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;
117 void print(raw_ostream &OS) const;
123 void RegSortData::print(raw_ostream &OS) const {
124 OS << "[NumUses=" << UsedByIndices.count() << ']';
127 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
128 void RegSortData::dump() const {
129 print(errs()); errs() << '\n';
135 /// RegUseTracker - Map register candidates to information about how they are
137 class RegUseTracker {
138 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
140 RegUsesTy RegUsesMap;
141 SmallVector<const SCEV *, 16> RegSequence;
144 void CountRegister(const SCEV *Reg, size_t LUIdx);
145 void DropRegister(const SCEV *Reg, size_t LUIdx);
146 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
148 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
150 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
154 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
155 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
156 iterator begin() { return RegSequence.begin(); }
157 iterator end() { return RegSequence.end(); }
158 const_iterator begin() const { return RegSequence.begin(); }
159 const_iterator end() const { return RegSequence.end(); }
165 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
166 std::pair<RegUsesTy::iterator, bool> Pair =
167 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
168 RegSortData &RSD = Pair.first->second;
170 RegSequence.push_back(Reg);
171 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
172 RSD.UsedByIndices.set(LUIdx);
176 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
177 RegUsesTy::iterator It = RegUsesMap.find(Reg);
178 assert(It != RegUsesMap.end());
179 RegSortData &RSD = It->second;
180 assert(RSD.UsedByIndices.size() > LUIdx);
181 RSD.UsedByIndices.reset(LUIdx);
185 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
186 assert(LUIdx <= LastLUIdx);
188 // Update RegUses. The data structure is not optimized for this purpose;
189 // we must iterate through it and update each of the bit vectors.
190 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
192 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
193 if (LUIdx < UsedByIndices.size())
194 UsedByIndices[LUIdx] =
195 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
196 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
201 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
202 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
203 if (I == RegUsesMap.end())
205 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
206 int i = UsedByIndices.find_first();
207 if (i == -1) return false;
208 if ((size_t)i != LUIdx) return true;
209 return UsedByIndices.find_next(i) != -1;
212 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
213 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
214 assert(I != RegUsesMap.end() && "Unknown register!");
215 return I->second.UsedByIndices;
218 void RegUseTracker::clear() {
225 /// Formula - This class holds information that describes a formula for
226 /// computing satisfying a use. It may include broken-out immediates and scaled
229 /// Global base address used for complex addressing.
232 /// Base offset for complex addressing.
235 /// Whether any complex addressing has a base register.
238 /// The scale of any complex addressing.
241 /// BaseRegs - The list of "base" registers for this use. When this is
242 /// non-empty. The canonical representation of a formula is
243 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
244 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
245 /// #1 enforces that the scaled register is always used when at least two
246 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
247 /// #2 enforces that 1 * reg is reg.
248 /// This invariant can be temporarly broken while building a formula.
249 /// However, every formula inserted into the LSRInstance must be in canonical
251 SmallVector<const SCEV *, 4> BaseRegs;
253 /// ScaledReg - The 'scaled' register for this use. This should be non-null
254 /// when Scale is not zero.
255 const SCEV *ScaledReg;
257 /// UnfoldedOffset - An additional constant offset which added near the
258 /// use. This requires a temporary register, but the offset itself can
259 /// live in an add immediate field rather than a register.
260 int64_t UnfoldedOffset;
263 : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
264 ScaledReg(nullptr), UnfoldedOffset(0) {}
266 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
268 bool isCanonical() const;
274 size_t getNumRegs() const;
275 Type *getType() const;
277 void DeleteBaseReg(const SCEV *&S);
279 bool referencesReg(const SCEV *S) const;
280 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
281 const RegUseTracker &RegUses) const;
283 void print(raw_ostream &OS) const;
289 /// DoInitialMatch - Recursion helper for InitialMatch.
290 static void DoInitialMatch(const SCEV *S, Loop *L,
291 SmallVectorImpl<const SCEV *> &Good,
292 SmallVectorImpl<const SCEV *> &Bad,
293 ScalarEvolution &SE) {
294 // Collect expressions which properly dominate the loop header.
295 if (SE.properlyDominates(S, L->getHeader())) {
300 // Look at add operands.
301 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
302 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
304 DoInitialMatch(*I, L, Good, Bad, SE);
308 // Look at addrec operands.
309 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
310 if (!AR->getStart()->isZero()) {
311 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
312 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
313 AR->getStepRecurrence(SE),
314 // FIXME: AR->getNoWrapFlags()
315 AR->getLoop(), SCEV::FlagAnyWrap),
320 // Handle a multiplication by -1 (negation) if it didn't fold.
321 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
322 if (Mul->getOperand(0)->isAllOnesValue()) {
323 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
324 const SCEV *NewMul = SE.getMulExpr(Ops);
326 SmallVector<const SCEV *, 4> MyGood;
327 SmallVector<const SCEV *, 4> MyBad;
328 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
329 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
330 SE.getEffectiveSCEVType(NewMul->getType())));
331 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
332 E = MyGood.end(); I != E; ++I)
333 Good.push_back(SE.getMulExpr(NegOne, *I));
334 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
335 E = MyBad.end(); I != E; ++I)
336 Bad.push_back(SE.getMulExpr(NegOne, *I));
340 // Ok, we can't do anything interesting. Just stuff the whole thing into a
341 // register and hope for the best.
345 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
346 /// attempting to keep all loop-invariant and loop-computable values in a
347 /// single base register.
348 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
349 SmallVector<const SCEV *, 4> Good;
350 SmallVector<const SCEV *, 4> Bad;
351 DoInitialMatch(S, L, Good, Bad, SE);
353 const SCEV *Sum = SE.getAddExpr(Good);
355 BaseRegs.push_back(Sum);
359 const SCEV *Sum = SE.getAddExpr(Bad);
361 BaseRegs.push_back(Sum);
367 /// \brief Check whether or not this formula statisfies the canonical
369 /// \see Formula::BaseRegs.
370 bool Formula::isCanonical() const {
372 return Scale != 1 || !BaseRegs.empty();
373 return BaseRegs.size() <= 1;
376 /// \brief Helper method to morph a formula into its canonical representation.
377 /// \see Formula::BaseRegs.
378 /// Every formula having more than one base register, must use the ScaledReg
379 /// field. Otherwise, we would have to do special cases everywhere in LSR
380 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
381 /// On the other hand, 1*reg should be canonicalized into reg.
382 void Formula::Canonicalize() {
385 // So far we did not need this case. This is easy to implement but it is
386 // useless to maintain dead code. Beside it could hurt compile time.
387 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
388 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
389 ScaledReg = BaseRegs.back();
392 size_t BaseRegsSize = BaseRegs.size();
394 // If ScaledReg is an invariant, try to find a variant expression.
395 while (Try < BaseRegsSize && !isa<SCEVAddRecExpr>(ScaledReg))
396 std::swap(ScaledReg, BaseRegs[Try++]);
399 /// \brief Get rid of the scale in the formula.
400 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
401 /// \return true if it was possible to get rid of the scale, false otherwise.
402 /// \note After this operation the formula may not be in the canonical form.
403 bool Formula::Unscale() {
407 BaseRegs.push_back(ScaledReg);
412 /// getNumRegs - Return the total number of register operands used by this
413 /// formula. This does not include register uses implied by non-constant
415 size_t Formula::getNumRegs() const {
416 return !!ScaledReg + BaseRegs.size();
419 /// getType - Return the type of this formula, if it has one, or null
420 /// otherwise. This type is meaningless except for the bit size.
421 Type *Formula::getType() const {
422 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
423 ScaledReg ? ScaledReg->getType() :
424 BaseGV ? BaseGV->getType() :
428 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
429 void Formula::DeleteBaseReg(const SCEV *&S) {
430 if (&S != &BaseRegs.back())
431 std::swap(S, BaseRegs.back());
435 /// referencesReg - Test if this formula references the given register.
436 bool Formula::referencesReg(const SCEV *S) const {
437 return S == ScaledReg ||
438 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
441 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
442 /// which are used by uses other than the use with the given index.
443 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
444 const RegUseTracker &RegUses) const {
446 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
448 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
449 E = BaseRegs.end(); I != E; ++I)
450 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
455 void Formula::print(raw_ostream &OS) const {
458 if (!First) OS << " + "; else First = false;
459 BaseGV->printAsOperand(OS, /*PrintType=*/false);
461 if (BaseOffset != 0) {
462 if (!First) OS << " + "; else First = false;
465 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
466 E = BaseRegs.end(); I != E; ++I) {
467 if (!First) OS << " + "; else First = false;
468 OS << "reg(" << **I << ')';
470 if (HasBaseReg && BaseRegs.empty()) {
471 if (!First) OS << " + "; else First = false;
472 OS << "**error: HasBaseReg**";
473 } else if (!HasBaseReg && !BaseRegs.empty()) {
474 if (!First) OS << " + "; else First = false;
475 OS << "**error: !HasBaseReg**";
478 if (!First) OS << " + "; else First = false;
479 OS << Scale << "*reg(";
486 if (UnfoldedOffset != 0) {
487 if (!First) OS << " + ";
488 OS << "imm(" << UnfoldedOffset << ')';
492 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
493 void Formula::dump() const {
494 print(errs()); errs() << '\n';
498 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
499 /// without changing its value.
500 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
502 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
503 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
506 /// isAddSExtable - Return true if the given add can be sign-extended
507 /// without changing its value.
508 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
510 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
511 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
514 /// isMulSExtable - Return true if the given mul can be sign-extended
515 /// without changing its value.
516 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
518 IntegerType::get(SE.getContext(),
519 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
520 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
523 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
524 /// and if the remainder is known to be zero, or null otherwise. If
525 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
526 /// to Y, ignoring that the multiplication may overflow, which is useful when
527 /// the result will be used in a context where the most significant bits are
529 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
531 bool IgnoreSignificantBits = false) {
532 // Handle the trivial case, which works for any SCEV type.
534 return SE.getConstant(LHS->getType(), 1);
536 // Handle a few RHS special cases.
537 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
539 const APInt &RA = RC->getValue()->getValue();
540 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
542 if (RA.isAllOnesValue())
543 return SE.getMulExpr(LHS, RC);
544 // Handle x /s 1 as x.
549 // Check for a division of a constant by a constant.
550 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
553 const APInt &LA = C->getValue()->getValue();
554 const APInt &RA = RC->getValue()->getValue();
555 if (LA.srem(RA) != 0)
557 return SE.getConstant(LA.sdiv(RA));
560 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
561 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
562 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
563 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
564 IgnoreSignificantBits);
565 if (!Step) return nullptr;
566 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
567 IgnoreSignificantBits);
568 if (!Start) return nullptr;
569 // FlagNW is independent of the start value, step direction, and is
570 // preserved with smaller magnitude steps.
571 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
572 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
577 // Distribute the sdiv over add operands, if the add doesn't overflow.
578 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
579 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
580 SmallVector<const SCEV *, 8> Ops;
581 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
583 const SCEV *Op = getExactSDiv(*I, RHS, SE,
584 IgnoreSignificantBits);
585 if (!Op) return nullptr;
588 return SE.getAddExpr(Ops);
593 // Check for a multiply operand that we can pull RHS out of.
594 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
595 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
596 SmallVector<const SCEV *, 4> Ops;
598 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
602 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
603 IgnoreSignificantBits)) {
609 return Found ? SE.getMulExpr(Ops) : nullptr;
614 // Otherwise we don't know.
618 /// ExtractImmediate - If S involves the addition of a constant integer value,
619 /// return that integer value, and mutate S to point to a new SCEV with that
621 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
622 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
623 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
624 S = SE.getConstant(C->getType(), 0);
625 return C->getValue()->getSExtValue();
627 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
628 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
629 int64_t Result = ExtractImmediate(NewOps.front(), SE);
631 S = SE.getAddExpr(NewOps);
633 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
634 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
635 int64_t Result = ExtractImmediate(NewOps.front(), SE);
637 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
638 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
645 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
646 /// return that symbol, and mutate S to point to a new SCEV with that
648 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
649 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
650 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
651 S = SE.getConstant(GV->getType(), 0);
654 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
655 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
656 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
658 S = SE.getAddExpr(NewOps);
660 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
661 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
662 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
664 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
665 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
672 /// isAddressUse - Returns true if the specified instruction is using the
673 /// specified value as an address.
674 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
675 bool isAddress = isa<LoadInst>(Inst);
676 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
677 if (SI->getOperand(1) == OperandVal)
679 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
680 // Addressing modes can also be folded into prefetches and a variety
682 switch (II->getIntrinsicID()) {
684 case Intrinsic::prefetch:
685 case Intrinsic::x86_sse_storeu_ps:
686 case Intrinsic::x86_sse2_storeu_pd:
687 case Intrinsic::x86_sse2_storeu_dq:
688 case Intrinsic::x86_sse2_storel_dq:
689 if (II->getArgOperand(0) == OperandVal)
697 /// getAccessType - Return the type of the memory being accessed.
698 static Type *getAccessType(const Instruction *Inst) {
699 Type *AccessTy = Inst->getType();
700 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
701 AccessTy = SI->getOperand(0)->getType();
702 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
703 // Addressing modes can also be folded into prefetches and a variety
705 switch (II->getIntrinsicID()) {
707 case Intrinsic::x86_sse_storeu_ps:
708 case Intrinsic::x86_sse2_storeu_pd:
709 case Intrinsic::x86_sse2_storeu_dq:
710 case Intrinsic::x86_sse2_storel_dq:
711 AccessTy = II->getArgOperand(0)->getType();
716 // All pointers have the same requirements, so canonicalize them to an
717 // arbitrary pointer type to minimize variation.
718 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
719 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
720 PTy->getAddressSpace());
725 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
726 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
727 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
728 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
729 if (SE.isSCEVable(PN->getType()) &&
730 (SE.getEffectiveSCEVType(PN->getType()) ==
731 SE.getEffectiveSCEVType(AR->getType())) &&
732 SE.getSCEV(PN) == AR)
738 /// Check if expanding this expression is likely to incur significant cost. This
739 /// is tricky because SCEV doesn't track which expressions are actually computed
740 /// by the current IR.
742 /// We currently allow expansion of IV increments that involve adds,
743 /// multiplication by constants, and AddRecs from existing phis.
745 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
746 /// obvious multiple of the UDivExpr.
747 static bool isHighCostExpansion(const SCEV *S,
748 SmallPtrSetImpl<const SCEV*> &Processed,
749 ScalarEvolution &SE) {
750 // Zero/One operand expressions
751 switch (S->getSCEVType()) {
756 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
759 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
762 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
766 if (!Processed.insert(S).second)
769 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
770 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
772 if (isHighCostExpansion(*I, Processed, SE))
778 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
779 if (Mul->getNumOperands() == 2) {
780 // Multiplication by a constant is ok
781 if (isa<SCEVConstant>(Mul->getOperand(0)))
782 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
784 // If we have the value of one operand, check if an existing
785 // multiplication already generates this expression.
786 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
787 Value *UVal = U->getValue();
788 for (User *UR : UVal->users()) {
789 // If U is a constant, it may be used by a ConstantExpr.
790 Instruction *UI = dyn_cast<Instruction>(UR);
791 if (UI && UI->getOpcode() == Instruction::Mul &&
792 SE.isSCEVable(UI->getType())) {
793 return SE.getSCEV(UI) == Mul;
800 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
801 if (isExistingPhi(AR, SE))
805 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
809 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
810 /// specified set are trivially dead, delete them and see if this makes any of
811 /// their operands subsequently dead.
813 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
814 bool Changed = false;
816 while (!DeadInsts.empty()) {
817 Value *V = DeadInsts.pop_back_val();
818 Instruction *I = dyn_cast_or_null<Instruction>(V);
820 if (!I || !isInstructionTriviallyDead(I))
823 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
824 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
827 DeadInsts.push_back(U);
830 I->eraseFromParent();
841 /// \brief Check if the addressing mode defined by \p F is completely
842 /// folded in \p LU at isel time.
843 /// This includes address-mode folding and special icmp tricks.
844 /// This function returns true if \p LU can accommodate what \p F
845 /// defines and up to 1 base + 1 scaled + offset.
846 /// In other words, if \p F has several base registers, this function may
847 /// still return true. Therefore, users still need to account for
848 /// additional base registers and/or unfolded offsets to derive an
849 /// accurate cost model.
850 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
851 const LSRUse &LU, const Formula &F);
852 // Get the cost of the scaling factor used in F for LU.
853 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
854 const LSRUse &LU, const Formula &F);
858 /// Cost - This class is used to measure and compare candidate formulae.
860 /// TODO: Some of these could be merged. Also, a lexical ordering
861 /// isn't always optimal.
865 unsigned NumBaseAdds;
872 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
873 SetupCost(0), ScaleCost(0) {}
875 bool operator<(const Cost &Other) const;
880 // Once any of the metrics loses, they must all remain losers.
882 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
883 | ImmCost | SetupCost | ScaleCost) != ~0u)
884 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
885 & ImmCost & SetupCost & ScaleCost) == ~0u);
890 assert(isValid() && "invalid cost");
891 return NumRegs == ~0u;
894 void RateFormula(const TargetTransformInfo &TTI,
896 SmallPtrSetImpl<const SCEV *> &Regs,
897 const DenseSet<const SCEV *> &VisitedRegs,
899 const SmallVectorImpl<int64_t> &Offsets,
900 ScalarEvolution &SE, DominatorTree &DT,
902 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
904 void print(raw_ostream &OS) const;
908 void RateRegister(const SCEV *Reg,
909 SmallPtrSetImpl<const SCEV *> &Regs,
911 ScalarEvolution &SE, DominatorTree &DT);
912 void RatePrimaryRegister(const SCEV *Reg,
913 SmallPtrSetImpl<const SCEV *> &Regs,
915 ScalarEvolution &SE, DominatorTree &DT,
916 SmallPtrSetImpl<const SCEV *> *LoserRegs);
921 /// RateRegister - Tally up interesting quantities from the given register.
922 void Cost::RateRegister(const SCEV *Reg,
923 SmallPtrSetImpl<const SCEV *> &Regs,
925 ScalarEvolution &SE, DominatorTree &DT) {
926 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
927 // If this is an addrec for another loop, don't second-guess its addrec phi
928 // nodes. LSR isn't currently smart enough to reason about more than one
929 // loop at a time. LSR has already run on inner loops, will not run on outer
930 // loops, and cannot be expected to change sibling loops.
931 if (AR->getLoop() != L) {
932 // If the AddRec exists, consider it's register free and leave it alone.
933 if (isExistingPhi(AR, SE))
936 // Otherwise, do not consider this formula at all.
940 AddRecCost += 1; /// TODO: This should be a function of the stride.
942 // Add the step value register, if it needs one.
943 // TODO: The non-affine case isn't precisely modeled here.
944 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
945 if (!Regs.count(AR->getOperand(1))) {
946 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
954 // Rough heuristic; favor registers which don't require extra setup
955 // instructions in the preheader.
956 if (!isa<SCEVUnknown>(Reg) &&
957 !isa<SCEVConstant>(Reg) &&
958 !(isa<SCEVAddRecExpr>(Reg) &&
959 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
960 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
963 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
964 SE.hasComputableLoopEvolution(Reg, L);
967 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
968 /// before, rate it. Optional LoserRegs provides a way to declare any formula
969 /// that refers to one of those regs an instant loser.
970 void Cost::RatePrimaryRegister(const SCEV *Reg,
971 SmallPtrSetImpl<const SCEV *> &Regs,
973 ScalarEvolution &SE, DominatorTree &DT,
974 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
975 if (LoserRegs && LoserRegs->count(Reg)) {
979 if (Regs.insert(Reg).second) {
980 RateRegister(Reg, Regs, L, SE, DT);
981 if (LoserRegs && isLoser())
982 LoserRegs->insert(Reg);
986 void Cost::RateFormula(const TargetTransformInfo &TTI,
988 SmallPtrSetImpl<const SCEV *> &Regs,
989 const DenseSet<const SCEV *> &VisitedRegs,
991 const SmallVectorImpl<int64_t> &Offsets,
992 ScalarEvolution &SE, DominatorTree &DT,
994 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
995 assert(F.isCanonical() && "Cost is accurate only for canonical formula");
996 // Tally up the registers.
997 if (const SCEV *ScaledReg = F.ScaledReg) {
998 if (VisitedRegs.count(ScaledReg)) {
1002 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1006 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
1007 E = F.BaseRegs.end(); I != E; ++I) {
1008 const SCEV *BaseReg = *I;
1009 if (VisitedRegs.count(BaseReg)) {
1013 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1018 // Determine how many (unfolded) adds we'll need inside the loop.
1019 size_t NumBaseParts = F.getNumRegs();
1020 if (NumBaseParts > 1)
1021 // Do not count the base and a possible second register if the target
1022 // allows to fold 2 registers.
1024 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1025 NumBaseAdds += (F.UnfoldedOffset != 0);
1027 // Accumulate non-free scaling amounts.
1028 ScaleCost += getScalingFactorCost(TTI, LU, F);
1030 // Tally up the non-zero immediates.
1031 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1032 E = Offsets.end(); I != E; ++I) {
1033 int64_t Offset = (uint64_t)*I + F.BaseOffset;
1035 ImmCost += 64; // Handle symbolic values conservatively.
1036 // TODO: This should probably be the pointer size.
1037 else if (Offset != 0)
1038 ImmCost += APInt(64, Offset, true).getMinSignedBits();
1040 assert(isValid() && "invalid cost");
1043 /// Lose - Set this cost to a losing value.
1054 /// operator< - Choose the lower cost.
1055 bool Cost::operator<(const Cost &Other) const {
1056 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
1057 ImmCost, SetupCost) <
1058 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
1059 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
1063 void Cost::print(raw_ostream &OS) const {
1064 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
1065 if (AddRecCost != 0)
1066 OS << ", with addrec cost " << AddRecCost;
1068 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
1069 if (NumBaseAdds != 0)
1070 OS << ", plus " << NumBaseAdds << " base add"
1071 << (NumBaseAdds == 1 ? "" : "s");
1073 OS << ", plus " << ScaleCost << " scale cost";
1075 OS << ", plus " << ImmCost << " imm cost";
1077 OS << ", plus " << SetupCost << " setup cost";
1080 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1081 void Cost::dump() const {
1082 print(errs()); errs() << '\n';
1088 /// LSRFixup - An operand value in an instruction which is to be replaced
1089 /// with some equivalent, possibly strength-reduced, replacement.
1091 /// UserInst - The instruction which will be updated.
1092 Instruction *UserInst;
1094 /// OperandValToReplace - The operand of the instruction which will
1095 /// be replaced. The operand may be used more than once; every instance
1096 /// will be replaced.
1097 Value *OperandValToReplace;
1099 /// PostIncLoops - If this user is to use the post-incremented value of an
1100 /// induction variable, this variable is non-null and holds the loop
1101 /// associated with the induction variable.
1102 PostIncLoopSet PostIncLoops;
1104 /// LUIdx - The index of the LSRUse describing the expression which
1105 /// this fixup needs, minus an offset (below).
1108 /// Offset - A constant offset to be added to the LSRUse expression.
1109 /// This allows multiple fixups to share the same LSRUse with different
1110 /// offsets, for example in an unrolled loop.
1113 bool isUseFullyOutsideLoop(const Loop *L) const;
1117 void print(raw_ostream &OS) const;
1123 LSRFixup::LSRFixup()
1124 : UserInst(nullptr), OperandValToReplace(nullptr), LUIdx(~size_t(0)),
1127 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1128 /// value outside of the given loop.
1129 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1130 // PHI nodes use their value in their incoming blocks.
1131 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1132 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1133 if (PN->getIncomingValue(i) == OperandValToReplace &&
1134 L->contains(PN->getIncomingBlock(i)))
1139 return !L->contains(UserInst);
1142 void LSRFixup::print(raw_ostream &OS) const {
1144 // Store is common and interesting enough to be worth special-casing.
1145 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1147 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1148 } else if (UserInst->getType()->isVoidTy())
1149 OS << UserInst->getOpcodeName();
1151 UserInst->printAsOperand(OS, /*PrintType=*/false);
1153 OS << ", OperandValToReplace=";
1154 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1156 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1157 E = PostIncLoops.end(); I != E; ++I) {
1158 OS << ", PostIncLoop=";
1159 (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1162 if (LUIdx != ~size_t(0))
1163 OS << ", LUIdx=" << LUIdx;
1166 OS << ", Offset=" << Offset;
1169 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1170 void LSRFixup::dump() const {
1171 print(errs()); errs() << '\n';
1177 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1178 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1179 struct UniquifierDenseMapInfo {
1180 static SmallVector<const SCEV *, 4> getEmptyKey() {
1181 SmallVector<const SCEV *, 4> V;
1182 V.push_back(reinterpret_cast<const SCEV *>(-1));
1186 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1187 SmallVector<const SCEV *, 4> V;
1188 V.push_back(reinterpret_cast<const SCEV *>(-2));
1192 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1193 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1196 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1197 const SmallVector<const SCEV *, 4> &RHS) {
1202 /// LSRUse - This class holds the state that LSR keeps for each use in
1203 /// IVUsers, as well as uses invented by LSR itself. It includes information
1204 /// about what kinds of things can be folded into the user, information about
1205 /// the user itself, and information about how the use may be satisfied.
1206 /// TODO: Represent multiple users of the same expression in common?
1208 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1211 /// KindType - An enum for a kind of use, indicating what types of
1212 /// scaled and immediate operands it might support.
1214 Basic, ///< A normal use, with no folding.
1215 Special, ///< A special case of basic, allowing -1 scales.
1216 Address, ///< An address use; folding according to TargetLowering
1217 ICmpZero ///< An equality icmp with both operands folded into one.
1218 // TODO: Add a generic icmp too?
1221 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1226 SmallVector<int64_t, 8> Offsets;
1230 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1231 /// LSRUse are outside of the loop, in which case some special-case heuristics
1233 bool AllFixupsOutsideLoop;
1235 /// RigidFormula is set to true to guarantee that this use will be associated
1236 /// with a single formula--the one that initially matched. Some SCEV
1237 /// expressions cannot be expanded. This allows LSR to consider the registers
1238 /// used by those expressions without the need to expand them later after
1239 /// changing the formula.
1242 /// WidestFixupType - This records the widest use type for any fixup using
1243 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1244 /// max fixup widths to be equivalent, because the narrower one may be relying
1245 /// on the implicit truncation to truncate away bogus bits.
1246 Type *WidestFixupType;
1248 /// Formulae - A list of ways to build a value that can satisfy this user.
1249 /// After the list is populated, one of these is selected heuristically and
1250 /// used to formulate a replacement for OperandValToReplace in UserInst.
1251 SmallVector<Formula, 12> Formulae;
1253 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1254 SmallPtrSet<const SCEV *, 4> Regs;
1256 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1257 MinOffset(INT64_MAX),
1258 MaxOffset(INT64_MIN),
1259 AllFixupsOutsideLoop(true),
1260 RigidFormula(false),
1261 WidestFixupType(nullptr) {}
1263 bool HasFormulaWithSameRegs(const Formula &F) const;
1264 bool InsertFormula(const Formula &F);
1265 void DeleteFormula(Formula &F);
1266 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1268 void print(raw_ostream &OS) const;
1274 /// HasFormula - Test whether this use as a formula which has the same
1275 /// registers as the given formula.
1276 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1277 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1278 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1279 // Unstable sort by host order ok, because this is only used for uniquifying.
1280 std::sort(Key.begin(), Key.end());
1281 return Uniquifier.count(Key);
1284 /// InsertFormula - If the given formula has not yet been inserted, add it to
1285 /// the list, and return true. Return false otherwise.
1286 /// The formula must be in canonical form.
1287 bool LSRUse::InsertFormula(const Formula &F) {
1288 assert(F.isCanonical() && "Invalid canonical representation");
1290 if (!Formulae.empty() && RigidFormula)
1293 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1294 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1295 // Unstable sort by host order ok, because this is only used for uniquifying.
1296 std::sort(Key.begin(), Key.end());
1298 if (!Uniquifier.insert(Key).second)
1301 // Using a register to hold the value of 0 is not profitable.
1302 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1303 "Zero allocated in a scaled register!");
1305 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1306 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1307 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1310 // Add the formula to the list.
1311 Formulae.push_back(F);
1313 // Record registers now being used by this use.
1314 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1316 Regs.insert(F.ScaledReg);
1321 /// DeleteFormula - Remove the given formula from this use's list.
1322 void LSRUse::DeleteFormula(Formula &F) {
1323 if (&F != &Formulae.back())
1324 std::swap(F, Formulae.back());
1325 Formulae.pop_back();
1328 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1329 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1330 // Now that we've filtered out some formulae, recompute the Regs set.
1331 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1333 for (const Formula &F : Formulae) {
1334 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1335 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1338 // Update the RegTracker.
1339 for (const SCEV *S : OldRegs)
1341 RegUses.DropRegister(S, LUIdx);
1344 void LSRUse::print(raw_ostream &OS) const {
1345 OS << "LSR Use: Kind=";
1347 case Basic: OS << "Basic"; break;
1348 case Special: OS << "Special"; break;
1349 case ICmpZero: OS << "ICmpZero"; break;
1351 OS << "Address of ";
1352 if (AccessTy->isPointerTy())
1353 OS << "pointer"; // the full pointer type could be really verbose
1358 OS << ", Offsets={";
1359 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1360 E = Offsets.end(); I != E; ++I) {
1362 if (std::next(I) != E)
1367 if (AllFixupsOutsideLoop)
1368 OS << ", all-fixups-outside-loop";
1370 if (WidestFixupType)
1371 OS << ", widest fixup type: " << *WidestFixupType;
1374 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1375 void LSRUse::dump() const {
1376 print(errs()); errs() << '\n';
1380 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1381 LSRUse::KindType Kind, Type *AccessTy,
1382 GlobalValue *BaseGV, int64_t BaseOffset,
1383 bool HasBaseReg, int64_t Scale) {
1385 case LSRUse::Address:
1386 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1388 // Otherwise, just guess that reg+reg addressing is legal.
1391 case LSRUse::ICmpZero:
1392 // There's not even a target hook for querying whether it would be legal to
1393 // fold a GV into an ICmp.
1397 // ICmp only has two operands; don't allow more than two non-trivial parts.
1398 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1401 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1402 // putting the scaled register in the other operand of the icmp.
1403 if (Scale != 0 && Scale != -1)
1406 // If we have low-level target information, ask the target if it can fold an
1407 // integer immediate on an icmp.
1408 if (BaseOffset != 0) {
1410 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1411 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1412 // Offs is the ICmp immediate.
1414 // The cast does the right thing with INT64_MIN.
1415 BaseOffset = -(uint64_t)BaseOffset;
1416 return TTI.isLegalICmpImmediate(BaseOffset);
1419 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1423 // Only handle single-register values.
1424 return !BaseGV && Scale == 0 && BaseOffset == 0;
1426 case LSRUse::Special:
1427 // Special case Basic to handle -1 scales.
1428 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1431 llvm_unreachable("Invalid LSRUse Kind!");
1434 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1435 int64_t MinOffset, int64_t MaxOffset,
1436 LSRUse::KindType Kind, Type *AccessTy,
1437 GlobalValue *BaseGV, int64_t BaseOffset,
1438 bool HasBaseReg, int64_t Scale) {
1439 // Check for overflow.
1440 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1443 MinOffset = (uint64_t)BaseOffset + MinOffset;
1444 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1447 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1449 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1450 HasBaseReg, Scale) &&
1451 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1455 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1456 int64_t MinOffset, int64_t MaxOffset,
1457 LSRUse::KindType Kind, Type *AccessTy,
1459 // For the purpose of isAMCompletelyFolded either having a canonical formula
1460 // or a scale not equal to zero is correct.
1461 // Problems may arise from non canonical formulae having a scale == 0.
1462 // Strictly speaking it would best to just rely on canonical formulae.
1463 // However, when we generate the scaled formulae, we first check that the
1464 // scaling factor is profitable before computing the actual ScaledReg for
1465 // compile time sake.
1466 assert((F.isCanonical() || F.Scale != 0));
1467 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1468 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1471 /// isLegalUse - Test whether we know how to expand the current formula.
1472 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1473 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1474 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1476 // We know how to expand completely foldable formulae.
1477 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1478 BaseOffset, HasBaseReg, Scale) ||
1479 // Or formulae that use a base register produced by a sum of base
1482 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1483 BaseGV, BaseOffset, true, 0));
1486 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1487 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1489 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1490 F.BaseOffset, F.HasBaseReg, F.Scale);
1493 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1494 const LSRUse &LU, const Formula &F) {
1495 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1496 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1500 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1501 const LSRUse &LU, const Formula &F) {
1505 // If the use is not completely folded in that instruction, we will have to
1506 // pay an extra cost only for scale != 1.
1507 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1509 return F.Scale != 1;
1512 case LSRUse::Address: {
1513 // Check the scaling factor cost with both the min and max offsets.
1514 int ScaleCostMinOffset =
1515 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1516 F.BaseOffset + LU.MinOffset,
1517 F.HasBaseReg, F.Scale);
1518 int ScaleCostMaxOffset =
1519 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1520 F.BaseOffset + LU.MaxOffset,
1521 F.HasBaseReg, F.Scale);
1523 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1524 "Legal addressing mode has an illegal cost!");
1525 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1527 case LSRUse::ICmpZero:
1529 case LSRUse::Special:
1530 // The use is completely folded, i.e., everything is folded into the
1535 llvm_unreachable("Invalid LSRUse Kind!");
1538 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1539 LSRUse::KindType Kind, Type *AccessTy,
1540 GlobalValue *BaseGV, int64_t BaseOffset,
1542 // Fast-path: zero is always foldable.
1543 if (BaseOffset == 0 && !BaseGV) return true;
1545 // Conservatively, create an address with an immediate and a
1546 // base and a scale.
1547 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1549 // Canonicalize a scale of 1 to a base register if the formula doesn't
1550 // already have a base register.
1551 if (!HasBaseReg && Scale == 1) {
1556 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1560 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1561 ScalarEvolution &SE, int64_t MinOffset,
1562 int64_t MaxOffset, LSRUse::KindType Kind,
1563 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1564 // Fast-path: zero is always foldable.
1565 if (S->isZero()) return true;
1567 // Conservatively, create an address with an immediate and a
1568 // base and a scale.
1569 int64_t BaseOffset = ExtractImmediate(S, SE);
1570 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1572 // If there's anything else involved, it's not foldable.
1573 if (!S->isZero()) return false;
1575 // Fast-path: zero is always foldable.
1576 if (BaseOffset == 0 && !BaseGV) return true;
1578 // Conservatively, create an address with an immediate and a
1579 // base and a scale.
1580 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1582 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1583 BaseOffset, HasBaseReg, Scale);
1588 /// IVInc - An individual increment in a Chain of IV increments.
1589 /// Relate an IV user to an expression that computes the IV it uses from the IV
1590 /// used by the previous link in the Chain.
1592 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1593 /// original IVOperand. The head of the chain's IVOperand is only valid during
1594 /// chain collection, before LSR replaces IV users. During chain generation,
1595 /// IncExpr can be used to find the new IVOperand that computes the same
1598 Instruction *UserInst;
1600 const SCEV *IncExpr;
1602 IVInc(Instruction *U, Value *O, const SCEV *E):
1603 UserInst(U), IVOperand(O), IncExpr(E) {}
1606 // IVChain - The list of IV increments in program order.
1607 // We typically add the head of a chain without finding subsequent links.
1609 SmallVector<IVInc,1> Incs;
1610 const SCEV *ExprBase;
1612 IVChain() : ExprBase(nullptr) {}
1614 IVChain(const IVInc &Head, const SCEV *Base)
1615 : Incs(1, Head), ExprBase(Base) {}
1617 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1619 // begin - return the first increment in the chain.
1620 const_iterator begin() const {
1621 assert(!Incs.empty());
1622 return std::next(Incs.begin());
1624 const_iterator end() const {
1628 // hasIncs - Returns true if this chain contains any increments.
1629 bool hasIncs() const { return Incs.size() >= 2; }
1631 // add - Add an IVInc to the end of this chain.
1632 void add(const IVInc &X) { Incs.push_back(X); }
1634 // tailUserInst - Returns the last UserInst in the chain.
1635 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1637 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1639 bool isProfitableIncrement(const SCEV *OperExpr,
1640 const SCEV *IncExpr,
1644 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1645 /// Distinguish between FarUsers that definitely cross IV increments and
1646 /// NearUsers that may be used between IV increments.
1648 SmallPtrSet<Instruction*, 4> FarUsers;
1649 SmallPtrSet<Instruction*, 4> NearUsers;
1652 /// LSRInstance - This class holds state for the main loop strength reduction
1656 ScalarEvolution &SE;
1659 const TargetTransformInfo &TTI;
1663 /// IVIncInsertPos - This is the insert position that the current loop's
1664 /// induction variable increment should be placed. In simple loops, this is
1665 /// the latch block's terminator. But in more complicated cases, this is a
1666 /// position which will dominate all the in-loop post-increment users.
1667 Instruction *IVIncInsertPos;
1669 /// Factors - Interesting factors between use strides.
1670 SmallSetVector<int64_t, 8> Factors;
1672 /// Types - Interesting use types, to facilitate truncation reuse.
1673 SmallSetVector<Type *, 4> Types;
1675 /// Fixups - The list of operands which are to be replaced.
1676 SmallVector<LSRFixup, 16> Fixups;
1678 /// Uses - The list of interesting uses.
1679 SmallVector<LSRUse, 16> Uses;
1681 /// RegUses - Track which uses use which register candidates.
1682 RegUseTracker RegUses;
1684 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1685 // have more than a few IV increment chains in a loop. Missing a Chain falls
1686 // back to normal LSR behavior for those uses.
1687 static const unsigned MaxChains = 8;
1689 /// IVChainVec - IV users can form a chain of IV increments.
1690 SmallVector<IVChain, MaxChains> IVChainVec;
1692 /// IVIncSet - IV users that belong to profitable IVChains.
1693 SmallPtrSet<Use*, MaxChains> IVIncSet;
1695 void OptimizeShadowIV();
1696 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1697 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1698 void OptimizeLoopTermCond();
1700 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1701 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1702 void FinalizeChain(IVChain &Chain);
1703 void CollectChains();
1704 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1705 SmallVectorImpl<WeakVH> &DeadInsts);
1707 void CollectInterestingTypesAndFactors();
1708 void CollectFixupsAndInitialFormulae();
1710 LSRFixup &getNewFixup() {
1711 Fixups.push_back(LSRFixup());
1712 return Fixups.back();
1715 // Support for sharing of LSRUses between LSRFixups.
1716 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1719 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1720 LSRUse::KindType Kind, Type *AccessTy);
1722 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1723 LSRUse::KindType Kind,
1726 void DeleteUse(LSRUse &LU, size_t LUIdx);
1728 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1730 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1731 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1732 void CountRegisters(const Formula &F, size_t LUIdx);
1733 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1735 void CollectLoopInvariantFixupsAndFormulae();
1737 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1738 unsigned Depth = 0);
1740 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1741 const Formula &Base, unsigned Depth,
1742 size_t Idx, bool IsScaledReg = false);
1743 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1744 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1745 const Formula &Base, size_t Idx,
1746 bool IsScaledReg = false);
1747 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1748 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1749 const Formula &Base,
1750 const SmallVectorImpl<int64_t> &Worklist,
1751 size_t Idx, bool IsScaledReg = false);
1752 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1753 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1754 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1755 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1756 void GenerateCrossUseConstantOffsets();
1757 void GenerateAllReuseFormulae();
1759 void FilterOutUndesirableDedicatedRegisters();
1761 size_t EstimateSearchSpaceComplexity() const;
1762 void NarrowSearchSpaceByDetectingSupersets();
1763 void NarrowSearchSpaceByCollapsingUnrolledCode();
1764 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1765 void NarrowSearchSpaceByPickingWinnerRegs();
1766 void NarrowSearchSpaceUsingHeuristics();
1768 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1770 SmallVectorImpl<const Formula *> &Workspace,
1771 const Cost &CurCost,
1772 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1773 DenseSet<const SCEV *> &VisitedRegs) const;
1774 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1776 BasicBlock::iterator
1777 HoistInsertPosition(BasicBlock::iterator IP,
1778 const SmallVectorImpl<Instruction *> &Inputs) const;
1779 BasicBlock::iterator
1780 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1783 SCEVExpander &Rewriter) const;
1785 Value *Expand(const LSRFixup &LF,
1787 BasicBlock::iterator IP,
1788 SCEVExpander &Rewriter,
1789 SmallVectorImpl<WeakVH> &DeadInsts) const;
1790 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1792 SCEVExpander &Rewriter,
1793 SmallVectorImpl<WeakVH> &DeadInsts,
1795 void Rewrite(const LSRFixup &LF,
1797 SCEVExpander &Rewriter,
1798 SmallVectorImpl<WeakVH> &DeadInsts,
1800 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1804 LSRInstance(Loop *L, Pass *P);
1806 bool getChanged() const { return Changed; }
1808 void print_factors_and_types(raw_ostream &OS) const;
1809 void print_fixups(raw_ostream &OS) const;
1810 void print_uses(raw_ostream &OS) const;
1811 void print(raw_ostream &OS) const;
1817 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1818 /// inside the loop then try to eliminate the cast operation.
1819 void LSRInstance::OptimizeShadowIV() {
1820 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1821 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1824 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1825 UI != E; /* empty */) {
1826 IVUsers::const_iterator CandidateUI = UI;
1828 Instruction *ShadowUse = CandidateUI->getUser();
1829 Type *DestTy = nullptr;
1830 bool IsSigned = false;
1832 /* If shadow use is a int->float cast then insert a second IV
1833 to eliminate this cast.
1835 for (unsigned i = 0; i < n; ++i)
1841 for (unsigned i = 0; i < n; ++i, ++d)
1844 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1846 DestTy = UCast->getDestTy();
1848 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1850 DestTy = SCast->getDestTy();
1852 if (!DestTy) continue;
1854 // If target does not support DestTy natively then do not apply
1855 // this transformation.
1856 if (!TTI.isTypeLegal(DestTy)) continue;
1858 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1860 if (PH->getNumIncomingValues() != 2) continue;
1862 Type *SrcTy = PH->getType();
1863 int Mantissa = DestTy->getFPMantissaWidth();
1864 if (Mantissa == -1) continue;
1865 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1868 unsigned Entry, Latch;
1869 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1877 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1878 if (!Init) continue;
1879 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1880 (double)Init->getSExtValue() :
1881 (double)Init->getZExtValue());
1883 BinaryOperator *Incr =
1884 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1885 if (!Incr) continue;
1886 if (Incr->getOpcode() != Instruction::Add
1887 && Incr->getOpcode() != Instruction::Sub)
1890 /* Initialize new IV, double d = 0.0 in above example. */
1891 ConstantInt *C = nullptr;
1892 if (Incr->getOperand(0) == PH)
1893 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1894 else if (Incr->getOperand(1) == PH)
1895 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1901 // Ignore negative constants, as the code below doesn't handle them
1902 // correctly. TODO: Remove this restriction.
1903 if (!C->getValue().isStrictlyPositive()) continue;
1905 /* Add new PHINode. */
1906 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1908 /* create new increment. '++d' in above example. */
1909 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1910 BinaryOperator *NewIncr =
1911 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1912 Instruction::FAdd : Instruction::FSub,
1913 NewPH, CFP, "IV.S.next.", Incr);
1915 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1916 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1918 /* Remove cast operation */
1919 ShadowUse->replaceAllUsesWith(NewPH);
1920 ShadowUse->eraseFromParent();
1926 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1927 /// set the IV user and stride information and return true, otherwise return
1929 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1930 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1931 if (UI->getUser() == Cond) {
1932 // NOTE: we could handle setcc instructions with multiple uses here, but
1933 // InstCombine does it as well for simple uses, it's not clear that it
1934 // occurs enough in real life to handle.
1941 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1942 /// a max computation.
1944 /// This is a narrow solution to a specific, but acute, problem. For loops
1950 /// } while (++i < n);
1952 /// the trip count isn't just 'n', because 'n' might not be positive. And
1953 /// unfortunately this can come up even for loops where the user didn't use
1954 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1955 /// will commonly be lowered like this:
1961 /// } while (++i < n);
1964 /// and then it's possible for subsequent optimization to obscure the if
1965 /// test in such a way that indvars can't find it.
1967 /// When indvars can't find the if test in loops like this, it creates a
1968 /// max expression, which allows it to give the loop a canonical
1969 /// induction variable:
1972 /// max = n < 1 ? 1 : n;
1975 /// } while (++i != max);
1977 /// Canonical induction variables are necessary because the loop passes
1978 /// are designed around them. The most obvious example of this is the
1979 /// LoopInfo analysis, which doesn't remember trip count values. It
1980 /// expects to be able to rediscover the trip count each time it is
1981 /// needed, and it does this using a simple analysis that only succeeds if
1982 /// the loop has a canonical induction variable.
1984 /// However, when it comes time to generate code, the maximum operation
1985 /// can be quite costly, especially if it's inside of an outer loop.
1987 /// This function solves this problem by detecting this type of loop and
1988 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1989 /// the instructions for the maximum computation.
1991 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1992 // Check that the loop matches the pattern we're looking for.
1993 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1994 Cond->getPredicate() != CmpInst::ICMP_NE)
1997 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1998 if (!Sel || !Sel->hasOneUse()) return Cond;
2000 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2001 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2003 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2005 // Add one to the backedge-taken count to get the trip count.
2006 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2007 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2009 // Check for a max calculation that matches the pattern. There's no check
2010 // for ICMP_ULE here because the comparison would be with zero, which
2011 // isn't interesting.
2012 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2013 const SCEVNAryExpr *Max = nullptr;
2014 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2015 Pred = ICmpInst::ICMP_SLE;
2017 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2018 Pred = ICmpInst::ICMP_SLT;
2020 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2021 Pred = ICmpInst::ICMP_ULT;
2028 // To handle a max with more than two operands, this optimization would
2029 // require additional checking and setup.
2030 if (Max->getNumOperands() != 2)
2033 const SCEV *MaxLHS = Max->getOperand(0);
2034 const SCEV *MaxRHS = Max->getOperand(1);
2036 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2037 // for a comparison with 1. For <= and >=, a comparison with zero.
2039 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2042 // Check the relevant induction variable for conformance to
2044 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2045 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2046 if (!AR || !AR->isAffine() ||
2047 AR->getStart() != One ||
2048 AR->getStepRecurrence(SE) != One)
2051 assert(AR->getLoop() == L &&
2052 "Loop condition operand is an addrec in a different loop!");
2054 // Check the right operand of the select, and remember it, as it will
2055 // be used in the new comparison instruction.
2056 Value *NewRHS = nullptr;
2057 if (ICmpInst::isTrueWhenEqual(Pred)) {
2058 // Look for n+1, and grab n.
2059 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2060 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2061 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2062 NewRHS = BO->getOperand(0);
2063 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2064 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2065 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2066 NewRHS = BO->getOperand(0);
2069 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2070 NewRHS = Sel->getOperand(1);
2071 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2072 NewRHS = Sel->getOperand(2);
2073 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2074 NewRHS = SU->getValue();
2076 // Max doesn't match expected pattern.
2079 // Determine the new comparison opcode. It may be signed or unsigned,
2080 // and the original comparison may be either equality or inequality.
2081 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2082 Pred = CmpInst::getInversePredicate(Pred);
2084 // Ok, everything looks ok to change the condition into an SLT or SGE and
2085 // delete the max calculation.
2087 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2089 // Delete the max calculation instructions.
2090 Cond->replaceAllUsesWith(NewCond);
2091 CondUse->setUser(NewCond);
2092 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2093 Cond->eraseFromParent();
2094 Sel->eraseFromParent();
2095 if (Cmp->use_empty())
2096 Cmp->eraseFromParent();
2100 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2101 /// postinc iv when possible.
2103 LSRInstance::OptimizeLoopTermCond() {
2104 SmallPtrSet<Instruction *, 4> PostIncs;
2106 BasicBlock *LatchBlock = L->getLoopLatch();
2107 SmallVector<BasicBlock*, 8> ExitingBlocks;
2108 L->getExitingBlocks(ExitingBlocks);
2110 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2111 BasicBlock *ExitingBlock = ExitingBlocks[i];
2113 // Get the terminating condition for the loop if possible. If we
2114 // can, we want to change it to use a post-incremented version of its
2115 // induction variable, to allow coalescing the live ranges for the IV into
2116 // one register value.
2118 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2121 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2122 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2125 // Search IVUsesByStride to find Cond's IVUse if there is one.
2126 IVStrideUse *CondUse = nullptr;
2127 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2128 if (!FindIVUserForCond(Cond, CondUse))
2131 // If the trip count is computed in terms of a max (due to ScalarEvolution
2132 // being unable to find a sufficient guard, for example), change the loop
2133 // comparison to use SLT or ULT instead of NE.
2134 // One consequence of doing this now is that it disrupts the count-down
2135 // optimization. That's not always a bad thing though, because in such
2136 // cases it may still be worthwhile to avoid a max.
2137 Cond = OptimizeMax(Cond, CondUse);
2139 // If this exiting block dominates the latch block, it may also use
2140 // the post-inc value if it won't be shared with other uses.
2141 // Check for dominance.
2142 if (!DT.dominates(ExitingBlock, LatchBlock))
2145 // Conservatively avoid trying to use the post-inc value in non-latch
2146 // exits if there may be pre-inc users in intervening blocks.
2147 if (LatchBlock != ExitingBlock)
2148 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2149 // Test if the use is reachable from the exiting block. This dominator
2150 // query is a conservative approximation of reachability.
2151 if (&*UI != CondUse &&
2152 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2153 // Conservatively assume there may be reuse if the quotient of their
2154 // strides could be a legal scale.
2155 const SCEV *A = IU.getStride(*CondUse, L);
2156 const SCEV *B = IU.getStride(*UI, L);
2157 if (!A || !B) continue;
2158 if (SE.getTypeSizeInBits(A->getType()) !=
2159 SE.getTypeSizeInBits(B->getType())) {
2160 if (SE.getTypeSizeInBits(A->getType()) >
2161 SE.getTypeSizeInBits(B->getType()))
2162 B = SE.getSignExtendExpr(B, A->getType());
2164 A = SE.getSignExtendExpr(A, B->getType());
2166 if (const SCEVConstant *D =
2167 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2168 const ConstantInt *C = D->getValue();
2169 // Stride of one or negative one can have reuse with non-addresses.
2170 if (C->isOne() || C->isAllOnesValue())
2171 goto decline_post_inc;
2172 // Avoid weird situations.
2173 if (C->getValue().getMinSignedBits() >= 64 ||
2174 C->getValue().isMinSignedValue())
2175 goto decline_post_inc;
2176 // Check for possible scaled-address reuse.
2177 Type *AccessTy = getAccessType(UI->getUser());
2178 int64_t Scale = C->getSExtValue();
2179 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ nullptr,
2181 /*HasBaseReg=*/ false, Scale))
2182 goto decline_post_inc;
2184 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ nullptr,
2186 /*HasBaseReg=*/ false, Scale))
2187 goto decline_post_inc;
2191 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2194 // It's possible for the setcc instruction to be anywhere in the loop, and
2195 // possible for it to have multiple users. If it is not immediately before
2196 // the exiting block branch, move it.
2197 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2198 if (Cond->hasOneUse()) {
2199 Cond->moveBefore(TermBr);
2201 // Clone the terminating condition and insert into the loopend.
2202 ICmpInst *OldCond = Cond;
2203 Cond = cast<ICmpInst>(Cond->clone());
2204 Cond->setName(L->getHeader()->getName() + ".termcond");
2205 ExitingBlock->getInstList().insert(TermBr, Cond);
2207 // Clone the IVUse, as the old use still exists!
2208 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2209 TermBr->replaceUsesOfWith(OldCond, Cond);
2213 // If we get to here, we know that we can transform the setcc instruction to
2214 // use the post-incremented version of the IV, allowing us to coalesce the
2215 // live ranges for the IV correctly.
2216 CondUse->transformToPostInc(L);
2219 PostIncs.insert(Cond);
2223 // Determine an insertion point for the loop induction variable increment. It
2224 // must dominate all the post-inc comparisons we just set up, and it must
2225 // dominate the loop latch edge.
2226 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2227 for (Instruction *Inst : PostIncs) {
2229 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2231 if (BB == Inst->getParent())
2232 IVIncInsertPos = Inst;
2233 else if (BB != IVIncInsertPos->getParent())
2234 IVIncInsertPos = BB->getTerminator();
2238 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2239 /// at the given offset and other details. If so, update the use and
2242 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2243 LSRUse::KindType Kind, Type *AccessTy) {
2244 int64_t NewMinOffset = LU.MinOffset;
2245 int64_t NewMaxOffset = LU.MaxOffset;
2246 Type *NewAccessTy = AccessTy;
2248 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2249 // something conservative, however this can pessimize in the case that one of
2250 // the uses will have all its uses outside the loop, for example.
2251 if (LU.Kind != Kind)
2254 // Check for a mismatched access type, and fall back conservatively as needed.
2255 // TODO: Be less conservative when the type is similar and can use the same
2256 // addressing modes.
2257 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2258 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2260 // Conservatively assume HasBaseReg is true for now.
2261 if (NewOffset < LU.MinOffset) {
2262 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2263 LU.MaxOffset - NewOffset, HasBaseReg))
2265 NewMinOffset = NewOffset;
2266 } else if (NewOffset > LU.MaxOffset) {
2267 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2268 NewOffset - LU.MinOffset, HasBaseReg))
2270 NewMaxOffset = NewOffset;
2274 LU.MinOffset = NewMinOffset;
2275 LU.MaxOffset = NewMaxOffset;
2276 LU.AccessTy = NewAccessTy;
2277 if (NewOffset != LU.Offsets.back())
2278 LU.Offsets.push_back(NewOffset);
2282 /// getUse - Return an LSRUse index and an offset value for a fixup which
2283 /// needs the given expression, with the given kind and optional access type.
2284 /// Either reuse an existing use or create a new one, as needed.
2285 std::pair<size_t, int64_t>
2286 LSRInstance::getUse(const SCEV *&Expr,
2287 LSRUse::KindType Kind, Type *AccessTy) {
2288 const SCEV *Copy = Expr;
2289 int64_t Offset = ExtractImmediate(Expr, SE);
2291 // Basic uses can't accept any offset, for example.
2292 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2293 Offset, /*HasBaseReg=*/ true)) {
2298 std::pair<UseMapTy::iterator, bool> P =
2299 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2301 // A use already existed with this base.
2302 size_t LUIdx = P.first->second;
2303 LSRUse &LU = Uses[LUIdx];
2304 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2306 return std::make_pair(LUIdx, Offset);
2309 // Create a new use.
2310 size_t LUIdx = Uses.size();
2311 P.first->second = LUIdx;
2312 Uses.push_back(LSRUse(Kind, AccessTy));
2313 LSRUse &LU = Uses[LUIdx];
2315 // We don't need to track redundant offsets, but we don't need to go out
2316 // of our way here to avoid them.
2317 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2318 LU.Offsets.push_back(Offset);
2320 LU.MinOffset = Offset;
2321 LU.MaxOffset = Offset;
2322 return std::make_pair(LUIdx, Offset);
2325 /// DeleteUse - Delete the given use from the Uses list.
2326 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2327 if (&LU != &Uses.back())
2328 std::swap(LU, Uses.back());
2332 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2335 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2336 /// a formula that has the same registers as the given formula.
2338 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2339 const LSRUse &OrigLU) {
2340 // Search all uses for the formula. This could be more clever.
2341 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2342 LSRUse &LU = Uses[LUIdx];
2343 // Check whether this use is close enough to OrigLU, to see whether it's
2344 // worthwhile looking through its formulae.
2345 // Ignore ICmpZero uses because they may contain formulae generated by
2346 // GenerateICmpZeroScales, in which case adding fixup offsets may
2348 if (&LU != &OrigLU &&
2349 LU.Kind != LSRUse::ICmpZero &&
2350 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2351 LU.WidestFixupType == OrigLU.WidestFixupType &&
2352 LU.HasFormulaWithSameRegs(OrigF)) {
2353 // Scan through this use's formulae.
2354 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2355 E = LU.Formulae.end(); I != E; ++I) {
2356 const Formula &F = *I;
2357 // Check to see if this formula has the same registers and symbols
2359 if (F.BaseRegs == OrigF.BaseRegs &&
2360 F.ScaledReg == OrigF.ScaledReg &&
2361 F.BaseGV == OrigF.BaseGV &&
2362 F.Scale == OrigF.Scale &&
2363 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2364 if (F.BaseOffset == 0)
2366 // This is the formula where all the registers and symbols matched;
2367 // there aren't going to be any others. Since we declined it, we
2368 // can skip the rest of the formulae and proceed to the next LSRUse.
2375 // Nothing looked good.
2379 void LSRInstance::CollectInterestingTypesAndFactors() {
2380 SmallSetVector<const SCEV *, 4> Strides;
2382 // Collect interesting types and strides.
2383 SmallVector<const SCEV *, 4> Worklist;
2384 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2385 const SCEV *Expr = IU.getExpr(*UI);
2387 // Collect interesting types.
2388 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2390 // Add strides for mentioned loops.
2391 Worklist.push_back(Expr);
2393 const SCEV *S = Worklist.pop_back_val();
2394 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2395 if (AR->getLoop() == L)
2396 Strides.insert(AR->getStepRecurrence(SE));
2397 Worklist.push_back(AR->getStart());
2398 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2399 Worklist.append(Add->op_begin(), Add->op_end());
2401 } while (!Worklist.empty());
2404 // Compute interesting factors from the set of interesting strides.
2405 for (SmallSetVector<const SCEV *, 4>::const_iterator
2406 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2407 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2408 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2409 const SCEV *OldStride = *I;
2410 const SCEV *NewStride = *NewStrideIter;
2412 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2413 SE.getTypeSizeInBits(NewStride->getType())) {
2414 if (SE.getTypeSizeInBits(OldStride->getType()) >
2415 SE.getTypeSizeInBits(NewStride->getType()))
2416 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2418 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2420 if (const SCEVConstant *Factor =
2421 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2423 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2424 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2425 } else if (const SCEVConstant *Factor =
2426 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2429 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2430 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2434 // If all uses use the same type, don't bother looking for truncation-based
2436 if (Types.size() == 1)
2439 DEBUG(print_factors_and_types(dbgs()));
2442 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2443 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2444 /// Instructions to IVStrideUses, we could partially skip this.
2445 static User::op_iterator
2446 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2447 Loop *L, ScalarEvolution &SE) {
2448 for(; OI != OE; ++OI) {
2449 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2450 if (!SE.isSCEVable(Oper->getType()))
2453 if (const SCEVAddRecExpr *AR =
2454 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2455 if (AR->getLoop() == L)
2463 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2464 /// operands, so wrap it in a convenient helper.
2465 static Value *getWideOperand(Value *Oper) {
2466 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2467 return Trunc->getOperand(0);
2471 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2473 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2474 Type *LType = LVal->getType();
2475 Type *RType = RVal->getType();
2476 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2479 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2480 /// NULL for any constant. Returning the expression itself is
2481 /// conservative. Returning a deeper subexpression is more precise and valid as
2482 /// long as it isn't less complex than another subexpression. For expressions
2483 /// involving multiple unscaled values, we need to return the pointer-type
2484 /// SCEVUnknown. This avoids forming chains across objects, such as:
2485 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2487 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2488 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2489 static const SCEV *getExprBase(const SCEV *S) {
2490 switch (S->getSCEVType()) {
2491 default: // uncluding scUnknown.
2496 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2498 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2500 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2502 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2503 // there's nothing more complex.
2504 // FIXME: not sure if we want to recognize negation.
2505 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2506 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2507 E(Add->op_begin()); I != E; ++I) {
2508 const SCEV *SubExpr = *I;
2509 if (SubExpr->getSCEVType() == scAddExpr)
2510 return getExprBase(SubExpr);
2512 if (SubExpr->getSCEVType() != scMulExpr)
2515 return S; // all operands are scaled, be conservative.
2518 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2522 /// Return true if the chain increment is profitable to expand into a loop
2523 /// invariant value, which may require its own register. A profitable chain
2524 /// increment will be an offset relative to the same base. We allow such offsets
2525 /// to potentially be used as chain increment as long as it's not obviously
2526 /// expensive to expand using real instructions.
2527 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2528 const SCEV *IncExpr,
2529 ScalarEvolution &SE) {
2530 // Aggressively form chains when -stress-ivchain.
2534 // Do not replace a constant offset from IV head with a nonconstant IV
2536 if (!isa<SCEVConstant>(IncExpr)) {
2537 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2538 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2542 SmallPtrSet<const SCEV*, 8> Processed;
2543 return !isHighCostExpansion(IncExpr, Processed, SE);
2546 /// Return true if the number of registers needed for the chain is estimated to
2547 /// be less than the number required for the individual IV users. First prohibit
2548 /// any IV users that keep the IV live across increments (the Users set should
2549 /// be empty). Next count the number and type of increments in the chain.
2551 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2552 /// effectively use postinc addressing modes. Only consider it profitable it the
2553 /// increments can be computed in fewer registers when chained.
2555 /// TODO: Consider IVInc free if it's already used in another chains.
2557 isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2558 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2562 if (!Chain.hasIncs())
2565 if (!Users.empty()) {
2566 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2567 for (Instruction *Inst : Users) {
2568 dbgs() << " " << *Inst << "\n";
2572 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2574 // The chain itself may require a register, so intialize cost to 1.
2577 // A complete chain likely eliminates the need for keeping the original IV in
2578 // a register. LSR does not currently know how to form a complete chain unless
2579 // the header phi already exists.
2580 if (isa<PHINode>(Chain.tailUserInst())
2581 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2584 const SCEV *LastIncExpr = nullptr;
2585 unsigned NumConstIncrements = 0;
2586 unsigned NumVarIncrements = 0;
2587 unsigned NumReusedIncrements = 0;
2588 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2591 if (I->IncExpr->isZero())
2594 // Incrementing by zero or some constant is neutral. We assume constants can
2595 // be folded into an addressing mode or an add's immediate operand.
2596 if (isa<SCEVConstant>(I->IncExpr)) {
2597 ++NumConstIncrements;
2601 if (I->IncExpr == LastIncExpr)
2602 ++NumReusedIncrements;
2606 LastIncExpr = I->IncExpr;
2608 // An IV chain with a single increment is handled by LSR's postinc
2609 // uses. However, a chain with multiple increments requires keeping the IV's
2610 // value live longer than it needs to be if chained.
2611 if (NumConstIncrements > 1)
2614 // Materializing increment expressions in the preheader that didn't exist in
2615 // the original code may cost a register. For example, sign-extended array
2616 // indices can produce ridiculous increments like this:
2617 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2618 cost += NumVarIncrements;
2620 // Reusing variable increments likely saves a register to hold the multiple of
2622 cost -= NumReusedIncrements;
2624 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2630 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2632 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2633 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2634 // When IVs are used as types of varying widths, they are generally converted
2635 // to a wider type with some uses remaining narrow under a (free) trunc.
2636 Value *const NextIV = getWideOperand(IVOper);
2637 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2638 const SCEV *const OperExprBase = getExprBase(OperExpr);
2640 // Visit all existing chains. Check if its IVOper can be computed as a
2641 // profitable loop invariant increment from the last link in the Chain.
2642 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2643 const SCEV *LastIncExpr = nullptr;
2644 for (; ChainIdx < NChains; ++ChainIdx) {
2645 IVChain &Chain = IVChainVec[ChainIdx];
2647 // Prune the solution space aggressively by checking that both IV operands
2648 // are expressions that operate on the same unscaled SCEVUnknown. This
2649 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2650 // first avoids creating extra SCEV expressions.
2651 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2654 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2655 if (!isCompatibleIVType(PrevIV, NextIV))
2658 // A phi node terminates a chain.
2659 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2662 // The increment must be loop-invariant so it can be kept in a register.
2663 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2664 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2665 if (!SE.isLoopInvariant(IncExpr, L))
2668 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2669 LastIncExpr = IncExpr;
2673 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2674 // bother for phi nodes, because they must be last in the chain.
2675 if (ChainIdx == NChains) {
2676 if (isa<PHINode>(UserInst))
2678 if (NChains >= MaxChains && !StressIVChain) {
2679 DEBUG(dbgs() << "IV Chain Limit\n");
2682 LastIncExpr = OperExpr;
2683 // IVUsers may have skipped over sign/zero extensions. We don't currently
2684 // attempt to form chains involving extensions unless they can be hoisted
2685 // into this loop's AddRec.
2686 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2689 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2691 ChainUsersVec.resize(NChains);
2692 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2693 << ") IV=" << *LastIncExpr << "\n");
2695 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2696 << ") IV+" << *LastIncExpr << "\n");
2697 // Add this IV user to the end of the chain.
2698 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2700 IVChain &Chain = IVChainVec[ChainIdx];
2702 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2703 // This chain's NearUsers become FarUsers.
2704 if (!LastIncExpr->isZero()) {
2705 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2710 // All other uses of IVOperand become near uses of the chain.
2711 // We currently ignore intermediate values within SCEV expressions, assuming
2712 // they will eventually be used be the current chain, or can be computed
2713 // from one of the chain increments. To be more precise we could
2714 // transitively follow its user and only add leaf IV users to the set.
2715 for (User *U : IVOper->users()) {
2716 Instruction *OtherUse = dyn_cast<Instruction>(U);
2719 // Uses in the chain will no longer be uses if the chain is formed.
2720 // Include the head of the chain in this iteration (not Chain.begin()).
2721 IVChain::const_iterator IncIter = Chain.Incs.begin();
2722 IVChain::const_iterator IncEnd = Chain.Incs.end();
2723 for( ; IncIter != IncEnd; ++IncIter) {
2724 if (IncIter->UserInst == OtherUse)
2727 if (IncIter != IncEnd)
2730 if (SE.isSCEVable(OtherUse->getType())
2731 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2732 && IU.isIVUserOrOperand(OtherUse)) {
2735 NearUsers.insert(OtherUse);
2738 // Since this user is part of the chain, it's no longer considered a use
2740 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2743 /// CollectChains - Populate the vector of Chains.
2745 /// This decreases ILP at the architecture level. Targets with ample registers,
2746 /// multiple memory ports, and no register renaming probably don't want
2747 /// this. However, such targets should probably disable LSR altogether.
2749 /// The job of LSR is to make a reasonable choice of induction variables across
2750 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2751 /// ILP *within the loop* if the target wants it.
2753 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2754 /// will not reorder memory operations, it will recognize this as a chain, but
2755 /// will generate redundant IV increments. Ideally this would be corrected later
2756 /// by a smart scheduler:
2762 /// TODO: Walk the entire domtree within this loop, not just the path to the
2763 /// loop latch. This will discover chains on side paths, but requires
2764 /// maintaining multiple copies of the Chains state.
2765 void LSRInstance::CollectChains() {
2766 DEBUG(dbgs() << "Collecting IV Chains.\n");
2767 SmallVector<ChainUsers, 8> ChainUsersVec;
2769 SmallVector<BasicBlock *,8> LatchPath;
2770 BasicBlock *LoopHeader = L->getHeader();
2771 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2772 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2773 LatchPath.push_back(Rung->getBlock());
2775 LatchPath.push_back(LoopHeader);
2777 // Walk the instruction stream from the loop header to the loop latch.
2778 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2779 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2780 BBIter != BBEnd; ++BBIter) {
2781 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2783 // Skip instructions that weren't seen by IVUsers analysis.
2784 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2787 // Ignore users that are part of a SCEV expression. This way we only
2788 // consider leaf IV Users. This effectively rediscovers a portion of
2789 // IVUsers analysis but in program order this time.
2790 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2793 // Remove this instruction from any NearUsers set it may be in.
2794 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2795 ChainIdx < NChains; ++ChainIdx) {
2796 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2798 // Search for operands that can be chained.
2799 SmallPtrSet<Instruction*, 4> UniqueOperands;
2800 User::op_iterator IVOpEnd = I->op_end();
2801 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2802 while (IVOpIter != IVOpEnd) {
2803 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2804 if (UniqueOperands.insert(IVOpInst).second)
2805 ChainInstruction(I, IVOpInst, ChainUsersVec);
2806 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2808 } // Continue walking down the instructions.
2809 } // Continue walking down the domtree.
2810 // Visit phi backedges to determine if the chain can generate the IV postinc.
2811 for (BasicBlock::iterator I = L->getHeader()->begin();
2812 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2813 if (!SE.isSCEVable(PN->getType()))
2817 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2819 ChainInstruction(PN, IncV, ChainUsersVec);
2821 // Remove any unprofitable chains.
2822 unsigned ChainIdx = 0;
2823 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2824 UsersIdx < NChains; ++UsersIdx) {
2825 if (!isProfitableChain(IVChainVec[UsersIdx],
2826 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2828 // Preserve the chain at UsesIdx.
2829 if (ChainIdx != UsersIdx)
2830 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2831 FinalizeChain(IVChainVec[ChainIdx]);
2834 IVChainVec.resize(ChainIdx);
2837 void LSRInstance::FinalizeChain(IVChain &Chain) {
2838 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2839 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2841 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2843 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2844 User::op_iterator UseI =
2845 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2846 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2847 IVIncSet.insert(UseI);
2851 /// Return true if the IVInc can be folded into an addressing mode.
2852 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2853 Value *Operand, const TargetTransformInfo &TTI) {
2854 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2855 if (!IncConst || !isAddressUse(UserInst, Operand))
2858 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2861 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2862 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2863 getAccessType(UserInst), /*BaseGV=*/ nullptr,
2864 IncOffset, /*HaseBaseReg=*/ false))
2870 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2871 /// materialize the IV user's operand from the previous IV user's operand.
2872 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2873 SmallVectorImpl<WeakVH> &DeadInsts) {
2874 // Find the new IVOperand for the head of the chain. It may have been replaced
2876 const IVInc &Head = Chain.Incs[0];
2877 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2878 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2879 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2881 Value *IVSrc = nullptr;
2882 while (IVOpIter != IVOpEnd) {
2883 IVSrc = getWideOperand(*IVOpIter);
2885 // If this operand computes the expression that the chain needs, we may use
2886 // it. (Check this after setting IVSrc which is used below.)
2888 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2889 // narrow for the chain, so we can no longer use it. We do allow using a
2890 // wider phi, assuming the LSR checked for free truncation. In that case we
2891 // should already have a truncate on this operand such that
2892 // getSCEV(IVSrc) == IncExpr.
2893 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2894 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2897 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2899 if (IVOpIter == IVOpEnd) {
2900 // Gracefully give up on this chain.
2901 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2905 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2906 Type *IVTy = IVSrc->getType();
2907 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2908 const SCEV *LeftOverExpr = nullptr;
2909 for (IVChain::const_iterator IncI = Chain.begin(),
2910 IncE = Chain.end(); IncI != IncE; ++IncI) {
2912 Instruction *InsertPt = IncI->UserInst;
2913 if (isa<PHINode>(InsertPt))
2914 InsertPt = L->getLoopLatch()->getTerminator();
2916 // IVOper will replace the current IV User's operand. IVSrc is the IV
2917 // value currently held in a register.
2918 Value *IVOper = IVSrc;
2919 if (!IncI->IncExpr->isZero()) {
2920 // IncExpr was the result of subtraction of two narrow values, so must
2922 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2923 LeftOverExpr = LeftOverExpr ?
2924 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2926 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2927 // Expand the IV increment.
2928 Rewriter.clearPostInc();
2929 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2930 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2931 SE.getUnknown(IncV));
2932 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2934 // If an IV increment can't be folded, use it as the next IV value.
2935 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2937 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2939 LeftOverExpr = nullptr;
2942 Type *OperTy = IncI->IVOperand->getType();
2943 if (IVTy != OperTy) {
2944 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2945 "cannot extend a chained IV");
2946 IRBuilder<> Builder(InsertPt);
2947 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2949 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2950 DeadInsts.push_back(IncI->IVOperand);
2952 // If LSR created a new, wider phi, we may also replace its postinc. We only
2953 // do this if we also found a wide value for the head of the chain.
2954 if (isa<PHINode>(Chain.tailUserInst())) {
2955 for (BasicBlock::iterator I = L->getHeader()->begin();
2956 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2957 if (!isCompatibleIVType(Phi, IVSrc))
2959 Instruction *PostIncV = dyn_cast<Instruction>(
2960 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2961 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2963 Value *IVOper = IVSrc;
2964 Type *PostIncTy = PostIncV->getType();
2965 if (IVTy != PostIncTy) {
2966 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2967 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2968 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2969 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2971 Phi->replaceUsesOfWith(PostIncV, IVOper);
2972 DeadInsts.push_back(PostIncV);
2977 void LSRInstance::CollectFixupsAndInitialFormulae() {
2978 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2979 Instruction *UserInst = UI->getUser();
2980 // Skip IV users that are part of profitable IV Chains.
2981 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2982 UI->getOperandValToReplace());
2983 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2984 if (IVIncSet.count(UseI))
2988 LSRFixup &LF = getNewFixup();
2989 LF.UserInst = UserInst;
2990 LF.OperandValToReplace = UI->getOperandValToReplace();
2991 LF.PostIncLoops = UI->getPostIncLoops();
2993 LSRUse::KindType Kind = LSRUse::Basic;
2994 Type *AccessTy = nullptr;
2995 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2996 Kind = LSRUse::Address;
2997 AccessTy = getAccessType(LF.UserInst);
3000 const SCEV *S = IU.getExpr(*UI);
3002 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3003 // (N - i == 0), and this allows (N - i) to be the expression that we work
3004 // with rather than just N or i, so we can consider the register
3005 // requirements for both N and i at the same time. Limiting this code to
3006 // equality icmps is not a problem because all interesting loops use
3007 // equality icmps, thanks to IndVarSimplify.
3008 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
3009 if (CI->isEquality()) {
3010 // Swap the operands if needed to put the OperandValToReplace on the
3011 // left, for consistency.
3012 Value *NV = CI->getOperand(1);
3013 if (NV == LF.OperandValToReplace) {
3014 CI->setOperand(1, CI->getOperand(0));
3015 CI->setOperand(0, NV);
3016 NV = CI->getOperand(1);
3020 // x == y --> x - y == 0
3021 const SCEV *N = SE.getSCEV(NV);
3022 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3023 // S is normalized, so normalize N before folding it into S
3024 // to keep the result normalized.
3025 N = TransformForPostIncUse(Normalize, N, CI, nullptr,
3026 LF.PostIncLoops, SE, DT);
3027 Kind = LSRUse::ICmpZero;
3028 S = SE.getMinusSCEV(N, S);
3031 // -1 and the negations of all interesting strides (except the negation
3032 // of -1) are now also interesting.
3033 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3034 if (Factors[i] != -1)
3035 Factors.insert(-(uint64_t)Factors[i]);
3039 // Set up the initial formula for this use.
3040 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3042 LF.Offset = P.second;
3043 LSRUse &LU = Uses[LF.LUIdx];
3044 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3045 if (!LU.WidestFixupType ||
3046 SE.getTypeSizeInBits(LU.WidestFixupType) <
3047 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3048 LU.WidestFixupType = LF.OperandValToReplace->getType();
3050 // If this is the first use of this LSRUse, give it a formula.
3051 if (LU.Formulae.empty()) {
3052 InsertInitialFormula(S, LU, LF.LUIdx);
3053 CountRegisters(LU.Formulae.back(), LF.LUIdx);
3057 DEBUG(print_fixups(dbgs()));
3060 /// InsertInitialFormula - Insert a formula for the given expression into
3061 /// the given use, separating out loop-variant portions from loop-invariant
3062 /// and loop-computable portions.
3064 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3065 // Mark uses whose expressions cannot be expanded.
3066 if (!isSafeToExpand(S, SE))
3067 LU.RigidFormula = true;
3070 F.InitialMatch(S, L, SE);
3071 bool Inserted = InsertFormula(LU, LUIdx, F);
3072 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3075 /// InsertSupplementalFormula - Insert a simple single-register formula for
3076 /// the given expression into the given use.
3078 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3079 LSRUse &LU, size_t LUIdx) {
3081 F.BaseRegs.push_back(S);
3082 F.HasBaseReg = true;
3083 bool Inserted = InsertFormula(LU, LUIdx, F);
3084 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3087 /// CountRegisters - Note which registers are used by the given formula,
3088 /// updating RegUses.
3089 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3091 RegUses.CountRegister(F.ScaledReg, LUIdx);
3092 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3093 E = F.BaseRegs.end(); I != E; ++I)
3094 RegUses.CountRegister(*I, LUIdx);
3097 /// InsertFormula - If the given formula has not yet been inserted, add it to
3098 /// the list, and return true. Return false otherwise.
3099 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3100 // Do not insert formula that we will not be able to expand.
3101 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3102 "Formula is illegal");
3103 if (!LU.InsertFormula(F))
3106 CountRegisters(F, LUIdx);
3110 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3111 /// loop-invariant values which we're tracking. These other uses will pin these
3112 /// values in registers, making them less profitable for elimination.
3113 /// TODO: This currently misses non-constant addrec step registers.
3114 /// TODO: Should this give more weight to users inside the loop?
3116 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3117 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3118 SmallPtrSet<const SCEV *, 32> Visited;
3120 while (!Worklist.empty()) {
3121 const SCEV *S = Worklist.pop_back_val();
3123 // Don't process the same SCEV twice
3124 if (!Visited.insert(S).second)
3127 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3128 Worklist.append(N->op_begin(), N->op_end());
3129 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3130 Worklist.push_back(C->getOperand());
3131 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3132 Worklist.push_back(D->getLHS());
3133 Worklist.push_back(D->getRHS());
3134 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3135 const Value *V = US->getValue();
3136 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3137 // Look for instructions defined outside the loop.
3138 if (L->contains(Inst)) continue;
3139 } else if (isa<UndefValue>(V))
3140 // Undef doesn't have a live range, so it doesn't matter.
3142 for (const Use &U : V->uses()) {
3143 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3144 // Ignore non-instructions.
3147 // Ignore instructions in other functions (as can happen with
3149 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3151 // Ignore instructions not dominated by the loop.
3152 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3153 UserInst->getParent() :
3154 cast<PHINode>(UserInst)->getIncomingBlock(
3155 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3156 if (!DT.dominates(L->getHeader(), UseBB))
3158 // Ignore uses which are part of other SCEV expressions, to avoid
3159 // analyzing them multiple times.
3160 if (SE.isSCEVable(UserInst->getType())) {
3161 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3162 // If the user is a no-op, look through to its uses.
3163 if (!isa<SCEVUnknown>(UserS))
3167 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3171 // Ignore icmp instructions which are already being analyzed.
3172 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3173 unsigned OtherIdx = !U.getOperandNo();
3174 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3175 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3179 LSRFixup &LF = getNewFixup();
3180 LF.UserInst = const_cast<Instruction *>(UserInst);
3181 LF.OperandValToReplace = U;
3182 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, nullptr);
3184 LF.Offset = P.second;
3185 LSRUse &LU = Uses[LF.LUIdx];
3186 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3187 if (!LU.WidestFixupType ||
3188 SE.getTypeSizeInBits(LU.WidestFixupType) <
3189 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3190 LU.WidestFixupType = LF.OperandValToReplace->getType();
3191 InsertSupplementalFormula(US, LU, LF.LUIdx);
3192 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3199 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3200 /// separate registers. If C is non-null, multiply each subexpression by C.
3202 /// Return remainder expression after factoring the subexpressions captured by
3203 /// Ops. If Ops is complete, return NULL.
3204 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3205 SmallVectorImpl<const SCEV *> &Ops,
3207 ScalarEvolution &SE,
3208 unsigned Depth = 0) {
3209 // Arbitrarily cap recursion to protect compile time.
3213 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3214 // Break out add operands.
3215 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3217 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3219 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3222 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3223 // Split a non-zero base out of an addrec.
3224 if (AR->getStart()->isZero())
3227 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3228 C, Ops, L, SE, Depth+1);
3229 // Split the non-zero AddRec unless it is part of a nested recurrence that
3230 // does not pertain to this loop.
3231 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3232 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3233 Remainder = nullptr;
3235 if (Remainder != AR->getStart()) {
3237 Remainder = SE.getConstant(AR->getType(), 0);
3238 return SE.getAddRecExpr(Remainder,
3239 AR->getStepRecurrence(SE),
3241 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3244 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3245 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3246 if (Mul->getNumOperands() != 2)
3248 if (const SCEVConstant *Op0 =
3249 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3250 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3251 const SCEV *Remainder =
3252 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3254 Ops.push_back(SE.getMulExpr(C, Remainder));
3261 /// \brief Helper function for LSRInstance::GenerateReassociations.
3262 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3263 const Formula &Base,
3264 unsigned Depth, size_t Idx,
3266 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3267 SmallVector<const SCEV *, 8> AddOps;
3268 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3270 AddOps.push_back(Remainder);
3272 if (AddOps.size() == 1)
3275 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3279 // Loop-variant "unknown" values are uninteresting; we won't be able to
3280 // do anything meaningful with them.
3281 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3284 // Don't pull a constant into a register if the constant could be folded
3285 // into an immediate field.
3286 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3287 LU.AccessTy, *J, Base.getNumRegs() > 1))
3290 // Collect all operands except *J.
3291 SmallVector<const SCEV *, 8> InnerAddOps(
3292 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3293 InnerAddOps.append(std::next(J),
3294 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3296 // Don't leave just a constant behind in a register if the constant could
3297 // be folded into an immediate field.
3298 if (InnerAddOps.size() == 1 &&
3299 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3300 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3303 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3304 if (InnerSum->isZero())
3308 // Add the remaining pieces of the add back into the new formula.
3309 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3310 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3311 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3312 InnerSumSC->getValue()->getZExtValue())) {
3314 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3316 F.ScaledReg = nullptr;
3318 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3319 } else if (IsScaledReg)
3320 F.ScaledReg = InnerSum;
3322 F.BaseRegs[Idx] = InnerSum;
3324 // Add J as its own register, or an unfolded immediate.
3325 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3326 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3327 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3328 SC->getValue()->getZExtValue()))
3330 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3332 F.BaseRegs.push_back(*J);
3333 // We may have changed the number of register in base regs, adjust the
3334 // formula accordingly.
3337 if (InsertFormula(LU, LUIdx, F))
3338 // If that formula hadn't been seen before, recurse to find more like
3340 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3344 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3346 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3347 Formula Base, unsigned Depth) {
3348 assert(Base.isCanonical() && "Input must be in the canonical form");
3349 // Arbitrarily cap recursion to protect compile time.
3353 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3354 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3356 if (Base.Scale == 1)
3357 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3358 /* Idx */ -1, /* IsScaledReg */ true);
3361 /// GenerateCombinations - Generate a formula consisting of all of the
3362 /// loop-dominating registers added into a single register.
3363 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3365 // This method is only interesting on a plurality of registers.
3366 if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3369 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3370 // processing the formula.
3374 SmallVector<const SCEV *, 4> Ops;
3375 for (SmallVectorImpl<const SCEV *>::const_iterator
3376 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3377 const SCEV *BaseReg = *I;
3378 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3379 !SE.hasComputableLoopEvolution(BaseReg, L))
3380 Ops.push_back(BaseReg);
3382 F.BaseRegs.push_back(BaseReg);
3384 if (Ops.size() > 1) {
3385 const SCEV *Sum = SE.getAddExpr(Ops);
3386 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3387 // opportunity to fold something. For now, just ignore such cases
3388 // rather than proceed with zero in a register.
3389 if (!Sum->isZero()) {
3390 F.BaseRegs.push_back(Sum);
3392 (void)InsertFormula(LU, LUIdx, F);
3397 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3398 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3399 const Formula &Base, size_t Idx,
3401 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3402 GlobalValue *GV = ExtractSymbol(G, SE);
3403 if (G->isZero() || !GV)
3407 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3412 F.BaseRegs[Idx] = G;
3413 (void)InsertFormula(LU, LUIdx, F);
3416 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3417 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3419 // We can't add a symbolic offset if the address already contains one.
3420 if (Base.BaseGV) return;
3422 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3423 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3424 if (Base.Scale == 1)
3425 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3426 /* IsScaledReg */ true);
3429 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3430 void LSRInstance::GenerateConstantOffsetsImpl(
3431 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3432 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3433 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3434 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3438 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3439 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3441 // Add the offset to the base register.
3442 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3443 // If it cancelled out, drop the base register, otherwise update it.
3444 if (NewG->isZero()) {
3447 F.ScaledReg = nullptr;
3449 F.DeleteBaseReg(F.BaseRegs[Idx]);
3451 } else if (IsScaledReg)
3454 F.BaseRegs[Idx] = NewG;
3456 (void)InsertFormula(LU, LUIdx, F);
3460 int64_t Imm = ExtractImmediate(G, SE);
3461 if (G->isZero() || Imm == 0)
3464 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3465 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3470 F.BaseRegs[Idx] = G;
3471 (void)InsertFormula(LU, LUIdx, F);
3474 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3475 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3477 // TODO: For now, just add the min and max offset, because it usually isn't
3478 // worthwhile looking at everything inbetween.
3479 SmallVector<int64_t, 2> Worklist;
3480 Worklist.push_back(LU.MinOffset);
3481 if (LU.MaxOffset != LU.MinOffset)
3482 Worklist.push_back(LU.MaxOffset);
3484 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3485 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3486 if (Base.Scale == 1)
3487 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3488 /* IsScaledReg */ true);
3491 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3492 /// the comparison. For example, x == y -> x*c == y*c.
3493 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3495 if (LU.Kind != LSRUse::ICmpZero) return;
3497 // Determine the integer type for the base formula.
3498 Type *IntTy = Base.getType();
3500 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3502 // Don't do this if there is more than one offset.
3503 if (LU.MinOffset != LU.MaxOffset) return;
3505 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3507 // Check each interesting stride.
3508 for (SmallSetVector<int64_t, 8>::const_iterator
3509 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3510 int64_t Factor = *I;
3512 // Check that the multiplication doesn't overflow.
3513 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3515 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3516 if (NewBaseOffset / Factor != Base.BaseOffset)
3518 // If the offset will be truncated at this use, check that it is in bounds.
3519 if (!IntTy->isPointerTy() &&
3520 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3523 // Check that multiplying with the use offset doesn't overflow.
3524 int64_t Offset = LU.MinOffset;
3525 if (Offset == INT64_MIN && Factor == -1)
3527 Offset = (uint64_t)Offset * Factor;
3528 if (Offset / Factor != LU.MinOffset)
3530 // If the offset will be truncated at this use, check that it is in bounds.
3531 if (!IntTy->isPointerTy() &&
3532 !ConstantInt::isValueValidForType(IntTy, Offset))
3536 F.BaseOffset = NewBaseOffset;
3538 // Check that this scale is legal.
3539 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3542 // Compensate for the use having MinOffset built into it.
3543 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3545 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3547 // Check that multiplying with each base register doesn't overflow.
3548 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3549 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3550 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3554 // Check that multiplying with the scaled register doesn't overflow.
3556 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3557 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3561 // Check that multiplying with the unfolded offset doesn't overflow.
3562 if (F.UnfoldedOffset != 0) {
3563 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3565 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3566 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3568 // If the offset will be truncated, check that it is in bounds.
3569 if (!IntTy->isPointerTy() &&
3570 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3574 // If we make it here and it's legal, add it.
3575 (void)InsertFormula(LU, LUIdx, F);
3580 /// GenerateScales - Generate stride factor reuse formulae by making use of
3581 /// scaled-offset address modes, for example.
3582 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3583 // Determine the integer type for the base formula.
3584 Type *IntTy = Base.getType();
3587 // If this Formula already has a scaled register, we can't add another one.
3588 // Try to unscale the formula to generate a better scale.
3589 if (Base.Scale != 0 && !Base.Unscale())
3592 assert(Base.Scale == 0 && "Unscale did not did its job!");
3594 // Check each interesting stride.
3595 for (SmallSetVector<int64_t, 8>::const_iterator
3596 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3597 int64_t Factor = *I;
3599 Base.Scale = Factor;
3600 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3601 // Check whether this scale is going to be legal.
3602 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3604 // As a special-case, handle special out-of-loop Basic users specially.
3605 // TODO: Reconsider this special case.
3606 if (LU.Kind == LSRUse::Basic &&
3607 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3608 LU.AccessTy, Base) &&
3609 LU.AllFixupsOutsideLoop)
3610 LU.Kind = LSRUse::Special;
3614 // For an ICmpZero, negating a solitary base register won't lead to
3616 if (LU.Kind == LSRUse::ICmpZero &&
3617 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3619 // For each addrec base reg, apply the scale, if possible.
3620 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3621 if (const SCEVAddRecExpr *AR =
3622 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3623 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3624 if (FactorS->isZero())
3626 // Divide out the factor, ignoring high bits, since we'll be
3627 // scaling the value back up in the end.
3628 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3629 // TODO: This could be optimized to avoid all the copying.
3631 F.ScaledReg = Quotient;
3632 F.DeleteBaseReg(F.BaseRegs[i]);
3633 // The canonical representation of 1*reg is reg, which is already in
3634 // Base. In that case, do not try to insert the formula, it will be
3636 if (F.Scale == 1 && F.BaseRegs.empty())
3638 (void)InsertFormula(LU, LUIdx, F);
3644 /// GenerateTruncates - Generate reuse formulae from different IV types.
3645 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3646 // Don't bother truncating symbolic values.
3647 if (Base.BaseGV) return;
3649 // Determine the integer type for the base formula.
3650 Type *DstTy = Base.getType();
3652 DstTy = SE.getEffectiveSCEVType(DstTy);
3654 for (SmallSetVector<Type *, 4>::const_iterator
3655 I = Types.begin(), E = Types.end(); I != E; ++I) {
3657 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3660 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3661 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3662 JE = F.BaseRegs.end(); J != JE; ++J)
3663 *J = SE.getAnyExtendExpr(*J, SrcTy);
3665 // TODO: This assumes we've done basic processing on all uses and
3666 // have an idea what the register usage is.
3667 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3670 (void)InsertFormula(LU, LUIdx, F);
3677 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3678 /// defer modifications so that the search phase doesn't have to worry about
3679 /// the data structures moving underneath it.
3683 const SCEV *OrigReg;
3685 WorkItem(size_t LI, int64_t I, const SCEV *R)
3686 : LUIdx(LI), Imm(I), OrigReg(R) {}
3688 void print(raw_ostream &OS) const;
3694 void WorkItem::print(raw_ostream &OS) const {
3695 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3696 << " , add offset " << Imm;
3699 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3700 void WorkItem::dump() const {
3701 print(errs()); errs() << '\n';
3705 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3706 /// distance apart and try to form reuse opportunities between them.
3707 void LSRInstance::GenerateCrossUseConstantOffsets() {
3708 // Group the registers by their value without any added constant offset.
3709 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3710 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3712 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3713 SmallVector<const SCEV *, 8> Sequence;
3714 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3716 const SCEV *Reg = *I;
3717 int64_t Imm = ExtractImmediate(Reg, SE);
3718 std::pair<RegMapTy::iterator, bool> Pair =
3719 Map.insert(std::make_pair(Reg, ImmMapTy()));
3721 Sequence.push_back(Reg);
3722 Pair.first->second.insert(std::make_pair(Imm, *I));
3723 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3726 // Now examine each set of registers with the same base value. Build up
3727 // a list of work to do and do the work in a separate step so that we're
3728 // not adding formulae and register counts while we're searching.
3729 SmallVector<WorkItem, 32> WorkItems;
3730 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3731 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3732 E = Sequence.end(); I != E; ++I) {
3733 const SCEV *Reg = *I;
3734 const ImmMapTy &Imms = Map.find(Reg)->second;
3736 // It's not worthwhile looking for reuse if there's only one offset.
3737 if (Imms.size() == 1)
3740 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3741 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3743 dbgs() << ' ' << J->first;
3746 // Examine each offset.
3747 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3749 const SCEV *OrigReg = J->second;
3751 int64_t JImm = J->first;
3752 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3754 if (!isa<SCEVConstant>(OrigReg) &&
3755 UsedByIndicesMap[Reg].count() == 1) {
3756 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3760 // Conservatively examine offsets between this orig reg a few selected
3762 ImmMapTy::const_iterator OtherImms[] = {
3763 Imms.begin(), std::prev(Imms.end()),
3764 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3767 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3768 ImmMapTy::const_iterator M = OtherImms[i];
3769 if (M == J || M == JE) continue;
3771 // Compute the difference between the two.
3772 int64_t Imm = (uint64_t)JImm - M->first;
3773 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3774 LUIdx = UsedByIndices.find_next(LUIdx))
3775 // Make a memo of this use, offset, and register tuple.
3776 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3777 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3784 UsedByIndicesMap.clear();
3785 UniqueItems.clear();
3787 // Now iterate through the worklist and add new formulae.
3788 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3789 E = WorkItems.end(); I != E; ++I) {
3790 const WorkItem &WI = *I;
3791 size_t LUIdx = WI.LUIdx;
3792 LSRUse &LU = Uses[LUIdx];
3793 int64_t Imm = WI.Imm;
3794 const SCEV *OrigReg = WI.OrigReg;
3796 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3797 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3798 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3800 // TODO: Use a more targeted data structure.
3801 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3802 Formula F = LU.Formulae[L];
3803 // FIXME: The code for the scaled and unscaled registers looks
3804 // very similar but slightly different. Investigate if they
3805 // could be merged. That way, we would not have to unscale the
3808 // Use the immediate in the scaled register.
3809 if (F.ScaledReg == OrigReg) {
3810 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3811 // Don't create 50 + reg(-50).
3812 if (F.referencesReg(SE.getSCEV(
3813 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3816 NewF.BaseOffset = Offset;
3817 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3820 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3822 // If the new scale is a constant in a register, and adding the constant
3823 // value to the immediate would produce a value closer to zero than the
3824 // immediate itself, then the formula isn't worthwhile.
3825 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3826 if (C->getValue()->isNegative() !=
3827 (NewF.BaseOffset < 0) &&
3828 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3829 .ule(std::abs(NewF.BaseOffset)))
3833 NewF.Canonicalize();
3834 (void)InsertFormula(LU, LUIdx, NewF);
3836 // Use the immediate in a base register.
3837 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3838 const SCEV *BaseReg = F.BaseRegs[N];
3839 if (BaseReg != OrigReg)
3842 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3843 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3844 LU.Kind, LU.AccessTy, NewF)) {
3845 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3848 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3850 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3852 // If the new formula has a constant in a register, and adding the
3853 // constant value to the immediate would produce a value closer to
3854 // zero than the immediate itself, then the formula isn't worthwhile.
3855 for (SmallVectorImpl<const SCEV *>::const_iterator
3856 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3858 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3859 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3860 std::abs(NewF.BaseOffset)) &&
3861 (C->getValue()->getValue() +
3862 NewF.BaseOffset).countTrailingZeros() >=
3863 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3867 NewF.Canonicalize();
3868 (void)InsertFormula(LU, LUIdx, NewF);
3877 /// GenerateAllReuseFormulae - Generate formulae for each use.
3879 LSRInstance::GenerateAllReuseFormulae() {
3880 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3881 // queries are more precise.
3882 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3883 LSRUse &LU = Uses[LUIdx];
3884 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3885 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3886 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3887 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3889 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3890 LSRUse &LU = Uses[LUIdx];
3891 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3892 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3893 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3894 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3895 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3896 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3897 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3898 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3900 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3901 LSRUse &LU = Uses[LUIdx];
3902 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3903 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3906 GenerateCrossUseConstantOffsets();
3908 DEBUG(dbgs() << "\n"
3909 "After generating reuse formulae:\n";
3910 print_uses(dbgs()));
3913 /// If there are multiple formulae with the same set of registers used
3914 /// by other uses, pick the best one and delete the others.
3915 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3916 DenseSet<const SCEV *> VisitedRegs;
3917 SmallPtrSet<const SCEV *, 16> Regs;
3918 SmallPtrSet<const SCEV *, 16> LoserRegs;
3920 bool ChangedFormulae = false;
3923 // Collect the best formula for each unique set of shared registers. This
3924 // is reset for each use.
3925 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3927 BestFormulaeTy BestFormulae;
3929 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3930 LSRUse &LU = Uses[LUIdx];
3931 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3934 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3935 FIdx != NumForms; ++FIdx) {
3936 Formula &F = LU.Formulae[FIdx];
3938 // Some formulas are instant losers. For example, they may depend on
3939 // nonexistent AddRecs from other loops. These need to be filtered
3940 // immediately, otherwise heuristics could choose them over others leading
3941 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3942 // avoids the need to recompute this information across formulae using the
3943 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3944 // the corresponding bad register from the Regs set.
3947 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3949 if (CostF.isLoser()) {
3950 // During initial formula generation, undesirable formulae are generated
3951 // by uses within other loops that have some non-trivial address mode or
3952 // use the postinc form of the IV. LSR needs to provide these formulae
3953 // as the basis of rediscovering the desired formula that uses an AddRec
3954 // corresponding to the existing phi. Once all formulae have been
3955 // generated, these initial losers may be pruned.
3956 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3960 SmallVector<const SCEV *, 4> Key;
3961 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3962 JE = F.BaseRegs.end(); J != JE; ++J) {
3963 const SCEV *Reg = *J;
3964 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3968 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3969 Key.push_back(F.ScaledReg);
3970 // Unstable sort by host order ok, because this is only used for
3972 std::sort(Key.begin(), Key.end());
3974 std::pair<BestFormulaeTy::const_iterator, bool> P =
3975 BestFormulae.insert(std::make_pair(Key, FIdx));
3979 Formula &Best = LU.Formulae[P.first->second];
3983 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3985 if (CostF < CostBest)
3987 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3989 " in favor of formula "; Best.print(dbgs());
3993 ChangedFormulae = true;
3995 LU.DeleteFormula(F);
4001 // Now that we've filtered out some formulae, recompute the Regs set.
4003 LU.RecomputeRegs(LUIdx, RegUses);
4005 // Reset this to prepare for the next use.
4006 BestFormulae.clear();
4009 DEBUG(if (ChangedFormulae) {
4011 "After filtering out undesirable candidates:\n";
4016 // This is a rough guess that seems to work fairly well.
4017 static const size_t ComplexityLimit = UINT16_MAX;
4019 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
4020 /// solutions the solver might have to consider. It almost never considers
4021 /// this many solutions because it prune the search space, but the pruning
4022 /// isn't always sufficient.
4023 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4025 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4026 E = Uses.end(); I != E; ++I) {
4027 size_t FSize = I->Formulae.size();
4028 if (FSize >= ComplexityLimit) {
4029 Power = ComplexityLimit;
4033 if (Power >= ComplexityLimit)
4039 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
4040 /// of the registers of another formula, it won't help reduce register
4041 /// pressure (though it may not necessarily hurt register pressure); remove
4042 /// it to simplify the system.
4043 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4044 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4045 DEBUG(dbgs() << "The search space is too complex.\n");
4047 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4048 "which use a superset of registers used by other "
4051 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4052 LSRUse &LU = Uses[LUIdx];
4054 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4055 Formula &F = LU.Formulae[i];
4056 // Look for a formula with a constant or GV in a register. If the use
4057 // also has a formula with that same value in an immediate field,
4058 // delete the one that uses a register.
4059 for (SmallVectorImpl<const SCEV *>::const_iterator
4060 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4061 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4063 NewF.BaseOffset += C->getValue()->getSExtValue();
4064 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4065 (I - F.BaseRegs.begin()));
4066 if (LU.HasFormulaWithSameRegs(NewF)) {
4067 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4068 LU.DeleteFormula(F);
4074 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4075 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4079 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4080 (I - F.BaseRegs.begin()));
4081 if (LU.HasFormulaWithSameRegs(NewF)) {
4082 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4084 LU.DeleteFormula(F);
4095 LU.RecomputeRegs(LUIdx, RegUses);
4098 DEBUG(dbgs() << "After pre-selection:\n";
4099 print_uses(dbgs()));
4103 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
4104 /// for expressions like A, A+1, A+2, etc., allocate a single register for
4106 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4107 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4110 DEBUG(dbgs() << "The search space is too complex.\n"
4111 "Narrowing the search space by assuming that uses separated "
4112 "by a constant offset will use the same registers.\n");
4114 // This is especially useful for unrolled loops.
4116 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4117 LSRUse &LU = Uses[LUIdx];
4118 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4119 E = LU.Formulae.end(); I != E; ++I) {
4120 const Formula &F = *I;
4121 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4124 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4128 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4129 LU.Kind, LU.AccessTy))
4132 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4134 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4136 // Update the relocs to reference the new use.
4137 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
4138 E = Fixups.end(); I != E; ++I) {
4139 LSRFixup &Fixup = *I;
4140 if (Fixup.LUIdx == LUIdx) {
4141 Fixup.LUIdx = LUThatHas - &Uses.front();
4142 Fixup.Offset += F.BaseOffset;
4143 // Add the new offset to LUThatHas' offset list.
4144 if (LUThatHas->Offsets.back() != Fixup.Offset) {
4145 LUThatHas->Offsets.push_back(Fixup.Offset);
4146 if (Fixup.Offset > LUThatHas->MaxOffset)
4147 LUThatHas->MaxOffset = Fixup.Offset;
4148 if (Fixup.Offset < LUThatHas->MinOffset)
4149 LUThatHas->MinOffset = Fixup.Offset;
4151 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4153 if (Fixup.LUIdx == NumUses-1)
4154 Fixup.LUIdx = LUIdx;
4157 // Delete formulae from the new use which are no longer legal.
4159 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4160 Formula &F = LUThatHas->Formulae[i];
4161 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4162 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4163 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4165 LUThatHas->DeleteFormula(F);
4173 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4175 // Delete the old use.
4176 DeleteUse(LU, LUIdx);
4183 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4186 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4187 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4188 /// we've done more filtering, as it may be able to find more formulae to
4190 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4191 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4192 DEBUG(dbgs() << "The search space is too complex.\n");
4194 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4195 "undesirable dedicated registers.\n");
4197 FilterOutUndesirableDedicatedRegisters();
4199 DEBUG(dbgs() << "After pre-selection:\n";
4200 print_uses(dbgs()));
4204 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4205 /// to be profitable, and then in any use which has any reference to that
4206 /// register, delete all formulae which do not reference that register.
4207 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4208 // With all other options exhausted, loop until the system is simple
4209 // enough to handle.
4210 SmallPtrSet<const SCEV *, 4> Taken;
4211 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4212 // Ok, we have too many of formulae on our hands to conveniently handle.
4213 // Use a rough heuristic to thin out the list.
4214 DEBUG(dbgs() << "The search space is too complex.\n");
4216 // Pick the register which is used by the most LSRUses, which is likely
4217 // to be a good reuse register candidate.
4218 const SCEV *Best = nullptr;
4219 unsigned BestNum = 0;
4220 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4222 const SCEV *Reg = *I;
4223 if (Taken.count(Reg))
4228 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4229 if (Count > BestNum) {
4236 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4237 << " will yield profitable reuse.\n");
4240 // In any use with formulae which references this register, delete formulae
4241 // which don't reference it.
4242 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4243 LSRUse &LU = Uses[LUIdx];
4244 if (!LU.Regs.count(Best)) continue;
4247 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4248 Formula &F = LU.Formulae[i];
4249 if (!F.referencesReg(Best)) {
4250 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4251 LU.DeleteFormula(F);
4255 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4261 LU.RecomputeRegs(LUIdx, RegUses);
4264 DEBUG(dbgs() << "After pre-selection:\n";
4265 print_uses(dbgs()));
4269 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4270 /// formulae to choose from, use some rough heuristics to prune down the number
4271 /// of formulae. This keeps the main solver from taking an extraordinary amount
4272 /// of time in some worst-case scenarios.
4273 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4274 NarrowSearchSpaceByDetectingSupersets();
4275 NarrowSearchSpaceByCollapsingUnrolledCode();
4276 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4277 NarrowSearchSpaceByPickingWinnerRegs();
4280 /// SolveRecurse - This is the recursive solver.
4281 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4283 SmallVectorImpl<const Formula *> &Workspace,
4284 const Cost &CurCost,
4285 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4286 DenseSet<const SCEV *> &VisitedRegs) const {
4289 // - use more aggressive filtering
4290 // - sort the formula so that the most profitable solutions are found first
4291 // - sort the uses too
4293 // - don't compute a cost, and then compare. compare while computing a cost
4295 // - track register sets with SmallBitVector
4297 const LSRUse &LU = Uses[Workspace.size()];
4299 // If this use references any register that's already a part of the
4300 // in-progress solution, consider it a requirement that a formula must
4301 // reference that register in order to be considered. This prunes out
4302 // unprofitable searching.
4303 SmallSetVector<const SCEV *, 4> ReqRegs;
4304 for (const SCEV *S : CurRegs)
4305 if (LU.Regs.count(S))
4308 SmallPtrSet<const SCEV *, 16> NewRegs;
4310 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4311 E = LU.Formulae.end(); I != E; ++I) {
4312 const Formula &F = *I;
4314 // Ignore formulae which may not be ideal in terms of register reuse of
4315 // ReqRegs. The formula should use all required registers before
4316 // introducing new ones.
4317 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4318 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4319 JE = ReqRegs.end(); J != JE; ++J) {
4320 const SCEV *Reg = *J;
4321 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4322 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) !=
4325 if (NumReqRegsToFind == 0)
4329 if (NumReqRegsToFind != 0) {
4330 // If none of the formulae satisfied the required registers, then we could
4331 // clear ReqRegs and try again. Currently, we simply give up in this case.
4335 // Evaluate the cost of the current formula. If it's already worse than
4336 // the current best, prune the search at that point.
4339 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4341 if (NewCost < SolutionCost) {
4342 Workspace.push_back(&F);
4343 if (Workspace.size() != Uses.size()) {
4344 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4345 NewRegs, VisitedRegs);
4346 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4347 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4349 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4350 dbgs() << ".\n Regs:";
4351 for (const SCEV *S : NewRegs)
4352 dbgs() << ' ' << *S;
4355 SolutionCost = NewCost;
4356 Solution = Workspace;
4358 Workspace.pop_back();
4363 /// Solve - Choose one formula from each use. Return the results in the given
4364 /// Solution vector.
4365 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4366 SmallVector<const Formula *, 8> Workspace;
4368 SolutionCost.Lose();
4370 SmallPtrSet<const SCEV *, 16> CurRegs;
4371 DenseSet<const SCEV *> VisitedRegs;
4372 Workspace.reserve(Uses.size());
4374 // SolveRecurse does all the work.
4375 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4376 CurRegs, VisitedRegs);
4377 if (Solution.empty()) {
4378 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4382 // Ok, we've now made all our decisions.
4383 DEBUG(dbgs() << "\n"
4384 "The chosen solution requires "; SolutionCost.print(dbgs());
4386 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4388 Uses[i].print(dbgs());
4391 Solution[i]->print(dbgs());
4395 assert(Solution.size() == Uses.size() && "Malformed solution!");
4398 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4399 /// the dominator tree far as we can go while still being dominated by the
4400 /// input positions. This helps canonicalize the insert position, which
4401 /// encourages sharing.
4402 BasicBlock::iterator
4403 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4404 const SmallVectorImpl<Instruction *> &Inputs)
4407 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4408 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4411 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4412 if (!Rung) return IP;
4413 Rung = Rung->getIDom();
4414 if (!Rung) return IP;
4415 IDom = Rung->getBlock();
4417 // Don't climb into a loop though.
4418 const Loop *IDomLoop = LI.getLoopFor(IDom);
4419 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4420 if (IDomDepth <= IPLoopDepth &&
4421 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4425 bool AllDominate = true;
4426 Instruction *BetterPos = nullptr;
4427 Instruction *Tentative = IDom->getTerminator();
4428 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4429 E = Inputs.end(); I != E; ++I) {
4430 Instruction *Inst = *I;
4431 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4432 AllDominate = false;
4435 // Attempt to find an insert position in the middle of the block,
4436 // instead of at the end, so that it can be used for other expansions.
4437 if (IDom == Inst->getParent() &&
4438 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4439 BetterPos = std::next(BasicBlock::iterator(Inst));
4452 /// AdjustInsertPositionForExpand - Determine an input position which will be
4453 /// dominated by the operands and which will dominate the result.
4454 BasicBlock::iterator
4455 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4458 SCEVExpander &Rewriter) const {
4459 // Collect some instructions which must be dominated by the
4460 // expanding replacement. These must be dominated by any operands that
4461 // will be required in the expansion.
4462 SmallVector<Instruction *, 4> Inputs;
4463 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4464 Inputs.push_back(I);
4465 if (LU.Kind == LSRUse::ICmpZero)
4466 if (Instruction *I =
4467 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4468 Inputs.push_back(I);
4469 if (LF.PostIncLoops.count(L)) {
4470 if (LF.isUseFullyOutsideLoop(L))
4471 Inputs.push_back(L->getLoopLatch()->getTerminator());
4473 Inputs.push_back(IVIncInsertPos);
4475 // The expansion must also be dominated by the increment positions of any
4476 // loops it for which it is using post-inc mode.
4477 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4478 E = LF.PostIncLoops.end(); I != E; ++I) {
4479 const Loop *PIL = *I;
4480 if (PIL == L) continue;
4482 // Be dominated by the loop exit.
4483 SmallVector<BasicBlock *, 4> ExitingBlocks;
4484 PIL->getExitingBlocks(ExitingBlocks);
4485 if (!ExitingBlocks.empty()) {
4486 BasicBlock *BB = ExitingBlocks[0];
4487 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4488 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4489 Inputs.push_back(BB->getTerminator());
4493 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4494 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4495 "Insertion point must be a normal instruction");
4497 // Then, climb up the immediate dominator tree as far as we can go while
4498 // still being dominated by the input positions.
4499 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4501 // Don't insert instructions before PHI nodes.
4502 while (isa<PHINode>(IP)) ++IP;
4504 // Ignore landingpad instructions.
4505 while (isa<LandingPadInst>(IP)) ++IP;
4507 // Ignore debug intrinsics.
4508 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4510 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4511 // IP consistent across expansions and allows the previously inserted
4512 // instructions to be reused by subsequent expansion.
4513 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4518 /// Expand - Emit instructions for the leading candidate expression for this
4519 /// LSRUse (this is called "expanding").
4520 Value *LSRInstance::Expand(const LSRFixup &LF,
4522 BasicBlock::iterator IP,
4523 SCEVExpander &Rewriter,
4524 SmallVectorImpl<WeakVH> &DeadInsts) const {
4525 const LSRUse &LU = Uses[LF.LUIdx];
4526 if (LU.RigidFormula)
4527 return LF.OperandValToReplace;
4529 // Determine an input position which will be dominated by the operands and
4530 // which will dominate the result.
4531 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4533 // Inform the Rewriter if we have a post-increment use, so that it can
4534 // perform an advantageous expansion.
4535 Rewriter.setPostInc(LF.PostIncLoops);
4537 // This is the type that the user actually needs.
4538 Type *OpTy = LF.OperandValToReplace->getType();
4539 // This will be the type that we'll initially expand to.
4540 Type *Ty = F.getType();
4542 // No type known; just expand directly to the ultimate type.
4544 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4545 // Expand directly to the ultimate type if it's the right size.
4547 // This is the type to do integer arithmetic in.
4548 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4550 // Build up a list of operands to add together to form the full base.
4551 SmallVector<const SCEV *, 8> Ops;
4553 // Expand the BaseRegs portion.
4554 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4555 E = F.BaseRegs.end(); I != E; ++I) {
4556 const SCEV *Reg = *I;
4557 assert(!Reg->isZero() && "Zero allocated in a base register!");
4559 // If we're expanding for a post-inc user, make the post-inc adjustment.
4560 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4561 Reg = TransformForPostIncUse(Denormalize, Reg,
4562 LF.UserInst, LF.OperandValToReplace,
4565 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr, IP)));
4568 // Expand the ScaledReg portion.
4569 Value *ICmpScaledV = nullptr;
4571 const SCEV *ScaledS = F.ScaledReg;
4573 // If we're expanding for a post-inc user, make the post-inc adjustment.
4574 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4575 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4576 LF.UserInst, LF.OperandValToReplace,
4579 if (LU.Kind == LSRUse::ICmpZero) {
4580 // Expand ScaleReg as if it was part of the base regs.
4583 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, IP)));
4585 // An interesting way of "folding" with an icmp is to use a negated
4586 // scale, which we'll implement by inserting it into the other operand
4588 assert(F.Scale == -1 &&
4589 "The only scale supported by ICmpZero uses is -1!");
4590 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr, IP);
4593 // Otherwise just expand the scaled register and an explicit scale,
4594 // which is expected to be matched as part of the address.
4596 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4597 // Unless the addressing mode will not be folded.
4598 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4599 isAMCompletelyFolded(TTI, LU, F)) {
4600 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4602 Ops.push_back(SE.getUnknown(FullV));
4604 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, IP));
4607 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4608 Ops.push_back(ScaledS);
4612 // Expand the GV portion.
4614 // Flush the operand list to suppress SCEVExpander hoisting.
4616 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4618 Ops.push_back(SE.getUnknown(FullV));
4620 Ops.push_back(SE.getUnknown(F.BaseGV));
4623 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4624 // unfolded offsets. LSR assumes they both live next to their uses.
4626 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4628 Ops.push_back(SE.getUnknown(FullV));
4631 // Expand the immediate portion.
4632 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4634 if (LU.Kind == LSRUse::ICmpZero) {
4635 // The other interesting way of "folding" with an ICmpZero is to use a
4636 // negated immediate.
4638 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4640 Ops.push_back(SE.getUnknown(ICmpScaledV));
4641 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4644 // Just add the immediate values. These again are expected to be matched
4645 // as part of the address.
4646 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4650 // Expand the unfolded offset portion.
4651 int64_t UnfoldedOffset = F.UnfoldedOffset;
4652 if (UnfoldedOffset != 0) {
4653 // Just add the immediate values.
4654 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4658 // Emit instructions summing all the operands.
4659 const SCEV *FullS = Ops.empty() ?
4660 SE.getConstant(IntTy, 0) :
4662 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4664 // We're done expanding now, so reset the rewriter.
4665 Rewriter.clearPostInc();
4667 // An ICmpZero Formula represents an ICmp which we're handling as a
4668 // comparison against zero. Now that we've expanded an expression for that
4669 // form, update the ICmp's other operand.
4670 if (LU.Kind == LSRUse::ICmpZero) {
4671 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4672 DeadInsts.push_back(CI->getOperand(1));
4673 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4674 "a scale at the same time!");
4675 if (F.Scale == -1) {
4676 if (ICmpScaledV->getType() != OpTy) {
4678 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4680 ICmpScaledV, OpTy, "tmp", CI);
4683 CI->setOperand(1, ICmpScaledV);
4685 // A scale of 1 means that the scale has been expanded as part of the
4687 assert((F.Scale == 0 || F.Scale == 1) &&
4688 "ICmp does not support folding a global value and "
4689 "a scale at the same time!");
4690 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4692 if (C->getType() != OpTy)
4693 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4697 CI->setOperand(1, C);
4704 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4705 /// of their operands effectively happens in their predecessor blocks, so the
4706 /// expression may need to be expanded in multiple places.
4707 void LSRInstance::RewriteForPHI(PHINode *PN,
4710 SCEVExpander &Rewriter,
4711 SmallVectorImpl<WeakVH> &DeadInsts,
4713 DenseMap<BasicBlock *, Value *> Inserted;
4714 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4715 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4716 BasicBlock *BB = PN->getIncomingBlock(i);
4718 // If this is a critical edge, split the edge so that we do not insert
4719 // the code on all predecessor/successor paths. We do this unless this
4720 // is the canonical backedge for this loop, which complicates post-inc
4722 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4723 !isa<IndirectBrInst>(BB->getTerminator())) {
4724 BasicBlock *Parent = PN->getParent();
4725 Loop *PNLoop = LI.getLoopFor(Parent);
4726 if (!PNLoop || Parent != PNLoop->getHeader()) {
4727 // Split the critical edge.
4728 BasicBlock *NewBB = nullptr;
4729 if (!Parent->isLandingPad()) {
4730 NewBB = SplitCriticalEdge(BB, Parent,
4731 CriticalEdgeSplittingOptions(&DT, &LI)
4732 .setMergeIdenticalEdges()
4733 .setDontDeleteUselessPHIs());
4735 SmallVector<BasicBlock*, 2> NewBBs;
4736 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs,
4737 /*AliasAnalysis*/ nullptr, &DT, &LI);
4740 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4741 // phi predecessors are identical. The simple thing to do is skip
4742 // splitting in this case rather than complicate the API.
4744 // If PN is outside of the loop and BB is in the loop, we want to
4745 // move the block to be immediately before the PHI block, not
4746 // immediately after BB.
4747 if (L->contains(BB) && !L->contains(PN))
4748 NewBB->moveBefore(PN->getParent());
4750 // Splitting the edge can reduce the number of PHI entries we have.
4751 e = PN->getNumIncomingValues();
4753 i = PN->getBasicBlockIndex(BB);
4758 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4759 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
4761 PN->setIncomingValue(i, Pair.first->second);
4763 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4765 // If this is reuse-by-noop-cast, insert the noop cast.
4766 Type *OpTy = LF.OperandValToReplace->getType();
4767 if (FullV->getType() != OpTy)
4769 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4771 FullV, LF.OperandValToReplace->getType(),
4772 "tmp", BB->getTerminator());
4774 PN->setIncomingValue(i, FullV);
4775 Pair.first->second = FullV;
4780 /// Rewrite - Emit instructions for the leading candidate expression for this
4781 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4782 /// the newly expanded value.
4783 void LSRInstance::Rewrite(const LSRFixup &LF,
4785 SCEVExpander &Rewriter,
4786 SmallVectorImpl<WeakVH> &DeadInsts,
4788 // First, find an insertion point that dominates UserInst. For PHI nodes,
4789 // find the nearest block which dominates all the relevant uses.
4790 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4791 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4793 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4795 // If this is reuse-by-noop-cast, insert the noop cast.
4796 Type *OpTy = LF.OperandValToReplace->getType();
4797 if (FullV->getType() != OpTy) {
4799 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4800 FullV, OpTy, "tmp", LF.UserInst);
4804 // Update the user. ICmpZero is handled specially here (for now) because
4805 // Expand may have updated one of the operands of the icmp already, and
4806 // its new value may happen to be equal to LF.OperandValToReplace, in
4807 // which case doing replaceUsesOfWith leads to replacing both operands
4808 // with the same value. TODO: Reorganize this.
4809 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4810 LF.UserInst->setOperand(0, FullV);
4812 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4815 DeadInsts.push_back(LF.OperandValToReplace);
4818 /// ImplementSolution - Rewrite all the fixup locations with new values,
4819 /// following the chosen solution.
4821 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4823 // Keep track of instructions we may have made dead, so that
4824 // we can remove them after we are done working.
4825 SmallVector<WeakVH, 16> DeadInsts;
4827 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
4830 Rewriter.setDebugType(DEBUG_TYPE);
4832 Rewriter.disableCanonicalMode();
4833 Rewriter.enableLSRMode();
4834 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4836 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4837 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4838 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4839 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4840 Rewriter.setChainedPhi(PN);
4843 // Expand the new value definitions and update the users.
4844 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4845 E = Fixups.end(); I != E; ++I) {
4846 const LSRFixup &Fixup = *I;
4848 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4853 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4854 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4855 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4858 // Clean up after ourselves. This must be done before deleting any
4862 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4865 LSRInstance::LSRInstance(Loop *L, Pass *P)
4866 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4867 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4868 LI(P->getAnalysis<LoopInfoWrapperPass>().getLoopInfo()),
4869 TTI(P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
4870 *L->getHeader()->getParent())),
4871 L(L), Changed(false), IVIncInsertPos(nullptr) {
4872 // If LoopSimplify form is not available, stay out of trouble.
4873 if (!L->isLoopSimplifyForm())
4876 // If there's no interesting work to be done, bail early.
4877 if (IU.empty()) return;
4879 // If there's too much analysis to be done, bail early. We won't be able to
4880 // model the problem anyway.
4881 unsigned NumUsers = 0;
4882 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4883 if (++NumUsers > MaxIVUsers) {
4884 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4891 // All dominating loops must have preheaders, or SCEVExpander may not be able
4892 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4894 // IVUsers analysis should only create users that are dominated by simple loop
4895 // headers. Since this loop should dominate all of its users, its user list
4896 // should be empty if this loop itself is not within a simple loop nest.
4897 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4898 Rung; Rung = Rung->getIDom()) {
4899 BasicBlock *BB = Rung->getBlock();
4900 const Loop *DomLoop = LI.getLoopFor(BB);
4901 if (DomLoop && DomLoop->getHeader() == BB) {
4902 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4907 DEBUG(dbgs() << "\nLSR on loop ";
4908 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4911 // First, perform some low-level loop optimizations.
4913 OptimizeLoopTermCond();
4915 // If loop preparation eliminates all interesting IV users, bail.
4916 if (IU.empty()) return;
4918 // Skip nested loops until we can model them better with formulae.
4920 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4924 // Start collecting data and preparing for the solver.
4926 CollectInterestingTypesAndFactors();
4927 CollectFixupsAndInitialFormulae();
4928 CollectLoopInvariantFixupsAndFormulae();
4930 assert(!Uses.empty() && "IVUsers reported at least one use");
4931 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4932 print_uses(dbgs()));
4934 // Now use the reuse data to generate a bunch of interesting ways
4935 // to formulate the values needed for the uses.
4936 GenerateAllReuseFormulae();
4938 FilterOutUndesirableDedicatedRegisters();
4939 NarrowSearchSpaceUsingHeuristics();
4941 SmallVector<const Formula *, 8> Solution;
4944 // Release memory that is no longer needed.
4949 if (Solution.empty())
4953 // Formulae should be legal.
4954 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4956 const LSRUse &LU = *I;
4957 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4958 JE = LU.Formulae.end();
4960 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4961 *J) && "Illegal formula generated!");
4965 // Now that we've decided what we want, make it so.
4966 ImplementSolution(Solution, P);
4969 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4970 if (Factors.empty() && Types.empty()) return;
4972 OS << "LSR has identified the following interesting factors and types: ";
4975 for (SmallSetVector<int64_t, 8>::const_iterator
4976 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4977 if (!First) OS << ", ";
4982 for (SmallSetVector<Type *, 4>::const_iterator
4983 I = Types.begin(), E = Types.end(); I != E; ++I) {
4984 if (!First) OS << ", ";
4986 OS << '(' << **I << ')';
4991 void LSRInstance::print_fixups(raw_ostream &OS) const {
4992 OS << "LSR is examining the following fixup sites:\n";
4993 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4994 E = Fixups.end(); I != E; ++I) {
5001 void LSRInstance::print_uses(raw_ostream &OS) const {
5002 OS << "LSR is examining the following uses:\n";
5003 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
5004 E = Uses.end(); I != E; ++I) {
5005 const LSRUse &LU = *I;
5009 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
5010 JE = LU.Formulae.end(); J != JE; ++J) {
5018 void LSRInstance::print(raw_ostream &OS) const {
5019 print_factors_and_types(OS);
5024 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5025 void LSRInstance::dump() const {
5026 print(errs()); errs() << '\n';
5032 class LoopStrengthReduce : public LoopPass {
5034 static char ID; // Pass ID, replacement for typeid
5035 LoopStrengthReduce();
5038 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5039 void getAnalysisUsage(AnalysisUsage &AU) const override;
5044 char LoopStrengthReduce::ID = 0;
5045 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5046 "Loop Strength Reduction", false, false)
5047 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5048 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5049 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5050 INITIALIZE_PASS_DEPENDENCY(IVUsers)
5051 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5052 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5053 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5054 "Loop Strength Reduction", false, false)
5057 Pass *llvm::createLoopStrengthReducePass() {
5058 return new LoopStrengthReduce();
5061 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5062 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5065 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5066 // We split critical edges, so we change the CFG. However, we do update
5067 // many analyses if they are around.
5068 AU.addPreservedID(LoopSimplifyID);
5070 AU.addRequired<LoopInfoWrapperPass>();
5071 AU.addPreserved<LoopInfoWrapperPass>();
5072 AU.addRequiredID(LoopSimplifyID);
5073 AU.addRequired<DominatorTreeWrapperPass>();
5074 AU.addPreserved<DominatorTreeWrapperPass>();
5075 AU.addRequired<ScalarEvolution>();
5076 AU.addPreserved<ScalarEvolution>();
5077 // Requiring LoopSimplify a second time here prevents IVUsers from running
5078 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5079 AU.addRequiredID(LoopSimplifyID);
5080 AU.addRequired<IVUsers>();
5081 AU.addPreserved<IVUsers>();
5082 AU.addRequired<TargetTransformInfoWrapperPass>();
5085 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5086 if (skipOptnoneFunction(L))
5089 bool Changed = false;
5091 // Run the main LSR transformation.
5092 Changed |= LSRInstance(L, this).getChanged();
5094 // Remove any extra phis created by processing inner loops.
5095 Changed |= DeleteDeadPHIs(L->getHeader());
5096 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5097 SmallVector<WeakVH, 16> DeadInsts;
5098 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5099 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), DL, "lsr");
5101 Rewriter.setDebugType(DEBUG_TYPE);
5103 unsigned numFolded = Rewriter.replaceCongruentIVs(
5104 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
5105 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5106 *L->getHeader()->getParent()));
5109 DeleteTriviallyDeadInstructions(DeadInsts);
5110 DeleteDeadPHIs(L->getHeader());