1853 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1854 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1855 void FinalizeChain(IVChain &Chain);
1856 void CollectChains();
1857 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1858 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1860 void CollectInterestingTypesAndFactors();
1861 void CollectFixupsAndInitialFormulae();
1863 // Support for sharing of LSRUses between LSRFixups.
1864 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1867 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1868 LSRUse::KindType Kind, MemAccessTy AccessTy);
1870 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1871 MemAccessTy AccessTy);
1873 void DeleteUse(LSRUse &LU, size_t LUIdx);
1875 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1877 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1878 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1879 void CountRegisters(const Formula &F, size_t LUIdx);
1880 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1882 void CollectLoopInvariantFixupsAndFormulae();
1884 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1885 unsigned Depth = 0);
1887 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1888 const Formula &Base, unsigned Depth,
1889 size_t Idx, bool IsScaledReg = false);
1890 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1891 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1892 const Formula &Base, size_t Idx,
1893 bool IsScaledReg = false);
1894 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1895 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1896 const Formula &Base,
1897 const SmallVectorImpl<int64_t> &Worklist,
1898 size_t Idx, bool IsScaledReg = false);
1899 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1900 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1901 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1902 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1903 void GenerateCrossUseConstantOffsets();
1904 void GenerateAllReuseFormulae();
1906 void FilterOutUndesirableDedicatedRegisters();
1908 size_t EstimateSearchSpaceComplexity() const;
1909 void NarrowSearchSpaceByDetectingSupersets();
1910 void NarrowSearchSpaceByCollapsingUnrolledCode();
1911 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1912 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
1913 void NarrowSearchSpaceByDeletingCostlyFormulas();
1914 void NarrowSearchSpaceByPickingWinnerRegs();
1915 void NarrowSearchSpaceUsingHeuristics();
1917 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1919 SmallVectorImpl<const Formula *> &Workspace,
1920 const Cost &CurCost,
1921 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1922 DenseSet<const SCEV *> &VisitedRegs) const;
1923 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1925 BasicBlock::iterator
1926 HoistInsertPosition(BasicBlock::iterator IP,
1927 const SmallVectorImpl<Instruction *> &Inputs) const;
1928 BasicBlock::iterator
1929 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1932 SCEVExpander &Rewriter) const;
1934 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1935 BasicBlock::iterator IP, SCEVExpander &Rewriter,
1936 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1937 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
1938 const Formula &F, SCEVExpander &Rewriter,
1939 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1940 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1941 SCEVExpander &Rewriter,
1942 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1943 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
1946 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
1947 LoopInfo &LI, const TargetTransformInfo &TTI);
1949 bool getChanged() const { return Changed; }
1951 void print_factors_and_types(raw_ostream &OS) const;
1952 void print_fixups(raw_ostream &OS) const;
1953 void print_uses(raw_ostream &OS) const;
1954 void print(raw_ostream &OS) const;
1958 } // end anonymous namespace
1960 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
1961 /// the cast operation.
1962 void LSRInstance::OptimizeShadowIV() {
1963 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1964 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1967 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1968 UI != E; /* empty */) {
1969 IVUsers::const_iterator CandidateUI = UI;
1971 Instruction *ShadowUse = CandidateUI->getUser();
1972 Type *DestTy = nullptr;
1973 bool IsSigned = false;
1975 /* If shadow use is a int->float cast then insert a second IV
1976 to eliminate this cast.
1978 for (unsigned i = 0; i < n; ++i)
1984 for (unsigned i = 0; i < n; ++i, ++d)
1987 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1989 DestTy = UCast->getDestTy();
1991 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1993 DestTy = SCast->getDestTy();
1995 if (!DestTy) continue;
1997 // If target does not support DestTy natively then do not apply
1998 // this transformation.
1999 if (!TTI.isTypeLegal(DestTy)) continue;
2001 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2003 if (PH->getNumIncomingValues() != 2) continue;
2005 Type *SrcTy = PH->getType();
2006 int Mantissa = DestTy->getFPMantissaWidth();
2007 if (Mantissa == -1) continue;
2008 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2011 unsigned Entry, Latch;
2012 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2020 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2021 if (!Init) continue;
2022 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2023 (double)Init->getSExtValue() :
2024 (double)Init->getZExtValue());
2026 BinaryOperator *Incr =
2027 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2028 if (!Incr) continue;
2029 if (Incr->getOpcode() != Instruction::Add
2030 && Incr->getOpcode() != Instruction::Sub)
2033 /* Initialize new IV, double d = 0.0 in above example. */
2034 ConstantInt *C = nullptr;
2035 if (Incr->getOperand(0) == PH)
2036 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2037 else if (Incr->getOperand(1) == PH)
2038 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2044 // Ignore negative constants, as the code below doesn't handle them
2045 // correctly. TODO: Remove this restriction.
2046 if (!C->getValue().isStrictlyPositive()) continue;
2048 /* Add new PHINode. */
2049 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2051 /* create new increment. '++d' in above example. */
2052 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2053 BinaryOperator *NewIncr =
2054 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2055 Instruction::FAdd : Instruction::FSub,
2056 NewPH, CFP, "IV.S.next.", Incr);
2058 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2059 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2061 /* Remove cast operation */
2062 ShadowUse->replaceAllUsesWith(NewPH);
2063 ShadowUse->eraseFromParent();
2069 /// If Cond has an operand that is an expression of an IV, set the IV user and
2070 /// stride information and return true, otherwise return false.
2071 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2072 for (IVStrideUse &U : IU)
2073 if (U.getUser() == Cond) {
2074 // NOTE: we could handle setcc instructions with multiple uses here, but
2075 // InstCombine does it as well for simple uses, it's not clear that it
2076 // occurs enough in real life to handle.
2083 /// Rewrite the loop's terminating condition if it uses a max computation.
2085 /// This is a narrow solution to a specific, but acute, problem. For loops
2091 /// } while (++i < n);
2093 /// the trip count isn't just 'n', because 'n' might not be positive. And
2094 /// unfortunately this can come up even for loops where the user didn't use
2095 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2096 /// will commonly be lowered like this:
2102 /// } while (++i < n);
2105 /// and then it's possible for subsequent optimization to obscure the if
2106 /// test in such a way that indvars can't find it.
2108 /// When indvars can't find the if test in loops like this, it creates a
2109 /// max expression, which allows it to give the loop a canonical
2110 /// induction variable:
2113 /// max = n < 1 ? 1 : n;
2116 /// } while (++i != max);
2118 /// Canonical induction variables are necessary because the loop passes
2119 /// are designed around them. The most obvious example of this is the
2120 /// LoopInfo analysis, which doesn't remember trip count values. It
2121 /// expects to be able to rediscover the trip count each time it is
2122 /// needed, and it does this using a simple analysis that only succeeds if
2123 /// the loop has a canonical induction variable.
2125 /// However, when it comes time to generate code, the maximum operation
2126 /// can be quite costly, especially if it's inside of an outer loop.
2128 /// This function solves this problem by detecting this type of loop and
2129 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2130 /// the instructions for the maximum computation.
2132 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2133 // Check that the loop matches the pattern we're looking for.
2134 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2135 Cond->getPredicate() != CmpInst::ICMP_NE)
2138 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2139 if (!Sel || !Sel->hasOneUse()) return Cond;
2141 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2142 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2144 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2146 // Add one to the backedge-taken count to get the trip count.
2147 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2148 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2150 // Check for a max calculation that matches the pattern. There's no check
2151 // for ICMP_ULE here because the comparison would be with zero, which
2152 // isn't interesting.
2153 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2154 const SCEVNAryExpr *Max = nullptr;
2155 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2156 Pred = ICmpInst::ICMP_SLE;
2158 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2159 Pred = ICmpInst::ICMP_SLT;
2161 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2162 Pred = ICmpInst::ICMP_ULT;
2169 // To handle a max with more than two operands, this optimization would
2170 // require additional checking and setup.
2171 if (Max->getNumOperands() != 2)
2174 const SCEV *MaxLHS = Max->getOperand(0);
2175 const SCEV *MaxRHS = Max->getOperand(1);
2177 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2178 // for a comparison with 1. For <= and >=, a comparison with zero.
2180 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2183 // Check the relevant induction variable for conformance to
2185 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2186 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2187 if (!AR || !AR->isAffine() ||
2188 AR->getStart() != One ||
2189 AR->getStepRecurrence(SE) != One)
2192 assert(AR->getLoop() == L &&
2193 "Loop condition operand is an addrec in a different loop!");
2195 // Check the right operand of the select, and remember it, as it will
2196 // be used in the new comparison instruction.
2197 Value *NewRHS = nullptr;
2198 if (ICmpInst::isTrueWhenEqual(Pred)) {
2199 // Look for n+1, and grab n.
2200 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2201 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2202 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2203 NewRHS = BO->getOperand(0);
2204 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2205 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2206 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2207 NewRHS = BO->getOperand(0);
2210 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2211 NewRHS = Sel->getOperand(1);
2212 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2213 NewRHS = Sel->getOperand(2);
2214 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2215 NewRHS = SU->getValue();
2217 // Max doesn't match expected pattern.
2220 // Determine the new comparison opcode. It may be signed or unsigned,
2221 // and the original comparison may be either equality or inequality.
2222 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2223 Pred = CmpInst::getInversePredicate(Pred);
2225 // Ok, everything looks ok to change the condition into an SLT or SGE and
2226 // delete the max calculation.
2228 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2230 // Delete the max calculation instructions.
2231 Cond->replaceAllUsesWith(NewCond);
2232 CondUse->setUser(NewCond);
2233 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2234 Cond->eraseFromParent();
2235 Sel->eraseFromParent();
2236 if (Cmp->use_empty())
2237 Cmp->eraseFromParent();
2241 /// Change loop terminating condition to use the postinc iv when possible.
2243 LSRInstance::OptimizeLoopTermCond() {
2244 SmallPtrSet<Instruction *, 4> PostIncs;
2246 // We need a different set of heuristics for rotated and non-rotated loops.
2247 // If a loop is rotated then the latch is also the backedge, so inserting
2248 // post-inc expressions just before the latch is ideal. To reduce live ranges
2249 // it also makes sense to rewrite terminating conditions to use post-inc
2252 // If the loop is not rotated then the latch is not a backedge; the latch
2253 // check is done in the loop head. Adding post-inc expressions before the
2254 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2255 // in the loop body. In this case we do *not* want to use post-inc expressions
2256 // in the latch check, and we want to insert post-inc expressions before
2258 BasicBlock *LatchBlock = L->getLoopLatch();
2259 SmallVector<BasicBlock*, 8> ExitingBlocks;
2260 L->getExitingBlocks(ExitingBlocks);
2261 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2262 return LatchBlock != BB;
2264 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2265 IVIncInsertPos = LatchBlock->getTerminator();
2269 // Otherwise treat this as a rotated loop.
2270 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2272 // Get the terminating condition for the loop if possible. If we
2273 // can, we want to change it to use a post-incremented version of its
2274 // induction variable, to allow coalescing the live ranges for the IV into
2275 // one register value.
2277 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2280 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2281 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2284 // Search IVUsesByStride to find Cond's IVUse if there is one.
2285 IVStrideUse *CondUse = nullptr;
2286 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2287 if (!FindIVUserForCond(Cond, CondUse))
2290 // If the trip count is computed in terms of a max (due to ScalarEvolution
2291 // being unable to find a sufficient guard, for example), change the loop
2292 // comparison to use SLT or ULT instead of NE.
2293 // One consequence of doing this now is that it disrupts the count-down
2294 // optimization. That's not always a bad thing though, because in such
2295 // cases it may still be worthwhile to avoid a max.
2296 Cond = OptimizeMax(Cond, CondUse);
2298 // If this exiting block dominates the latch block, it may also use
2299 // the post-inc value if it won't be shared with other uses.
2300 // Check for dominance.
2301 if (!DT.dominates(ExitingBlock, LatchBlock))
2304 // Conservatively avoid trying to use the post-inc value in non-latch
2305 // exits if there may be pre-inc users in intervening blocks.
2306 if (LatchBlock != ExitingBlock)
2307 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2308 // Test if the use is reachable from the exiting block. This dominator
2309 // query is a conservative approximation of reachability.
2310 if (&*UI != CondUse &&
2311 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2312 // Conservatively assume there may be reuse if the quotient of their
2313 // strides could be a legal scale.
2314 const SCEV *A = IU.getStride(*CondUse, L);
2315 const SCEV *B = IU.getStride(*UI, L);
2316 if (!A || !B) continue;
2317 if (SE.getTypeSizeInBits(A->getType()) !=
2318 SE.getTypeSizeInBits(B->getType())) {
2319 if (SE.getTypeSizeInBits(A->getType()) >
2320 SE.getTypeSizeInBits(B->getType()))
2321 B = SE.getSignExtendExpr(B, A->getType());
2323 A = SE.getSignExtendExpr(A, B->getType());
2325 if (const SCEVConstant *D =
2326 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2327 const ConstantInt *C = D->getValue();
2328 // Stride of one or negative one can have reuse with non-addresses.
2329 if (C->isOne() || C->isMinusOne())
2330 goto decline_post_inc;
2331 // Avoid weird situations.
2332 if (C->getValue().getMinSignedBits() >= 64 ||
2333 C->getValue().isMinSignedValue())
2334 goto decline_post_inc;
2335 // Check for possible scaled-address reuse.
2336 MemAccessTy AccessTy = getAccessType(UI->getUser());
2337 int64_t Scale = C->getSExtValue();
2338 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2340 /*HasBaseReg=*/false, Scale,
2341 AccessTy.AddrSpace))
2342 goto decline_post_inc;
2344 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2346 /*HasBaseReg=*/false, Scale,
2347 AccessTy.AddrSpace))
2348 goto decline_post_inc;
2352 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2355 // It's possible for the setcc instruction to be anywhere in the loop, and
2356 // possible for it to have multiple users. If it is not immediately before
2357 // the exiting block branch, move it.
2358 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2359 if (Cond->hasOneUse()) {
2360 Cond->moveBefore(TermBr);
2362 // Clone the terminating condition and insert into the loopend.
2363 ICmpInst *OldCond = Cond;
2364 Cond = cast<ICmpInst>(Cond->clone());
2365 Cond->setName(L->getHeader()->getName() + ".termcond");
2366 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2368 // Clone the IVUse, as the old use still exists!
2369 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2370 TermBr->replaceUsesOfWith(OldCond, Cond);
2374 // If we get to here, we know that we can transform the setcc instruction to
2375 // use the post-incremented version of the IV, allowing us to coalesce the
2376 // live ranges for the IV correctly.
2377 CondUse->transformToPostInc(L);
2380 PostIncs.insert(Cond);
2384 // Determine an insertion point for the loop induction variable increment. It
2385 // must dominate all the post-inc comparisons we just set up, and it must
2386 // dominate the loop latch edge.
2387 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2388 for (Instruction *Inst : PostIncs) {
2390 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2392 if (BB == Inst->getParent())
2393 IVIncInsertPos = Inst;
2394 else if (BB != IVIncInsertPos->getParent())
2395 IVIncInsertPos = BB->getTerminator();
2399 /// Determine if the given use can accommodate a fixup at the given offset and
2400 /// other details. If so, update the use and return true.
2401 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2402 bool HasBaseReg, LSRUse::KindType Kind,
2403 MemAccessTy AccessTy) {
2404 int64_t NewMinOffset = LU.MinOffset;
2405 int64_t NewMaxOffset = LU.MaxOffset;
2406 MemAccessTy NewAccessTy = AccessTy;
2408 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2409 // something conservative, however this can pessimize in the case that one of
2410 // the uses will have all its uses outside the loop, for example.
2411 if (LU.Kind != Kind)
2414 // Check for a mismatched access type, and fall back conservatively as needed.
2415 // TODO: Be less conservative when the type is similar and can use the same
2416 // addressing modes.
2417 if (Kind == LSRUse::Address) {
2418 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2419 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2420 AccessTy.AddrSpace);
2424 // Conservatively assume HasBaseReg is true for now.
2425 if (NewOffset < LU.MinOffset) {
2426 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2427 LU.MaxOffset - NewOffset, HasBaseReg))
2429 NewMinOffset = NewOffset;
2430 } else if (NewOffset > LU.MaxOffset) {
2431 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2432 NewOffset - LU.MinOffset, HasBaseReg))
2434 NewMaxOffset = NewOffset;
2438 LU.MinOffset = NewMinOffset;
2439 LU.MaxOffset = NewMaxOffset;
2440 LU.AccessTy = NewAccessTy;
2444 /// Return an LSRUse index and an offset value for a fixup which needs the given
2445 /// expression, with the given kind and optional access type. Either reuse an
2446 /// existing use or create a new one, as needed.
2447 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2448 LSRUse::KindType Kind,
2449 MemAccessTy AccessTy) {
2450 const SCEV *Copy = Expr;
2451 int64_t Offset = ExtractImmediate(Expr, SE);
2453 // Basic uses can't accept any offset, for example.
2454 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2455 Offset, /*HasBaseReg=*/ true)) {
2460 std::pair<UseMapTy::iterator, bool> P =
2461 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2463 // A use already existed with this base.
2464 size_t LUIdx = P.first->second;
2465 LSRUse &LU = Uses[LUIdx];
2466 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2468 return std::make_pair(LUIdx, Offset);
2471 // Create a new use.
2472 size_t LUIdx = Uses.size();
2473 P.first->second = LUIdx;
2474 Uses.push_back(LSRUse(Kind, AccessTy));
2475 LSRUse &LU = Uses[LUIdx];
2477 LU.MinOffset = Offset;
2478 LU.MaxOffset = Offset;
2479 return std::make_pair(LUIdx, Offset);
2482 /// Delete the given use from the Uses list.
2483 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2484 if (&LU != &Uses.back())
2485 std::swap(LU, Uses.back());
2489 RegUses.swapAndDropUse(LUIdx, Uses.size());
2492 /// Look for a use distinct from OrigLU which is has a formula that has the same
2493 /// registers as the given formula.
2495 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2496 const LSRUse &OrigLU) {
2497 // Search all uses for the formula. This could be more clever.
2498 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2499 LSRUse &LU = Uses[LUIdx];
2500 // Check whether this use is close enough to OrigLU, to see whether it's
2501 // worthwhile looking through its formulae.
2502 // Ignore ICmpZero uses because they may contain formulae generated by
2503 // GenerateICmpZeroScales, in which case adding fixup offsets may
2505 if (&LU != &OrigLU &&
2506 LU.Kind != LSRUse::ICmpZero &&
2507 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2508 LU.WidestFixupType == OrigLU.WidestFixupType &&
2509 LU.HasFormulaWithSameRegs(OrigF)) {
2510 // Scan through this use's formulae.
2511 for (const Formula &F : LU.Formulae) {
2512 // Check to see if this formula has the same registers and symbols
2514 if (F.BaseRegs == OrigF.BaseRegs &&
2515 F.ScaledReg == OrigF.ScaledReg &&
2516 F.BaseGV == OrigF.BaseGV &&
2517 F.Scale == OrigF.Scale &&
2518 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2519 if (F.BaseOffset == 0)
2521 // This is the formula where all the registers and symbols matched;
2522 // there aren't going to be any others. Since we declined it, we
2523 // can skip the rest of the formulae and proceed to the next LSRUse.
2530 // Nothing looked good.
2534 void LSRInstance::CollectInterestingTypesAndFactors() {
2535 SmallSetVector<const SCEV *, 4> Strides;
2537 // Collect interesting types and strides.
2538 SmallVector<const SCEV *, 4> Worklist;
2539 for (const IVStrideUse &U : IU) {
2540 const SCEV *Expr = IU.getExpr(U);
2542 // Collect interesting types.
2543 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2545 // Add strides for mentioned loops.
2546 Worklist.push_back(Expr);
2548 const SCEV *S = Worklist.pop_back_val();
2549 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2550 if (AR->getLoop() == L)
2551 Strides.insert(AR->getStepRecurrence(SE));
2552 Worklist.push_back(AR->getStart());
2553 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2554 Worklist.append(Add->op_begin(), Add->op_end());
2556 } while (!Worklist.empty());
2559 // Compute interesting factors from the set of interesting strides.
2560 for (SmallSetVector<const SCEV *, 4>::const_iterator
2561 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2562 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2563 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2564 const SCEV *OldStride = *I;
2565 const SCEV *NewStride = *NewStrideIter;
2567 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2568 SE.getTypeSizeInBits(NewStride->getType())) {
2569 if (SE.getTypeSizeInBits(OldStride->getType()) >
2570 SE.getTypeSizeInBits(NewStride->getType()))
2571 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2573 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2575 if (const SCEVConstant *Factor =
2576 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2578 if (Factor->getAPInt().getMinSignedBits() <= 64)
2579 Factors.insert(Factor->getAPInt().getSExtValue());
2580 } else if (const SCEVConstant *Factor =
2581 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2584 if (Factor->getAPInt().getMinSignedBits() <= 64)
2585 Factors.insert(Factor->getAPInt().getSExtValue());
2589 // If all uses use the same type, don't bother looking for truncation-based
2591 if (Types.size() == 1)
2594 DEBUG(print_factors_and_types(dbgs()));
2597 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2598 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2599 /// IVStrideUses, we could partially skip this.
2600 static User::op_iterator
2601 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2602 Loop *L, ScalarEvolution &SE) {
2603 for(; OI != OE; ++OI) {
2604 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2605 if (!SE.isSCEVable(Oper->getType()))
2608 if (const SCEVAddRecExpr *AR =
2609 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2610 if (AR->getLoop() == L)
2618 /// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2619 /// a convenient helper.
2620 static Value *getWideOperand(Value *Oper) {
2621 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2622 return Trunc->getOperand(0);
2626 /// Return true if we allow an IV chain to include both types.
2627 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2628 Type *LType = LVal->getType();
2629 Type *RType = RVal->getType();
2630 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2631 // Different address spaces means (possibly)
2632 // different types of the pointer implementation,
2633 // e.g. i16 vs i32 so disallow that.
2634 (LType->getPointerAddressSpace() ==
2635 RType->getPointerAddressSpace()));
2638 /// Return an approximation of this SCEV expression's "base", or NULL for any
2639 /// constant. Returning the expression itself is conservative. Returning a
2640 /// deeper subexpression is more precise and valid as long as it isn't less
2641 /// complex than another subexpression. For expressions involving multiple
2642 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2643 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2646 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2647 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2648 static const SCEV *getExprBase(const SCEV *S) {
2649 switch (S->getSCEVType()) {
2650 default: // uncluding scUnknown.
2655 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2657 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2659 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2661 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2662 // there's nothing more complex.
2663 // FIXME: not sure if we want to recognize negation.
2664 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2665 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2666 E(Add->op_begin()); I != E; ++I) {
2667 const SCEV *SubExpr = *I;
2668 if (SubExpr->getSCEVType() == scAddExpr)
2669 return getExprBase(SubExpr);
2671 if (SubExpr->getSCEVType() != scMulExpr)
2674 return S; // all operands are scaled, be conservative.
2677 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2681 /// Return true if the chain increment is profitable to expand into a loop
2682 /// invariant value, which may require its own register. A profitable chain
2683 /// increment will be an offset relative to the same base. We allow such offsets
2684 /// to potentially be used as chain increment as long as it's not obviously
2685 /// expensive to expand using real instructions.
2686 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2687 const SCEV *IncExpr,
2688 ScalarEvolution &SE) {
2689 // Aggressively form chains when -stress-ivchain.
2693 // Do not replace a constant offset from IV head with a nonconstant IV
2695 if (!isa<SCEVConstant>(IncExpr)) {
2696 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2697 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2701 SmallPtrSet<const SCEV*, 8> Processed;
2702 return !isHighCostExpansion(IncExpr, Processed, SE);
2705 /// Return true if the number of registers needed for the chain is estimated to
2706 /// be less than the number required for the individual IV users. First prohibit
2707 /// any IV users that keep the IV live across increments (the Users set should
2708 /// be empty). Next count the number and type of increments in the chain.
2710 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2711 /// effectively use postinc addressing modes. Only consider it profitable it the
2712 /// increments can be computed in fewer registers when chained.
2714 /// TODO: Consider IVInc free if it's already used in another chains.
2716 isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2717 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2721 if (!Chain.hasIncs())
2724 if (!Users.empty()) {
2725 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2726 for (Instruction *Inst : Users) {
2727 dbgs() << " " << *Inst << "\n";
2731 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2733 // The chain itself may require a register, so intialize cost to 1.
2736 // A complete chain likely eliminates the need for keeping the original IV in
2737 // a register. LSR does not currently know how to form a complete chain unless
2738 // the header phi already exists.
2739 if (isa<PHINode>(Chain.tailUserInst())
2740 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2743 const SCEV *LastIncExpr = nullptr;
2744 unsigned NumConstIncrements = 0;
2745 unsigned NumVarIncrements = 0;
2746 unsigned NumReusedIncrements = 0;
2747 for (const IVInc &Inc : Chain) {
2748 if (Inc.IncExpr->isZero())
2751 // Incrementing by zero or some constant is neutral. We assume constants can
2752 // be folded into an addressing mode or an add's immediate operand.
2753 if (isa<SCEVConstant>(Inc.IncExpr)) {
2754 ++NumConstIncrements;
2758 if (Inc.IncExpr == LastIncExpr)
2759 ++NumReusedIncrements;
2763 LastIncExpr = Inc.IncExpr;
2765 // An IV chain with a single increment is handled by LSR's postinc
2766 // uses. However, a chain with multiple increments requires keeping the IV's
2767 // value live longer than it needs to be if chained.
2768 if (NumConstIncrements > 1)
2771 // Materializing increment expressions in the preheader that didn't exist in
2772 // the original code may cost a register. For example, sign-extended array
2773 // indices can produce ridiculous increments like this:
2774 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2775 cost += NumVarIncrements;
2777 // Reusing variable increments likely saves a register to hold the multiple of
2779 cost -= NumReusedIncrements;
2781 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2787 /// Add this IV user to an existing chain or make it the head of a new chain.
2788 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2789 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2790 // When IVs are used as types of varying widths, they are generally converted
2791 // to a wider type with some uses remaining narrow under a (free) trunc.
2792 Value *const NextIV = getWideOperand(IVOper);
2793 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2794 const SCEV *const OperExprBase = getExprBase(OperExpr);
2796 // Visit all existing chains. Check if its IVOper can be computed as a
2797 // profitable loop invariant increment from the last link in the Chain.
2798 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2799 const SCEV *LastIncExpr = nullptr;
2800 for (; ChainIdx < NChains; ++ChainIdx) {
2801 IVChain &Chain = IVChainVec[ChainIdx];
2803 // Prune the solution space aggressively by checking that both IV operands
2804 // are expressions that operate on the same unscaled SCEVUnknown. This
2805 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2806 // first avoids creating extra SCEV expressions.
2807 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2810 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2811 if (!isCompatibleIVType(PrevIV, NextIV))
2814 // A phi node terminates a chain.
2815 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2818 // The increment must be loop-invariant so it can be kept in a register.
2819 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2820 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2821 if (!SE.isLoopInvariant(IncExpr, L))
2824 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2825 LastIncExpr = IncExpr;
2829 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2830 // bother for phi nodes, because they must be last in the chain.
2831 if (ChainIdx == NChains) {
2832 if (isa<PHINode>(UserInst))
2834 if (NChains >= MaxChains && !StressIVChain) {
2835 DEBUG(dbgs() << "IV Chain Limit\n");
2838 LastIncExpr = OperExpr;
2839 // IVUsers may have skipped over sign/zero extensions. We don't currently
2840 // attempt to form chains involving extensions unless they can be hoisted
2841 // into this loop's AddRec.
2842 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2845 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2847 ChainUsersVec.resize(NChains);
2848 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2849 << ") IV=" << *LastIncExpr << "\n");
2851 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2852 << ") IV+" << *LastIncExpr << "\n");
2853 // Add this IV user to the end of the chain.
2854 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2856 IVChain &Chain = IVChainVec[ChainIdx];
2858 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2859 // This chain's NearUsers become FarUsers.
2860 if (!LastIncExpr->isZero()) {
2861 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2866 // All other uses of IVOperand become near uses of the chain.
2867 // We currently ignore intermediate values within SCEV expressions, assuming
2868 // they will eventually be used be the current chain, or can be computed
2869 // from one of the chain increments. To be more precise we could
2870 // transitively follow its user and only add leaf IV users to the set.
2871 for (User *U : IVOper->users()) {
2872 Instruction *OtherUse = dyn_cast<Instruction>(U);
2875 // Uses in the chain will no longer be uses if the chain is formed.
2876 // Include the head of the chain in this iteration (not Chain.begin()).
2877 IVChain::const_iterator IncIter = Chain.Incs.begin();
2878 IVChain::const_iterator IncEnd = Chain.Incs.end();
2879 for( ; IncIter != IncEnd; ++IncIter) {
2880 if (IncIter->UserInst == OtherUse)
2883 if (IncIter != IncEnd)
2886 if (SE.isSCEVable(OtherUse->getType())
2887 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2888 && IU.isIVUserOrOperand(OtherUse)) {
2891 NearUsers.insert(OtherUse);
2894 // Since this user is part of the chain, it's no longer considered a use
2896 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2899 /// Populate the vector of Chains.
2901 /// This decreases ILP at the architecture level. Targets with ample registers,
2902 /// multiple memory ports, and no register renaming probably don't want
2903 /// this. However, such targets should probably disable LSR altogether.
2905 /// The job of LSR is to make a reasonable choice of induction variables across
2906 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2907 /// ILP *within the loop* if the target wants it.
2909 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2910 /// will not reorder memory operations, it will recognize this as a chain, but
2911 /// will generate redundant IV increments. Ideally this would be corrected later
2912 /// by a smart scheduler:
2918 /// TODO: Walk the entire domtree within this loop, not just the path to the
2919 /// loop latch. This will discover chains on side paths, but requires
2920 /// maintaining multiple copies of the Chains state.
2921 void LSRInstance::CollectChains() {
2922 DEBUG(dbgs() << "Collecting IV Chains.\n");
2923 SmallVector<ChainUsers, 8> ChainUsersVec;
2925 SmallVector<BasicBlock *,8> LatchPath;
2926 BasicBlock *LoopHeader = L->getHeader();
2927 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2928 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2929 LatchPath.push_back(Rung->getBlock());
2931 LatchPath.push_back(LoopHeader);
2933 // Walk the instruction stream from the loop header to the loop latch.
2934 for (BasicBlock *BB : reverse(LatchPath)) {
2935 for (Instruction &I : *BB) {
2936 // Skip instructions that weren't seen by IVUsers analysis.
2937 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
2940 // Ignore users that are part of a SCEV expression. This way we only
2941 // consider leaf IV Users. This effectively rediscovers a portion of
2942 // IVUsers analysis but in program order this time.
2943 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
2946 // Remove this instruction from any NearUsers set it may be in.
2947 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2948 ChainIdx < NChains; ++ChainIdx) {
2949 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
2951 // Search for operands that can be chained.
2952 SmallPtrSet<Instruction*, 4> UniqueOperands;
2953 User::op_iterator IVOpEnd = I.op_end();
2954 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
2955 while (IVOpIter != IVOpEnd) {
2956 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2957 if (UniqueOperands.insert(IVOpInst).second)
2958 ChainInstruction(&I, IVOpInst, ChainUsersVec);
2959 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2961 } // Continue walking down the instructions.
2962 } // Continue walking down the domtree.
2963 // Visit phi backedges to determine if the chain can generate the IV postinc.
2964 for (BasicBlock::iterator I = L->getHeader()->begin();
2965 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2966 if (!SE.isSCEVable(PN->getType()))
2970 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2972 ChainInstruction(PN, IncV, ChainUsersVec);
2974 // Remove any unprofitable chains.
2975 unsigned ChainIdx = 0;
2976 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2977 UsersIdx < NChains; ++UsersIdx) {
2978 if (!isProfitableChain(IVChainVec[UsersIdx],
2979 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2981 // Preserve the chain at UsesIdx.
2982 if (ChainIdx != UsersIdx)
2983 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2984 FinalizeChain(IVChainVec[ChainIdx]);
2987 IVChainVec.resize(ChainIdx);
2990 void LSRInstance::FinalizeChain(IVChain &Chain) {
2991 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2992 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2994 for (const IVInc &Inc : Chain) {
2995 DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
2996 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
2997 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
2998 IVIncSet.insert(UseI);
3002 /// Return true if the IVInc can be folded into an addressing mode.
3003 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3004 Value *Operand, const TargetTransformInfo &TTI) {
3005 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3006 if (!IncConst || !isAddressUse(UserInst, Operand))
3009 if (IncConst->getAPInt().getMinSignedBits() > 64)
3012 MemAccessTy AccessTy = getAccessType(UserInst);
3013 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3014 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3015 IncOffset, /*HaseBaseReg=*/false))
3021 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3022 /// user's operand from the previous IV user's operand.
3023 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3024 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3025 // Find the new IVOperand for the head of the chain. It may have been replaced
3027 const IVInc &Head = Chain.Incs[0];
3028 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3029 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3030 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3032 Value *IVSrc = nullptr;
3033 while (IVOpIter != IVOpEnd) {
3034 IVSrc = getWideOperand(*IVOpIter);
3036 // If this operand computes the expression that the chain needs, we may use
3037 // it. (Check this after setting IVSrc which is used below.)
3039 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3040 // narrow for the chain, so we can no longer use it. We do allow using a
3041 // wider phi, assuming the LSR checked for free truncation. In that case we
3042 // should already have a truncate on this operand such that
3043 // getSCEV(IVSrc) == IncExpr.
3044 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3045 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3048 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3050 if (IVOpIter == IVOpEnd) {
3051 // Gracefully give up on this chain.
3052 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3056 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3057 Type *IVTy = IVSrc->getType();
3058 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3059 const SCEV *LeftOverExpr = nullptr;
3060 for (const IVInc &Inc : Chain) {
3061 Instruction *InsertPt = Inc.UserInst;
3062 if (isa<PHINode>(InsertPt))
3063 InsertPt = L->getLoopLatch()->getTerminator();
3065 // IVOper will replace the current IV User's operand. IVSrc is the IV
3066 // value currently held in a register.
3067 Value *IVOper = IVSrc;
3068 if (!Inc.IncExpr->isZero()) {
3069 // IncExpr was the result of subtraction of two narrow values, so must
3071 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3072 LeftOverExpr = LeftOverExpr ?
3073 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3075 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3076 // Expand the IV increment.
3077 Rewriter.clearPostInc();
3078 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3079 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3080 SE.getUnknown(IncV));
3081 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3083 // If an IV increment can't be folded, use it as the next IV value.
3084 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3085 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3087 LeftOverExpr = nullptr;
3090 Type *OperTy = Inc.IVOperand->getType();
3091 if (IVTy != OperTy) {
3092 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3093 "cannot extend a chained IV");
3094 IRBuilder<> Builder(InsertPt);
3095 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3097 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3098 DeadInsts.emplace_back(Inc.IVOperand);
3100 // If LSR created a new, wider phi, we may also replace its postinc. We only
3101 // do this if we also found a wide value for the head of the chain.
3102 if (isa<PHINode>(Chain.tailUserInst())) {
3103 for (BasicBlock::iterator I = L->getHeader()->begin();
3104 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
3105 if (!isCompatibleIVType(Phi, IVSrc))
3107 Instruction *PostIncV = dyn_cast<Instruction>(
3108 Phi->getIncomingValueForBlock(L->getLoopLatch()));
3109 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3111 Value *IVOper = IVSrc;
3112 Type *PostIncTy = PostIncV->getType();
3113 if (IVTy != PostIncTy) {
3114 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3115 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3116 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3117 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3119 Phi->replaceUsesOfWith(PostIncV, IVOper);
3120 DeadInsts.emplace_back(PostIncV);
3125 void LSRInstance::CollectFixupsAndInitialFormulae() {
3126 for (const IVStrideUse &U : IU) {
3127 Instruction *UserInst = U.getUser();
3128 // Skip IV users that are part of profitable IV Chains.
3129 User::op_iterator UseI =
3130 find(UserInst->operands(), U.getOperandValToReplace());
3131 assert(UseI != UserInst->op_end() && "cannot find IV operand");
3132 if (IVIncSet.count(UseI)) {
3133 DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3137 LSRUse::KindType Kind = LSRUse::Basic;
3138 MemAccessTy AccessTy;
3139 if (isAddressUse(UserInst, U.getOperandValToReplace())) {
3140 Kind = LSRUse::Address;
3141 AccessTy = getAccessType(UserInst);
3144 const SCEV *S = IU.getExpr(U);
3145 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3147 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3148 // (N - i == 0), and this allows (N - i) to be the expression that we work
3149 // with rather than just N or i, so we can consider the register
3150 // requirements for both N and i at the same time. Limiting this code to
3151 // equality icmps is not a problem because all interesting loops use
3152 // equality icmps, thanks to IndVarSimplify.
3153 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3154 if (CI->isEquality()) {
3155 // Swap the operands if needed to put the OperandValToReplace on the
3156 // left, for consistency.
3157 Value *NV = CI->getOperand(1);
3158 if (NV == U.getOperandValToReplace()) {
3159 CI->setOperand(1, CI->getOperand(0));
3160 CI->setOperand(0, NV);
3161 NV = CI->getOperand(1);
3165 // x == y --> x - y == 0
3166 const SCEV *N = SE.getSCEV(NV);
3167 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3168 // S is normalized, so normalize N before folding it into S
3169 // to keep the result normalized.
3170 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3171 Kind = LSRUse::ICmpZero;
3172 S = SE.getMinusSCEV(N, S);
3175 // -1 and the negations of all interesting strides (except the negation
3176 // of -1) are now also interesting.
3177 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3178 if (Factors[i] != -1)
3179 Factors.insert(-(uint64_t)Factors[i]);
3183 // Get or create an LSRUse.
3184 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3185 size_t LUIdx = P.first;
3186 int64_t Offset = P.second;
3187 LSRUse &LU = Uses[LUIdx];
3189 // Record the fixup.
3190 LSRFixup &LF = LU.getNewFixup();
3191 LF.UserInst = UserInst;
3192 LF.OperandValToReplace = U.getOperandValToReplace();
3193 LF.PostIncLoops = TmpPostIncLoops;
3195 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3197 if (!LU.WidestFixupType ||
3198 SE.getTypeSizeInBits(LU.WidestFixupType) <
3199 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3200 LU.WidestFixupType = LF.OperandValToReplace->getType();
3202 // If this is the first use of this LSRUse, give it a formula.
3203 if (LU.Formulae.empty()) {
3204 InsertInitialFormula(S, LU, LUIdx);
3205 CountRegisters(LU.Formulae.back(), LUIdx);
3209 DEBUG(print_fixups(dbgs()));
3212 /// Insert a formula for the given expression into the given use, separating out
3213 /// loop-variant portions from loop-invariant and loop-computable portions.
3215 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3216 // Mark uses whose expressions cannot be expanded.
3217 if (!isSafeToExpand(S, SE))
3218 LU.RigidFormula = true;
3221 F.initialMatch(S, L, SE);
3222 bool Inserted = InsertFormula(LU, LUIdx, F);
3223 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3226 /// Insert a simple single-register formula for the given expression into the
3229 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3230 LSRUse &LU, size_t LUIdx) {
3232 F.BaseRegs.push_back(S);
3233 F.HasBaseReg = true;
3234 bool Inserted = InsertFormula(LU, LUIdx, F);
3235 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3238 /// Note which registers are used by the given formula, updating RegUses.
3239 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3241 RegUses.countRegister(F.ScaledReg, LUIdx);
3242 for (const SCEV *BaseReg : F.BaseRegs)
3243 RegUses.countRegister(BaseReg, LUIdx);
3246 /// If the given formula has not yet been inserted, add it to the list, and
3247 /// return true. Return false otherwise.
3248 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3249 // Do not insert formula that we will not be able to expand.
3250 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3251 "Formula is illegal");
3253 if (!LU.InsertFormula(F, *L))
3256 CountRegisters(F, LUIdx);
3260 /// Check for other uses of loop-invariant values which we're tracking. These
3261 /// other uses will pin these values in registers, making them less profitable
3262 /// for elimination.
3263 /// TODO: This currently misses non-constant addrec step registers.
3264 /// TODO: Should this give more weight to users inside the loop?
3266 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3267 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3268 SmallPtrSet<const SCEV *, 32> Visited;
3270 while (!Worklist.empty()) {
3271 const SCEV *S = Worklist.pop_back_val();
3273 // Don't process the same SCEV twice
3274 if (!Visited.insert(S).second)
3277 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3278 Worklist.append(N->op_begin(), N->op_end());
3279 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3280 Worklist.push_back(C->getOperand());
3281 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3282 Worklist.push_back(D->getLHS());
3283 Worklist.push_back(D->getRHS());
3284 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3285 const Value *V = US->getValue();
3286 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3287 // Look for instructions defined outside the loop.
3288 if (L->contains(Inst)) continue;
3289 } else if (isa<UndefValue>(V))
3290 // Undef doesn't have a live range, so it doesn't matter.
3292 for (const Use &U : V->uses()) {
3293 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3294 // Ignore non-instructions.
3297 // Ignore instructions in other functions (as can happen with
3299 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3301 // Ignore instructions not dominated by the loop.
3302 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3303 UserInst->getParent() :
3304 cast<PHINode>(UserInst)->getIncomingBlock(
3305 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3306 if (!DT.dominates(L->getHeader(), UseBB))
3308 // Don't bother if the instruction is in a BB which ends in an EHPad.
3309 if (UseBB->getTerminator()->isEHPad())
3311 // Don't bother rewriting PHIs in catchswitch blocks.
3312 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3314 // Ignore uses which are part of other SCEV expressions, to avoid
3315 // analyzing them multiple times.
3316 if (SE.isSCEVable(UserInst->getType())) {
3317 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3318 // If the user is a no-op, look through to its uses.
3319 if (!isa<SCEVUnknown>(UserS))
3323 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3327 // Ignore icmp instructions which are already being analyzed.
3328 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3329 unsigned OtherIdx = !U.getOperandNo();
3330 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3331 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3335 std::pair<size_t, int64_t> P = getUse(
3336 S, LSRUse::Basic, MemAccessTy());
3337 size_t LUIdx = P.first;
3338 int64_t Offset = P.second;
3339 LSRUse &LU = Uses[LUIdx];
3340 LSRFixup &LF = LU.getNewFixup();
3341 LF.UserInst = const_cast<Instruction *>(UserInst);
3342 LF.OperandValToReplace = U;
3344 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3345 if (!LU.WidestFixupType ||
3346 SE.getTypeSizeInBits(LU.WidestFixupType) <
3347 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3348 LU.WidestFixupType = LF.OperandValToReplace->getType();
3349 InsertSupplementalFormula(US, LU, LUIdx);
3350 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3357 /// Split S into subexpressions which can be pulled out into separate
3358 /// registers. If C is non-null, multiply each subexpression by C.
3360 /// Return remainder expression after factoring the subexpressions captured by
3361 /// Ops. If Ops is complete, return NULL.
3362 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3363 SmallVectorImpl<const SCEV *> &Ops,
3365 ScalarEvolution &SE,
3366 unsigned Depth = 0) {
3367 // Arbitrarily cap recursion to protect compile time.
3371 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3372 // Break out add operands.
3373 for (const SCEV *S : Add->operands()) {
3374 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3376 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3379 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3380 // Split a non-zero base out of an addrec.
3381 if (AR->getStart()->isZero() || !AR->isAffine())
3384 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3385 C, Ops, L, SE, Depth+1);
3386 // Split the non-zero AddRec unless it is part of a nested recurrence that
3387 // does not pertain to this loop.
3388 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3389 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3390 Remainder = nullptr;
3392 if (Remainder != AR->getStart()) {
3394 Remainder = SE.getConstant(AR->getType(), 0);
3395 return SE.getAddRecExpr(Remainder,
3396 AR->getStepRecurrence(SE),
3398 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3401 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3402 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3403 if (Mul->getNumOperands() != 2)
3405 if (const SCEVConstant *Op0 =
3406 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3407 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3408 const SCEV *Remainder =
3409 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3411 Ops.push_back(SE.getMulExpr(C, Remainder));
3418 /// \brief Helper function for LSRInstance::GenerateReassociations.
3419 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3420 const Formula &Base,
3421 unsigned Depth, size_t Idx,
3423 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3424 SmallVector<const SCEV *, 8> AddOps;
3425 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3427 AddOps.push_back(Remainder);
3429 if (AddOps.size() == 1)
3432 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3436 // Loop-variant "unknown" values are uninteresting; we won't be able to
3437 // do anything meaningful with them.
3438 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3441 // Don't pull a constant into a register if the constant could be folded
3442 // into an immediate field.
3443 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3444 LU.AccessTy, *J, Base.getNumRegs() > 1))
3447 // Collect all operands except *J.
3448 SmallVector<const SCEV *, 8> InnerAddOps(
3449 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3450 InnerAddOps.append(std::next(J),
3451 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3453 // Don't leave just a constant behind in a register if the constant could
3454 // be folded into an immediate field.
3455 if (InnerAddOps.size() == 1 &&
3456 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3457 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3460 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3461 if (InnerSum->isZero())
3465 // Add the remaining pieces of the add back into the new formula.
3466 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3467 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3468 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3469 InnerSumSC->getValue()->getZExtValue())) {
3471 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3473 F.ScaledReg = nullptr;
3475 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3476 } else if (IsScaledReg)
3477 F.ScaledReg = InnerSum;
3479 F.BaseRegs[Idx] = InnerSum;
3481 // Add J as its own register, or an unfolded immediate.
3482 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3483 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3484 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3485 SC->getValue()->getZExtValue()))
3487 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3489 F.BaseRegs.push_back(*J);
3490 // We may have changed the number of register in base regs, adjust the
3491 // formula accordingly.
3494 if (InsertFormula(LU, LUIdx, F))
3495 // If that formula hadn't been seen before, recurse to find more like
3497 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3501 /// Split out subexpressions from adds and the bases of addrecs.
3502 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3503 Formula Base, unsigned Depth) {
3504 assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3505 // Arbitrarily cap recursion to protect compile time.
3509 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3510 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3512 if (Base.Scale == 1)
3513 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3514 /* Idx */ -1, /* IsScaledReg */ true);
3517 /// Generate a formula consisting of all of the loop-dominating registers added
3518 /// into a single register.
3519 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3521 // This method is only interesting on a plurality of registers.
3522 if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3525 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3526 // processing the formula.
3530 SmallVector<const SCEV *, 4> Ops;
3531 for (const SCEV *BaseReg : Base.BaseRegs) {
3532 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3533 !SE.hasComputableLoopEvolution(BaseReg, L))
3534 Ops.push_back(BaseReg);
3536 F.BaseRegs.push_back(BaseReg);
3538 if (Ops.size() > 1) {
3539 const SCEV *Sum = SE.getAddExpr(Ops);
3540 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3541 // opportunity to fold something. For now, just ignore such cases
3542 // rather than proceed with zero in a register.
3543 if (!Sum->isZero()) {
3544 F.BaseRegs.push_back(Sum);
3546 (void)InsertFormula(LU, LUIdx, F);
3551 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3552 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3553 const Formula &Base, size_t Idx,
3555 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3556 GlobalValue *GV = ExtractSymbol(G, SE);
3557 if (G->isZero() || !GV)
3561 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3566 F.BaseRegs[Idx] = G;
3567 (void)InsertFormula(LU, LUIdx, F);
3570 /// Generate reuse formulae using symbolic offsets.
3571 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3573 // We can't add a symbolic offset if the address already contains one.
3574 if (Base.BaseGV) return;
3576 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3577 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3578 if (Base.Scale == 1)
3579 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3580 /* IsScaledReg */ true);
3583 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3584 void LSRInstance::GenerateConstantOffsetsImpl(
3585 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3586 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3587 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3588 for (int64_t Offset : Worklist) {
3590 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3591 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3593 // Add the offset to the base register.
3594 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3595 // If it cancelled out, drop the base register, otherwise update it.
3596 if (NewG->isZero()) {
3599 F.ScaledReg = nullptr;
3601 F.deleteBaseReg(F.BaseRegs[Idx]);
3603 } else if (IsScaledReg)
3606 F.BaseRegs[Idx] = NewG;
3608 (void)InsertFormula(LU, LUIdx, F);
3612 int64_t Imm = ExtractImmediate(G, SE);
3613 if (G->isZero() || Imm == 0)
3616 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3617 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3622 F.BaseRegs[Idx] = G;
3623 (void)InsertFormula(LU, LUIdx, F);
3626 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3627 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3629 // TODO: For now, just add the min and max offset, because it usually isn't
3630 // worthwhile looking at everything inbetween.
3631 SmallVector<int64_t, 2> Worklist;
3632 Worklist.push_back(LU.MinOffset);
3633 if (LU.MaxOffset != LU.MinOffset)
3634 Worklist.push_back(LU.MaxOffset);
3636 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3637 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3638 if (Base.Scale == 1)
3639 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3640 /* IsScaledReg */ true);
3643 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3644 /// == y -> x*c == y*c.
3645 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3647 if (LU.Kind != LSRUse::ICmpZero) return;
3649 // Determine the integer type for the base formula.
3650 Type *IntTy = Base.getType();
3652 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3654 // Don't do this if there is more than one offset.
3655 if (LU.MinOffset != LU.MaxOffset) return;
3657 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3659 // Check each interesting stride.
3660 for (int64_t Factor : Factors) {
3661 // Check that the multiplication doesn't overflow.
3662 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3664 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3665 if (NewBaseOffset / Factor != Base.BaseOffset)
3667 // If the offset will be truncated at this use, check that it is in bounds.
3668 if (!IntTy->isPointerTy() &&
3669 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3672 // Check that multiplying with the use offset doesn't overflow.
3673 int64_t Offset = LU.MinOffset;
3674 if (Offset == INT64_MIN && Factor == -1)
3676 Offset = (uint64_t)Offset * Factor;
3677 if (Offset / Factor != LU.MinOffset)
3679 // If the offset will be truncated at this use, check that it is in bounds.
3680 if (!IntTy->isPointerTy() &&
3681 !ConstantInt::isValueValidForType(IntTy, Offset))
3685 F.BaseOffset = NewBaseOffset;
3687 // Check that this scale is legal.
3688 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3691 // Compensate for the use having MinOffset built into it.
3692 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3694 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3696 // Check that multiplying with each base register doesn't overflow.
3697 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3698 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3699 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3703 // Check that multiplying with the scaled register doesn't overflow.
3705 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3706 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3710 // Check that multiplying with the unfolded offset doesn't overflow.
3711 if (F.UnfoldedOffset != 0) {
3712 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3714 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3715 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3717 // If the offset will be truncated, check that it is in bounds.
3718 if (!IntTy->isPointerTy() &&
3719 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3723 // If we make it here and it's legal, add it.
3724 (void)InsertFormula(LU, LUIdx, F);
3729 /// Generate stride factor reuse formulae by making use of scaled-offset address
3730 /// modes, for example.
3731 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3732 // Determine the integer type for the base formula.
3733 Type *IntTy = Base.getType();
3736 // If this Formula already has a scaled register, we can't add another one.
3737 // Try to unscale the formula to generate a better scale.
3738 if (Base.Scale != 0 && !Base.unscale())
3741 assert(Base.Scale == 0 && "unscale did not did its job!");
3743 // Check each interesting stride.
3744 for (int64_t Factor : Factors) {
3745 Base.Scale = Factor;
3746 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3747 // Check whether this scale is going to be legal.
3748 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3750 // As a special-case, handle special out-of-loop Basic users specially.
3751 // TODO: Reconsider this special case.
3752 if (LU.Kind == LSRUse::Basic &&
3753 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3754 LU.AccessTy, Base) &&
3755 LU.AllFixupsOutsideLoop)
3756 LU.Kind = LSRUse::Special;
3760 // For an ICmpZero, negating a solitary base register won't lead to
3762 if (LU.Kind == LSRUse::ICmpZero &&
3763 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3765 // For each addrec base reg, if its loop is current loop, apply the scale.
3766 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3767 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3768 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3769 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3770 if (FactorS->isZero())
3772 // Divide out the factor, ignoring high bits, since we'll be
3773 // scaling the value back up in the end.
3774 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3775 // TODO: This could be optimized to avoid all the copying.
3777 F.ScaledReg = Quotient;
3778 F.deleteBaseReg(F.BaseRegs[i]);
3779 // The canonical representation of 1*reg is reg, which is already in
3780 // Base. In that case, do not try to insert the formula, it will be
3782 if (F.Scale == 1 && (F.BaseRegs.empty() ||
3783 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3785 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3786 // non canonical Formula with ScaledReg's loop not being L.
3787 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
3789 (void)InsertFormula(LU, LUIdx, F);
3796 /// Generate reuse formulae from different IV types.
3797 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3798 // Don't bother truncating symbolic values.
3799 if (Base.BaseGV) return;
3801 // Determine the integer type for the base formula.
3802 Type *DstTy = Base.getType();
3804 DstTy = SE.getEffectiveSCEVType(DstTy);
3806 for (Type *SrcTy : Types) {
3807 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3810 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3811 for (const SCEV *&BaseReg : F.BaseRegs)
3812 BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3814 // TODO: This assumes we've done basic processing on all uses and
3815 // have an idea what the register usage is.
3816 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3820 (void)InsertFormula(LU, LUIdx, F);
3827 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3828 /// modifications so that the search phase doesn't have to worry about the data
3829 /// structures moving underneath it.
3833 const SCEV *OrigReg;
3835 WorkItem(size_t LI, int64_t I, const SCEV *R)
3836 : LUIdx(LI), Imm(I), OrigReg(R) {}
3838 void print(raw_ostream &OS) const;
3842 } // end anonymous namespace
3844 void WorkItem::print(raw_ostream &OS) const {
3845 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3846 << " , add offset " << Imm;
3849 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3850 LLVM_DUMP_METHOD void WorkItem::dump() const {
3851 print(errs()); errs() << '\n';
3855 /// Look for registers which are a constant distance apart and try to form reuse
3856 /// opportunities between them.
3857 void LSRInstance::GenerateCrossUseConstantOffsets() {
3858 // Group the registers by their value without any added constant offset.
3859 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3860 DenseMap<const SCEV *, ImmMapTy> Map;
3861 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3862 SmallVector<const SCEV *, 8> Sequence;
3863 for (const SCEV *Use : RegUses) {
3864 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
3865 int64_t Imm = ExtractImmediate(Reg, SE);
3866 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
3868 Sequence.push_back(Reg);
3869 Pair.first->second.insert(std::make_pair(Imm, Use));
3870 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
3873 // Now examine each set of registers with the same base value. Build up
3874 // a list of work to do and do the work in a separate step so that we're
3875 // not adding formulae and register counts while we're searching.
3876 SmallVector<WorkItem, 32> WorkItems;
3877 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3878 for (const SCEV *Reg : Sequence) {
3879 const ImmMapTy &Imms = Map.find(Reg)->second;
3881 // It's not worthwhile looking for reuse if there's only one offset.
3882 if (Imms.size() == 1)
3885 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3886 for (const auto &Entry : Imms)
3887 dbgs() << ' ' << Entry.first;
3890 // Examine each offset.
3891 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3893 const SCEV *OrigReg = J->second;
3895 int64_t JImm = J->first;
3896 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3898 if (!isa<SCEVConstant>(OrigReg) &&
3899 UsedByIndicesMap[Reg].count() == 1) {
3900 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3904 // Conservatively examine offsets between this orig reg a few selected
3906 ImmMapTy::const_iterator OtherImms[] = {
3907 Imms.begin(), std::prev(Imms.end()),
3908 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3911 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3912 ImmMapTy::const_iterator M = OtherImms[i];
3913 if (M == J || M == JE) continue;
3915 // Compute the difference between the two.
3916 int64_t Imm = (uint64_t)JImm - M->first;
3917 for (unsigned LUIdx : UsedByIndices.set_bits())
3918 // Make a memo of this use, offset, and register tuple.
3919 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3920 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3927 UsedByIndicesMap.clear();
3928 UniqueItems.clear();
3930 // Now iterate through the worklist and add new formulae.
3931 for (const WorkItem &WI : WorkItems) {
3932 size_t LUIdx = WI.LUIdx;
3933 LSRUse &LU = Uses[LUIdx];
3934 int64_t Imm = WI.Imm;
3935 const SCEV *OrigReg = WI.OrigReg;
3937 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3938 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3939 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3941 // TODO: Use a more targeted data structure.
3942 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3943 Formula F = LU.Formulae[L];
3944 // FIXME: The code for the scaled and unscaled registers looks
3945 // very similar but slightly different. Investigate if they
3946 // could be merged. That way, we would not have to unscale the
3949 // Use the immediate in the scaled register.
3950 if (F.ScaledReg == OrigReg) {
3951 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3952 // Don't create 50 + reg(-50).
3953 if (F.referencesReg(SE.getSCEV(
3954 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3957 NewF.BaseOffset = Offset;
3958 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3961 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3963 // If the new scale is a constant in a register, and adding the constant
3964 // value to the immediate would produce a value closer to zero than the
3965 // immediate itself, then the formula isn't worthwhile.
3966 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3967 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
3968 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
3969 .ule(std::abs(NewF.BaseOffset)))
3973 NewF.canonicalize(*this->L);
3974 (void)InsertFormula(LU, LUIdx, NewF);
3976 // Use the immediate in a base register.
3977 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3978 const SCEV *BaseReg = F.BaseRegs[N];
3979 if (BaseReg != OrigReg)
3982 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3983 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3984 LU.Kind, LU.AccessTy, NewF)) {
3985 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3988 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3990 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3992 // If the new formula has a constant in a register, and adding the
3993 // constant value to the immediate would produce a value closer to
3994 // zero than the immediate itself, then the formula isn't worthwhile.
3995 for (const SCEV *NewReg : NewF.BaseRegs)
3996 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
3997 if ((C->getAPInt() + NewF.BaseOffset)
3999 .slt(std::abs(NewF.BaseOffset)) &&
4000 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4001 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4005 NewF.canonicalize(*this->L);
4006 (void)InsertFormula(LU, LUIdx, NewF);
4015 /// Generate formulae for each use.
4017 LSRInstance::GenerateAllReuseFormulae() {
4018 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4019 // queries are more precise.
4020 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4021 LSRUse &LU = Uses[LUIdx];
4022 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4023 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4024 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4025 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4027 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4028 LSRUse &LU = Uses[LUIdx];
4029 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4030 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4031 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4032 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4033 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4034 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4035 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4036 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4038 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4039 LSRUse &LU = Uses[LUIdx];
4040 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4041 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4044 GenerateCrossUseConstantOffsets();
4046 DEBUG(dbgs() << "\n"
4047 "After generating reuse formulae:\n";
4048 print_uses(dbgs()));
4051 /// If there are multiple formulae with the same set of registers used
4052 /// by other uses, pick the best one and delete the others.
4053 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4054 DenseSet<const SCEV *> VisitedRegs;
4055 SmallPtrSet<const SCEV *, 16> Regs;
4056 SmallPtrSet<const SCEV *, 16> LoserRegs;
4058 bool ChangedFormulae = false;
4061 // Collect the best formula for each unique set of shared registers. This
4062 // is reset for each use.
4063 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
4065 BestFormulaeTy BestFormulae;
4067 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4068 LSRUse &LU = Uses[LUIdx];
4069 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4072 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4073 FIdx != NumForms; ++FIdx) {
4074 Formula &F = LU.Formulae[FIdx];
4076 // Some formulas are instant losers. For example, they may depend on
4077 // nonexistent AddRecs from other loops. These need to be filtered
4078 // immediately, otherwise heuristics could choose them over others leading
4079 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4080 // avoids the need to recompute this information across formulae using the
4081 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4082 // the corresponding bad register from the Regs set.
4085 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, SE, DT, LU, &LoserRegs);
4086 if (CostF.isLoser()) {
4087 // During initial formula generation, undesirable formulae are generated
4088 // by uses within other loops that have some non-trivial address mode or
4089 // use the postinc form of the IV. LSR needs to provide these formulae
4090 // as the basis of rediscovering the desired formula that uses an AddRec
4091 // corresponding to the existing phi. Once all formulae have been
4092 // generated, these initial losers may be pruned.
4093 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4097 SmallVector<const SCEV *, 4> Key;
4098 for (const SCEV *Reg : F.BaseRegs) {
4099 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4103 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4104 Key.push_back(F.ScaledReg);
4105 // Unstable sort by host order ok, because this is only used for
4107 std::sort(Key.begin(), Key.end());
4109 std::pair<BestFormulaeTy::const_iterator, bool> P =
4110 BestFormulae.insert(std::make_pair(Key, FIdx));
4114 Formula &Best = LU.Formulae[P.first->second];
4118 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, SE, DT, LU);
4119 if (CostF.isLess(CostBest, TTI))
4121 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4123 " in favor of formula "; Best.print(dbgs());
4127 ChangedFormulae = true;
4129 LU.DeleteFormula(F);
4135 // Now that we've filtered out some formulae, recompute the Regs set.
4137 LU.RecomputeRegs(LUIdx, RegUses);
4139 // Reset this to prepare for the next use.
4140 BestFormulae.clear();
4143 DEBUG(if (ChangedFormulae) {
4145 "After filtering out undesirable candidates:\n";
4150 // This is a rough guess that seems to work fairly well.
4151 static const size_t ComplexityLimit = UINT16_MAX;
4153 /// Estimate the worst-case number of solutions the solver might have to
4154 /// consider. It almost never considers this many solutions because it prune the
4155 /// search space, but the pruning isn't always sufficient.
4156 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4158 for (const LSRUse &LU : Uses) {
4159 size_t FSize = LU.Formulae.size();
4160 if (FSize >= ComplexityLimit) {
4161 Power = ComplexityLimit;
4165 if (Power >= ComplexityLimit)
4171 /// When one formula uses a superset of the registers of another formula, it
4172 /// won't help reduce register pressure (though it may not necessarily hurt
4173 /// register pressure); remove it to simplify the system.
4174 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4175 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4176 DEBUG(dbgs() << "The search space is too complex.\n");
4178 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4179 "which use a superset of registers used by other "
4182 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4183 LSRUse &LU = Uses[LUIdx];
4185 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4186 Formula &F = LU.Formulae[i];
4187 // Look for a formula with a constant or GV in a register. If the use
4188 // also has a formula with that same value in an immediate field,
4189 // delete the one that uses a register.
4190 for (SmallVectorImpl<const SCEV *>::const_iterator
4191 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4192 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4194 NewF.BaseOffset += C->getValue()->getSExtValue();
4195 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4196 (I - F.BaseRegs.begin()));
4197 if (LU.HasFormulaWithSameRegs(NewF)) {
4198 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4199 LU.DeleteFormula(F);
4205 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4206 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4210 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4211 (I - F.BaseRegs.begin()));
4212 if (LU.HasFormulaWithSameRegs(NewF)) {
4213 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4215 LU.DeleteFormula(F);
4226 LU.RecomputeRegs(LUIdx, RegUses);
4229 DEBUG(dbgs() << "After pre-selection:\n";
4230 print_uses(dbgs()));
4234 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4235 /// allocate a single register for them.
4236 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4237 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4240 DEBUG(dbgs() << "The search space is too complex.\n"
4241 "Narrowing the search space by assuming that uses separated "
4242 "by a constant offset will use the same registers.\n");
4244 // This is especially useful for unrolled loops.
4246 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4247 LSRUse &LU = Uses[LUIdx];
4248 for (const Formula &F : LU.Formulae) {
4249 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4252 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4256 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4257 LU.Kind, LU.AccessTy))
4260 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4262 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4264 // Transfer the fixups of LU to LUThatHas.
4265 for (LSRFixup &Fixup : LU.Fixups) {
4266 Fixup.Offset += F.BaseOffset;
4267 LUThatHas->pushFixup(Fixup);
4268 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4271 // Delete formulae from the new use which are no longer legal.
4273 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4274 Formula &F = LUThatHas->Formulae[i];
4275 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4276 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4277 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4279 LUThatHas->DeleteFormula(F);
4287 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4289 // Delete the old use.
4290 DeleteUse(LU, LUIdx);
4297 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4300 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4301 /// we've done more filtering, as it may be able to find more formulae to
4303 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4304 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4305 DEBUG(dbgs() << "The search space is too complex.\n");
4307 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4308 "undesirable dedicated registers.\n");
4310 FilterOutUndesirableDedicatedRegisters();
4312 DEBUG(dbgs() << "After pre-selection:\n";
4313 print_uses(dbgs()));
4317 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4318 /// Pick the best one and delete the others.
4319 /// This narrowing heuristic is to keep as many formulae with different
4320 /// Scale and ScaledReg pair as possible while narrowing the search space.
4321 /// The benefit is that it is more likely to find out a better solution
4322 /// from a formulae set with more Scale and ScaledReg variations than
4323 /// a formulae set with the same Scale and ScaledReg. The picking winner
4324 /// reg heurstic will often keep the formulae with the same Scale and
4325 /// ScaledReg and filter others, and we want to avoid that if possible.
4326 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4327 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4330 DEBUG(dbgs() << "The search space is too complex.\n"
4331 "Narrowing the search space by choosing the best Formula "
4332 "from the Formulae with the same Scale and ScaledReg.\n");
4334 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4335 typedef DenseMap<std::pair<const SCEV *, int64_t>, size_t> BestFormulaeTy;
4336 BestFormulaeTy BestFormulae;
4338 bool ChangedFormulae = false;
4340 DenseSet<const SCEV *> VisitedRegs;
4341 SmallPtrSet<const SCEV *, 16> Regs;
4343 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4344 LSRUse &LU = Uses[LUIdx];
4345 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4347 // Return true if Formula FA is better than Formula FB.
4348 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4349 // First we will try to choose the Formula with fewer new registers.
4350 // For a register used by current Formula, the more the register is
4351 // shared among LSRUses, the less we increase the register number
4352 // counter of the formula.
4353 size_t FARegNum = 0;
4354 for (const SCEV *Reg : FA.BaseRegs) {
4355 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4356 FARegNum += (NumUses - UsedByIndices.count() + 1);
4358 size_t FBRegNum = 0;
4359 for (const SCEV *Reg : FB.BaseRegs) {
4360 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4361 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4363 if (FARegNum != FBRegNum)
4364 return FARegNum < FBRegNum;
4366 // If the new register numbers are the same, choose the Formula with
4368 Cost CostFA, CostFB;
4370 CostFA.RateFormula(TTI, FA, Regs, VisitedRegs, L, SE, DT, LU);
4372 CostFB.RateFormula(TTI, FB, Regs, VisitedRegs, L, SE, DT, LU);
4373 return CostFA.isLess(CostFB, TTI);
4377 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4379 Formula &F = LU.Formulae[FIdx];
4382 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4386 Formula &Best = LU.Formulae[P.first->second];
4387 if (IsBetterThan(F, Best))
4389 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4391 " in favor of formula ";
4392 Best.print(dbgs()); dbgs() << '\n');
4394 ChangedFormulae = true;
4396 LU.DeleteFormula(F);
4402 LU.RecomputeRegs(LUIdx, RegUses);
4404 // Reset this to prepare for the next use.
4405 BestFormulae.clear();
4408 DEBUG(if (ChangedFormulae) {
4410 "After filtering out undesirable candidates:\n";
4415 /// The function delete formulas with high registers number expectation.
4416 /// Assuming we don't know the value of each formula (already delete
4417 /// all inefficient), generate probability of not selecting for each
4421 /// reg(a) + reg({0,+,1})
4422 /// reg(a) + reg({-1,+,1}) + 1
4425 /// reg(b) + reg({0,+,1})
4426 /// reg(b) + reg({-1,+,1}) + 1
4429 /// reg(c) + reg(b) + reg({0,+,1})
4430 /// reg(c) + reg({b,+,1})
4432 /// Probability of not selecting
4434 /// reg(a) (1/3) * 1 * 1
4435 /// reg(b) 1 * (1/3) * (1/2)
4436 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4437 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4438 /// reg({a,+,1}) (2/3) * 1 * 1
4439 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4440 /// reg(c) 1 * 1 * 0
4442 /// Now count registers number mathematical expectation for each formula:
4443 /// Note that for each use we exclude probability if not selecting for the use.
4444 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4445 /// probabilty 1/3 of not selecting for Use1).
4447 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4448 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4451 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4452 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4453 /// reg({b,+,1}) 2/3
4455 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4456 /// reg(c) + reg({b,+,1}) 1 + 2/3
4458 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4459 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4461 // Ok, we have too many of formulae on our hands to conveniently handle.
4462 // Use a rough heuristic to thin out the list.
4464 // Set of Regs wich will be 100% used in final solution.
4465 // Used in each formula of a solution (in example above this is reg(c)).
4466 // We can skip them in calculations.
4467 SmallPtrSet<const SCEV *, 4> UniqRegs;
4468 DEBUG(dbgs() << "The search space is too complex.\n");
4470 // Map each register to probability of not selecting
4471 DenseMap <const SCEV *, float> RegNumMap;
4472 for (const SCEV *Reg : RegUses) {
4473 if (UniqRegs.count(Reg))
4476 for (const LSRUse &LU : Uses) {
4477 if (!LU.Regs.count(Reg))
4479 float P = LU.getNotSelectedProbability(Reg);
4483 UniqRegs.insert(Reg);
4485 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4488 DEBUG(dbgs() << "Narrowing the search space by deleting costly formulas\n");
4490 // Delete formulas where registers number expectation is high.
4491 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4492 LSRUse &LU = Uses[LUIdx];
4493 // If nothing to delete - continue.
4494 if (LU.Formulae.size() < 2)
4496 // This is temporary solution to test performance. Float should be
4497 // replaced with round independent type (based on integers) to avoid
4498 // different results for different target builds.
4499 float FMinRegNum = LU.Formulae[0].getNumRegs();
4500 float FMinARegNum = LU.Formulae[0].getNumRegs();
4502 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4503 Formula &F = LU.Formulae[i];
4506 for (const SCEV *BaseReg : F.BaseRegs) {
4507 if (UniqRegs.count(BaseReg))
4509 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4510 if (isa<SCEVAddRecExpr>(BaseReg))
4512 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4514 if (const SCEV *ScaledReg = F.ScaledReg) {
4515 if (!UniqRegs.count(ScaledReg)) {
4517 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4518 if (isa<SCEVAddRecExpr>(ScaledReg))
4520 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4523 if (FMinRegNum > FRegNum ||
4524 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4525 FMinRegNum = FRegNum;
4526 FMinARegNum = FARegNum;
4530 DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4531 dbgs() << " with min reg num " << FMinRegNum << '\n');
4533 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4534 while (LU.Formulae.size() != 1) {
4535 DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4537 LU.Formulae.pop_back();
4539 LU.RecomputeRegs(LUIdx, RegUses);
4540 assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4541 Formula &F = LU.Formulae[0];
4542 DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4543 // When we choose the formula, the regs become unique.
4544 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4546 UniqRegs.insert(F.ScaledReg);
4548 DEBUG(dbgs() << "After pre-selection:\n";
4549 print_uses(dbgs()));
4553 /// Pick a register which seems likely to be profitable, and then in any use
4554 /// which has any reference to that register, delete all formulae which do not
4555 /// reference that register.
4556 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4557 // With all other options exhausted, loop until the system is simple
4558 // enough to handle.
4559 SmallPtrSet<const SCEV *, 4> Taken;
4560 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4561 // Ok, we have too many of formulae on our hands to conveniently handle.
4562 // Use a rough heuristic to thin out the list.
4563 DEBUG(dbgs() << "The search space is too complex.\n");
4565 // Pick the register which is used by the most LSRUses, which is likely
4566 // to be a good reuse register candidate.
4567 const SCEV *Best = nullptr;
4568 unsigned BestNum = 0;
4569 for (const SCEV *Reg : RegUses) {
4570 if (Taken.count(Reg))
4574 BestNum = RegUses.getUsedByIndices(Reg).count();
4576 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4577 if (Count > BestNum) {
4584 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4585 << " will yield profitable reuse.\n");
4588 // In any use with formulae which references this register, delete formulae
4589 // which don't reference it.
4590 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4591 LSRUse &LU = Uses[LUIdx];
4592 if (!LU.Regs.count(Best)) continue;
4595 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4596 Formula &F = LU.Formulae[i];
4597 if (!F.referencesReg(Best)) {
4598 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4599 LU.DeleteFormula(F);
4603 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4609 LU.RecomputeRegs(LUIdx, RegUses);
4612 DEBUG(dbgs() << "After pre-selection:\n";
4613 print_uses(dbgs()));
4617 /// If there are an extraordinary number of formulae to choose from, use some
4618 /// rough heuristics to prune down the number of formulae. This keeps the main
4619 /// solver from taking an extraordinary amount of time in some worst-case
4621 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4622 NarrowSearchSpaceByDetectingSupersets();
4623 NarrowSearchSpaceByCollapsingUnrolledCode();
4624 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4625 if (FilterSameScaledReg)
4626 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4628 NarrowSearchSpaceByDeletingCostlyFormulas();
4630 NarrowSearchSpaceByPickingWinnerRegs();
4633 /// This is the recursive solver.
4634 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4636 SmallVectorImpl<const Formula *> &Workspace,
4637 const Cost &CurCost,
4638 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4639 DenseSet<const SCEV *> &VisitedRegs) const {
4642 // - use more aggressive filtering
4643 // - sort the formula so that the most profitable solutions are found first
4644 // - sort the uses too
4646 // - don't compute a cost, and then compare. compare while computing a cost
4648 // - track register sets with SmallBitVector
4650 const LSRUse &LU = Uses[Workspace.size()];
4652 // If this use references any register that's already a part of the
4653 // in-progress solution, consider it a requirement that a formula must
4654 // reference that register in order to be considered. This prunes out
4655 // unprofitable searching.
4656 SmallSetVector<const SCEV *, 4> ReqRegs;
4657 for (const SCEV *S : CurRegs)
4658 if (LU.Regs.count(S))
4661 SmallPtrSet<const SCEV *, 16> NewRegs;
4663 for (const Formula &F : LU.Formulae) {
4664 // Ignore formulae which may not be ideal in terms of register reuse of
4665 // ReqRegs. The formula should use all required registers before
4666 // introducing new ones.
4667 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4668 for (const SCEV *Reg : ReqRegs) {
4669 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4670 is_contained(F.BaseRegs, Reg)) {
4672 if (NumReqRegsToFind == 0)
4676 if (NumReqRegsToFind != 0) {
4677 // If none of the formulae satisfied the required registers, then we could
4678 // clear ReqRegs and try again. Currently, we simply give up in this case.
4682 // Evaluate the cost of the current formula. If it's already worse than
4683 // the current best, prune the search at that point.
4686 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, SE, DT, LU);
4687 if (NewCost.isLess(SolutionCost, TTI)) {
4688 Workspace.push_back(&F);
4689 if (Workspace.size() != Uses.size()) {
4690 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4691 NewRegs, VisitedRegs);
4692 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4693 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4695 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4696 dbgs() << ".\n Regs:";
4697 for (const SCEV *S : NewRegs)
4698 dbgs() << ' ' << *S;
4701 SolutionCost = NewCost;
4702 Solution = Workspace;
4704 Workspace.pop_back();
4709 /// Choose one formula from each use. Return the results in the given Solution
4711 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4712 SmallVector<const Formula *, 8> Workspace;
4714 SolutionCost.Lose();
4716 SmallPtrSet<const SCEV *, 16> CurRegs;
4717 DenseSet<const SCEV *> VisitedRegs;
4718 Workspace.reserve(Uses.size());
4720 // SolveRecurse does all the work.
4721 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4722 CurRegs, VisitedRegs);
4723 if (Solution.empty()) {
4724 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4728 // Ok, we've now made all our decisions.
4729 DEBUG(dbgs() << "\n"
4730 "The chosen solution requires "; SolutionCost.print(dbgs());
4732 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4734 Uses[i].print(dbgs());
4737 Solution[i]->print(dbgs());
4741 assert(Solution.size() == Uses.size() && "Malformed solution!");
4744 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4745 /// we can go while still being dominated by the input positions. This helps
4746 /// canonicalize the insert position, which encourages sharing.
4747 BasicBlock::iterator
4748 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4749 const SmallVectorImpl<Instruction *> &Inputs)
4751 Instruction *Tentative = &*IP;
4753 bool AllDominate = true;
4754 Instruction *BetterPos = nullptr;
4755 // Don't bother attempting to insert before a catchswitch, their basic block
4756 // cannot have other non-PHI instructions.
4757 if (isa<CatchSwitchInst>(Tentative))
4760 for (Instruction *Inst : Inputs) {
4761 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4762 AllDominate = false;
4765 // Attempt to find an insert position in the middle of the block,
4766 // instead of at the end, so that it can be used for other expansions.
4767 if (Tentative->getParent() == Inst->getParent() &&
4768 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4769 BetterPos = &*std::next(BasicBlock::iterator(Inst));
4774 IP = BetterPos->getIterator();
4776 IP = Tentative->getIterator();
4778 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4779 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4782 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4783 if (!Rung) return IP;
4784 Rung = Rung->getIDom();
4785 if (!Rung) return IP;
4786 IDom = Rung->getBlock();
4788 // Don't climb into a loop though.
4789 const Loop *IDomLoop = LI.getLoopFor(IDom);
4790 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4791 if (IDomDepth <= IPLoopDepth &&
4792 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4796 Tentative = IDom->getTerminator();
4802 /// Determine an input position which will be dominated by the operands and
4803 /// which will dominate the result.
4804 BasicBlock::iterator
4805 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4808 SCEVExpander &Rewriter) const {
4809 // Collect some instructions which must be dominated by the
4810 // expanding replacement. These must be dominated by any operands that
4811 // will be required in the expansion.
4812 SmallVector<Instruction *, 4> Inputs;
4813 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4814 Inputs.push_back(I);
4815 if (LU.Kind == LSRUse::ICmpZero)
4816 if (Instruction *I =
4817 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4818 Inputs.push_back(I);
4819 if (LF.PostIncLoops.count(L)) {
4820 if (LF.isUseFullyOutsideLoop(L))
4821 Inputs.push_back(L->getLoopLatch()->getTerminator());
4823 Inputs.push_back(IVIncInsertPos);
4825 // The expansion must also be dominated by the increment positions of any
4826 // loops it for which it is using post-inc mode.
4827 for (const Loop *PIL : LF.PostIncLoops) {
4828 if (PIL == L) continue;
4830 // Be dominated by the loop exit.
4831 SmallVector<BasicBlock *, 4> ExitingBlocks;
4832 PIL->getExitingBlocks(ExitingBlocks);
4833 if (!ExitingBlocks.empty()) {
4834 BasicBlock *BB = ExitingBlocks[0];
4835 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4836 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4837 Inputs.push_back(BB->getTerminator());
4841 assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
4842 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4843 "Insertion point must be a normal instruction");
4845 // Then, climb up the immediate dominator tree as far as we can go while
4846 // still being dominated by the input positions.
4847 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4849 // Don't insert instructions before PHI nodes.
4850 while (isa<PHINode>(IP)) ++IP;
4852 // Ignore landingpad instructions.
4853 while (IP->isEHPad()) ++IP;
4855 // Ignore debug intrinsics.
4856 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4858 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4859 // IP consistent across expansions and allows the previously inserted
4860 // instructions to be reused by subsequent expansion.
4861 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
4867 /// Emit instructions for the leading candidate expression for this LSRUse (this
4868 /// is called "expanding").
4869 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
4870 const Formula &F, BasicBlock::iterator IP,
4871 SCEVExpander &Rewriter,
4872 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
4873 if (LU.RigidFormula)
4874 return LF.OperandValToReplace;
4876 // Determine an input position which will be dominated by the operands and
4877 // which will dominate the result.
4878 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4879 Rewriter.setInsertPoint(&*IP);
4881 // Inform the Rewriter if we have a post-increment use, so that it can
4882 // perform an advantageous expansion.
4883 Rewriter.setPostInc(LF.PostIncLoops);
4885 // This is the type that the user actually needs.
4886 Type *OpTy = LF.OperandValToReplace->getType();
4887 // This will be the type that we'll initially expand to.
4888 Type *Ty = F.getType();
4890 // No type known; just expand directly to the ultimate type.
4892 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4893 // Expand directly to the ultimate type if it's the right size.
4895 // This is the type to do integer arithmetic in.
4896 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4898 // Build up a list of operands to add together to form the full base.
4899 SmallVector<const SCEV *, 8> Ops;
4901 // Expand the BaseRegs portion.
4902 for (const SCEV *Reg : F.BaseRegs) {
4903 assert(!Reg->isZero() && "Zero allocated in a base register!");
4905 // If we're expanding for a post-inc user, make the post-inc adjustment.
4906 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
4907 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
4910 // Expand the ScaledReg portion.
4911 Value *ICmpScaledV = nullptr;
4913 const SCEV *ScaledS = F.ScaledReg;
4915 // If we're expanding for a post-inc user, make the post-inc adjustment.
4916 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4917 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
4919 if (LU.Kind == LSRUse::ICmpZero) {
4920 // Expand ScaleReg as if it was part of the base regs.
4923 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
4925 // An interesting way of "folding" with an icmp is to use a negated
4926 // scale, which we'll implement by inserting it into the other operand
4928 assert(F.Scale == -1 &&
4929 "The only scale supported by ICmpZero uses is -1!");
4930 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
4933 // Otherwise just expand the scaled register and an explicit scale,
4934 // which is expected to be matched as part of the address.
4936 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4937 // Unless the addressing mode will not be folded.
4938 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4939 isAMCompletelyFolded(TTI, LU, F)) {
4940 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4942 Ops.push_back(SE.getUnknown(FullV));
4944 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
4947 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4948 Ops.push_back(ScaledS);
4952 // Expand the GV portion.
4954 // Flush the operand list to suppress SCEVExpander hoisting.
4956 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4958 Ops.push_back(SE.getUnknown(FullV));
4960 Ops.push_back(SE.getUnknown(F.BaseGV));
4963 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4964 // unfolded offsets. LSR assumes they both live next to their uses.
4966 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4968 Ops.push_back(SE.getUnknown(FullV));
4971 // Expand the immediate portion.
4972 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4974 if (LU.Kind == LSRUse::ICmpZero) {
4975 // The other interesting way of "folding" with an ICmpZero is to use a
4976 // negated immediate.
4978 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4980 Ops.push_back(SE.getUnknown(ICmpScaledV));
4981 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4984 // Just add the immediate values. These again are expected to be matched
4985 // as part of the address.
4986 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4990 // Expand the unfolded offset portion.
4991 int64_t UnfoldedOffset = F.UnfoldedOffset;
4992 if (UnfoldedOffset != 0) {
4993 // Just add the immediate values.
4994 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4998 // Emit instructions summing all the operands.
4999 const SCEV *FullS = Ops.empty() ?
5000 SE.getConstant(IntTy, 0) :
5002 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5004 // We're done expanding now, so reset the rewriter.
5005 Rewriter.clearPostInc();
5007 // An ICmpZero Formula represents an ICmp which we're handling as a
5008 // comparison against zero. Now that we've expanded an expression for that
5009 // form, update the ICmp's other operand.
5010 if (LU.Kind == LSRUse::ICmpZero) {
5011 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5012 DeadInsts.emplace_back(CI->getOperand(1));
5013 assert(!F.BaseGV && "ICmp does not support folding a global value and "
5014 "a scale at the same time!");
5015 if (F.Scale == -1) {
5016 if (ICmpScaledV->getType() != OpTy) {
5018 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5020 ICmpScaledV, OpTy, "tmp", CI);
5023 CI->setOperand(1, ICmpScaledV);
5025 // A scale of 1 means that the scale has been expanded as part of the
5027 assert((F.Scale == 0 || F.Scale == 1) &&
5028 "ICmp does not support folding a global value and "
5029 "a scale at the same time!");
5030 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5032 if (C->getType() != OpTy)
5033 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5037 CI->setOperand(1, C);
5044 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5045 /// effectively happens in their predecessor blocks, so the expression may need
5046 /// to be expanded in multiple places.
5047 void LSRInstance::RewriteForPHI(
5048 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5049 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5050 DenseMap<BasicBlock *, Value *> Inserted;
5051 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5052 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5053 BasicBlock *BB = PN->getIncomingBlock(i);
5055 // If this is a critical edge, split the edge so that we do not insert
5056 // the code on all predecessor/successor paths. We do this unless this
5057 // is the canonical backedge for this loop, which complicates post-inc
5059 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5060 !isa<IndirectBrInst>(BB->getTerminator()) &&
5061 !isa<CatchSwitchInst>(BB->getTerminator())) {
5062 BasicBlock *Parent = PN->getParent();
5063 Loop *PNLoop = LI.getLoopFor(Parent);
5064 if (!PNLoop || Parent != PNLoop->getHeader()) {
5065 // Split the critical edge.
5066 BasicBlock *NewBB = nullptr;
5067 if (!Parent->isLandingPad()) {
5068 NewBB = SplitCriticalEdge(BB, Parent,
5069 CriticalEdgeSplittingOptions(&DT, &LI)
5070 .setMergeIdenticalEdges()
5071 .setDontDeleteUselessPHIs());
5073 SmallVector<BasicBlock*, 2> NewBBs;
5074 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5077 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5078 // phi predecessors are identical. The simple thing to do is skip
5079 // splitting in this case rather than complicate the API.
5081 // If PN is outside of the loop and BB is in the loop, we want to
5082 // move the block to be immediately before the PHI block, not
5083 // immediately after BB.
5084 if (L->contains(BB) && !L->contains(PN))
5085 NewBB->moveBefore(PN->getParent());
5087 // Splitting the edge can reduce the number of PHI entries we have.
5088 e = PN->getNumIncomingValues();
5090 i = PN->getBasicBlockIndex(BB);
5095 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5096 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5098 PN->setIncomingValue(i, Pair.first->second);
5100 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5101 Rewriter, DeadInsts);
5103 // If this is reuse-by-noop-cast, insert the noop cast.
5104 Type *OpTy = LF.OperandValToReplace->getType();
5105 if (FullV->getType() != OpTy)
5107 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5109 FullV, LF.OperandValToReplace->getType(),
5110 "tmp", BB->getTerminator());
5112 PN->setIncomingValue(i, FullV);
5113 Pair.first->second = FullV;
5118 /// Emit instructions for the leading candidate expression for this LSRUse (this
5119 /// is called "expanding"), and update the UserInst to reference the newly
5121 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5122 const Formula &F, SCEVExpander &Rewriter,
5123 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5124 // First, find an insertion point that dominates UserInst. For PHI nodes,
5125 // find the nearest block which dominates all the relevant uses.
5126 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5127 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5130 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5132 // If this is reuse-by-noop-cast, insert the noop cast.
5133 Type *OpTy = LF.OperandValToReplace->getType();
5134 if (FullV->getType() != OpTy) {
5136 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5137 FullV, OpTy, "tmp", LF.UserInst);
5141 // Update the user. ICmpZero is handled specially here (for now) because
5142 // Expand may have updated one of the operands of the icmp already, and
5143 // its new value may happen to be equal to LF.OperandValToReplace, in
5144 // which case doing replaceUsesOfWith leads to replacing both operands
5145 // with the same value. TODO: Reorganize this.
5146 if (LU.Kind == LSRUse::ICmpZero)
5147 LF.UserInst->setOperand(0, FullV);
5149 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5152 DeadInsts.emplace_back(LF.OperandValToReplace);
5155 /// Rewrite all the fixup locations with new values, following the chosen
5157 void LSRInstance::ImplementSolution(
5158 const SmallVectorImpl<const Formula *> &Solution) {
5159 // Keep track of instructions we may have made dead, so that
5160 // we can remove them after we are done working.
5161 SmallVector<WeakTrackingVH, 16> DeadInsts;
5163 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
5166 Rewriter.setDebugType(DEBUG_TYPE);
5168 Rewriter.disableCanonicalMode();
5169 Rewriter.enableLSRMode();
5170 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5172 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5173 for (const IVChain &Chain : IVChainVec) {
5174 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5175 Rewriter.setChainedPhi(PN);
5178 // Expand the new value definitions and update the users.
5179 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5180 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5181 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5185 for (const IVChain &Chain : IVChainVec) {
5186 GenerateIVChain(Chain, Rewriter, DeadInsts);
5189 // Clean up after ourselves. This must be done before deleting any
5193 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5196 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5197 DominatorTree &DT, LoopInfo &LI,
5198 const TargetTransformInfo &TTI)
5199 : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L), Changed(false),
5200 IVIncInsertPos(nullptr) {
5201 // If LoopSimplify form is not available, stay out of trouble.
5202 if (!L->isLoopSimplifyForm())
5205 // If there's no interesting work to be done, bail early.
5206 if (IU.empty()) return;
5208 // If there's too much analysis to be done, bail early. We won't be able to
5209 // model the problem anyway.
5210 unsigned NumUsers = 0;
5211 for (const IVStrideUse &U : IU) {
5212 if (++NumUsers > MaxIVUsers) {
5214 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n");
5217 // Bail out if we have a PHI on an EHPad that gets a value from a
5218 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5219 // no good place to stick any instructions.
5220 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5221 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5222 if (isa<FuncletPadInst>(FirstNonPHI) ||
5223 isa<CatchSwitchInst>(FirstNonPHI))
5224 for (BasicBlock *PredBB : PN->blocks())
5225 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5231 // All dominating loops must have preheaders, or SCEVExpander may not be able
5232 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5234 // IVUsers analysis should only create users that are dominated by simple loop
5235 // headers. Since this loop should dominate all of its users, its user list
5236 // should be empty if this loop itself is not within a simple loop nest.
5237 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5238 Rung; Rung = Rung->getIDom()) {
5239 BasicBlock *BB = Rung->getBlock();
5240 const Loop *DomLoop = LI.getLoopFor(BB);
5241 if (DomLoop && DomLoop->getHeader() == BB) {
5242 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
5247 DEBUG(dbgs() << "\nLSR on loop ";
5248 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
5251 // First, perform some low-level loop optimizations.
5253 OptimizeLoopTermCond();
5255 // If loop preparation eliminates all interesting IV users, bail.
5256 if (IU.empty()) return;
5258 // Skip nested loops until we can model them better with formulae.
5260 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
5264 // Start collecting data and preparing for the solver.
5266 CollectInterestingTypesAndFactors();
5267 CollectFixupsAndInitialFormulae();
5268 CollectLoopInvariantFixupsAndFormulae();
5270 assert(!Uses.empty() && "IVUsers reported at least one use");
5271 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
5272 print_uses(dbgs()));
5274 // Now use the reuse data to generate a bunch of interesting ways
5275 // to formulate the values needed for the uses.
5276 GenerateAllReuseFormulae();
5278 FilterOutUndesirableDedicatedRegisters();
5279 NarrowSearchSpaceUsingHeuristics();
5281 SmallVector<const Formula *, 8> Solution;
5284 // Release memory that is no longer needed.
5289 if (Solution.empty())
5293 // Formulae should be legal.
5294 for (const LSRUse &LU : Uses) {
5295 for (const Formula &F : LU.Formulae)
5296 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
5297 F) && "Illegal formula generated!");
5301 // Now that we've decided what we want, make it so.
5302 ImplementSolution(Solution);
5305 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5306 if (Factors.empty() && Types.empty()) return;
5308 OS << "LSR has identified the following interesting factors and types: ";
5311 for (int64_t Factor : Factors) {
5312 if (!First) OS << ", ";
5314 OS << '*' << Factor;
5317 for (Type *Ty : Types) {
5318 if (!First) OS << ", ";
5320 OS << '(' << *Ty << ')';
5325 void LSRInstance::print_fixups(raw_ostream &OS) const {
5326 OS << "LSR is examining the following fixup sites:\n";
5327 for (const LSRUse &LU : Uses)
5328 for (const LSRFixup &LF : LU.Fixups) {
5335 void LSRInstance::print_uses(raw_ostream &OS) const {
5336 OS << "LSR is examining the following uses:\n";
5337 for (const LSRUse &LU : Uses) {
5341 for (const Formula &F : LU.Formulae) {
5349 void LSRInstance::print(raw_ostream &OS) const {
5350 print_factors_and_types(OS);
5355 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5356 LLVM_DUMP_METHOD void LSRInstance::dump() const {
5357 print(errs()); errs() << '\n';
5363 class LoopStrengthReduce : public LoopPass {
5365 static char ID; // Pass ID, replacement for typeid
5367 LoopStrengthReduce();
5370 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5371 void getAnalysisUsage(AnalysisUsage &AU) const override;
5374 } // end anonymous namespace
5376 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5377 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5380 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5381 // We split critical edges, so we change the CFG. However, we do update
5382 // many analyses if they are around.
5383 AU.addPreservedID(LoopSimplifyID);
5385 AU.addRequired<LoopInfoWrapperPass>();
5386 AU.addPreserved<LoopInfoWrapperPass>();
5387 AU.addRequiredID(LoopSimplifyID);
5388 AU.addRequired<DominatorTreeWrapperPass>();
5389 AU.addPreserved<DominatorTreeWrapperPass>();
5390 AU.addRequired<ScalarEvolutionWrapperPass>();
5391 AU.addPreserved<ScalarEvolutionWrapperPass>();
5392 // Requiring LoopSimplify a second time here prevents IVUsers from running
5393 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5394 AU.addRequiredID(LoopSimplifyID);
5395 AU.addRequired<IVUsersWrapperPass>();
5396 AU.addPreserved<IVUsersWrapperPass>();
5397 AU.addRequired<TargetTransformInfoWrapperPass>();
5400 static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5401 DominatorTree &DT, LoopInfo &LI,
5402 const TargetTransformInfo &TTI) {
5403 bool Changed = false;
5405 // Run the main LSR transformation.
5406 Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5408 // Remove any extra phis created by processing inner loops.
5409 Changed |= DeleteDeadPHIs(L->getHeader());
5410 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5411 SmallVector<WeakTrackingVH, 16> DeadInsts;
5412 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5413 SCEVExpander Rewriter(SE, DL, "lsr");
5415 Rewriter.setDebugType(DEBUG_TYPE);
5417 unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5420 DeleteTriviallyDeadInstructions(DeadInsts);
5421 DeleteDeadPHIs(L->getHeader());
5427 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5431 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5432 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5433 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5434 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5435 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5436 *L->getHeader()->getParent());
5437 return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5440 PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
5441 LoopStandardAnalysisResults &AR,
5443 if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5444 AR.DT, AR.LI, AR.TTI))
5445 return PreservedAnalyses::all();
5447 return getLoopPassPreservedAnalyses();
5450 char LoopStrengthReduce::ID = 0;
5451 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5452 "Loop Strength Reduction", false, false)
5453 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5454 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5455 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5456 INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)
5457 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5458 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5459 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5460 "Loop Strength Reduction", false, false)
5462 Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }