1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #include "llvm/Transforms/Scalar.h"
57 #include "llvm/ADT/DenseSet.h"
58 #include "llvm/ADT/Hashing.h"
59 #include "llvm/ADT/STLExtras.h"
60 #include "llvm/ADT/SetVector.h"
61 #include "llvm/ADT/SmallBitVector.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/LoopPass.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/IR/Constants.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/Instructions.h"
70 #include "llvm/IR/IntrinsicInst.h"
71 #include "llvm/IR/Module.h"
72 #include "llvm/IR/ValueHandle.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/Debug.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
81 #define DEBUG_TYPE "loop-reduce"
83 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
84 /// bail out. This threshold is far beyond the number of users that LSR can
85 /// conceivably solve, so it should not affect generated code, but catches the
86 /// worst cases before LSR burns too much compile time and stack space.
87 static const unsigned MaxIVUsers = 200;
89 // Temporary flag to cleanup congruent phis after LSR phi expansion.
90 // It's currently disabled until we can determine whether it's truly useful or
91 // not. The flag should be removed after the v3.0 release.
92 // This is now needed for ivchains.
93 static cl::opt<bool> EnablePhiElim(
94 "enable-lsr-phielim", cl::Hidden, cl::init(true),
95 cl::desc("Enable LSR phi elimination"));
98 // Stress test IV chain generation.
99 static cl::opt<bool> StressIVChain(
100 "stress-ivchain", cl::Hidden, cl::init(false),
101 cl::desc("Stress test LSR IV chains"));
103 static bool StressIVChain = false;
108 /// RegSortData - This class holds data which is used to order reuse candidates.
111 /// UsedByIndices - This represents the set of LSRUse indices which reference
112 /// a particular register.
113 SmallBitVector UsedByIndices;
115 void print(raw_ostream &OS) const;
121 void RegSortData::print(raw_ostream &OS) const {
122 OS << "[NumUses=" << UsedByIndices.count() << ']';
125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
126 void RegSortData::dump() const {
127 print(errs()); errs() << '\n';
133 /// RegUseTracker - Map register candidates to information about how they are
135 class RegUseTracker {
136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
138 RegUsesTy RegUsesMap;
139 SmallVector<const SCEV *, 16> RegSequence;
142 void CountRegister(const SCEV *Reg, size_t LUIdx);
143 void DropRegister(const SCEV *Reg, size_t LUIdx);
144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
152 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
154 iterator begin() { return RegSequence.begin(); }
155 iterator end() { return RegSequence.end(); }
156 const_iterator begin() const { return RegSequence.begin(); }
157 const_iterator end() const { return RegSequence.end(); }
163 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
164 std::pair<RegUsesTy::iterator, bool> Pair =
165 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
166 RegSortData &RSD = Pair.first->second;
168 RegSequence.push_back(Reg);
169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
170 RSD.UsedByIndices.set(LUIdx);
174 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
175 RegUsesTy::iterator It = RegUsesMap.find(Reg);
176 assert(It != RegUsesMap.end());
177 RegSortData &RSD = It->second;
178 assert(RSD.UsedByIndices.size() > LUIdx);
179 RSD.UsedByIndices.reset(LUIdx);
183 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
184 assert(LUIdx <= LastLUIdx);
186 // Update RegUses. The data structure is not optimized for this purpose;
187 // we must iterate through it and update each of the bit vectors.
188 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
190 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
191 if (LUIdx < UsedByIndices.size())
192 UsedByIndices[LUIdx] =
193 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
194 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
199 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
200 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
201 if (I == RegUsesMap.end())
203 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
204 int i = UsedByIndices.find_first();
205 if (i == -1) return false;
206 if ((size_t)i != LUIdx) return true;
207 return UsedByIndices.find_next(i) != -1;
210 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
211 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
212 assert(I != RegUsesMap.end() && "Unknown register!");
213 return I->second.UsedByIndices;
216 void RegUseTracker::clear() {
223 /// Formula - This class holds information that describes a formula for
224 /// computing satisfying a use. It may include broken-out immediates and scaled
227 /// Global base address used for complex addressing.
230 /// Base offset for complex addressing.
233 /// Whether any complex addressing has a base register.
236 /// The scale of any complex addressing.
239 /// BaseRegs - The list of "base" registers for this use. When this is
240 /// non-empty. The canonical representation of a formula is
241 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
242 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
243 /// #1 enforces that the scaled register is always used when at least two
244 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
245 /// #2 enforces that 1 * reg is reg.
246 /// This invariant can be temporarly broken while building a formula.
247 /// However, every formula inserted into the LSRInstance must be in canonical
249 SmallVector<const SCEV *, 4> BaseRegs;
251 /// ScaledReg - The 'scaled' register for this use. This should be non-null
252 /// when Scale is not zero.
253 const SCEV *ScaledReg;
255 /// UnfoldedOffset - An additional constant offset which added near the
256 /// use. This requires a temporary register, but the offset itself can
257 /// live in an add immediate field rather than a register.
258 int64_t UnfoldedOffset;
261 : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
262 ScaledReg(nullptr), UnfoldedOffset(0) {}
264 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
266 bool isCanonical() const;
272 size_t getNumRegs() const;
273 Type *getType() const;
275 void DeleteBaseReg(const SCEV *&S);
277 bool referencesReg(const SCEV *S) const;
278 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
279 const RegUseTracker &RegUses) const;
281 void print(raw_ostream &OS) const;
287 /// DoInitialMatch - Recursion helper for InitialMatch.
288 static void DoInitialMatch(const SCEV *S, Loop *L,
289 SmallVectorImpl<const SCEV *> &Good,
290 SmallVectorImpl<const SCEV *> &Bad,
291 ScalarEvolution &SE) {
292 // Collect expressions which properly dominate the loop header.
293 if (SE.properlyDominates(S, L->getHeader())) {
298 // Look at add operands.
299 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
300 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
302 DoInitialMatch(*I, L, Good, Bad, SE);
306 // Look at addrec operands.
307 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
308 if (!AR->getStart()->isZero()) {
309 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
310 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
311 AR->getStepRecurrence(SE),
312 // FIXME: AR->getNoWrapFlags()
313 AR->getLoop(), SCEV::FlagAnyWrap),
318 // Handle a multiplication by -1 (negation) if it didn't fold.
319 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
320 if (Mul->getOperand(0)->isAllOnesValue()) {
321 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
322 const SCEV *NewMul = SE.getMulExpr(Ops);
324 SmallVector<const SCEV *, 4> MyGood;
325 SmallVector<const SCEV *, 4> MyBad;
326 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
327 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
328 SE.getEffectiveSCEVType(NewMul->getType())));
329 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
330 E = MyGood.end(); I != E; ++I)
331 Good.push_back(SE.getMulExpr(NegOne, *I));
332 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
333 E = MyBad.end(); I != E; ++I)
334 Bad.push_back(SE.getMulExpr(NegOne, *I));
338 // Ok, we can't do anything interesting. Just stuff the whole thing into a
339 // register and hope for the best.
343 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
344 /// attempting to keep all loop-invariant and loop-computable values in a
345 /// single base register.
346 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
347 SmallVector<const SCEV *, 4> Good;
348 SmallVector<const SCEV *, 4> Bad;
349 DoInitialMatch(S, L, Good, Bad, SE);
351 const SCEV *Sum = SE.getAddExpr(Good);
353 BaseRegs.push_back(Sum);
357 const SCEV *Sum = SE.getAddExpr(Bad);
359 BaseRegs.push_back(Sum);
365 /// \brief Check whether or not this formula statisfies the canonical
367 /// \see Formula::BaseRegs.
368 bool Formula::isCanonical() const {
370 return Scale != 1 || !BaseRegs.empty();
371 return BaseRegs.size() <= 1;
374 /// \brief Helper method to morph a formula into its canonical representation.
375 /// \see Formula::BaseRegs.
376 /// Every formula having more than one base register, must use the ScaledReg
377 /// field. Otherwise, we would have to do special cases everywhere in LSR
378 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
379 /// On the other hand, 1*reg should be canonicalized into reg.
380 void Formula::Canonicalize() {
383 // So far we did not need this case. This is easy to implement but it is
384 // useless to maintain dead code. Beside it could hurt compile time.
385 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
386 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
387 ScaledReg = BaseRegs.back();
390 size_t BaseRegsSize = BaseRegs.size();
392 // If ScaledReg is an invariant, try to find a variant expression.
393 while (Try < BaseRegsSize && !isa<SCEVAddRecExpr>(ScaledReg))
394 std::swap(ScaledReg, BaseRegs[Try++]);
397 /// \brief Get rid of the scale in the formula.
398 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
399 /// \return true if it was possible to get rid of the scale, false otherwise.
400 /// \note After this operation the formula may not be in the canonical form.
401 bool Formula::Unscale() {
405 BaseRegs.push_back(ScaledReg);
410 /// getNumRegs - Return the total number of register operands used by this
411 /// formula. This does not include register uses implied by non-constant
413 size_t Formula::getNumRegs() const {
414 return !!ScaledReg + BaseRegs.size();
417 /// getType - Return the type of this formula, if it has one, or null
418 /// otherwise. This type is meaningless except for the bit size.
419 Type *Formula::getType() const {
420 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
421 ScaledReg ? ScaledReg->getType() :
422 BaseGV ? BaseGV->getType() :
426 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
427 void Formula::DeleteBaseReg(const SCEV *&S) {
428 if (&S != &BaseRegs.back())
429 std::swap(S, BaseRegs.back());
433 /// referencesReg - Test if this formula references the given register.
434 bool Formula::referencesReg(const SCEV *S) const {
435 return S == ScaledReg ||
436 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
439 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
440 /// which are used by uses other than the use with the given index.
441 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
442 const RegUseTracker &RegUses) const {
444 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
446 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
447 E = BaseRegs.end(); I != E; ++I)
448 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
453 void Formula::print(raw_ostream &OS) const {
456 if (!First) OS << " + "; else First = false;
457 BaseGV->printAsOperand(OS, /*PrintType=*/false);
459 if (BaseOffset != 0) {
460 if (!First) OS << " + "; else First = false;
463 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
464 E = BaseRegs.end(); I != E; ++I) {
465 if (!First) OS << " + "; else First = false;
466 OS << "reg(" << **I << ')';
468 if (HasBaseReg && BaseRegs.empty()) {
469 if (!First) OS << " + "; else First = false;
470 OS << "**error: HasBaseReg**";
471 } else if (!HasBaseReg && !BaseRegs.empty()) {
472 if (!First) OS << " + "; else First = false;
473 OS << "**error: !HasBaseReg**";
476 if (!First) OS << " + "; else First = false;
477 OS << Scale << "*reg(";
484 if (UnfoldedOffset != 0) {
485 if (!First) OS << " + ";
486 OS << "imm(" << UnfoldedOffset << ')';
490 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
491 void Formula::dump() const {
492 print(errs()); errs() << '\n';
496 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
497 /// without changing its value.
498 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
500 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
501 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
504 /// isAddSExtable - Return true if the given add can be sign-extended
505 /// without changing its value.
506 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
508 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
509 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
512 /// isMulSExtable - Return true if the given mul can be sign-extended
513 /// without changing its value.
514 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
516 IntegerType::get(SE.getContext(),
517 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
518 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
521 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
522 /// and if the remainder is known to be zero, or null otherwise. If
523 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
524 /// to Y, ignoring that the multiplication may overflow, which is useful when
525 /// the result will be used in a context where the most significant bits are
527 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
529 bool IgnoreSignificantBits = false) {
530 // Handle the trivial case, which works for any SCEV type.
532 return SE.getConstant(LHS->getType(), 1);
534 // Handle a few RHS special cases.
535 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
537 const APInt &RA = RC->getValue()->getValue();
538 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
540 if (RA.isAllOnesValue())
541 return SE.getMulExpr(LHS, RC);
542 // Handle x /s 1 as x.
547 // Check for a division of a constant by a constant.
548 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
551 const APInt &LA = C->getValue()->getValue();
552 const APInt &RA = RC->getValue()->getValue();
553 if (LA.srem(RA) != 0)
555 return SE.getConstant(LA.sdiv(RA));
558 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
559 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
560 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
561 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
562 IgnoreSignificantBits);
563 if (!Step) return nullptr;
564 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
565 IgnoreSignificantBits);
566 if (!Start) return nullptr;
567 // FlagNW is independent of the start value, step direction, and is
568 // preserved with smaller magnitude steps.
569 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
570 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
575 // Distribute the sdiv over add operands, if the add doesn't overflow.
576 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
577 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
578 SmallVector<const SCEV *, 8> Ops;
579 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
581 const SCEV *Op = getExactSDiv(*I, RHS, SE,
582 IgnoreSignificantBits);
583 if (!Op) return nullptr;
586 return SE.getAddExpr(Ops);
591 // Check for a multiply operand that we can pull RHS out of.
592 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
593 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
594 SmallVector<const SCEV *, 4> Ops;
596 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
600 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
601 IgnoreSignificantBits)) {
607 return Found ? SE.getMulExpr(Ops) : nullptr;
612 // Otherwise we don't know.
616 /// ExtractImmediate - If S involves the addition of a constant integer value,
617 /// return that integer value, and mutate S to point to a new SCEV with that
619 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
620 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
621 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
622 S = SE.getConstant(C->getType(), 0);
623 return C->getValue()->getSExtValue();
625 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
626 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
627 int64_t Result = ExtractImmediate(NewOps.front(), SE);
629 S = SE.getAddExpr(NewOps);
631 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
632 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
633 int64_t Result = ExtractImmediate(NewOps.front(), SE);
635 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
636 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
643 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
644 /// return that symbol, and mutate S to point to a new SCEV with that
646 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
647 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
648 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
649 S = SE.getConstant(GV->getType(), 0);
652 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
653 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
654 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
656 S = SE.getAddExpr(NewOps);
658 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
659 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
660 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
662 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
663 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
670 /// isAddressUse - Returns true if the specified instruction is using the
671 /// specified value as an address.
672 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
673 bool isAddress = isa<LoadInst>(Inst);
674 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
675 if (SI->getOperand(1) == OperandVal)
677 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
678 // Addressing modes can also be folded into prefetches and a variety
680 switch (II->getIntrinsicID()) {
682 case Intrinsic::prefetch:
683 case Intrinsic::x86_sse_storeu_ps:
684 case Intrinsic::x86_sse2_storeu_pd:
685 case Intrinsic::x86_sse2_storeu_dq:
686 case Intrinsic::x86_sse2_storel_dq:
687 if (II->getArgOperand(0) == OperandVal)
695 /// getAccessType - Return the type of the memory being accessed.
696 static Type *getAccessType(const Instruction *Inst) {
697 Type *AccessTy = Inst->getType();
698 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
699 AccessTy = SI->getOperand(0)->getType();
700 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
701 // Addressing modes can also be folded into prefetches and a variety
703 switch (II->getIntrinsicID()) {
705 case Intrinsic::x86_sse_storeu_ps:
706 case Intrinsic::x86_sse2_storeu_pd:
707 case Intrinsic::x86_sse2_storeu_dq:
708 case Intrinsic::x86_sse2_storel_dq:
709 AccessTy = II->getArgOperand(0)->getType();
714 // All pointers have the same requirements, so canonicalize them to an
715 // arbitrary pointer type to minimize variation.
716 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
717 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
718 PTy->getAddressSpace());
723 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
724 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
725 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
726 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
727 if (SE.isSCEVable(PN->getType()) &&
728 (SE.getEffectiveSCEVType(PN->getType()) ==
729 SE.getEffectiveSCEVType(AR->getType())) &&
730 SE.getSCEV(PN) == AR)
736 /// Check if expanding this expression is likely to incur significant cost. This
737 /// is tricky because SCEV doesn't track which expressions are actually computed
738 /// by the current IR.
740 /// We currently allow expansion of IV increments that involve adds,
741 /// multiplication by constants, and AddRecs from existing phis.
743 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
744 /// obvious multiple of the UDivExpr.
745 static bool isHighCostExpansion(const SCEV *S,
746 SmallPtrSetImpl<const SCEV*> &Processed,
747 ScalarEvolution &SE) {
748 // Zero/One operand expressions
749 switch (S->getSCEVType()) {
754 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
757 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
760 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
764 if (!Processed.insert(S).second)
767 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
768 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
770 if (isHighCostExpansion(*I, Processed, SE))
776 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
777 if (Mul->getNumOperands() == 2) {
778 // Multiplication by a constant is ok
779 if (isa<SCEVConstant>(Mul->getOperand(0)))
780 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
782 // If we have the value of one operand, check if an existing
783 // multiplication already generates this expression.
784 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
785 Value *UVal = U->getValue();
786 for (User *UR : UVal->users()) {
787 // If U is a constant, it may be used by a ConstantExpr.
788 Instruction *UI = dyn_cast<Instruction>(UR);
789 if (UI && UI->getOpcode() == Instruction::Mul &&
790 SE.isSCEVable(UI->getType())) {
791 return SE.getSCEV(UI) == Mul;
798 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
799 if (isExistingPhi(AR, SE))
803 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
807 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
808 /// specified set are trivially dead, delete them and see if this makes any of
809 /// their operands subsequently dead.
811 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
812 bool Changed = false;
814 while (!DeadInsts.empty()) {
815 Value *V = DeadInsts.pop_back_val();
816 Instruction *I = dyn_cast_or_null<Instruction>(V);
818 if (!I || !isInstructionTriviallyDead(I))
821 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
822 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
825 DeadInsts.push_back(U);
828 I->eraseFromParent();
839 /// \brief Check if the addressing mode defined by \p F is completely
840 /// folded in \p LU at isel time.
841 /// This includes address-mode folding and special icmp tricks.
842 /// This function returns true if \p LU can accommodate what \p F
843 /// defines and up to 1 base + 1 scaled + offset.
844 /// In other words, if \p F has several base registers, this function may
845 /// still return true. Therefore, users still need to account for
846 /// additional base registers and/or unfolded offsets to derive an
847 /// accurate cost model.
848 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
849 const LSRUse &LU, const Formula &F);
850 // Get the cost of the scaling factor used in F for LU.
851 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
852 const LSRUse &LU, const Formula &F);
856 /// Cost - This class is used to measure and compare candidate formulae.
858 /// TODO: Some of these could be merged. Also, a lexical ordering
859 /// isn't always optimal.
863 unsigned NumBaseAdds;
870 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
871 SetupCost(0), ScaleCost(0) {}
873 bool operator<(const Cost &Other) const;
878 // Once any of the metrics loses, they must all remain losers.
880 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
881 | ImmCost | SetupCost | ScaleCost) != ~0u)
882 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
883 & ImmCost & SetupCost & ScaleCost) == ~0u);
888 assert(isValid() && "invalid cost");
889 return NumRegs == ~0u;
892 void RateFormula(const TargetTransformInfo &TTI,
894 SmallPtrSetImpl<const SCEV *> &Regs,
895 const DenseSet<const SCEV *> &VisitedRegs,
897 const SmallVectorImpl<int64_t> &Offsets,
898 ScalarEvolution &SE, DominatorTree &DT,
900 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
902 void print(raw_ostream &OS) const;
906 void RateRegister(const SCEV *Reg,
907 SmallPtrSetImpl<const SCEV *> &Regs,
909 ScalarEvolution &SE, DominatorTree &DT);
910 void RatePrimaryRegister(const SCEV *Reg,
911 SmallPtrSetImpl<const SCEV *> &Regs,
913 ScalarEvolution &SE, DominatorTree &DT,
914 SmallPtrSetImpl<const SCEV *> *LoserRegs);
919 /// RateRegister - Tally up interesting quantities from the given register.
920 void Cost::RateRegister(const SCEV *Reg,
921 SmallPtrSetImpl<const SCEV *> &Regs,
923 ScalarEvolution &SE, DominatorTree &DT) {
924 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
925 // If this is an addrec for another loop, don't second-guess its addrec phi
926 // nodes. LSR isn't currently smart enough to reason about more than one
927 // loop at a time. LSR has already run on inner loops, will not run on outer
928 // loops, and cannot be expected to change sibling loops.
929 if (AR->getLoop() != L) {
930 // If the AddRec exists, consider it's register free and leave it alone.
931 if (isExistingPhi(AR, SE))
934 // Otherwise, do not consider this formula at all.
938 AddRecCost += 1; /// TODO: This should be a function of the stride.
940 // Add the step value register, if it needs one.
941 // TODO: The non-affine case isn't precisely modeled here.
942 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
943 if (!Regs.count(AR->getOperand(1))) {
944 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
952 // Rough heuristic; favor registers which don't require extra setup
953 // instructions in the preheader.
954 if (!isa<SCEVUnknown>(Reg) &&
955 !isa<SCEVConstant>(Reg) &&
956 !(isa<SCEVAddRecExpr>(Reg) &&
957 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
958 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
961 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
962 SE.hasComputableLoopEvolution(Reg, L);
965 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
966 /// before, rate it. Optional LoserRegs provides a way to declare any formula
967 /// that refers to one of those regs an instant loser.
968 void Cost::RatePrimaryRegister(const SCEV *Reg,
969 SmallPtrSetImpl<const SCEV *> &Regs,
971 ScalarEvolution &SE, DominatorTree &DT,
972 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
973 if (LoserRegs && LoserRegs->count(Reg)) {
977 if (Regs.insert(Reg).second) {
978 RateRegister(Reg, Regs, L, SE, DT);
979 if (LoserRegs && isLoser())
980 LoserRegs->insert(Reg);
984 void Cost::RateFormula(const TargetTransformInfo &TTI,
986 SmallPtrSetImpl<const SCEV *> &Regs,
987 const DenseSet<const SCEV *> &VisitedRegs,
989 const SmallVectorImpl<int64_t> &Offsets,
990 ScalarEvolution &SE, DominatorTree &DT,
992 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
993 assert(F.isCanonical() && "Cost is accurate only for canonical formula");
994 // Tally up the registers.
995 if (const SCEV *ScaledReg = F.ScaledReg) {
996 if (VisitedRegs.count(ScaledReg)) {
1000 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1004 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
1005 E = F.BaseRegs.end(); I != E; ++I) {
1006 const SCEV *BaseReg = *I;
1007 if (VisitedRegs.count(BaseReg)) {
1011 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1016 // Determine how many (unfolded) adds we'll need inside the loop.
1017 size_t NumBaseParts = F.getNumRegs();
1018 if (NumBaseParts > 1)
1019 // Do not count the base and a possible second register if the target
1020 // allows to fold 2 registers.
1022 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1023 NumBaseAdds += (F.UnfoldedOffset != 0);
1025 // Accumulate non-free scaling amounts.
1026 ScaleCost += getScalingFactorCost(TTI, LU, F);
1028 // Tally up the non-zero immediates.
1029 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1030 E = Offsets.end(); I != E; ++I) {
1031 int64_t Offset = (uint64_t)*I + F.BaseOffset;
1033 ImmCost += 64; // Handle symbolic values conservatively.
1034 // TODO: This should probably be the pointer size.
1035 else if (Offset != 0)
1036 ImmCost += APInt(64, Offset, true).getMinSignedBits();
1038 assert(isValid() && "invalid cost");
1041 /// Lose - Set this cost to a losing value.
1052 /// operator< - Choose the lower cost.
1053 bool Cost::operator<(const Cost &Other) const {
1054 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
1055 ImmCost, SetupCost) <
1056 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
1057 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
1061 void Cost::print(raw_ostream &OS) const {
1062 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
1063 if (AddRecCost != 0)
1064 OS << ", with addrec cost " << AddRecCost;
1066 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
1067 if (NumBaseAdds != 0)
1068 OS << ", plus " << NumBaseAdds << " base add"
1069 << (NumBaseAdds == 1 ? "" : "s");
1071 OS << ", plus " << ScaleCost << " scale cost";
1073 OS << ", plus " << ImmCost << " imm cost";
1075 OS << ", plus " << SetupCost << " setup cost";
1078 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1079 void Cost::dump() const {
1080 print(errs()); errs() << '\n';
1086 /// LSRFixup - An operand value in an instruction which is to be replaced
1087 /// with some equivalent, possibly strength-reduced, replacement.
1089 /// UserInst - The instruction which will be updated.
1090 Instruction *UserInst;
1092 /// OperandValToReplace - The operand of the instruction which will
1093 /// be replaced. The operand may be used more than once; every instance
1094 /// will be replaced.
1095 Value *OperandValToReplace;
1097 /// PostIncLoops - If this user is to use the post-incremented value of an
1098 /// induction variable, this variable is non-null and holds the loop
1099 /// associated with the induction variable.
1100 PostIncLoopSet PostIncLoops;
1102 /// LUIdx - The index of the LSRUse describing the expression which
1103 /// this fixup needs, minus an offset (below).
1106 /// Offset - A constant offset to be added to the LSRUse expression.
1107 /// This allows multiple fixups to share the same LSRUse with different
1108 /// offsets, for example in an unrolled loop.
1111 bool isUseFullyOutsideLoop(const Loop *L) const;
1115 void print(raw_ostream &OS) const;
1121 LSRFixup::LSRFixup()
1122 : UserInst(nullptr), OperandValToReplace(nullptr), LUIdx(~size_t(0)),
1125 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1126 /// value outside of the given loop.
1127 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1128 // PHI nodes use their value in their incoming blocks.
1129 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1130 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1131 if (PN->getIncomingValue(i) == OperandValToReplace &&
1132 L->contains(PN->getIncomingBlock(i)))
1137 return !L->contains(UserInst);
1140 void LSRFixup::print(raw_ostream &OS) const {
1142 // Store is common and interesting enough to be worth special-casing.
1143 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1145 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1146 } else if (UserInst->getType()->isVoidTy())
1147 OS << UserInst->getOpcodeName();
1149 UserInst->printAsOperand(OS, /*PrintType=*/false);
1151 OS << ", OperandValToReplace=";
1152 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1154 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1155 E = PostIncLoops.end(); I != E; ++I) {
1156 OS << ", PostIncLoop=";
1157 (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1160 if (LUIdx != ~size_t(0))
1161 OS << ", LUIdx=" << LUIdx;
1164 OS << ", Offset=" << Offset;
1167 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1168 void LSRFixup::dump() const {
1169 print(errs()); errs() << '\n';
1175 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1176 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1177 struct UniquifierDenseMapInfo {
1178 static SmallVector<const SCEV *, 4> getEmptyKey() {
1179 SmallVector<const SCEV *, 4> V;
1180 V.push_back(reinterpret_cast<const SCEV *>(-1));
1184 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1185 SmallVector<const SCEV *, 4> V;
1186 V.push_back(reinterpret_cast<const SCEV *>(-2));
1190 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1191 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1194 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1195 const SmallVector<const SCEV *, 4> &RHS) {
1200 /// LSRUse - This class holds the state that LSR keeps for each use in
1201 /// IVUsers, as well as uses invented by LSR itself. It includes information
1202 /// about what kinds of things can be folded into the user, information about
1203 /// the user itself, and information about how the use may be satisfied.
1204 /// TODO: Represent multiple users of the same expression in common?
1206 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1209 /// KindType - An enum for a kind of use, indicating what types of
1210 /// scaled and immediate operands it might support.
1212 Basic, ///< A normal use, with no folding.
1213 Special, ///< A special case of basic, allowing -1 scales.
1214 Address, ///< An address use; folding according to TargetLowering
1215 ICmpZero ///< An equality icmp with both operands folded into one.
1216 // TODO: Add a generic icmp too?
1219 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1224 SmallVector<int64_t, 8> Offsets;
1228 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1229 /// LSRUse are outside of the loop, in which case some special-case heuristics
1231 bool AllFixupsOutsideLoop;
1233 /// RigidFormula is set to true to guarantee that this use will be associated
1234 /// with a single formula--the one that initially matched. Some SCEV
1235 /// expressions cannot be expanded. This allows LSR to consider the registers
1236 /// used by those expressions without the need to expand them later after
1237 /// changing the formula.
1240 /// WidestFixupType - This records the widest use type for any fixup using
1241 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1242 /// max fixup widths to be equivalent, because the narrower one may be relying
1243 /// on the implicit truncation to truncate away bogus bits.
1244 Type *WidestFixupType;
1246 /// Formulae - A list of ways to build a value that can satisfy this user.
1247 /// After the list is populated, one of these is selected heuristically and
1248 /// used to formulate a replacement for OperandValToReplace in UserInst.
1249 SmallVector<Formula, 12> Formulae;
1251 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1252 SmallPtrSet<const SCEV *, 4> Regs;
1254 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1255 MinOffset(INT64_MAX),
1256 MaxOffset(INT64_MIN),
1257 AllFixupsOutsideLoop(true),
1258 RigidFormula(false),
1259 WidestFixupType(nullptr) {}
1261 bool HasFormulaWithSameRegs(const Formula &F) const;
1262 bool InsertFormula(const Formula &F);
1263 void DeleteFormula(Formula &F);
1264 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1266 void print(raw_ostream &OS) const;
1272 /// HasFormula - Test whether this use as a formula which has the same
1273 /// registers as the given formula.
1274 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1275 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1276 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1277 // Unstable sort by host order ok, because this is only used for uniquifying.
1278 std::sort(Key.begin(), Key.end());
1279 return Uniquifier.count(Key);
1282 /// InsertFormula - If the given formula has not yet been inserted, add it to
1283 /// the list, and return true. Return false otherwise.
1284 /// The formula must be in canonical form.
1285 bool LSRUse::InsertFormula(const Formula &F) {
1286 assert(F.isCanonical() && "Invalid canonical representation");
1288 if (!Formulae.empty() && RigidFormula)
1291 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1292 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1293 // Unstable sort by host order ok, because this is only used for uniquifying.
1294 std::sort(Key.begin(), Key.end());
1296 if (!Uniquifier.insert(Key).second)
1299 // Using a register to hold the value of 0 is not profitable.
1300 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1301 "Zero allocated in a scaled register!");
1303 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1304 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1305 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1308 // Add the formula to the list.
1309 Formulae.push_back(F);
1311 // Record registers now being used by this use.
1312 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1314 Regs.insert(F.ScaledReg);
1319 /// DeleteFormula - Remove the given formula from this use's list.
1320 void LSRUse::DeleteFormula(Formula &F) {
1321 if (&F != &Formulae.back())
1322 std::swap(F, Formulae.back());
1323 Formulae.pop_back();
1326 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1327 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1328 // Now that we've filtered out some formulae, recompute the Regs set.
1329 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1331 for (const Formula &F : Formulae) {
1332 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1333 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1336 // Update the RegTracker.
1337 for (const SCEV *S : OldRegs)
1339 RegUses.DropRegister(S, LUIdx);
1342 void LSRUse::print(raw_ostream &OS) const {
1343 OS << "LSR Use: Kind=";
1345 case Basic: OS << "Basic"; break;
1346 case Special: OS << "Special"; break;
1347 case ICmpZero: OS << "ICmpZero"; break;
1349 OS << "Address of ";
1350 if (AccessTy->isPointerTy())
1351 OS << "pointer"; // the full pointer type could be really verbose
1356 OS << ", Offsets={";
1357 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1358 E = Offsets.end(); I != E; ++I) {
1360 if (std::next(I) != E)
1365 if (AllFixupsOutsideLoop)
1366 OS << ", all-fixups-outside-loop";
1368 if (WidestFixupType)
1369 OS << ", widest fixup type: " << *WidestFixupType;
1372 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1373 void LSRUse::dump() const {
1374 print(errs()); errs() << '\n';
1378 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1379 LSRUse::KindType Kind, Type *AccessTy,
1380 GlobalValue *BaseGV, int64_t BaseOffset,
1381 bool HasBaseReg, int64_t Scale) {
1383 case LSRUse::Address:
1384 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1386 case LSRUse::ICmpZero:
1387 // There's not even a target hook for querying whether it would be legal to
1388 // fold a GV into an ICmp.
1392 // ICmp only has two operands; don't allow more than two non-trivial parts.
1393 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1396 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1397 // putting the scaled register in the other operand of the icmp.
1398 if (Scale != 0 && Scale != -1)
1401 // If we have low-level target information, ask the target if it can fold an
1402 // integer immediate on an icmp.
1403 if (BaseOffset != 0) {
1405 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1406 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1407 // Offs is the ICmp immediate.
1409 // The cast does the right thing with INT64_MIN.
1410 BaseOffset = -(uint64_t)BaseOffset;
1411 return TTI.isLegalICmpImmediate(BaseOffset);
1414 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1418 // Only handle single-register values.
1419 return !BaseGV && Scale == 0 && BaseOffset == 0;
1421 case LSRUse::Special:
1422 // Special case Basic to handle -1 scales.
1423 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1426 llvm_unreachable("Invalid LSRUse Kind!");
1429 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1430 int64_t MinOffset, int64_t MaxOffset,
1431 LSRUse::KindType Kind, Type *AccessTy,
1432 GlobalValue *BaseGV, int64_t BaseOffset,
1433 bool HasBaseReg, int64_t Scale) {
1434 // Check for overflow.
1435 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1438 MinOffset = (uint64_t)BaseOffset + MinOffset;
1439 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1442 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1444 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1445 HasBaseReg, Scale) &&
1446 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1450 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1451 int64_t MinOffset, int64_t MaxOffset,
1452 LSRUse::KindType Kind, Type *AccessTy,
1454 // For the purpose of isAMCompletelyFolded either having a canonical formula
1455 // or a scale not equal to zero is correct.
1456 // Problems may arise from non canonical formulae having a scale == 0.
1457 // Strictly speaking it would best to just rely on canonical formulae.
1458 // However, when we generate the scaled formulae, we first check that the
1459 // scaling factor is profitable before computing the actual ScaledReg for
1460 // compile time sake.
1461 assert((F.isCanonical() || F.Scale != 0));
1462 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1463 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1466 /// isLegalUse - Test whether we know how to expand the current formula.
1467 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1468 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1469 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1471 // We know how to expand completely foldable formulae.
1472 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1473 BaseOffset, HasBaseReg, Scale) ||
1474 // Or formulae that use a base register produced by a sum of base
1477 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1478 BaseGV, BaseOffset, true, 0));
1481 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1482 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1484 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1485 F.BaseOffset, F.HasBaseReg, F.Scale);
1488 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1489 const LSRUse &LU, const Formula &F) {
1490 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1491 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1495 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1496 const LSRUse &LU, const Formula &F) {
1500 // If the use is not completely folded in that instruction, we will have to
1501 // pay an extra cost only for scale != 1.
1502 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1504 return F.Scale != 1;
1507 case LSRUse::Address: {
1508 // Check the scaling factor cost with both the min and max offsets.
1509 int ScaleCostMinOffset =
1510 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1511 F.BaseOffset + LU.MinOffset,
1512 F.HasBaseReg, F.Scale);
1513 int ScaleCostMaxOffset =
1514 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1515 F.BaseOffset + LU.MaxOffset,
1516 F.HasBaseReg, F.Scale);
1518 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1519 "Legal addressing mode has an illegal cost!");
1520 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1522 case LSRUse::ICmpZero:
1524 case LSRUse::Special:
1525 // The use is completely folded, i.e., everything is folded into the
1530 llvm_unreachable("Invalid LSRUse Kind!");
1533 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1534 LSRUse::KindType Kind, Type *AccessTy,
1535 GlobalValue *BaseGV, int64_t BaseOffset,
1537 // Fast-path: zero is always foldable.
1538 if (BaseOffset == 0 && !BaseGV) return true;
1540 // Conservatively, create an address with an immediate and a
1541 // base and a scale.
1542 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1544 // Canonicalize a scale of 1 to a base register if the formula doesn't
1545 // already have a base register.
1546 if (!HasBaseReg && Scale == 1) {
1551 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1555 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1556 ScalarEvolution &SE, int64_t MinOffset,
1557 int64_t MaxOffset, LSRUse::KindType Kind,
1558 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1559 // Fast-path: zero is always foldable.
1560 if (S->isZero()) return true;
1562 // Conservatively, create an address with an immediate and a
1563 // base and a scale.
1564 int64_t BaseOffset = ExtractImmediate(S, SE);
1565 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1567 // If there's anything else involved, it's not foldable.
1568 if (!S->isZero()) return false;
1570 // Fast-path: zero is always foldable.
1571 if (BaseOffset == 0 && !BaseGV) return true;
1573 // Conservatively, create an address with an immediate and a
1574 // base and a scale.
1575 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1577 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1578 BaseOffset, HasBaseReg, Scale);
1583 /// IVInc - An individual increment in a Chain of IV increments.
1584 /// Relate an IV user to an expression that computes the IV it uses from the IV
1585 /// used by the previous link in the Chain.
1587 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1588 /// original IVOperand. The head of the chain's IVOperand is only valid during
1589 /// chain collection, before LSR replaces IV users. During chain generation,
1590 /// IncExpr can be used to find the new IVOperand that computes the same
1593 Instruction *UserInst;
1595 const SCEV *IncExpr;
1597 IVInc(Instruction *U, Value *O, const SCEV *E):
1598 UserInst(U), IVOperand(O), IncExpr(E) {}
1601 // IVChain - The list of IV increments in program order.
1602 // We typically add the head of a chain without finding subsequent links.
1604 SmallVector<IVInc,1> Incs;
1605 const SCEV *ExprBase;
1607 IVChain() : ExprBase(nullptr) {}
1609 IVChain(const IVInc &Head, const SCEV *Base)
1610 : Incs(1, Head), ExprBase(Base) {}
1612 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1614 // begin - return the first increment in the chain.
1615 const_iterator begin() const {
1616 assert(!Incs.empty());
1617 return std::next(Incs.begin());
1619 const_iterator end() const {
1623 // hasIncs - Returns true if this chain contains any increments.
1624 bool hasIncs() const { return Incs.size() >= 2; }
1626 // add - Add an IVInc to the end of this chain.
1627 void add(const IVInc &X) { Incs.push_back(X); }
1629 // tailUserInst - Returns the last UserInst in the chain.
1630 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1632 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1634 bool isProfitableIncrement(const SCEV *OperExpr,
1635 const SCEV *IncExpr,
1639 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1640 /// Distinguish between FarUsers that definitely cross IV increments and
1641 /// NearUsers that may be used between IV increments.
1643 SmallPtrSet<Instruction*, 4> FarUsers;
1644 SmallPtrSet<Instruction*, 4> NearUsers;
1647 /// LSRInstance - This class holds state for the main loop strength reduction
1651 ScalarEvolution &SE;
1654 const TargetTransformInfo &TTI;
1658 /// IVIncInsertPos - This is the insert position that the current loop's
1659 /// induction variable increment should be placed. In simple loops, this is
1660 /// the latch block's terminator. But in more complicated cases, this is a
1661 /// position which will dominate all the in-loop post-increment users.
1662 Instruction *IVIncInsertPos;
1664 /// Factors - Interesting factors between use strides.
1665 SmallSetVector<int64_t, 8> Factors;
1667 /// Types - Interesting use types, to facilitate truncation reuse.
1668 SmallSetVector<Type *, 4> Types;
1670 /// Fixups - The list of operands which are to be replaced.
1671 SmallVector<LSRFixup, 16> Fixups;
1673 /// Uses - The list of interesting uses.
1674 SmallVector<LSRUse, 16> Uses;
1676 /// RegUses - Track which uses use which register candidates.
1677 RegUseTracker RegUses;
1679 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1680 // have more than a few IV increment chains in a loop. Missing a Chain falls
1681 // back to normal LSR behavior for those uses.
1682 static const unsigned MaxChains = 8;
1684 /// IVChainVec - IV users can form a chain of IV increments.
1685 SmallVector<IVChain, MaxChains> IVChainVec;
1687 /// IVIncSet - IV users that belong to profitable IVChains.
1688 SmallPtrSet<Use*, MaxChains> IVIncSet;
1690 void OptimizeShadowIV();
1691 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1692 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1693 void OptimizeLoopTermCond();
1695 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1696 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1697 void FinalizeChain(IVChain &Chain);
1698 void CollectChains();
1699 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1700 SmallVectorImpl<WeakVH> &DeadInsts);
1702 void CollectInterestingTypesAndFactors();
1703 void CollectFixupsAndInitialFormulae();
1705 LSRFixup &getNewFixup() {
1706 Fixups.push_back(LSRFixup());
1707 return Fixups.back();
1710 // Support for sharing of LSRUses between LSRFixups.
1711 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1714 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1715 LSRUse::KindType Kind, Type *AccessTy);
1717 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1718 LSRUse::KindType Kind,
1721 void DeleteUse(LSRUse &LU, size_t LUIdx);
1723 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1725 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1726 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1727 void CountRegisters(const Formula &F, size_t LUIdx);
1728 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1730 void CollectLoopInvariantFixupsAndFormulae();
1732 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1733 unsigned Depth = 0);
1735 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1736 const Formula &Base, unsigned Depth,
1737 size_t Idx, bool IsScaledReg = false);
1738 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1739 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1740 const Formula &Base, size_t Idx,
1741 bool IsScaledReg = false);
1742 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1743 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1744 const Formula &Base,
1745 const SmallVectorImpl<int64_t> &Worklist,
1746 size_t Idx, bool IsScaledReg = false);
1747 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1748 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1749 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1750 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1751 void GenerateCrossUseConstantOffsets();
1752 void GenerateAllReuseFormulae();
1754 void FilterOutUndesirableDedicatedRegisters();
1756 size_t EstimateSearchSpaceComplexity() const;
1757 void NarrowSearchSpaceByDetectingSupersets();
1758 void NarrowSearchSpaceByCollapsingUnrolledCode();
1759 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1760 void NarrowSearchSpaceByPickingWinnerRegs();
1761 void NarrowSearchSpaceUsingHeuristics();
1763 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1765 SmallVectorImpl<const Formula *> &Workspace,
1766 const Cost &CurCost,
1767 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1768 DenseSet<const SCEV *> &VisitedRegs) const;
1769 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1771 BasicBlock::iterator
1772 HoistInsertPosition(BasicBlock::iterator IP,
1773 const SmallVectorImpl<Instruction *> &Inputs) const;
1774 BasicBlock::iterator
1775 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1778 SCEVExpander &Rewriter) const;
1780 Value *Expand(const LSRFixup &LF,
1782 BasicBlock::iterator IP,
1783 SCEVExpander &Rewriter,
1784 SmallVectorImpl<WeakVH> &DeadInsts) const;
1785 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1787 SCEVExpander &Rewriter,
1788 SmallVectorImpl<WeakVH> &DeadInsts,
1790 void Rewrite(const LSRFixup &LF,
1792 SCEVExpander &Rewriter,
1793 SmallVectorImpl<WeakVH> &DeadInsts,
1795 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1799 LSRInstance(Loop *L, Pass *P);
1801 bool getChanged() const { return Changed; }
1803 void print_factors_and_types(raw_ostream &OS) const;
1804 void print_fixups(raw_ostream &OS) const;
1805 void print_uses(raw_ostream &OS) const;
1806 void print(raw_ostream &OS) const;
1812 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1813 /// inside the loop then try to eliminate the cast operation.
1814 void LSRInstance::OptimizeShadowIV() {
1815 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1816 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1819 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1820 UI != E; /* empty */) {
1821 IVUsers::const_iterator CandidateUI = UI;
1823 Instruction *ShadowUse = CandidateUI->getUser();
1824 Type *DestTy = nullptr;
1825 bool IsSigned = false;
1827 /* If shadow use is a int->float cast then insert a second IV
1828 to eliminate this cast.
1830 for (unsigned i = 0; i < n; ++i)
1836 for (unsigned i = 0; i < n; ++i, ++d)
1839 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1841 DestTy = UCast->getDestTy();
1843 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1845 DestTy = SCast->getDestTy();
1847 if (!DestTy) continue;
1849 // If target does not support DestTy natively then do not apply
1850 // this transformation.
1851 if (!TTI.isTypeLegal(DestTy)) continue;
1853 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1855 if (PH->getNumIncomingValues() != 2) continue;
1857 Type *SrcTy = PH->getType();
1858 int Mantissa = DestTy->getFPMantissaWidth();
1859 if (Mantissa == -1) continue;
1860 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1863 unsigned Entry, Latch;
1864 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1872 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1873 if (!Init) continue;
1874 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1875 (double)Init->getSExtValue() :
1876 (double)Init->getZExtValue());
1878 BinaryOperator *Incr =
1879 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1880 if (!Incr) continue;
1881 if (Incr->getOpcode() != Instruction::Add
1882 && Incr->getOpcode() != Instruction::Sub)
1885 /* Initialize new IV, double d = 0.0 in above example. */
1886 ConstantInt *C = nullptr;
1887 if (Incr->getOperand(0) == PH)
1888 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1889 else if (Incr->getOperand(1) == PH)
1890 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1896 // Ignore negative constants, as the code below doesn't handle them
1897 // correctly. TODO: Remove this restriction.
1898 if (!C->getValue().isStrictlyPositive()) continue;
1900 /* Add new PHINode. */
1901 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1903 /* create new increment. '++d' in above example. */
1904 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1905 BinaryOperator *NewIncr =
1906 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1907 Instruction::FAdd : Instruction::FSub,
1908 NewPH, CFP, "IV.S.next.", Incr);
1910 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1911 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1913 /* Remove cast operation */
1914 ShadowUse->replaceAllUsesWith(NewPH);
1915 ShadowUse->eraseFromParent();
1921 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1922 /// set the IV user and stride information and return true, otherwise return
1924 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1925 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1926 if (UI->getUser() == Cond) {
1927 // NOTE: we could handle setcc instructions with multiple uses here, but
1928 // InstCombine does it as well for simple uses, it's not clear that it
1929 // occurs enough in real life to handle.
1936 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1937 /// a max computation.
1939 /// This is a narrow solution to a specific, but acute, problem. For loops
1945 /// } while (++i < n);
1947 /// the trip count isn't just 'n', because 'n' might not be positive. And
1948 /// unfortunately this can come up even for loops where the user didn't use
1949 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1950 /// will commonly be lowered like this:
1956 /// } while (++i < n);
1959 /// and then it's possible for subsequent optimization to obscure the if
1960 /// test in such a way that indvars can't find it.
1962 /// When indvars can't find the if test in loops like this, it creates a
1963 /// max expression, which allows it to give the loop a canonical
1964 /// induction variable:
1967 /// max = n < 1 ? 1 : n;
1970 /// } while (++i != max);
1972 /// Canonical induction variables are necessary because the loop passes
1973 /// are designed around them. The most obvious example of this is the
1974 /// LoopInfo analysis, which doesn't remember trip count values. It
1975 /// expects to be able to rediscover the trip count each time it is
1976 /// needed, and it does this using a simple analysis that only succeeds if
1977 /// the loop has a canonical induction variable.
1979 /// However, when it comes time to generate code, the maximum operation
1980 /// can be quite costly, especially if it's inside of an outer loop.
1982 /// This function solves this problem by detecting this type of loop and
1983 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1984 /// the instructions for the maximum computation.
1986 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1987 // Check that the loop matches the pattern we're looking for.
1988 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1989 Cond->getPredicate() != CmpInst::ICMP_NE)
1992 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1993 if (!Sel || !Sel->hasOneUse()) return Cond;
1995 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1996 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1998 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2000 // Add one to the backedge-taken count to get the trip count.
2001 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2002 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2004 // Check for a max calculation that matches the pattern. There's no check
2005 // for ICMP_ULE here because the comparison would be with zero, which
2006 // isn't interesting.
2007 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2008 const SCEVNAryExpr *Max = nullptr;
2009 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2010 Pred = ICmpInst::ICMP_SLE;
2012 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2013 Pred = ICmpInst::ICMP_SLT;
2015 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2016 Pred = ICmpInst::ICMP_ULT;
2023 // To handle a max with more than two operands, this optimization would
2024 // require additional checking and setup.
2025 if (Max->getNumOperands() != 2)
2028 const SCEV *MaxLHS = Max->getOperand(0);
2029 const SCEV *MaxRHS = Max->getOperand(1);
2031 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2032 // for a comparison with 1. For <= and >=, a comparison with zero.
2034 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2037 // Check the relevant induction variable for conformance to
2039 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2040 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2041 if (!AR || !AR->isAffine() ||
2042 AR->getStart() != One ||
2043 AR->getStepRecurrence(SE) != One)
2046 assert(AR->getLoop() == L &&
2047 "Loop condition operand is an addrec in a different loop!");
2049 // Check the right operand of the select, and remember it, as it will
2050 // be used in the new comparison instruction.
2051 Value *NewRHS = nullptr;
2052 if (ICmpInst::isTrueWhenEqual(Pred)) {
2053 // Look for n+1, and grab n.
2054 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2055 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2056 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2057 NewRHS = BO->getOperand(0);
2058 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2059 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2060 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2061 NewRHS = BO->getOperand(0);
2064 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2065 NewRHS = Sel->getOperand(1);
2066 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2067 NewRHS = Sel->getOperand(2);
2068 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2069 NewRHS = SU->getValue();
2071 // Max doesn't match expected pattern.
2074 // Determine the new comparison opcode. It may be signed or unsigned,
2075 // and the original comparison may be either equality or inequality.
2076 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2077 Pred = CmpInst::getInversePredicate(Pred);
2079 // Ok, everything looks ok to change the condition into an SLT or SGE and
2080 // delete the max calculation.
2082 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2084 // Delete the max calculation instructions.
2085 Cond->replaceAllUsesWith(NewCond);
2086 CondUse->setUser(NewCond);
2087 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2088 Cond->eraseFromParent();
2089 Sel->eraseFromParent();
2090 if (Cmp->use_empty())
2091 Cmp->eraseFromParent();
2095 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2096 /// postinc iv when possible.
2098 LSRInstance::OptimizeLoopTermCond() {
2099 SmallPtrSet<Instruction *, 4> PostIncs;
2101 BasicBlock *LatchBlock = L->getLoopLatch();
2102 SmallVector<BasicBlock*, 8> ExitingBlocks;
2103 L->getExitingBlocks(ExitingBlocks);
2105 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2106 BasicBlock *ExitingBlock = ExitingBlocks[i];
2108 // Get the terminating condition for the loop if possible. If we
2109 // can, we want to change it to use a post-incremented version of its
2110 // induction variable, to allow coalescing the live ranges for the IV into
2111 // one register value.
2113 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2116 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2117 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2120 // Search IVUsesByStride to find Cond's IVUse if there is one.
2121 IVStrideUse *CondUse = nullptr;
2122 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2123 if (!FindIVUserForCond(Cond, CondUse))
2126 // If the trip count is computed in terms of a max (due to ScalarEvolution
2127 // being unable to find a sufficient guard, for example), change the loop
2128 // comparison to use SLT or ULT instead of NE.
2129 // One consequence of doing this now is that it disrupts the count-down
2130 // optimization. That's not always a bad thing though, because in such
2131 // cases it may still be worthwhile to avoid a max.
2132 Cond = OptimizeMax(Cond, CondUse);
2134 // If this exiting block dominates the latch block, it may also use
2135 // the post-inc value if it won't be shared with other uses.
2136 // Check for dominance.
2137 if (!DT.dominates(ExitingBlock, LatchBlock))
2140 // Conservatively avoid trying to use the post-inc value in non-latch
2141 // exits if there may be pre-inc users in intervening blocks.
2142 if (LatchBlock != ExitingBlock)
2143 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2144 // Test if the use is reachable from the exiting block. This dominator
2145 // query is a conservative approximation of reachability.
2146 if (&*UI != CondUse &&
2147 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2148 // Conservatively assume there may be reuse if the quotient of their
2149 // strides could be a legal scale.
2150 const SCEV *A = IU.getStride(*CondUse, L);
2151 const SCEV *B = IU.getStride(*UI, L);
2152 if (!A || !B) continue;
2153 if (SE.getTypeSizeInBits(A->getType()) !=
2154 SE.getTypeSizeInBits(B->getType())) {
2155 if (SE.getTypeSizeInBits(A->getType()) >
2156 SE.getTypeSizeInBits(B->getType()))
2157 B = SE.getSignExtendExpr(B, A->getType());
2159 A = SE.getSignExtendExpr(A, B->getType());
2161 if (const SCEVConstant *D =
2162 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2163 const ConstantInt *C = D->getValue();
2164 // Stride of one or negative one can have reuse with non-addresses.
2165 if (C->isOne() || C->isAllOnesValue())
2166 goto decline_post_inc;
2167 // Avoid weird situations.
2168 if (C->getValue().getMinSignedBits() >= 64 ||
2169 C->getValue().isMinSignedValue())
2170 goto decline_post_inc;
2171 // Check for possible scaled-address reuse.
2172 Type *AccessTy = getAccessType(UI->getUser());
2173 int64_t Scale = C->getSExtValue();
2174 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ nullptr,
2176 /*HasBaseReg=*/ false, Scale))
2177 goto decline_post_inc;
2179 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ nullptr,
2181 /*HasBaseReg=*/ false, Scale))
2182 goto decline_post_inc;
2186 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2189 // It's possible for the setcc instruction to be anywhere in the loop, and
2190 // possible for it to have multiple users. If it is not immediately before
2191 // the exiting block branch, move it.
2192 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2193 if (Cond->hasOneUse()) {
2194 Cond->moveBefore(TermBr);
2196 // Clone the terminating condition and insert into the loopend.
2197 ICmpInst *OldCond = Cond;
2198 Cond = cast<ICmpInst>(Cond->clone());
2199 Cond->setName(L->getHeader()->getName() + ".termcond");
2200 ExitingBlock->getInstList().insert(TermBr, Cond);
2202 // Clone the IVUse, as the old use still exists!
2203 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2204 TermBr->replaceUsesOfWith(OldCond, Cond);
2208 // If we get to here, we know that we can transform the setcc instruction to
2209 // use the post-incremented version of the IV, allowing us to coalesce the
2210 // live ranges for the IV correctly.
2211 CondUse->transformToPostInc(L);
2214 PostIncs.insert(Cond);
2218 // Determine an insertion point for the loop induction variable increment. It
2219 // must dominate all the post-inc comparisons we just set up, and it must
2220 // dominate the loop latch edge.
2221 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2222 for (Instruction *Inst : PostIncs) {
2224 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2226 if (BB == Inst->getParent())
2227 IVIncInsertPos = Inst;
2228 else if (BB != IVIncInsertPos->getParent())
2229 IVIncInsertPos = BB->getTerminator();
2233 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2234 /// at the given offset and other details. If so, update the use and
2237 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2238 LSRUse::KindType Kind, Type *AccessTy) {
2239 int64_t NewMinOffset = LU.MinOffset;
2240 int64_t NewMaxOffset = LU.MaxOffset;
2241 Type *NewAccessTy = AccessTy;
2243 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2244 // something conservative, however this can pessimize in the case that one of
2245 // the uses will have all its uses outside the loop, for example.
2246 if (LU.Kind != Kind)
2249 // Check for a mismatched access type, and fall back conservatively as needed.
2250 // TODO: Be less conservative when the type is similar and can use the same
2251 // addressing modes.
2252 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2253 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2255 // Conservatively assume HasBaseReg is true for now.
2256 if (NewOffset < LU.MinOffset) {
2257 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2258 LU.MaxOffset - NewOffset, HasBaseReg))
2260 NewMinOffset = NewOffset;
2261 } else if (NewOffset > LU.MaxOffset) {
2262 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2263 NewOffset - LU.MinOffset, HasBaseReg))
2265 NewMaxOffset = NewOffset;
2269 LU.MinOffset = NewMinOffset;
2270 LU.MaxOffset = NewMaxOffset;
2271 LU.AccessTy = NewAccessTy;
2272 if (NewOffset != LU.Offsets.back())
2273 LU.Offsets.push_back(NewOffset);
2277 /// getUse - Return an LSRUse index and an offset value for a fixup which
2278 /// needs the given expression, with the given kind and optional access type.
2279 /// Either reuse an existing use or create a new one, as needed.
2280 std::pair<size_t, int64_t>
2281 LSRInstance::getUse(const SCEV *&Expr,
2282 LSRUse::KindType Kind, Type *AccessTy) {
2283 const SCEV *Copy = Expr;
2284 int64_t Offset = ExtractImmediate(Expr, SE);
2286 // Basic uses can't accept any offset, for example.
2287 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2288 Offset, /*HasBaseReg=*/ true)) {
2293 std::pair<UseMapTy::iterator, bool> P =
2294 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2296 // A use already existed with this base.
2297 size_t LUIdx = P.first->second;
2298 LSRUse &LU = Uses[LUIdx];
2299 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2301 return std::make_pair(LUIdx, Offset);
2304 // Create a new use.
2305 size_t LUIdx = Uses.size();
2306 P.first->second = LUIdx;
2307 Uses.push_back(LSRUse(Kind, AccessTy));
2308 LSRUse &LU = Uses[LUIdx];
2310 // We don't need to track redundant offsets, but we don't need to go out
2311 // of our way here to avoid them.
2312 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2313 LU.Offsets.push_back(Offset);
2315 LU.MinOffset = Offset;
2316 LU.MaxOffset = Offset;
2317 return std::make_pair(LUIdx, Offset);
2320 /// DeleteUse - Delete the given use from the Uses list.
2321 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2322 if (&LU != &Uses.back())
2323 std::swap(LU, Uses.back());
2327 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2330 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2331 /// a formula that has the same registers as the given formula.
2333 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2334 const LSRUse &OrigLU) {
2335 // Search all uses for the formula. This could be more clever.
2336 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2337 LSRUse &LU = Uses[LUIdx];
2338 // Check whether this use is close enough to OrigLU, to see whether it's
2339 // worthwhile looking through its formulae.
2340 // Ignore ICmpZero uses because they may contain formulae generated by
2341 // GenerateICmpZeroScales, in which case adding fixup offsets may
2343 if (&LU != &OrigLU &&
2344 LU.Kind != LSRUse::ICmpZero &&
2345 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2346 LU.WidestFixupType == OrigLU.WidestFixupType &&
2347 LU.HasFormulaWithSameRegs(OrigF)) {
2348 // Scan through this use's formulae.
2349 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2350 E = LU.Formulae.end(); I != E; ++I) {
2351 const Formula &F = *I;
2352 // Check to see if this formula has the same registers and symbols
2354 if (F.BaseRegs == OrigF.BaseRegs &&
2355 F.ScaledReg == OrigF.ScaledReg &&
2356 F.BaseGV == OrigF.BaseGV &&
2357 F.Scale == OrigF.Scale &&
2358 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2359 if (F.BaseOffset == 0)
2361 // This is the formula where all the registers and symbols matched;
2362 // there aren't going to be any others. Since we declined it, we
2363 // can skip the rest of the formulae and proceed to the next LSRUse.
2370 // Nothing looked good.
2374 void LSRInstance::CollectInterestingTypesAndFactors() {
2375 SmallSetVector<const SCEV *, 4> Strides;
2377 // Collect interesting types and strides.
2378 SmallVector<const SCEV *, 4> Worklist;
2379 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2380 const SCEV *Expr = IU.getExpr(*UI);
2382 // Collect interesting types.
2383 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2385 // Add strides for mentioned loops.
2386 Worklist.push_back(Expr);
2388 const SCEV *S = Worklist.pop_back_val();
2389 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2390 if (AR->getLoop() == L)
2391 Strides.insert(AR->getStepRecurrence(SE));
2392 Worklist.push_back(AR->getStart());
2393 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2394 Worklist.append(Add->op_begin(), Add->op_end());
2396 } while (!Worklist.empty());
2399 // Compute interesting factors from the set of interesting strides.
2400 for (SmallSetVector<const SCEV *, 4>::const_iterator
2401 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2402 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2403 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2404 const SCEV *OldStride = *I;
2405 const SCEV *NewStride = *NewStrideIter;
2407 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2408 SE.getTypeSizeInBits(NewStride->getType())) {
2409 if (SE.getTypeSizeInBits(OldStride->getType()) >
2410 SE.getTypeSizeInBits(NewStride->getType()))
2411 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2413 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2415 if (const SCEVConstant *Factor =
2416 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2418 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2419 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2420 } else if (const SCEVConstant *Factor =
2421 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2424 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2425 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2429 // If all uses use the same type, don't bother looking for truncation-based
2431 if (Types.size() == 1)
2434 DEBUG(print_factors_and_types(dbgs()));
2437 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2438 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2439 /// Instructions to IVStrideUses, we could partially skip this.
2440 static User::op_iterator
2441 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2442 Loop *L, ScalarEvolution &SE) {
2443 for(; OI != OE; ++OI) {
2444 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2445 if (!SE.isSCEVable(Oper->getType()))
2448 if (const SCEVAddRecExpr *AR =
2449 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2450 if (AR->getLoop() == L)
2458 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2459 /// operands, so wrap it in a convenient helper.
2460 static Value *getWideOperand(Value *Oper) {
2461 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2462 return Trunc->getOperand(0);
2466 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2468 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2469 Type *LType = LVal->getType();
2470 Type *RType = RVal->getType();
2471 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2474 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2475 /// NULL for any constant. Returning the expression itself is
2476 /// conservative. Returning a deeper subexpression is more precise and valid as
2477 /// long as it isn't less complex than another subexpression. For expressions
2478 /// involving multiple unscaled values, we need to return the pointer-type
2479 /// SCEVUnknown. This avoids forming chains across objects, such as:
2480 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2482 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2483 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2484 static const SCEV *getExprBase(const SCEV *S) {
2485 switch (S->getSCEVType()) {
2486 default: // uncluding scUnknown.
2491 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2493 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2495 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2497 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2498 // there's nothing more complex.
2499 // FIXME: not sure if we want to recognize negation.
2500 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2501 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2502 E(Add->op_begin()); I != E; ++I) {
2503 const SCEV *SubExpr = *I;
2504 if (SubExpr->getSCEVType() == scAddExpr)
2505 return getExprBase(SubExpr);
2507 if (SubExpr->getSCEVType() != scMulExpr)
2510 return S; // all operands are scaled, be conservative.
2513 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2517 /// Return true if the chain increment is profitable to expand into a loop
2518 /// invariant value, which may require its own register. A profitable chain
2519 /// increment will be an offset relative to the same base. We allow such offsets
2520 /// to potentially be used as chain increment as long as it's not obviously
2521 /// expensive to expand using real instructions.
2522 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2523 const SCEV *IncExpr,
2524 ScalarEvolution &SE) {
2525 // Aggressively form chains when -stress-ivchain.
2529 // Do not replace a constant offset from IV head with a nonconstant IV
2531 if (!isa<SCEVConstant>(IncExpr)) {
2532 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2533 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2537 SmallPtrSet<const SCEV*, 8> Processed;
2538 return !isHighCostExpansion(IncExpr, Processed, SE);
2541 /// Return true if the number of registers needed for the chain is estimated to
2542 /// be less than the number required for the individual IV users. First prohibit
2543 /// any IV users that keep the IV live across increments (the Users set should
2544 /// be empty). Next count the number and type of increments in the chain.
2546 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2547 /// effectively use postinc addressing modes. Only consider it profitable it the
2548 /// increments can be computed in fewer registers when chained.
2550 /// TODO: Consider IVInc free if it's already used in another chains.
2552 isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2553 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2557 if (!Chain.hasIncs())
2560 if (!Users.empty()) {
2561 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2562 for (Instruction *Inst : Users) {
2563 dbgs() << " " << *Inst << "\n";
2567 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2569 // The chain itself may require a register, so intialize cost to 1.
2572 // A complete chain likely eliminates the need for keeping the original IV in
2573 // a register. LSR does not currently know how to form a complete chain unless
2574 // the header phi already exists.
2575 if (isa<PHINode>(Chain.tailUserInst())
2576 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2579 const SCEV *LastIncExpr = nullptr;
2580 unsigned NumConstIncrements = 0;
2581 unsigned NumVarIncrements = 0;
2582 unsigned NumReusedIncrements = 0;
2583 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2586 if (I->IncExpr->isZero())
2589 // Incrementing by zero or some constant is neutral. We assume constants can
2590 // be folded into an addressing mode or an add's immediate operand.
2591 if (isa<SCEVConstant>(I->IncExpr)) {
2592 ++NumConstIncrements;
2596 if (I->IncExpr == LastIncExpr)
2597 ++NumReusedIncrements;
2601 LastIncExpr = I->IncExpr;
2603 // An IV chain with a single increment is handled by LSR's postinc
2604 // uses. However, a chain with multiple increments requires keeping the IV's
2605 // value live longer than it needs to be if chained.
2606 if (NumConstIncrements > 1)
2609 // Materializing increment expressions in the preheader that didn't exist in
2610 // the original code may cost a register. For example, sign-extended array
2611 // indices can produce ridiculous increments like this:
2612 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2613 cost += NumVarIncrements;
2615 // Reusing variable increments likely saves a register to hold the multiple of
2617 cost -= NumReusedIncrements;
2619 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2625 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2627 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2628 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2629 // When IVs are used as types of varying widths, they are generally converted
2630 // to a wider type with some uses remaining narrow under a (free) trunc.
2631 Value *const NextIV = getWideOperand(IVOper);
2632 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2633 const SCEV *const OperExprBase = getExprBase(OperExpr);
2635 // Visit all existing chains. Check if its IVOper can be computed as a
2636 // profitable loop invariant increment from the last link in the Chain.
2637 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2638 const SCEV *LastIncExpr = nullptr;
2639 for (; ChainIdx < NChains; ++ChainIdx) {
2640 IVChain &Chain = IVChainVec[ChainIdx];
2642 // Prune the solution space aggressively by checking that both IV operands
2643 // are expressions that operate on the same unscaled SCEVUnknown. This
2644 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2645 // first avoids creating extra SCEV expressions.
2646 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2649 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2650 if (!isCompatibleIVType(PrevIV, NextIV))
2653 // A phi node terminates a chain.
2654 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2657 // The increment must be loop-invariant so it can be kept in a register.
2658 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2659 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2660 if (!SE.isLoopInvariant(IncExpr, L))
2663 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2664 LastIncExpr = IncExpr;
2668 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2669 // bother for phi nodes, because they must be last in the chain.
2670 if (ChainIdx == NChains) {
2671 if (isa<PHINode>(UserInst))
2673 if (NChains >= MaxChains && !StressIVChain) {
2674 DEBUG(dbgs() << "IV Chain Limit\n");
2677 LastIncExpr = OperExpr;
2678 // IVUsers may have skipped over sign/zero extensions. We don't currently
2679 // attempt to form chains involving extensions unless they can be hoisted
2680 // into this loop's AddRec.
2681 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2684 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2686 ChainUsersVec.resize(NChains);
2687 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2688 << ") IV=" << *LastIncExpr << "\n");
2690 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2691 << ") IV+" << *LastIncExpr << "\n");
2692 // Add this IV user to the end of the chain.
2693 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2695 IVChain &Chain = IVChainVec[ChainIdx];
2697 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2698 // This chain's NearUsers become FarUsers.
2699 if (!LastIncExpr->isZero()) {
2700 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2705 // All other uses of IVOperand become near uses of the chain.
2706 // We currently ignore intermediate values within SCEV expressions, assuming
2707 // they will eventually be used be the current chain, or can be computed
2708 // from one of the chain increments. To be more precise we could
2709 // transitively follow its user and only add leaf IV users to the set.
2710 for (User *U : IVOper->users()) {
2711 Instruction *OtherUse = dyn_cast<Instruction>(U);
2714 // Uses in the chain will no longer be uses if the chain is formed.
2715 // Include the head of the chain in this iteration (not Chain.begin()).
2716 IVChain::const_iterator IncIter = Chain.Incs.begin();
2717 IVChain::const_iterator IncEnd = Chain.Incs.end();
2718 for( ; IncIter != IncEnd; ++IncIter) {
2719 if (IncIter->UserInst == OtherUse)
2722 if (IncIter != IncEnd)
2725 if (SE.isSCEVable(OtherUse->getType())
2726 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2727 && IU.isIVUserOrOperand(OtherUse)) {
2730 NearUsers.insert(OtherUse);
2733 // Since this user is part of the chain, it's no longer considered a use
2735 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2738 /// CollectChains - Populate the vector of Chains.
2740 /// This decreases ILP at the architecture level. Targets with ample registers,
2741 /// multiple memory ports, and no register renaming probably don't want
2742 /// this. However, such targets should probably disable LSR altogether.
2744 /// The job of LSR is to make a reasonable choice of induction variables across
2745 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2746 /// ILP *within the loop* if the target wants it.
2748 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2749 /// will not reorder memory operations, it will recognize this as a chain, but
2750 /// will generate redundant IV increments. Ideally this would be corrected later
2751 /// by a smart scheduler:
2757 /// TODO: Walk the entire domtree within this loop, not just the path to the
2758 /// loop latch. This will discover chains on side paths, but requires
2759 /// maintaining multiple copies of the Chains state.
2760 void LSRInstance::CollectChains() {
2761 DEBUG(dbgs() << "Collecting IV Chains.\n");
2762 SmallVector<ChainUsers, 8> ChainUsersVec;
2764 SmallVector<BasicBlock *,8> LatchPath;
2765 BasicBlock *LoopHeader = L->getHeader();
2766 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2767 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2768 LatchPath.push_back(Rung->getBlock());
2770 LatchPath.push_back(LoopHeader);
2772 // Walk the instruction stream from the loop header to the loop latch.
2773 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2774 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2775 BBIter != BBEnd; ++BBIter) {
2776 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2778 // Skip instructions that weren't seen by IVUsers analysis.
2779 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2782 // Ignore users that are part of a SCEV expression. This way we only
2783 // consider leaf IV Users. This effectively rediscovers a portion of
2784 // IVUsers analysis but in program order this time.
2785 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2788 // Remove this instruction from any NearUsers set it may be in.
2789 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2790 ChainIdx < NChains; ++ChainIdx) {
2791 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2793 // Search for operands that can be chained.
2794 SmallPtrSet<Instruction*, 4> UniqueOperands;
2795 User::op_iterator IVOpEnd = I->op_end();
2796 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2797 while (IVOpIter != IVOpEnd) {
2798 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2799 if (UniqueOperands.insert(IVOpInst).second)
2800 ChainInstruction(I, IVOpInst, ChainUsersVec);
2801 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2803 } // Continue walking down the instructions.
2804 } // Continue walking down the domtree.
2805 // Visit phi backedges to determine if the chain can generate the IV postinc.
2806 for (BasicBlock::iterator I = L->getHeader()->begin();
2807 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2808 if (!SE.isSCEVable(PN->getType()))
2812 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2814 ChainInstruction(PN, IncV, ChainUsersVec);
2816 // Remove any unprofitable chains.
2817 unsigned ChainIdx = 0;
2818 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2819 UsersIdx < NChains; ++UsersIdx) {
2820 if (!isProfitableChain(IVChainVec[UsersIdx],
2821 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2823 // Preserve the chain at UsesIdx.
2824 if (ChainIdx != UsersIdx)
2825 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2826 FinalizeChain(IVChainVec[ChainIdx]);
2829 IVChainVec.resize(ChainIdx);
2832 void LSRInstance::FinalizeChain(IVChain &Chain) {
2833 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2834 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2836 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2838 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2839 User::op_iterator UseI =
2840 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2841 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2842 IVIncSet.insert(UseI);
2846 /// Return true if the IVInc can be folded into an addressing mode.
2847 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2848 Value *Operand, const TargetTransformInfo &TTI) {
2849 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2850 if (!IncConst || !isAddressUse(UserInst, Operand))
2853 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2856 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2857 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2858 getAccessType(UserInst), /*BaseGV=*/ nullptr,
2859 IncOffset, /*HaseBaseReg=*/ false))
2865 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2866 /// materialize the IV user's operand from the previous IV user's operand.
2867 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2868 SmallVectorImpl<WeakVH> &DeadInsts) {
2869 // Find the new IVOperand for the head of the chain. It may have been replaced
2871 const IVInc &Head = Chain.Incs[0];
2872 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2873 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2874 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2876 Value *IVSrc = nullptr;
2877 while (IVOpIter != IVOpEnd) {
2878 IVSrc = getWideOperand(*IVOpIter);
2880 // If this operand computes the expression that the chain needs, we may use
2881 // it. (Check this after setting IVSrc which is used below.)
2883 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2884 // narrow for the chain, so we can no longer use it. We do allow using a
2885 // wider phi, assuming the LSR checked for free truncation. In that case we
2886 // should already have a truncate on this operand such that
2887 // getSCEV(IVSrc) == IncExpr.
2888 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2889 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2892 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2894 if (IVOpIter == IVOpEnd) {
2895 // Gracefully give up on this chain.
2896 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2900 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2901 Type *IVTy = IVSrc->getType();
2902 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2903 const SCEV *LeftOverExpr = nullptr;
2904 for (IVChain::const_iterator IncI = Chain.begin(),
2905 IncE = Chain.end(); IncI != IncE; ++IncI) {
2907 Instruction *InsertPt = IncI->UserInst;
2908 if (isa<PHINode>(InsertPt))
2909 InsertPt = L->getLoopLatch()->getTerminator();
2911 // IVOper will replace the current IV User's operand. IVSrc is the IV
2912 // value currently held in a register.
2913 Value *IVOper = IVSrc;
2914 if (!IncI->IncExpr->isZero()) {
2915 // IncExpr was the result of subtraction of two narrow values, so must
2917 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2918 LeftOverExpr = LeftOverExpr ?
2919 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2921 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2922 // Expand the IV increment.
2923 Rewriter.clearPostInc();
2924 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2925 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2926 SE.getUnknown(IncV));
2927 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2929 // If an IV increment can't be folded, use it as the next IV value.
2930 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2932 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2934 LeftOverExpr = nullptr;
2937 Type *OperTy = IncI->IVOperand->getType();
2938 if (IVTy != OperTy) {
2939 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2940 "cannot extend a chained IV");
2941 IRBuilder<> Builder(InsertPt);
2942 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2944 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2945 DeadInsts.push_back(IncI->IVOperand);
2947 // If LSR created a new, wider phi, we may also replace its postinc. We only
2948 // do this if we also found a wide value for the head of the chain.
2949 if (isa<PHINode>(Chain.tailUserInst())) {
2950 for (BasicBlock::iterator I = L->getHeader()->begin();
2951 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2952 if (!isCompatibleIVType(Phi, IVSrc))
2954 Instruction *PostIncV = dyn_cast<Instruction>(
2955 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2956 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2958 Value *IVOper = IVSrc;
2959 Type *PostIncTy = PostIncV->getType();
2960 if (IVTy != PostIncTy) {
2961 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2962 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2963 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2964 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2966 Phi->replaceUsesOfWith(PostIncV, IVOper);
2967 DeadInsts.push_back(PostIncV);
2972 void LSRInstance::CollectFixupsAndInitialFormulae() {
2973 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2974 Instruction *UserInst = UI->getUser();
2975 // Skip IV users that are part of profitable IV Chains.
2976 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2977 UI->getOperandValToReplace());
2978 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2979 if (IVIncSet.count(UseI))
2983 LSRFixup &LF = getNewFixup();
2984 LF.UserInst = UserInst;
2985 LF.OperandValToReplace = UI->getOperandValToReplace();
2986 LF.PostIncLoops = UI->getPostIncLoops();
2988 LSRUse::KindType Kind = LSRUse::Basic;
2989 Type *AccessTy = nullptr;
2990 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2991 Kind = LSRUse::Address;
2992 AccessTy = getAccessType(LF.UserInst);
2995 const SCEV *S = IU.getExpr(*UI);
2997 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2998 // (N - i == 0), and this allows (N - i) to be the expression that we work
2999 // with rather than just N or i, so we can consider the register
3000 // requirements for both N and i at the same time. Limiting this code to
3001 // equality icmps is not a problem because all interesting loops use
3002 // equality icmps, thanks to IndVarSimplify.
3003 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
3004 if (CI->isEquality()) {
3005 // Swap the operands if needed to put the OperandValToReplace on the
3006 // left, for consistency.
3007 Value *NV = CI->getOperand(1);
3008 if (NV == LF.OperandValToReplace) {
3009 CI->setOperand(1, CI->getOperand(0));
3010 CI->setOperand(0, NV);
3011 NV = CI->getOperand(1);
3015 // x == y --> x - y == 0
3016 const SCEV *N = SE.getSCEV(NV);
3017 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3018 // S is normalized, so normalize N before folding it into S
3019 // to keep the result normalized.
3020 N = TransformForPostIncUse(Normalize, N, CI, nullptr,
3021 LF.PostIncLoops, SE, DT);
3022 Kind = LSRUse::ICmpZero;
3023 S = SE.getMinusSCEV(N, S);
3026 // -1 and the negations of all interesting strides (except the negation
3027 // of -1) are now also interesting.
3028 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3029 if (Factors[i] != -1)
3030 Factors.insert(-(uint64_t)Factors[i]);
3034 // Set up the initial formula for this use.
3035 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3037 LF.Offset = P.second;
3038 LSRUse &LU = Uses[LF.LUIdx];
3039 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3040 if (!LU.WidestFixupType ||
3041 SE.getTypeSizeInBits(LU.WidestFixupType) <
3042 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3043 LU.WidestFixupType = LF.OperandValToReplace->getType();
3045 // If this is the first use of this LSRUse, give it a formula.
3046 if (LU.Formulae.empty()) {
3047 InsertInitialFormula(S, LU, LF.LUIdx);
3048 CountRegisters(LU.Formulae.back(), LF.LUIdx);
3052 DEBUG(print_fixups(dbgs()));
3055 /// InsertInitialFormula - Insert a formula for the given expression into
3056 /// the given use, separating out loop-variant portions from loop-invariant
3057 /// and loop-computable portions.
3059 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3060 // Mark uses whose expressions cannot be expanded.
3061 if (!isSafeToExpand(S, SE))
3062 LU.RigidFormula = true;
3065 F.InitialMatch(S, L, SE);
3066 bool Inserted = InsertFormula(LU, LUIdx, F);
3067 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3070 /// InsertSupplementalFormula - Insert a simple single-register formula for
3071 /// the given expression into the given use.
3073 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3074 LSRUse &LU, size_t LUIdx) {
3076 F.BaseRegs.push_back(S);
3077 F.HasBaseReg = true;
3078 bool Inserted = InsertFormula(LU, LUIdx, F);
3079 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3082 /// CountRegisters - Note which registers are used by the given formula,
3083 /// updating RegUses.
3084 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3086 RegUses.CountRegister(F.ScaledReg, LUIdx);
3087 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3088 E = F.BaseRegs.end(); I != E; ++I)
3089 RegUses.CountRegister(*I, LUIdx);
3092 /// InsertFormula - If the given formula has not yet been inserted, add it to
3093 /// the list, and return true. Return false otherwise.
3094 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3095 // Do not insert formula that we will not be able to expand.
3096 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3097 "Formula is illegal");
3098 if (!LU.InsertFormula(F))
3101 CountRegisters(F, LUIdx);
3105 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3106 /// loop-invariant values which we're tracking. These other uses will pin these
3107 /// values in registers, making them less profitable for elimination.
3108 /// TODO: This currently misses non-constant addrec step registers.
3109 /// TODO: Should this give more weight to users inside the loop?
3111 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3112 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3113 SmallPtrSet<const SCEV *, 32> Visited;
3115 while (!Worklist.empty()) {
3116 const SCEV *S = Worklist.pop_back_val();
3118 // Don't process the same SCEV twice
3119 if (!Visited.insert(S).second)
3122 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3123 Worklist.append(N->op_begin(), N->op_end());
3124 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3125 Worklist.push_back(C->getOperand());
3126 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3127 Worklist.push_back(D->getLHS());
3128 Worklist.push_back(D->getRHS());
3129 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3130 const Value *V = US->getValue();
3131 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3132 // Look for instructions defined outside the loop.
3133 if (L->contains(Inst)) continue;
3134 } else if (isa<UndefValue>(V))
3135 // Undef doesn't have a live range, so it doesn't matter.
3137 for (const Use &U : V->uses()) {
3138 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3139 // Ignore non-instructions.
3142 // Ignore instructions in other functions (as can happen with
3144 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3146 // Ignore instructions not dominated by the loop.
3147 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3148 UserInst->getParent() :
3149 cast<PHINode>(UserInst)->getIncomingBlock(
3150 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3151 if (!DT.dominates(L->getHeader(), UseBB))
3153 // Ignore uses which are part of other SCEV expressions, to avoid
3154 // analyzing them multiple times.
3155 if (SE.isSCEVable(UserInst->getType())) {
3156 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3157 // If the user is a no-op, look through to its uses.
3158 if (!isa<SCEVUnknown>(UserS))
3162 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3166 // Ignore icmp instructions which are already being analyzed.
3167 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3168 unsigned OtherIdx = !U.getOperandNo();
3169 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3170 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3174 LSRFixup &LF = getNewFixup();
3175 LF.UserInst = const_cast<Instruction *>(UserInst);
3176 LF.OperandValToReplace = U;
3177 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, nullptr);
3179 LF.Offset = P.second;
3180 LSRUse &LU = Uses[LF.LUIdx];
3181 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3182 if (!LU.WidestFixupType ||
3183 SE.getTypeSizeInBits(LU.WidestFixupType) <
3184 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3185 LU.WidestFixupType = LF.OperandValToReplace->getType();
3186 InsertSupplementalFormula(US, LU, LF.LUIdx);
3187 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3194 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3195 /// separate registers. If C is non-null, multiply each subexpression by C.
3197 /// Return remainder expression after factoring the subexpressions captured by
3198 /// Ops. If Ops is complete, return NULL.
3199 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3200 SmallVectorImpl<const SCEV *> &Ops,
3202 ScalarEvolution &SE,
3203 unsigned Depth = 0) {
3204 // Arbitrarily cap recursion to protect compile time.
3208 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3209 // Break out add operands.
3210 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3212 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3214 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3217 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3218 // Split a non-zero base out of an addrec.
3219 if (AR->getStart()->isZero())
3222 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3223 C, Ops, L, SE, Depth+1);
3224 // Split the non-zero AddRec unless it is part of a nested recurrence that
3225 // does not pertain to this loop.
3226 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3227 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3228 Remainder = nullptr;
3230 if (Remainder != AR->getStart()) {
3232 Remainder = SE.getConstant(AR->getType(), 0);
3233 return SE.getAddRecExpr(Remainder,
3234 AR->getStepRecurrence(SE),
3236 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3239 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3240 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3241 if (Mul->getNumOperands() != 2)
3243 if (const SCEVConstant *Op0 =
3244 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3245 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3246 const SCEV *Remainder =
3247 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3249 Ops.push_back(SE.getMulExpr(C, Remainder));
3256 /// \brief Helper function for LSRInstance::GenerateReassociations.
3257 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3258 const Formula &Base,
3259 unsigned Depth, size_t Idx,
3261 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3262 SmallVector<const SCEV *, 8> AddOps;
3263 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3265 AddOps.push_back(Remainder);
3267 if (AddOps.size() == 1)
3270 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3274 // Loop-variant "unknown" values are uninteresting; we won't be able to
3275 // do anything meaningful with them.
3276 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3279 // Don't pull a constant into a register if the constant could be folded
3280 // into an immediate field.
3281 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3282 LU.AccessTy, *J, Base.getNumRegs() > 1))
3285 // Collect all operands except *J.
3286 SmallVector<const SCEV *, 8> InnerAddOps(
3287 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3288 InnerAddOps.append(std::next(J),
3289 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3291 // Don't leave just a constant behind in a register if the constant could
3292 // be folded into an immediate field.
3293 if (InnerAddOps.size() == 1 &&
3294 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3295 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3298 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3299 if (InnerSum->isZero())
3303 // Add the remaining pieces of the add back into the new formula.
3304 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3305 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3306 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3307 InnerSumSC->getValue()->getZExtValue())) {
3309 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3311 F.ScaledReg = nullptr;
3313 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3314 } else if (IsScaledReg)
3315 F.ScaledReg = InnerSum;
3317 F.BaseRegs[Idx] = InnerSum;
3319 // Add J as its own register, or an unfolded immediate.
3320 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3321 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3322 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3323 SC->getValue()->getZExtValue()))
3325 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3327 F.BaseRegs.push_back(*J);
3328 // We may have changed the number of register in base regs, adjust the
3329 // formula accordingly.
3332 if (InsertFormula(LU, LUIdx, F))
3333 // If that formula hadn't been seen before, recurse to find more like
3335 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3339 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3341 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3342 Formula Base, unsigned Depth) {
3343 assert(Base.isCanonical() && "Input must be in the canonical form");
3344 // Arbitrarily cap recursion to protect compile time.
3348 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3349 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3351 if (Base.Scale == 1)
3352 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3353 /* Idx */ -1, /* IsScaledReg */ true);
3356 /// GenerateCombinations - Generate a formula consisting of all of the
3357 /// loop-dominating registers added into a single register.
3358 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3360 // This method is only interesting on a plurality of registers.
3361 if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3364 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3365 // processing the formula.
3369 SmallVector<const SCEV *, 4> Ops;
3370 for (SmallVectorImpl<const SCEV *>::const_iterator
3371 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3372 const SCEV *BaseReg = *I;
3373 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3374 !SE.hasComputableLoopEvolution(BaseReg, L))
3375 Ops.push_back(BaseReg);
3377 F.BaseRegs.push_back(BaseReg);
3379 if (Ops.size() > 1) {
3380 const SCEV *Sum = SE.getAddExpr(Ops);
3381 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3382 // opportunity to fold something. For now, just ignore such cases
3383 // rather than proceed with zero in a register.
3384 if (!Sum->isZero()) {
3385 F.BaseRegs.push_back(Sum);
3387 (void)InsertFormula(LU, LUIdx, F);
3392 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3393 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3394 const Formula &Base, size_t Idx,
3396 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3397 GlobalValue *GV = ExtractSymbol(G, SE);
3398 if (G->isZero() || !GV)
3402 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3407 F.BaseRegs[Idx] = G;
3408 (void)InsertFormula(LU, LUIdx, F);
3411 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3412 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3414 // We can't add a symbolic offset if the address already contains one.
3415 if (Base.BaseGV) return;
3417 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3418 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3419 if (Base.Scale == 1)
3420 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3421 /* IsScaledReg */ true);
3424 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3425 void LSRInstance::GenerateConstantOffsetsImpl(
3426 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3427 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3428 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3429 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3433 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3434 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3436 // Add the offset to the base register.
3437 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3438 // If it cancelled out, drop the base register, otherwise update it.
3439 if (NewG->isZero()) {
3442 F.ScaledReg = nullptr;
3444 F.DeleteBaseReg(F.BaseRegs[Idx]);
3446 } else if (IsScaledReg)
3449 F.BaseRegs[Idx] = NewG;
3451 (void)InsertFormula(LU, LUIdx, F);
3455 int64_t Imm = ExtractImmediate(G, SE);
3456 if (G->isZero() || Imm == 0)
3459 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3460 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3465 F.BaseRegs[Idx] = G;
3466 (void)InsertFormula(LU, LUIdx, F);
3469 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3470 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3472 // TODO: For now, just add the min and max offset, because it usually isn't
3473 // worthwhile looking at everything inbetween.
3474 SmallVector<int64_t, 2> Worklist;
3475 Worklist.push_back(LU.MinOffset);
3476 if (LU.MaxOffset != LU.MinOffset)
3477 Worklist.push_back(LU.MaxOffset);
3479 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3480 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3481 if (Base.Scale == 1)
3482 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3483 /* IsScaledReg */ true);
3486 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3487 /// the comparison. For example, x == y -> x*c == y*c.
3488 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3490 if (LU.Kind != LSRUse::ICmpZero) return;
3492 // Determine the integer type for the base formula.
3493 Type *IntTy = Base.getType();
3495 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3497 // Don't do this if there is more than one offset.
3498 if (LU.MinOffset != LU.MaxOffset) return;
3500 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3502 // Check each interesting stride.
3503 for (SmallSetVector<int64_t, 8>::const_iterator
3504 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3505 int64_t Factor = *I;
3507 // Check that the multiplication doesn't overflow.
3508 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3510 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3511 if (NewBaseOffset / Factor != Base.BaseOffset)
3513 // If the offset will be truncated at this use, check that it is in bounds.
3514 if (!IntTy->isPointerTy() &&
3515 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3518 // Check that multiplying with the use offset doesn't overflow.
3519 int64_t Offset = LU.MinOffset;
3520 if (Offset == INT64_MIN && Factor == -1)
3522 Offset = (uint64_t)Offset * Factor;
3523 if (Offset / Factor != LU.MinOffset)
3525 // If the offset will be truncated at this use, check that it is in bounds.
3526 if (!IntTy->isPointerTy() &&
3527 !ConstantInt::isValueValidForType(IntTy, Offset))
3531 F.BaseOffset = NewBaseOffset;
3533 // Check that this scale is legal.
3534 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3537 // Compensate for the use having MinOffset built into it.
3538 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3540 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3542 // Check that multiplying with each base register doesn't overflow.
3543 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3544 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3545 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3549 // Check that multiplying with the scaled register doesn't overflow.
3551 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3552 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3556 // Check that multiplying with the unfolded offset doesn't overflow.
3557 if (F.UnfoldedOffset != 0) {
3558 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3560 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3561 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3563 // If the offset will be truncated, check that it is in bounds.
3564 if (!IntTy->isPointerTy() &&
3565 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3569 // If we make it here and it's legal, add it.
3570 (void)InsertFormula(LU, LUIdx, F);
3575 /// GenerateScales - Generate stride factor reuse formulae by making use of
3576 /// scaled-offset address modes, for example.
3577 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3578 // Determine the integer type for the base formula.
3579 Type *IntTy = Base.getType();
3582 // If this Formula already has a scaled register, we can't add another one.
3583 // Try to unscale the formula to generate a better scale.
3584 if (Base.Scale != 0 && !Base.Unscale())
3587 assert(Base.Scale == 0 && "Unscale did not did its job!");
3589 // Check each interesting stride.
3590 for (SmallSetVector<int64_t, 8>::const_iterator
3591 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3592 int64_t Factor = *I;
3594 Base.Scale = Factor;
3595 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3596 // Check whether this scale is going to be legal.
3597 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3599 // As a special-case, handle special out-of-loop Basic users specially.
3600 // TODO: Reconsider this special case.
3601 if (LU.Kind == LSRUse::Basic &&
3602 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3603 LU.AccessTy, Base) &&
3604 LU.AllFixupsOutsideLoop)
3605 LU.Kind = LSRUse::Special;
3609 // For an ICmpZero, negating a solitary base register won't lead to
3611 if (LU.Kind == LSRUse::ICmpZero &&
3612 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3614 // For each addrec base reg, apply the scale, if possible.
3615 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3616 if (const SCEVAddRecExpr *AR =
3617 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3618 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3619 if (FactorS->isZero())
3621 // Divide out the factor, ignoring high bits, since we'll be
3622 // scaling the value back up in the end.
3623 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3624 // TODO: This could be optimized to avoid all the copying.
3626 F.ScaledReg = Quotient;
3627 F.DeleteBaseReg(F.BaseRegs[i]);
3628 // The canonical representation of 1*reg is reg, which is already in
3629 // Base. In that case, do not try to insert the formula, it will be
3631 if (F.Scale == 1 && F.BaseRegs.empty())
3633 (void)InsertFormula(LU, LUIdx, F);
3639 /// GenerateTruncates - Generate reuse formulae from different IV types.
3640 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3641 // Don't bother truncating symbolic values.
3642 if (Base.BaseGV) return;
3644 // Determine the integer type for the base formula.
3645 Type *DstTy = Base.getType();
3647 DstTy = SE.getEffectiveSCEVType(DstTy);
3649 for (SmallSetVector<Type *, 4>::const_iterator
3650 I = Types.begin(), E = Types.end(); I != E; ++I) {
3652 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3655 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3656 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3657 JE = F.BaseRegs.end(); J != JE; ++J)
3658 *J = SE.getAnyExtendExpr(*J, SrcTy);
3660 // TODO: This assumes we've done basic processing on all uses and
3661 // have an idea what the register usage is.
3662 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3665 (void)InsertFormula(LU, LUIdx, F);
3672 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3673 /// defer modifications so that the search phase doesn't have to worry about
3674 /// the data structures moving underneath it.
3678 const SCEV *OrigReg;
3680 WorkItem(size_t LI, int64_t I, const SCEV *R)
3681 : LUIdx(LI), Imm(I), OrigReg(R) {}
3683 void print(raw_ostream &OS) const;
3689 void WorkItem::print(raw_ostream &OS) const {
3690 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3691 << " , add offset " << Imm;
3694 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3695 void WorkItem::dump() const {
3696 print(errs()); errs() << '\n';
3700 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3701 /// distance apart and try to form reuse opportunities between them.
3702 void LSRInstance::GenerateCrossUseConstantOffsets() {
3703 // Group the registers by their value without any added constant offset.
3704 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3705 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3707 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3708 SmallVector<const SCEV *, 8> Sequence;
3709 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3711 const SCEV *Reg = *I;
3712 int64_t Imm = ExtractImmediate(Reg, SE);
3713 std::pair<RegMapTy::iterator, bool> Pair =
3714 Map.insert(std::make_pair(Reg, ImmMapTy()));
3716 Sequence.push_back(Reg);
3717 Pair.first->second.insert(std::make_pair(Imm, *I));
3718 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3721 // Now examine each set of registers with the same base value. Build up
3722 // a list of work to do and do the work in a separate step so that we're
3723 // not adding formulae and register counts while we're searching.
3724 SmallVector<WorkItem, 32> WorkItems;
3725 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3726 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3727 E = Sequence.end(); I != E; ++I) {
3728 const SCEV *Reg = *I;
3729 const ImmMapTy &Imms = Map.find(Reg)->second;
3731 // It's not worthwhile looking for reuse if there's only one offset.
3732 if (Imms.size() == 1)
3735 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3736 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3738 dbgs() << ' ' << J->first;
3741 // Examine each offset.
3742 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3744 const SCEV *OrigReg = J->second;
3746 int64_t JImm = J->first;
3747 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3749 if (!isa<SCEVConstant>(OrigReg) &&
3750 UsedByIndicesMap[Reg].count() == 1) {
3751 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3755 // Conservatively examine offsets between this orig reg a few selected
3757 ImmMapTy::const_iterator OtherImms[] = {
3758 Imms.begin(), std::prev(Imms.end()),
3759 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3762 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3763 ImmMapTy::const_iterator M = OtherImms[i];
3764 if (M == J || M == JE) continue;
3766 // Compute the difference between the two.
3767 int64_t Imm = (uint64_t)JImm - M->first;
3768 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3769 LUIdx = UsedByIndices.find_next(LUIdx))
3770 // Make a memo of this use, offset, and register tuple.
3771 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3772 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3779 UsedByIndicesMap.clear();
3780 UniqueItems.clear();
3782 // Now iterate through the worklist and add new formulae.
3783 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3784 E = WorkItems.end(); I != E; ++I) {
3785 const WorkItem &WI = *I;
3786 size_t LUIdx = WI.LUIdx;
3787 LSRUse &LU = Uses[LUIdx];
3788 int64_t Imm = WI.Imm;
3789 const SCEV *OrigReg = WI.OrigReg;
3791 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3792 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3793 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3795 // TODO: Use a more targeted data structure.
3796 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3797 Formula F = LU.Formulae[L];
3798 // FIXME: The code for the scaled and unscaled registers looks
3799 // very similar but slightly different. Investigate if they
3800 // could be merged. That way, we would not have to unscale the
3803 // Use the immediate in the scaled register.
3804 if (F.ScaledReg == OrigReg) {
3805 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3806 // Don't create 50 + reg(-50).
3807 if (F.referencesReg(SE.getSCEV(
3808 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3811 NewF.BaseOffset = Offset;
3812 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3815 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3817 // If the new scale is a constant in a register, and adding the constant
3818 // value to the immediate would produce a value closer to zero than the
3819 // immediate itself, then the formula isn't worthwhile.
3820 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3821 if (C->getValue()->isNegative() !=
3822 (NewF.BaseOffset < 0) &&
3823 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3824 .ule(std::abs(NewF.BaseOffset)))
3828 NewF.Canonicalize();
3829 (void)InsertFormula(LU, LUIdx, NewF);
3831 // Use the immediate in a base register.
3832 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3833 const SCEV *BaseReg = F.BaseRegs[N];
3834 if (BaseReg != OrigReg)
3837 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3838 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3839 LU.Kind, LU.AccessTy, NewF)) {
3840 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3843 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3845 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3847 // If the new formula has a constant in a register, and adding the
3848 // constant value to the immediate would produce a value closer to
3849 // zero than the immediate itself, then the formula isn't worthwhile.
3850 for (SmallVectorImpl<const SCEV *>::const_iterator
3851 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3853 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3854 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3855 std::abs(NewF.BaseOffset)) &&
3856 (C->getValue()->getValue() +
3857 NewF.BaseOffset).countTrailingZeros() >=
3858 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3862 NewF.Canonicalize();
3863 (void)InsertFormula(LU, LUIdx, NewF);
3872 /// GenerateAllReuseFormulae - Generate formulae for each use.
3874 LSRInstance::GenerateAllReuseFormulae() {
3875 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3876 // queries are more precise.
3877 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3878 LSRUse &LU = Uses[LUIdx];
3879 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3880 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3881 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3882 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3884 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3885 LSRUse &LU = Uses[LUIdx];
3886 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3887 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3888 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3889 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3890 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3891 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3892 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3893 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3895 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3896 LSRUse &LU = Uses[LUIdx];
3897 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3898 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3901 GenerateCrossUseConstantOffsets();
3903 DEBUG(dbgs() << "\n"
3904 "After generating reuse formulae:\n";
3905 print_uses(dbgs()));
3908 /// If there are multiple formulae with the same set of registers used
3909 /// by other uses, pick the best one and delete the others.
3910 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3911 DenseSet<const SCEV *> VisitedRegs;
3912 SmallPtrSet<const SCEV *, 16> Regs;
3913 SmallPtrSet<const SCEV *, 16> LoserRegs;
3915 bool ChangedFormulae = false;
3918 // Collect the best formula for each unique set of shared registers. This
3919 // is reset for each use.
3920 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3922 BestFormulaeTy BestFormulae;
3924 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3925 LSRUse &LU = Uses[LUIdx];
3926 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3929 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3930 FIdx != NumForms; ++FIdx) {
3931 Formula &F = LU.Formulae[FIdx];
3933 // Some formulas are instant losers. For example, they may depend on
3934 // nonexistent AddRecs from other loops. These need to be filtered
3935 // immediately, otherwise heuristics could choose them over others leading
3936 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3937 // avoids the need to recompute this information across formulae using the
3938 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3939 // the corresponding bad register from the Regs set.
3942 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3944 if (CostF.isLoser()) {
3945 // During initial formula generation, undesirable formulae are generated
3946 // by uses within other loops that have some non-trivial address mode or
3947 // use the postinc form of the IV. LSR needs to provide these formulae
3948 // as the basis of rediscovering the desired formula that uses an AddRec
3949 // corresponding to the existing phi. Once all formulae have been
3950 // generated, these initial losers may be pruned.
3951 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3955 SmallVector<const SCEV *, 4> Key;
3956 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3957 JE = F.BaseRegs.end(); J != JE; ++J) {
3958 const SCEV *Reg = *J;
3959 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3963 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3964 Key.push_back(F.ScaledReg);
3965 // Unstable sort by host order ok, because this is only used for
3967 std::sort(Key.begin(), Key.end());
3969 std::pair<BestFormulaeTy::const_iterator, bool> P =
3970 BestFormulae.insert(std::make_pair(Key, FIdx));
3974 Formula &Best = LU.Formulae[P.first->second];
3978 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3980 if (CostF < CostBest)
3982 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3984 " in favor of formula "; Best.print(dbgs());
3988 ChangedFormulae = true;
3990 LU.DeleteFormula(F);
3996 // Now that we've filtered out some formulae, recompute the Regs set.
3998 LU.RecomputeRegs(LUIdx, RegUses);
4000 // Reset this to prepare for the next use.
4001 BestFormulae.clear();
4004 DEBUG(if (ChangedFormulae) {
4006 "After filtering out undesirable candidates:\n";
4011 // This is a rough guess that seems to work fairly well.
4012 static const size_t ComplexityLimit = UINT16_MAX;
4014 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
4015 /// solutions the solver might have to consider. It almost never considers
4016 /// this many solutions because it prune the search space, but the pruning
4017 /// isn't always sufficient.
4018 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4020 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4021 E = Uses.end(); I != E; ++I) {
4022 size_t FSize = I->Formulae.size();
4023 if (FSize >= ComplexityLimit) {
4024 Power = ComplexityLimit;
4028 if (Power >= ComplexityLimit)
4034 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
4035 /// of the registers of another formula, it won't help reduce register
4036 /// pressure (though it may not necessarily hurt register pressure); remove
4037 /// it to simplify the system.
4038 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4039 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4040 DEBUG(dbgs() << "The search space is too complex.\n");
4042 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4043 "which use a superset of registers used by other "
4046 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4047 LSRUse &LU = Uses[LUIdx];
4049 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4050 Formula &F = LU.Formulae[i];
4051 // Look for a formula with a constant or GV in a register. If the use
4052 // also has a formula with that same value in an immediate field,
4053 // delete the one that uses a register.
4054 for (SmallVectorImpl<const SCEV *>::const_iterator
4055 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4056 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4058 NewF.BaseOffset += C->getValue()->getSExtValue();
4059 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4060 (I - F.BaseRegs.begin()));
4061 if (LU.HasFormulaWithSameRegs(NewF)) {
4062 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4063 LU.DeleteFormula(F);
4069 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4070 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4074 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4075 (I - F.BaseRegs.begin()));
4076 if (LU.HasFormulaWithSameRegs(NewF)) {
4077 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4079 LU.DeleteFormula(F);
4090 LU.RecomputeRegs(LUIdx, RegUses);
4093 DEBUG(dbgs() << "After pre-selection:\n";
4094 print_uses(dbgs()));
4098 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
4099 /// for expressions like A, A+1, A+2, etc., allocate a single register for
4101 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4102 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4105 DEBUG(dbgs() << "The search space is too complex.\n"
4106 "Narrowing the search space by assuming that uses separated "
4107 "by a constant offset will use the same registers.\n");
4109 // This is especially useful for unrolled loops.
4111 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4112 LSRUse &LU = Uses[LUIdx];
4113 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4114 E = LU.Formulae.end(); I != E; ++I) {
4115 const Formula &F = *I;
4116 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4119 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4123 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4124 LU.Kind, LU.AccessTy))
4127 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4129 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4131 // Update the relocs to reference the new use.
4132 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
4133 E = Fixups.end(); I != E; ++I) {
4134 LSRFixup &Fixup = *I;
4135 if (Fixup.LUIdx == LUIdx) {
4136 Fixup.LUIdx = LUThatHas - &Uses.front();
4137 Fixup.Offset += F.BaseOffset;
4138 // Add the new offset to LUThatHas' offset list.
4139 if (LUThatHas->Offsets.back() != Fixup.Offset) {
4140 LUThatHas->Offsets.push_back(Fixup.Offset);
4141 if (Fixup.Offset > LUThatHas->MaxOffset)
4142 LUThatHas->MaxOffset = Fixup.Offset;
4143 if (Fixup.Offset < LUThatHas->MinOffset)
4144 LUThatHas->MinOffset = Fixup.Offset;
4146 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4148 if (Fixup.LUIdx == NumUses-1)
4149 Fixup.LUIdx = LUIdx;
4152 // Delete formulae from the new use which are no longer legal.
4154 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4155 Formula &F = LUThatHas->Formulae[i];
4156 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4157 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4158 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4160 LUThatHas->DeleteFormula(F);
4168 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4170 // Delete the old use.
4171 DeleteUse(LU, LUIdx);
4178 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4181 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4182 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4183 /// we've done more filtering, as it may be able to find more formulae to
4185 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4186 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4187 DEBUG(dbgs() << "The search space is too complex.\n");
4189 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4190 "undesirable dedicated registers.\n");
4192 FilterOutUndesirableDedicatedRegisters();
4194 DEBUG(dbgs() << "After pre-selection:\n";
4195 print_uses(dbgs()));
4199 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4200 /// to be profitable, and then in any use which has any reference to that
4201 /// register, delete all formulae which do not reference that register.
4202 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4203 // With all other options exhausted, loop until the system is simple
4204 // enough to handle.
4205 SmallPtrSet<const SCEV *, 4> Taken;
4206 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4207 // Ok, we have too many of formulae on our hands to conveniently handle.
4208 // Use a rough heuristic to thin out the list.
4209 DEBUG(dbgs() << "The search space is too complex.\n");
4211 // Pick the register which is used by the most LSRUses, which is likely
4212 // to be a good reuse register candidate.
4213 const SCEV *Best = nullptr;
4214 unsigned BestNum = 0;
4215 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4217 const SCEV *Reg = *I;
4218 if (Taken.count(Reg))
4223 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4224 if (Count > BestNum) {
4231 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4232 << " will yield profitable reuse.\n");
4235 // In any use with formulae which references this register, delete formulae
4236 // which don't reference it.
4237 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4238 LSRUse &LU = Uses[LUIdx];
4239 if (!LU.Regs.count(Best)) continue;
4242 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4243 Formula &F = LU.Formulae[i];
4244 if (!F.referencesReg(Best)) {
4245 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4246 LU.DeleteFormula(F);
4250 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4256 LU.RecomputeRegs(LUIdx, RegUses);
4259 DEBUG(dbgs() << "After pre-selection:\n";
4260 print_uses(dbgs()));
4264 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4265 /// formulae to choose from, use some rough heuristics to prune down the number
4266 /// of formulae. This keeps the main solver from taking an extraordinary amount
4267 /// of time in some worst-case scenarios.
4268 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4269 NarrowSearchSpaceByDetectingSupersets();
4270 NarrowSearchSpaceByCollapsingUnrolledCode();
4271 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4272 NarrowSearchSpaceByPickingWinnerRegs();
4275 /// SolveRecurse - This is the recursive solver.
4276 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4278 SmallVectorImpl<const Formula *> &Workspace,
4279 const Cost &CurCost,
4280 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4281 DenseSet<const SCEV *> &VisitedRegs) const {
4284 // - use more aggressive filtering
4285 // - sort the formula so that the most profitable solutions are found first
4286 // - sort the uses too
4288 // - don't compute a cost, and then compare. compare while computing a cost
4290 // - track register sets with SmallBitVector
4292 const LSRUse &LU = Uses[Workspace.size()];
4294 // If this use references any register that's already a part of the
4295 // in-progress solution, consider it a requirement that a formula must
4296 // reference that register in order to be considered. This prunes out
4297 // unprofitable searching.
4298 SmallSetVector<const SCEV *, 4> ReqRegs;
4299 for (const SCEV *S : CurRegs)
4300 if (LU.Regs.count(S))
4303 SmallPtrSet<const SCEV *, 16> NewRegs;
4305 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4306 E = LU.Formulae.end(); I != E; ++I) {
4307 const Formula &F = *I;
4309 // Ignore formulae which may not be ideal in terms of register reuse of
4310 // ReqRegs. The formula should use all required registers before
4311 // introducing new ones.
4312 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4313 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4314 JE = ReqRegs.end(); J != JE; ++J) {
4315 const SCEV *Reg = *J;
4316 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4317 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) !=
4320 if (NumReqRegsToFind == 0)
4324 if (NumReqRegsToFind != 0) {
4325 // If none of the formulae satisfied the required registers, then we could
4326 // clear ReqRegs and try again. Currently, we simply give up in this case.
4330 // Evaluate the cost of the current formula. If it's already worse than
4331 // the current best, prune the search at that point.
4334 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4336 if (NewCost < SolutionCost) {
4337 Workspace.push_back(&F);
4338 if (Workspace.size() != Uses.size()) {
4339 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4340 NewRegs, VisitedRegs);
4341 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4342 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4344 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4345 dbgs() << ".\n Regs:";
4346 for (const SCEV *S : NewRegs)
4347 dbgs() << ' ' << *S;
4350 SolutionCost = NewCost;
4351 Solution = Workspace;
4353 Workspace.pop_back();
4358 /// Solve - Choose one formula from each use. Return the results in the given
4359 /// Solution vector.
4360 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4361 SmallVector<const Formula *, 8> Workspace;
4363 SolutionCost.Lose();
4365 SmallPtrSet<const SCEV *, 16> CurRegs;
4366 DenseSet<const SCEV *> VisitedRegs;
4367 Workspace.reserve(Uses.size());
4369 // SolveRecurse does all the work.
4370 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4371 CurRegs, VisitedRegs);
4372 if (Solution.empty()) {
4373 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4377 // Ok, we've now made all our decisions.
4378 DEBUG(dbgs() << "\n"
4379 "The chosen solution requires "; SolutionCost.print(dbgs());
4381 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4383 Uses[i].print(dbgs());
4386 Solution[i]->print(dbgs());
4390 assert(Solution.size() == Uses.size() && "Malformed solution!");
4393 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4394 /// the dominator tree far as we can go while still being dominated by the
4395 /// input positions. This helps canonicalize the insert position, which
4396 /// encourages sharing.
4397 BasicBlock::iterator
4398 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4399 const SmallVectorImpl<Instruction *> &Inputs)
4402 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4403 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4406 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4407 if (!Rung) return IP;
4408 Rung = Rung->getIDom();
4409 if (!Rung) return IP;
4410 IDom = Rung->getBlock();
4412 // Don't climb into a loop though.
4413 const Loop *IDomLoop = LI.getLoopFor(IDom);
4414 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4415 if (IDomDepth <= IPLoopDepth &&
4416 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4420 bool AllDominate = true;
4421 Instruction *BetterPos = nullptr;
4422 Instruction *Tentative = IDom->getTerminator();
4423 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4424 E = Inputs.end(); I != E; ++I) {
4425 Instruction *Inst = *I;
4426 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4427 AllDominate = false;
4430 // Attempt to find an insert position in the middle of the block,
4431 // instead of at the end, so that it can be used for other expansions.
4432 if (IDom == Inst->getParent() &&
4433 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4434 BetterPos = std::next(BasicBlock::iterator(Inst));
4447 /// AdjustInsertPositionForExpand - Determine an input position which will be
4448 /// dominated by the operands and which will dominate the result.
4449 BasicBlock::iterator
4450 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4453 SCEVExpander &Rewriter) const {
4454 // Collect some instructions which must be dominated by the
4455 // expanding replacement. These must be dominated by any operands that
4456 // will be required in the expansion.
4457 SmallVector<Instruction *, 4> Inputs;
4458 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4459 Inputs.push_back(I);
4460 if (LU.Kind == LSRUse::ICmpZero)
4461 if (Instruction *I =
4462 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4463 Inputs.push_back(I);
4464 if (LF.PostIncLoops.count(L)) {
4465 if (LF.isUseFullyOutsideLoop(L))
4466 Inputs.push_back(L->getLoopLatch()->getTerminator());
4468 Inputs.push_back(IVIncInsertPos);
4470 // The expansion must also be dominated by the increment positions of any
4471 // loops it for which it is using post-inc mode.
4472 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4473 E = LF.PostIncLoops.end(); I != E; ++I) {
4474 const Loop *PIL = *I;
4475 if (PIL == L) continue;
4477 // Be dominated by the loop exit.
4478 SmallVector<BasicBlock *, 4> ExitingBlocks;
4479 PIL->getExitingBlocks(ExitingBlocks);
4480 if (!ExitingBlocks.empty()) {
4481 BasicBlock *BB = ExitingBlocks[0];
4482 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4483 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4484 Inputs.push_back(BB->getTerminator());
4488 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4489 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4490 "Insertion point must be a normal instruction");
4492 // Then, climb up the immediate dominator tree as far as we can go while
4493 // still being dominated by the input positions.
4494 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4496 // Don't insert instructions before PHI nodes.
4497 while (isa<PHINode>(IP)) ++IP;
4499 // Ignore landingpad instructions.
4500 while (isa<LandingPadInst>(IP)) ++IP;
4502 // Ignore debug intrinsics.
4503 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4505 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4506 // IP consistent across expansions and allows the previously inserted
4507 // instructions to be reused by subsequent expansion.
4508 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4513 /// Expand - Emit instructions for the leading candidate expression for this
4514 /// LSRUse (this is called "expanding").
4515 Value *LSRInstance::Expand(const LSRFixup &LF,
4517 BasicBlock::iterator IP,
4518 SCEVExpander &Rewriter,
4519 SmallVectorImpl<WeakVH> &DeadInsts) const {
4520 const LSRUse &LU = Uses[LF.LUIdx];
4521 if (LU.RigidFormula)
4522 return LF.OperandValToReplace;
4524 // Determine an input position which will be dominated by the operands and
4525 // which will dominate the result.
4526 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4528 // Inform the Rewriter if we have a post-increment use, so that it can
4529 // perform an advantageous expansion.
4530 Rewriter.setPostInc(LF.PostIncLoops);
4532 // This is the type that the user actually needs.
4533 Type *OpTy = LF.OperandValToReplace->getType();
4534 // This will be the type that we'll initially expand to.
4535 Type *Ty = F.getType();
4537 // No type known; just expand directly to the ultimate type.
4539 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4540 // Expand directly to the ultimate type if it's the right size.
4542 // This is the type to do integer arithmetic in.
4543 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4545 // Build up a list of operands to add together to form the full base.
4546 SmallVector<const SCEV *, 8> Ops;
4548 // Expand the BaseRegs portion.
4549 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4550 E = F.BaseRegs.end(); I != E; ++I) {
4551 const SCEV *Reg = *I;
4552 assert(!Reg->isZero() && "Zero allocated in a base register!");
4554 // If we're expanding for a post-inc user, make the post-inc adjustment.
4555 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4556 Reg = TransformForPostIncUse(Denormalize, Reg,
4557 LF.UserInst, LF.OperandValToReplace,
4560 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr, IP)));
4563 // Expand the ScaledReg portion.
4564 Value *ICmpScaledV = nullptr;
4566 const SCEV *ScaledS = F.ScaledReg;
4568 // If we're expanding for a post-inc user, make the post-inc adjustment.
4569 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4570 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4571 LF.UserInst, LF.OperandValToReplace,
4574 if (LU.Kind == LSRUse::ICmpZero) {
4575 // Expand ScaleReg as if it was part of the base regs.
4578 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, IP)));
4580 // An interesting way of "folding" with an icmp is to use a negated
4581 // scale, which we'll implement by inserting it into the other operand
4583 assert(F.Scale == -1 &&
4584 "The only scale supported by ICmpZero uses is -1!");
4585 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr, IP);
4588 // Otherwise just expand the scaled register and an explicit scale,
4589 // which is expected to be matched as part of the address.
4591 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4592 // Unless the addressing mode will not be folded.
4593 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4594 isAMCompletelyFolded(TTI, LU, F)) {
4595 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4597 Ops.push_back(SE.getUnknown(FullV));
4599 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, IP));
4602 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4603 Ops.push_back(ScaledS);
4607 // Expand the GV portion.
4609 // Flush the operand list to suppress SCEVExpander hoisting.
4611 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4613 Ops.push_back(SE.getUnknown(FullV));
4615 Ops.push_back(SE.getUnknown(F.BaseGV));
4618 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4619 // unfolded offsets. LSR assumes they both live next to their uses.
4621 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4623 Ops.push_back(SE.getUnknown(FullV));
4626 // Expand the immediate portion.
4627 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4629 if (LU.Kind == LSRUse::ICmpZero) {
4630 // The other interesting way of "folding" with an ICmpZero is to use a
4631 // negated immediate.
4633 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4635 Ops.push_back(SE.getUnknown(ICmpScaledV));
4636 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4639 // Just add the immediate values. These again are expected to be matched
4640 // as part of the address.
4641 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4645 // Expand the unfolded offset portion.
4646 int64_t UnfoldedOffset = F.UnfoldedOffset;
4647 if (UnfoldedOffset != 0) {
4648 // Just add the immediate values.
4649 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4653 // Emit instructions summing all the operands.
4654 const SCEV *FullS = Ops.empty() ?
4655 SE.getConstant(IntTy, 0) :
4657 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4659 // We're done expanding now, so reset the rewriter.
4660 Rewriter.clearPostInc();
4662 // An ICmpZero Formula represents an ICmp which we're handling as a
4663 // comparison against zero. Now that we've expanded an expression for that
4664 // form, update the ICmp's other operand.
4665 if (LU.Kind == LSRUse::ICmpZero) {
4666 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4667 DeadInsts.push_back(CI->getOperand(1));
4668 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4669 "a scale at the same time!");
4670 if (F.Scale == -1) {
4671 if (ICmpScaledV->getType() != OpTy) {
4673 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4675 ICmpScaledV, OpTy, "tmp", CI);
4678 CI->setOperand(1, ICmpScaledV);
4680 // A scale of 1 means that the scale has been expanded as part of the
4682 assert((F.Scale == 0 || F.Scale == 1) &&
4683 "ICmp does not support folding a global value and "
4684 "a scale at the same time!");
4685 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4687 if (C->getType() != OpTy)
4688 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4692 CI->setOperand(1, C);
4699 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4700 /// of their operands effectively happens in their predecessor blocks, so the
4701 /// expression may need to be expanded in multiple places.
4702 void LSRInstance::RewriteForPHI(PHINode *PN,
4705 SCEVExpander &Rewriter,
4706 SmallVectorImpl<WeakVH> &DeadInsts,
4708 DenseMap<BasicBlock *, Value *> Inserted;
4709 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4710 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4711 BasicBlock *BB = PN->getIncomingBlock(i);
4713 // If this is a critical edge, split the edge so that we do not insert
4714 // the code on all predecessor/successor paths. We do this unless this
4715 // is the canonical backedge for this loop, which complicates post-inc
4717 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4718 !isa<IndirectBrInst>(BB->getTerminator())) {
4719 BasicBlock *Parent = PN->getParent();
4720 Loop *PNLoop = LI.getLoopFor(Parent);
4721 if (!PNLoop || Parent != PNLoop->getHeader()) {
4722 // Split the critical edge.
4723 BasicBlock *NewBB = nullptr;
4724 if (!Parent->isLandingPad()) {
4725 NewBB = SplitCriticalEdge(BB, Parent,
4726 CriticalEdgeSplittingOptions(&DT, &LI)
4727 .setMergeIdenticalEdges()
4728 .setDontDeleteUselessPHIs());
4730 SmallVector<BasicBlock*, 2> NewBBs;
4731 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs,
4732 /*AliasAnalysis*/ nullptr, &DT, &LI);
4735 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4736 // phi predecessors are identical. The simple thing to do is skip
4737 // splitting in this case rather than complicate the API.
4739 // If PN is outside of the loop and BB is in the loop, we want to
4740 // move the block to be immediately before the PHI block, not
4741 // immediately after BB.
4742 if (L->contains(BB) && !L->contains(PN))
4743 NewBB->moveBefore(PN->getParent());
4745 // Splitting the edge can reduce the number of PHI entries we have.
4746 e = PN->getNumIncomingValues();
4748 i = PN->getBasicBlockIndex(BB);
4753 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4754 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
4756 PN->setIncomingValue(i, Pair.first->second);
4758 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4760 // If this is reuse-by-noop-cast, insert the noop cast.
4761 Type *OpTy = LF.OperandValToReplace->getType();
4762 if (FullV->getType() != OpTy)
4764 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4766 FullV, LF.OperandValToReplace->getType(),
4767 "tmp", BB->getTerminator());
4769 PN->setIncomingValue(i, FullV);
4770 Pair.first->second = FullV;
4775 /// Rewrite - Emit instructions for the leading candidate expression for this
4776 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4777 /// the newly expanded value.
4778 void LSRInstance::Rewrite(const LSRFixup &LF,
4780 SCEVExpander &Rewriter,
4781 SmallVectorImpl<WeakVH> &DeadInsts,
4783 // First, find an insertion point that dominates UserInst. For PHI nodes,
4784 // find the nearest block which dominates all the relevant uses.
4785 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4786 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4788 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4790 // If this is reuse-by-noop-cast, insert the noop cast.
4791 Type *OpTy = LF.OperandValToReplace->getType();
4792 if (FullV->getType() != OpTy) {
4794 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4795 FullV, OpTy, "tmp", LF.UserInst);
4799 // Update the user. ICmpZero is handled specially here (for now) because
4800 // Expand may have updated one of the operands of the icmp already, and
4801 // its new value may happen to be equal to LF.OperandValToReplace, in
4802 // which case doing replaceUsesOfWith leads to replacing both operands
4803 // with the same value. TODO: Reorganize this.
4804 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4805 LF.UserInst->setOperand(0, FullV);
4807 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4810 DeadInsts.push_back(LF.OperandValToReplace);
4813 /// ImplementSolution - Rewrite all the fixup locations with new values,
4814 /// following the chosen solution.
4816 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4818 // Keep track of instructions we may have made dead, so that
4819 // we can remove them after we are done working.
4820 SmallVector<WeakVH, 16> DeadInsts;
4822 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
4825 Rewriter.setDebugType(DEBUG_TYPE);
4827 Rewriter.disableCanonicalMode();
4828 Rewriter.enableLSRMode();
4829 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4831 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4832 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4833 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4834 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4835 Rewriter.setChainedPhi(PN);
4838 // Expand the new value definitions and update the users.
4839 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4840 E = Fixups.end(); I != E; ++I) {
4841 const LSRFixup &Fixup = *I;
4843 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4848 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4849 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4850 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4853 // Clean up after ourselves. This must be done before deleting any
4857 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4860 LSRInstance::LSRInstance(Loop *L, Pass *P)
4861 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4862 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4863 LI(P->getAnalysis<LoopInfoWrapperPass>().getLoopInfo()),
4864 TTI(P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
4865 *L->getHeader()->getParent())),
4866 L(L), Changed(false), IVIncInsertPos(nullptr) {
4867 // If LoopSimplify form is not available, stay out of trouble.
4868 if (!L->isLoopSimplifyForm())
4871 // If there's no interesting work to be done, bail early.
4872 if (IU.empty()) return;
4874 // If there's too much analysis to be done, bail early. We won't be able to
4875 // model the problem anyway.
4876 unsigned NumUsers = 0;
4877 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4878 if (++NumUsers > MaxIVUsers) {
4879 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4886 // All dominating loops must have preheaders, or SCEVExpander may not be able
4887 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4889 // IVUsers analysis should only create users that are dominated by simple loop
4890 // headers. Since this loop should dominate all of its users, its user list
4891 // should be empty if this loop itself is not within a simple loop nest.
4892 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4893 Rung; Rung = Rung->getIDom()) {
4894 BasicBlock *BB = Rung->getBlock();
4895 const Loop *DomLoop = LI.getLoopFor(BB);
4896 if (DomLoop && DomLoop->getHeader() == BB) {
4897 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4902 DEBUG(dbgs() << "\nLSR on loop ";
4903 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4906 // First, perform some low-level loop optimizations.
4908 OptimizeLoopTermCond();
4910 // If loop preparation eliminates all interesting IV users, bail.
4911 if (IU.empty()) return;
4913 // Skip nested loops until we can model them better with formulae.
4915 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4919 // Start collecting data and preparing for the solver.
4921 CollectInterestingTypesAndFactors();
4922 CollectFixupsAndInitialFormulae();
4923 CollectLoopInvariantFixupsAndFormulae();
4925 assert(!Uses.empty() && "IVUsers reported at least one use");
4926 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4927 print_uses(dbgs()));
4929 // Now use the reuse data to generate a bunch of interesting ways
4930 // to formulate the values needed for the uses.
4931 GenerateAllReuseFormulae();
4933 FilterOutUndesirableDedicatedRegisters();
4934 NarrowSearchSpaceUsingHeuristics();
4936 SmallVector<const Formula *, 8> Solution;
4939 // Release memory that is no longer needed.
4944 if (Solution.empty())
4948 // Formulae should be legal.
4949 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4951 const LSRUse &LU = *I;
4952 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4953 JE = LU.Formulae.end();
4955 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4956 *J) && "Illegal formula generated!");
4960 // Now that we've decided what we want, make it so.
4961 ImplementSolution(Solution, P);
4964 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4965 if (Factors.empty() && Types.empty()) return;
4967 OS << "LSR has identified the following interesting factors and types: ";
4970 for (SmallSetVector<int64_t, 8>::const_iterator
4971 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4972 if (!First) OS << ", ";
4977 for (SmallSetVector<Type *, 4>::const_iterator
4978 I = Types.begin(), E = Types.end(); I != E; ++I) {
4979 if (!First) OS << ", ";
4981 OS << '(' << **I << ')';
4986 void LSRInstance::print_fixups(raw_ostream &OS) const {
4987 OS << "LSR is examining the following fixup sites:\n";
4988 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4989 E = Fixups.end(); I != E; ++I) {
4996 void LSRInstance::print_uses(raw_ostream &OS) const {
4997 OS << "LSR is examining the following uses:\n";
4998 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4999 E = Uses.end(); I != E; ++I) {
5000 const LSRUse &LU = *I;
5004 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
5005 JE = LU.Formulae.end(); J != JE; ++J) {
5013 void LSRInstance::print(raw_ostream &OS) const {
5014 print_factors_and_types(OS);
5019 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5020 void LSRInstance::dump() const {
5021 print(errs()); errs() << '\n';
5027 class LoopStrengthReduce : public LoopPass {
5029 static char ID; // Pass ID, replacement for typeid
5030 LoopStrengthReduce();
5033 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5034 void getAnalysisUsage(AnalysisUsage &AU) const override;
5039 char LoopStrengthReduce::ID = 0;
5040 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5041 "Loop Strength Reduction", false, false)
5042 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5043 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5044 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5045 INITIALIZE_PASS_DEPENDENCY(IVUsers)
5046 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5047 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5048 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5049 "Loop Strength Reduction", false, false)
5052 Pass *llvm::createLoopStrengthReducePass() {
5053 return new LoopStrengthReduce();
5056 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5057 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5060 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5061 // We split critical edges, so we change the CFG. However, we do update
5062 // many analyses if they are around.
5063 AU.addPreservedID(LoopSimplifyID);
5065 AU.addRequired<LoopInfoWrapperPass>();
5066 AU.addPreserved<LoopInfoWrapperPass>();
5067 AU.addRequiredID(LoopSimplifyID);
5068 AU.addRequired<DominatorTreeWrapperPass>();
5069 AU.addPreserved<DominatorTreeWrapperPass>();
5070 AU.addRequired<ScalarEvolution>();
5071 AU.addPreserved<ScalarEvolution>();
5072 // Requiring LoopSimplify a second time here prevents IVUsers from running
5073 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5074 AU.addRequiredID(LoopSimplifyID);
5075 AU.addRequired<IVUsers>();
5076 AU.addPreserved<IVUsers>();
5077 AU.addRequired<TargetTransformInfoWrapperPass>();
5080 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5081 if (skipOptnoneFunction(L))
5084 bool Changed = false;
5086 // Run the main LSR transformation.
5087 Changed |= LSRInstance(L, this).getChanged();
5089 // Remove any extra phis created by processing inner loops.
5090 Changed |= DeleteDeadPHIs(L->getHeader());
5091 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5092 SmallVector<WeakVH, 16> DeadInsts;
5093 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5094 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), DL, "lsr");
5096 Rewriter.setDebugType(DEBUG_TYPE);
5098 unsigned numFolded = Rewriter.replaceCongruentIVs(
5099 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
5100 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5101 *L->getHeader()->getParent()));
5104 DeleteTriviallyDeadInstructions(DeadInsts);
5105 DeleteDeadPHIs(L->getHeader());