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/ValueHandle.h"
72 #include "llvm/Support/CommandLine.h"
73 #include "llvm/Support/Debug.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
76 #include "llvm/Transforms/Utils/Local.h"
80 #define DEBUG_TYPE "loop-reduce"
82 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
83 /// bail out. This threshold is far beyond the number of users that LSR can
84 /// conceivably solve, so it should not affect generated code, but catches the
85 /// worst cases before LSR burns too much compile time and stack space.
86 static const unsigned MaxIVUsers = 200;
88 // Temporary flag to cleanup congruent phis after LSR phi expansion.
89 // It's currently disabled until we can determine whether it's truly useful or
90 // not. The flag should be removed after the v3.0 release.
91 // This is now needed for ivchains.
92 static cl::opt<bool> EnablePhiElim(
93 "enable-lsr-phielim", cl::Hidden, cl::init(true),
94 cl::desc("Enable LSR phi elimination"));
97 // Stress test IV chain generation.
98 static cl::opt<bool> StressIVChain(
99 "stress-ivchain", cl::Hidden, cl::init(false),
100 cl::desc("Stress test LSR IV chains"));
102 static bool StressIVChain = false;
107 /// RegSortData - This class holds data which is used to order reuse candidates.
110 /// UsedByIndices - This represents the set of LSRUse indices which reference
111 /// a particular register.
112 SmallBitVector UsedByIndices;
116 void print(raw_ostream &OS) const;
122 void RegSortData::print(raw_ostream &OS) const {
123 OS << "[NumUses=" << UsedByIndices.count() << ']';
126 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
127 void RegSortData::dump() const {
128 print(errs()); errs() << '\n';
134 /// RegUseTracker - Map register candidates to information about how they are
136 class RegUseTracker {
137 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
139 RegUsesTy RegUsesMap;
140 SmallVector<const SCEV *, 16> RegSequence;
143 void CountRegister(const SCEV *Reg, size_t LUIdx);
144 void DropRegister(const SCEV *Reg, size_t LUIdx);
145 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
147 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
149 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
153 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
154 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
155 iterator begin() { return RegSequence.begin(); }
156 iterator end() { return RegSequence.end(); }
157 const_iterator begin() const { return RegSequence.begin(); }
158 const_iterator end() const { return RegSequence.end(); }
164 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
165 std::pair<RegUsesTy::iterator, bool> Pair =
166 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
167 RegSortData &RSD = Pair.first->second;
169 RegSequence.push_back(Reg);
170 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
171 RSD.UsedByIndices.set(LUIdx);
175 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
176 RegUsesTy::iterator It = RegUsesMap.find(Reg);
177 assert(It != RegUsesMap.end());
178 RegSortData &RSD = It->second;
179 assert(RSD.UsedByIndices.size() > LUIdx);
180 RSD.UsedByIndices.reset(LUIdx);
184 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
185 assert(LUIdx <= LastLUIdx);
187 // Update RegUses. The data structure is not optimized for this purpose;
188 // we must iterate through it and update each of the bit vectors.
189 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
191 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
192 if (LUIdx < UsedByIndices.size())
193 UsedByIndices[LUIdx] =
194 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
195 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
200 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
201 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
202 if (I == RegUsesMap.end())
204 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
205 int i = UsedByIndices.find_first();
206 if (i == -1) return false;
207 if ((size_t)i != LUIdx) return true;
208 return UsedByIndices.find_next(i) != -1;
211 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
212 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
213 assert(I != RegUsesMap.end() && "Unknown register!");
214 return I->second.UsedByIndices;
217 void RegUseTracker::clear() {
224 /// Formula - This class holds information that describes a formula for
225 /// computing satisfying a use. It may include broken-out immediates and scaled
228 /// Global base address used for complex addressing.
231 /// Base offset for complex addressing.
234 /// Whether any complex addressing has a base register.
237 /// The scale of any complex addressing.
240 /// BaseRegs - The list of "base" registers for this use. When this is
242 SmallVector<const SCEV *, 4> BaseRegs;
244 /// ScaledReg - The 'scaled' register for this use. This should be non-null
245 /// when Scale is not zero.
246 const SCEV *ScaledReg;
248 /// UnfoldedOffset - An additional constant offset which added near the
249 /// use. This requires a temporary register, but the offset itself can
250 /// live in an add immediate field rather than a register.
251 int64_t UnfoldedOffset;
254 : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
255 ScaledReg(nullptr), UnfoldedOffset(0) {}
257 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
259 size_t getNumRegs() const;
260 Type *getType() const;
262 void DeleteBaseReg(const SCEV *&S);
264 bool referencesReg(const SCEV *S) const;
265 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
266 const RegUseTracker &RegUses) const;
268 void print(raw_ostream &OS) const;
274 /// DoInitialMatch - Recursion helper for InitialMatch.
275 static void DoInitialMatch(const SCEV *S, Loop *L,
276 SmallVectorImpl<const SCEV *> &Good,
277 SmallVectorImpl<const SCEV *> &Bad,
278 ScalarEvolution &SE) {
279 // Collect expressions which properly dominate the loop header.
280 if (SE.properlyDominates(S, L->getHeader())) {
285 // Look at add operands.
286 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
287 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
289 DoInitialMatch(*I, L, Good, Bad, SE);
293 // Look at addrec operands.
294 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
295 if (!AR->getStart()->isZero()) {
296 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
297 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
298 AR->getStepRecurrence(SE),
299 // FIXME: AR->getNoWrapFlags()
300 AR->getLoop(), SCEV::FlagAnyWrap),
305 // Handle a multiplication by -1 (negation) if it didn't fold.
306 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
307 if (Mul->getOperand(0)->isAllOnesValue()) {
308 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
309 const SCEV *NewMul = SE.getMulExpr(Ops);
311 SmallVector<const SCEV *, 4> MyGood;
312 SmallVector<const SCEV *, 4> MyBad;
313 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
314 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
315 SE.getEffectiveSCEVType(NewMul->getType())));
316 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
317 E = MyGood.end(); I != E; ++I)
318 Good.push_back(SE.getMulExpr(NegOne, *I));
319 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
320 E = MyBad.end(); I != E; ++I)
321 Bad.push_back(SE.getMulExpr(NegOne, *I));
325 // Ok, we can't do anything interesting. Just stuff the whole thing into a
326 // register and hope for the best.
330 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
331 /// attempting to keep all loop-invariant and loop-computable values in a
332 /// single base register.
333 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
334 SmallVector<const SCEV *, 4> Good;
335 SmallVector<const SCEV *, 4> Bad;
336 DoInitialMatch(S, L, Good, Bad, SE);
338 const SCEV *Sum = SE.getAddExpr(Good);
340 BaseRegs.push_back(Sum);
344 const SCEV *Sum = SE.getAddExpr(Bad);
346 BaseRegs.push_back(Sum);
351 /// getNumRegs - Return the total number of register operands used by this
352 /// formula. This does not include register uses implied by non-constant
354 size_t Formula::getNumRegs() const {
355 return !!ScaledReg + BaseRegs.size();
358 /// getType - Return the type of this formula, if it has one, or null
359 /// otherwise. This type is meaningless except for the bit size.
360 Type *Formula::getType() const {
361 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
362 ScaledReg ? ScaledReg->getType() :
363 BaseGV ? BaseGV->getType() :
367 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
368 void Formula::DeleteBaseReg(const SCEV *&S) {
369 if (&S != &BaseRegs.back())
370 std::swap(S, BaseRegs.back());
374 /// referencesReg - Test if this formula references the given register.
375 bool Formula::referencesReg(const SCEV *S) const {
376 return S == ScaledReg ||
377 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
380 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
381 /// which are used by uses other than the use with the given index.
382 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
383 const RegUseTracker &RegUses) const {
385 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
387 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
388 E = BaseRegs.end(); I != E; ++I)
389 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
394 void Formula::print(raw_ostream &OS) const {
397 if (!First) OS << " + "; else First = false;
398 BaseGV->printAsOperand(OS, /*PrintType=*/false);
400 if (BaseOffset != 0) {
401 if (!First) OS << " + "; else First = false;
404 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
405 E = BaseRegs.end(); I != E; ++I) {
406 if (!First) OS << " + "; else First = false;
407 OS << "reg(" << **I << ')';
409 if (HasBaseReg && BaseRegs.empty()) {
410 if (!First) OS << " + "; else First = false;
411 OS << "**error: HasBaseReg**";
412 } else if (!HasBaseReg && !BaseRegs.empty()) {
413 if (!First) OS << " + "; else First = false;
414 OS << "**error: !HasBaseReg**";
417 if (!First) OS << " + "; else First = false;
418 OS << Scale << "*reg(";
425 if (UnfoldedOffset != 0) {
426 if (!First) OS << " + ";
427 OS << "imm(" << UnfoldedOffset << ')';
431 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
432 void Formula::dump() const {
433 print(errs()); errs() << '\n';
437 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
438 /// without changing its value.
439 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
441 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
442 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
445 /// isAddSExtable - Return true if the given add can be sign-extended
446 /// without changing its value.
447 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
449 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
450 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
453 /// isMulSExtable - Return true if the given mul can be sign-extended
454 /// without changing its value.
455 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
457 IntegerType::get(SE.getContext(),
458 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
459 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
462 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
463 /// and if the remainder is known to be zero, or null otherwise. If
464 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
465 /// to Y, ignoring that the multiplication may overflow, which is useful when
466 /// the result will be used in a context where the most significant bits are
468 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
470 bool IgnoreSignificantBits = false) {
471 // Handle the trivial case, which works for any SCEV type.
473 return SE.getConstant(LHS->getType(), 1);
475 // Handle a few RHS special cases.
476 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
478 const APInt &RA = RC->getValue()->getValue();
479 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
481 if (RA.isAllOnesValue())
482 return SE.getMulExpr(LHS, RC);
483 // Handle x /s 1 as x.
488 // Check for a division of a constant by a constant.
489 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
492 const APInt &LA = C->getValue()->getValue();
493 const APInt &RA = RC->getValue()->getValue();
494 if (LA.srem(RA) != 0)
496 return SE.getConstant(LA.sdiv(RA));
499 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
500 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
501 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
502 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
503 IgnoreSignificantBits);
504 if (!Step) return nullptr;
505 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
506 IgnoreSignificantBits);
507 if (!Start) return nullptr;
508 // FlagNW is independent of the start value, step direction, and is
509 // preserved with smaller magnitude steps.
510 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
511 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
516 // Distribute the sdiv over add operands, if the add doesn't overflow.
517 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
518 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
519 SmallVector<const SCEV *, 8> Ops;
520 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
522 const SCEV *Op = getExactSDiv(*I, RHS, SE,
523 IgnoreSignificantBits);
524 if (!Op) return nullptr;
527 return SE.getAddExpr(Ops);
532 // Check for a multiply operand that we can pull RHS out of.
533 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
534 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
535 SmallVector<const SCEV *, 4> Ops;
537 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
541 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
542 IgnoreSignificantBits)) {
548 return Found ? SE.getMulExpr(Ops) : nullptr;
553 // Otherwise we don't know.
557 /// ExtractImmediate - If S involves the addition of a constant integer value,
558 /// return that integer value, and mutate S to point to a new SCEV with that
560 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
561 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
562 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
563 S = SE.getConstant(C->getType(), 0);
564 return C->getValue()->getSExtValue();
566 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
567 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
568 int64_t Result = ExtractImmediate(NewOps.front(), SE);
570 S = SE.getAddExpr(NewOps);
572 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
573 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
574 int64_t Result = ExtractImmediate(NewOps.front(), SE);
576 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
577 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
584 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
585 /// return that symbol, and mutate S to point to a new SCEV with that
587 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
588 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
589 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
590 S = SE.getConstant(GV->getType(), 0);
593 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
594 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
595 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
597 S = SE.getAddExpr(NewOps);
599 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
600 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
601 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
603 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
604 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
611 /// isAddressUse - Returns true if the specified instruction is using the
612 /// specified value as an address.
613 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
614 bool isAddress = isa<LoadInst>(Inst);
615 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
616 if (SI->getOperand(1) == OperandVal)
618 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
619 // Addressing modes can also be folded into prefetches and a variety
621 switch (II->getIntrinsicID()) {
623 case Intrinsic::prefetch:
624 case Intrinsic::x86_sse_storeu_ps:
625 case Intrinsic::x86_sse2_storeu_pd:
626 case Intrinsic::x86_sse2_storeu_dq:
627 case Intrinsic::x86_sse2_storel_dq:
628 if (II->getArgOperand(0) == OperandVal)
636 /// getAccessType - Return the type of the memory being accessed.
637 static Type *getAccessType(const Instruction *Inst) {
638 Type *AccessTy = Inst->getType();
639 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
640 AccessTy = SI->getOperand(0)->getType();
641 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
642 // Addressing modes can also be folded into prefetches and a variety
644 switch (II->getIntrinsicID()) {
646 case Intrinsic::x86_sse_storeu_ps:
647 case Intrinsic::x86_sse2_storeu_pd:
648 case Intrinsic::x86_sse2_storeu_dq:
649 case Intrinsic::x86_sse2_storel_dq:
650 AccessTy = II->getArgOperand(0)->getType();
655 // All pointers have the same requirements, so canonicalize them to an
656 // arbitrary pointer type to minimize variation.
657 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
658 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
659 PTy->getAddressSpace());
664 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
665 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
666 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
667 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
668 if (SE.isSCEVable(PN->getType()) &&
669 (SE.getEffectiveSCEVType(PN->getType()) ==
670 SE.getEffectiveSCEVType(AR->getType())) &&
671 SE.getSCEV(PN) == AR)
677 /// Check if expanding this expression is likely to incur significant cost. This
678 /// is tricky because SCEV doesn't track which expressions are actually computed
679 /// by the current IR.
681 /// We currently allow expansion of IV increments that involve adds,
682 /// multiplication by constants, and AddRecs from existing phis.
684 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
685 /// obvious multiple of the UDivExpr.
686 static bool isHighCostExpansion(const SCEV *S,
687 SmallPtrSet<const SCEV*, 8> &Processed,
688 ScalarEvolution &SE) {
689 // Zero/One operand expressions
690 switch (S->getSCEVType()) {
695 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
698 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
701 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
705 if (!Processed.insert(S))
708 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
709 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
711 if (isHighCostExpansion(*I, Processed, SE))
717 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
718 if (Mul->getNumOperands() == 2) {
719 // Multiplication by a constant is ok
720 if (isa<SCEVConstant>(Mul->getOperand(0)))
721 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
723 // If we have the value of one operand, check if an existing
724 // multiplication already generates this expression.
725 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
726 Value *UVal = U->getValue();
727 for (User *UR : UVal->users()) {
728 // If U is a constant, it may be used by a ConstantExpr.
729 Instruction *UI = dyn_cast<Instruction>(UR);
730 if (UI && UI->getOpcode() == Instruction::Mul &&
731 SE.isSCEVable(UI->getType())) {
732 return SE.getSCEV(UI) == Mul;
739 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
740 if (isExistingPhi(AR, SE))
744 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
748 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
749 /// specified set are trivially dead, delete them and see if this makes any of
750 /// their operands subsequently dead.
752 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
753 bool Changed = false;
755 while (!DeadInsts.empty()) {
756 Value *V = DeadInsts.pop_back_val();
757 Instruction *I = dyn_cast_or_null<Instruction>(V);
759 if (!I || !isInstructionTriviallyDead(I))
762 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
763 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
766 DeadInsts.push_back(U);
769 I->eraseFromParent();
779 // Check if it is legal to fold 2 base registers.
780 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
782 // Get the cost of the scaling factor used in F for LU.
783 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
784 const LSRUse &LU, const Formula &F);
788 /// Cost - This class is used to measure and compare candidate formulae.
790 /// TODO: Some of these could be merged. Also, a lexical ordering
791 /// isn't always optimal.
795 unsigned NumBaseAdds;
802 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
803 SetupCost(0), ScaleCost(0) {}
805 bool operator<(const Cost &Other) const;
810 // Once any of the metrics loses, they must all remain losers.
812 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
813 | ImmCost | SetupCost | ScaleCost) != ~0u)
814 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
815 & ImmCost & SetupCost & ScaleCost) == ~0u);
820 assert(isValid() && "invalid cost");
821 return NumRegs == ~0u;
824 void RateFormula(const TargetTransformInfo &TTI,
826 SmallPtrSet<const SCEV *, 16> &Regs,
827 const DenseSet<const SCEV *> &VisitedRegs,
829 const SmallVectorImpl<int64_t> &Offsets,
830 ScalarEvolution &SE, DominatorTree &DT,
832 SmallPtrSet<const SCEV *, 16> *LoserRegs = nullptr);
834 void print(raw_ostream &OS) const;
838 void RateRegister(const SCEV *Reg,
839 SmallPtrSet<const SCEV *, 16> &Regs,
841 ScalarEvolution &SE, DominatorTree &DT);
842 void RatePrimaryRegister(const SCEV *Reg,
843 SmallPtrSet<const SCEV *, 16> &Regs,
845 ScalarEvolution &SE, DominatorTree &DT,
846 SmallPtrSet<const SCEV *, 16> *LoserRegs);
851 /// RateRegister - Tally up interesting quantities from the given register.
852 void Cost::RateRegister(const SCEV *Reg,
853 SmallPtrSet<const SCEV *, 16> &Regs,
855 ScalarEvolution &SE, DominatorTree &DT) {
856 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
857 // If this is an addrec for another loop, don't second-guess its addrec phi
858 // nodes. LSR isn't currently smart enough to reason about more than one
859 // loop at a time. LSR has already run on inner loops, will not run on outer
860 // loops, and cannot be expected to change sibling loops.
861 if (AR->getLoop() != L) {
862 // If the AddRec exists, consider it's register free and leave it alone.
863 if (isExistingPhi(AR, SE))
866 // Otherwise, do not consider this formula at all.
870 AddRecCost += 1; /// TODO: This should be a function of the stride.
872 // Add the step value register, if it needs one.
873 // TODO: The non-affine case isn't precisely modeled here.
874 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
875 if (!Regs.count(AR->getOperand(1))) {
876 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
884 // Rough heuristic; favor registers which don't require extra setup
885 // instructions in the preheader.
886 if (!isa<SCEVUnknown>(Reg) &&
887 !isa<SCEVConstant>(Reg) &&
888 !(isa<SCEVAddRecExpr>(Reg) &&
889 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
890 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
893 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
894 SE.hasComputableLoopEvolution(Reg, L);
897 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
898 /// before, rate it. Optional LoserRegs provides a way to declare any formula
899 /// that refers to one of those regs an instant loser.
900 void Cost::RatePrimaryRegister(const SCEV *Reg,
901 SmallPtrSet<const SCEV *, 16> &Regs,
903 ScalarEvolution &SE, DominatorTree &DT,
904 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
905 if (LoserRegs && LoserRegs->count(Reg)) {
909 if (Regs.insert(Reg)) {
910 RateRegister(Reg, Regs, L, SE, DT);
911 if (LoserRegs && isLoser())
912 LoserRegs->insert(Reg);
916 void Cost::RateFormula(const TargetTransformInfo &TTI,
918 SmallPtrSet<const SCEV *, 16> &Regs,
919 const DenseSet<const SCEV *> &VisitedRegs,
921 const SmallVectorImpl<int64_t> &Offsets,
922 ScalarEvolution &SE, DominatorTree &DT,
924 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
925 // Tally up the registers.
926 if (const SCEV *ScaledReg = F.ScaledReg) {
927 if (VisitedRegs.count(ScaledReg)) {
931 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
935 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
936 E = F.BaseRegs.end(); I != E; ++I) {
937 const SCEV *BaseReg = *I;
938 if (VisitedRegs.count(BaseReg)) {
942 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
947 // Determine how many (unfolded) adds we'll need inside the loop.
948 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
949 if (NumBaseParts > 1)
950 // Do not count the base and a possible second register if the target
951 // allows to fold 2 registers.
952 NumBaseAdds += NumBaseParts - (1 + isLegal2RegAMUse(TTI, LU, F));
954 // Accumulate non-free scaling amounts.
955 ScaleCost += getScalingFactorCost(TTI, LU, F);
957 // Tally up the non-zero immediates.
958 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
959 E = Offsets.end(); I != E; ++I) {
960 int64_t Offset = (uint64_t)*I + F.BaseOffset;
962 ImmCost += 64; // Handle symbolic values conservatively.
963 // TODO: This should probably be the pointer size.
964 else if (Offset != 0)
965 ImmCost += APInt(64, Offset, true).getMinSignedBits();
967 assert(isValid() && "invalid cost");
970 /// Lose - Set this cost to a losing value.
981 /// operator< - Choose the lower cost.
982 bool Cost::operator<(const Cost &Other) const {
983 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
984 ImmCost, SetupCost) <
985 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
986 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
990 void Cost::print(raw_ostream &OS) const {
991 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
993 OS << ", with addrec cost " << AddRecCost;
995 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
996 if (NumBaseAdds != 0)
997 OS << ", plus " << NumBaseAdds << " base add"
998 << (NumBaseAdds == 1 ? "" : "s");
1000 OS << ", plus " << ScaleCost << " scale cost";
1002 OS << ", plus " << ImmCost << " imm cost";
1004 OS << ", plus " << SetupCost << " setup cost";
1007 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1008 void Cost::dump() const {
1009 print(errs()); errs() << '\n';
1015 /// LSRFixup - An operand value in an instruction which is to be replaced
1016 /// with some equivalent, possibly strength-reduced, replacement.
1018 /// UserInst - The instruction which will be updated.
1019 Instruction *UserInst;
1021 /// OperandValToReplace - The operand of the instruction which will
1022 /// be replaced. The operand may be used more than once; every instance
1023 /// will be replaced.
1024 Value *OperandValToReplace;
1026 /// PostIncLoops - If this user is to use the post-incremented value of an
1027 /// induction variable, this variable is non-null and holds the loop
1028 /// associated with the induction variable.
1029 PostIncLoopSet PostIncLoops;
1031 /// LUIdx - The index of the LSRUse describing the expression which
1032 /// this fixup needs, minus an offset (below).
1035 /// Offset - A constant offset to be added to the LSRUse expression.
1036 /// This allows multiple fixups to share the same LSRUse with different
1037 /// offsets, for example in an unrolled loop.
1040 bool isUseFullyOutsideLoop(const Loop *L) const;
1044 void print(raw_ostream &OS) const;
1050 LSRFixup::LSRFixup()
1051 : UserInst(nullptr), OperandValToReplace(nullptr), LUIdx(~size_t(0)),
1054 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1055 /// value outside of the given loop.
1056 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1057 // PHI nodes use their value in their incoming blocks.
1058 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1059 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1060 if (PN->getIncomingValue(i) == OperandValToReplace &&
1061 L->contains(PN->getIncomingBlock(i)))
1066 return !L->contains(UserInst);
1069 void LSRFixup::print(raw_ostream &OS) const {
1071 // Store is common and interesting enough to be worth special-casing.
1072 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1074 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1075 } else if (UserInst->getType()->isVoidTy())
1076 OS << UserInst->getOpcodeName();
1078 UserInst->printAsOperand(OS, /*PrintType=*/false);
1080 OS << ", OperandValToReplace=";
1081 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1083 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1084 E = PostIncLoops.end(); I != E; ++I) {
1085 OS << ", PostIncLoop=";
1086 (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1089 if (LUIdx != ~size_t(0))
1090 OS << ", LUIdx=" << LUIdx;
1093 OS << ", Offset=" << Offset;
1096 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1097 void LSRFixup::dump() const {
1098 print(errs()); errs() << '\n';
1104 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1105 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1106 struct UniquifierDenseMapInfo {
1107 static SmallVector<const SCEV *, 4> getEmptyKey() {
1108 SmallVector<const SCEV *, 4> V;
1109 V.push_back(reinterpret_cast<const SCEV *>(-1));
1113 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1114 SmallVector<const SCEV *, 4> V;
1115 V.push_back(reinterpret_cast<const SCEV *>(-2));
1119 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1120 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1123 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1124 const SmallVector<const SCEV *, 4> &RHS) {
1129 /// LSRUse - This class holds the state that LSR keeps for each use in
1130 /// IVUsers, as well as uses invented by LSR itself. It includes information
1131 /// about what kinds of things can be folded into the user, information about
1132 /// the user itself, and information about how the use may be satisfied.
1133 /// TODO: Represent multiple users of the same expression in common?
1135 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1138 /// KindType - An enum for a kind of use, indicating what types of
1139 /// scaled and immediate operands it might support.
1141 Basic, ///< A normal use, with no folding.
1142 Special, ///< A special case of basic, allowing -1 scales.
1143 Address, ///< An address use; folding according to TargetLowering
1144 ICmpZero ///< An equality icmp with both operands folded into one.
1145 // TODO: Add a generic icmp too?
1148 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1153 SmallVector<int64_t, 8> Offsets;
1157 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1158 /// LSRUse are outside of the loop, in which case some special-case heuristics
1160 bool AllFixupsOutsideLoop;
1162 /// RigidFormula is set to true to guarantee that this use will be associated
1163 /// with a single formula--the one that initially matched. Some SCEV
1164 /// expressions cannot be expanded. This allows LSR to consider the registers
1165 /// used by those expressions without the need to expand them later after
1166 /// changing the formula.
1169 /// WidestFixupType - This records the widest use type for any fixup using
1170 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1171 /// max fixup widths to be equivalent, because the narrower one may be relying
1172 /// on the implicit truncation to truncate away bogus bits.
1173 Type *WidestFixupType;
1175 /// Formulae - A list of ways to build a value that can satisfy this user.
1176 /// After the list is populated, one of these is selected heuristically and
1177 /// used to formulate a replacement for OperandValToReplace in UserInst.
1178 SmallVector<Formula, 12> Formulae;
1180 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1181 SmallPtrSet<const SCEV *, 4> Regs;
1183 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1184 MinOffset(INT64_MAX),
1185 MaxOffset(INT64_MIN),
1186 AllFixupsOutsideLoop(true),
1187 RigidFormula(false),
1188 WidestFixupType(nullptr) {}
1190 bool HasFormulaWithSameRegs(const Formula &F) const;
1191 bool InsertFormula(const Formula &F);
1192 void DeleteFormula(Formula &F);
1193 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1195 void print(raw_ostream &OS) const;
1201 /// HasFormula - Test whether this use as a formula which has the same
1202 /// registers as the given formula.
1203 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1204 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1205 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1206 // Unstable sort by host order ok, because this is only used for uniquifying.
1207 std::sort(Key.begin(), Key.end());
1208 return Uniquifier.count(Key);
1211 /// InsertFormula - If the given formula has not yet been inserted, add it to
1212 /// the list, and return true. Return false otherwise.
1213 bool LSRUse::InsertFormula(const Formula &F) {
1214 if (!Formulae.empty() && RigidFormula)
1217 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1218 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1219 // Unstable sort by host order ok, because this is only used for uniquifying.
1220 std::sort(Key.begin(), Key.end());
1222 if (!Uniquifier.insert(Key).second)
1225 // Using a register to hold the value of 0 is not profitable.
1226 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1227 "Zero allocated in a scaled register!");
1229 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1230 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1231 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1234 // Add the formula to the list.
1235 Formulae.push_back(F);
1237 // Record registers now being used by this use.
1238 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1243 /// DeleteFormula - Remove the given formula from this use's list.
1244 void LSRUse::DeleteFormula(Formula &F) {
1245 if (&F != &Formulae.back())
1246 std::swap(F, Formulae.back());
1247 Formulae.pop_back();
1250 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1251 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1252 // Now that we've filtered out some formulae, recompute the Regs set.
1253 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1255 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1256 E = Formulae.end(); I != E; ++I) {
1257 const Formula &F = *I;
1258 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1259 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1262 // Update the RegTracker.
1263 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1264 E = OldRegs.end(); I != E; ++I)
1265 if (!Regs.count(*I))
1266 RegUses.DropRegister(*I, LUIdx);
1269 void LSRUse::print(raw_ostream &OS) const {
1270 OS << "LSR Use: Kind=";
1272 case Basic: OS << "Basic"; break;
1273 case Special: OS << "Special"; break;
1274 case ICmpZero: OS << "ICmpZero"; break;
1276 OS << "Address of ";
1277 if (AccessTy->isPointerTy())
1278 OS << "pointer"; // the full pointer type could be really verbose
1283 OS << ", Offsets={";
1284 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1285 E = Offsets.end(); I != E; ++I) {
1287 if (std::next(I) != E)
1292 if (AllFixupsOutsideLoop)
1293 OS << ", all-fixups-outside-loop";
1295 if (WidestFixupType)
1296 OS << ", widest fixup type: " << *WidestFixupType;
1299 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1300 void LSRUse::dump() const {
1301 print(errs()); errs() << '\n';
1305 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1306 /// be completely folded into the user instruction at isel time. This includes
1307 /// address-mode folding and special icmp tricks.
1308 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1309 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1310 bool HasBaseReg, int64_t Scale) {
1312 case LSRUse::Address:
1313 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1315 // Otherwise, just guess that reg+reg addressing is legal.
1318 case LSRUse::ICmpZero:
1319 // There's not even a target hook for querying whether it would be legal to
1320 // fold a GV into an ICmp.
1324 // ICmp only has two operands; don't allow more than two non-trivial parts.
1325 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1328 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1329 // putting the scaled register in the other operand of the icmp.
1330 if (Scale != 0 && Scale != -1)
1333 // If we have low-level target information, ask the target if it can fold an
1334 // integer immediate on an icmp.
1335 if (BaseOffset != 0) {
1337 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1338 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1339 // Offs is the ICmp immediate.
1341 // The cast does the right thing with INT64_MIN.
1342 BaseOffset = -(uint64_t)BaseOffset;
1343 return TTI.isLegalICmpImmediate(BaseOffset);
1346 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1350 // Only handle single-register values.
1351 return !BaseGV && Scale == 0 && BaseOffset == 0;
1353 case LSRUse::Special:
1354 // Special case Basic to handle -1 scales.
1355 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1358 llvm_unreachable("Invalid LSRUse Kind!");
1361 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1362 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1363 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1365 // Check for overflow.
1366 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1369 MinOffset = (uint64_t)BaseOffset + MinOffset;
1370 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1373 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1375 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1377 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1380 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1381 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1383 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1384 F.BaseOffset, F.HasBaseReg, F.Scale);
1387 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
1389 // If F is used as an Addressing Mode, it may fold one Base plus one
1390 // scaled register. If the scaled register is nil, do as if another
1391 // element of the base regs is a 1-scaled register.
1392 // This is possible if BaseRegs has at least 2 registers.
1394 // If this is not an address calculation, this is not an addressing mode
1396 if (LU.Kind != LSRUse::Address)
1399 // F is already scaled.
1403 // We need to keep one register for the base and one to scale.
1404 if (F.BaseRegs.size() < 2)
1407 return isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
1408 F.BaseGV, F.BaseOffset, F.HasBaseReg, 1);
1411 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1412 const LSRUse &LU, const Formula &F) {
1415 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1416 LU.AccessTy, F) && "Illegal formula in use.");
1419 case LSRUse::Address: {
1420 // Check the scaling factor cost with both the min and max offsets.
1421 int ScaleCostMinOffset =
1422 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1423 F.BaseOffset + LU.MinOffset,
1424 F.HasBaseReg, F.Scale);
1425 int ScaleCostMaxOffset =
1426 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1427 F.BaseOffset + LU.MaxOffset,
1428 F.HasBaseReg, F.Scale);
1430 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1431 "Legal addressing mode has an illegal cost!");
1432 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1434 case LSRUse::ICmpZero:
1435 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
1436 // Therefore, return 0 in case F.Scale == -1.
1437 return F.Scale != -1;
1440 case LSRUse::Special:
1444 llvm_unreachable("Invalid LSRUse Kind!");
1447 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1448 LSRUse::KindType Kind, Type *AccessTy,
1449 GlobalValue *BaseGV, int64_t BaseOffset,
1451 // Fast-path: zero is always foldable.
1452 if (BaseOffset == 0 && !BaseGV) return true;
1454 // Conservatively, create an address with an immediate and a
1455 // base and a scale.
1456 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1458 // Canonicalize a scale of 1 to a base register if the formula doesn't
1459 // already have a base register.
1460 if (!HasBaseReg && Scale == 1) {
1465 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1468 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1469 ScalarEvolution &SE, int64_t MinOffset,
1470 int64_t MaxOffset, LSRUse::KindType Kind,
1471 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1472 // Fast-path: zero is always foldable.
1473 if (S->isZero()) return true;
1475 // Conservatively, create an address with an immediate and a
1476 // base and a scale.
1477 int64_t BaseOffset = ExtractImmediate(S, SE);
1478 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1480 // If there's anything else involved, it's not foldable.
1481 if (!S->isZero()) return false;
1483 // Fast-path: zero is always foldable.
1484 if (BaseOffset == 0 && !BaseGV) return true;
1486 // Conservatively, create an address with an immediate and a
1487 // base and a scale.
1488 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1490 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1491 BaseOffset, HasBaseReg, Scale);
1496 /// IVInc - An individual increment in a Chain of IV increments.
1497 /// Relate an IV user to an expression that computes the IV it uses from the IV
1498 /// used by the previous link in the Chain.
1500 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1501 /// original IVOperand. The head of the chain's IVOperand is only valid during
1502 /// chain collection, before LSR replaces IV users. During chain generation,
1503 /// IncExpr can be used to find the new IVOperand that computes the same
1506 Instruction *UserInst;
1508 const SCEV *IncExpr;
1510 IVInc(Instruction *U, Value *O, const SCEV *E):
1511 UserInst(U), IVOperand(O), IncExpr(E) {}
1514 // IVChain - The list of IV increments in program order.
1515 // We typically add the head of a chain without finding subsequent links.
1517 SmallVector<IVInc,1> Incs;
1518 const SCEV *ExprBase;
1520 IVChain() : ExprBase(nullptr) {}
1522 IVChain(const IVInc &Head, const SCEV *Base)
1523 : Incs(1, Head), ExprBase(Base) {}
1525 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1527 // begin - return the first increment in the chain.
1528 const_iterator begin() const {
1529 assert(!Incs.empty());
1530 return std::next(Incs.begin());
1532 const_iterator end() const {
1536 // hasIncs - Returns true if this chain contains any increments.
1537 bool hasIncs() const { return Incs.size() >= 2; }
1539 // add - Add an IVInc to the end of this chain.
1540 void add(const IVInc &X) { Incs.push_back(X); }
1542 // tailUserInst - Returns the last UserInst in the chain.
1543 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1545 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1547 bool isProfitableIncrement(const SCEV *OperExpr,
1548 const SCEV *IncExpr,
1552 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1553 /// Distinguish between FarUsers that definitely cross IV increments and
1554 /// NearUsers that may be used between IV increments.
1556 SmallPtrSet<Instruction*, 4> FarUsers;
1557 SmallPtrSet<Instruction*, 4> NearUsers;
1560 /// LSRInstance - This class holds state for the main loop strength reduction
1564 ScalarEvolution &SE;
1567 const TargetTransformInfo &TTI;
1571 /// IVIncInsertPos - This is the insert position that the current loop's
1572 /// induction variable increment should be placed. In simple loops, this is
1573 /// the latch block's terminator. But in more complicated cases, this is a
1574 /// position which will dominate all the in-loop post-increment users.
1575 Instruction *IVIncInsertPos;
1577 /// Factors - Interesting factors between use strides.
1578 SmallSetVector<int64_t, 8> Factors;
1580 /// Types - Interesting use types, to facilitate truncation reuse.
1581 SmallSetVector<Type *, 4> Types;
1583 /// Fixups - The list of operands which are to be replaced.
1584 SmallVector<LSRFixup, 16> Fixups;
1586 /// Uses - The list of interesting uses.
1587 SmallVector<LSRUse, 16> Uses;
1589 /// RegUses - Track which uses use which register candidates.
1590 RegUseTracker RegUses;
1592 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1593 // have more than a few IV increment chains in a loop. Missing a Chain falls
1594 // back to normal LSR behavior for those uses.
1595 static const unsigned MaxChains = 8;
1597 /// IVChainVec - IV users can form a chain of IV increments.
1598 SmallVector<IVChain, MaxChains> IVChainVec;
1600 /// IVIncSet - IV users that belong to profitable IVChains.
1601 SmallPtrSet<Use*, MaxChains> IVIncSet;
1603 void OptimizeShadowIV();
1604 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1605 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1606 void OptimizeLoopTermCond();
1608 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1609 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1610 void FinalizeChain(IVChain &Chain);
1611 void CollectChains();
1612 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1613 SmallVectorImpl<WeakVH> &DeadInsts);
1615 void CollectInterestingTypesAndFactors();
1616 void CollectFixupsAndInitialFormulae();
1618 LSRFixup &getNewFixup() {
1619 Fixups.push_back(LSRFixup());
1620 return Fixups.back();
1623 // Support for sharing of LSRUses between LSRFixups.
1624 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1627 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1628 LSRUse::KindType Kind, Type *AccessTy);
1630 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1631 LSRUse::KindType Kind,
1634 void DeleteUse(LSRUse &LU, size_t LUIdx);
1636 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1638 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1639 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1640 void CountRegisters(const Formula &F, size_t LUIdx);
1641 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1643 void CollectLoopInvariantFixupsAndFormulae();
1645 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1646 unsigned Depth = 0);
1647 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1648 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1649 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1650 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1651 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1652 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1653 void GenerateCrossUseConstantOffsets();
1654 void GenerateAllReuseFormulae();
1656 void FilterOutUndesirableDedicatedRegisters();
1658 size_t EstimateSearchSpaceComplexity() const;
1659 void NarrowSearchSpaceByDetectingSupersets();
1660 void NarrowSearchSpaceByCollapsingUnrolledCode();
1661 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1662 void NarrowSearchSpaceByPickingWinnerRegs();
1663 void NarrowSearchSpaceUsingHeuristics();
1665 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1667 SmallVectorImpl<const Formula *> &Workspace,
1668 const Cost &CurCost,
1669 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1670 DenseSet<const SCEV *> &VisitedRegs) const;
1671 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1673 BasicBlock::iterator
1674 HoistInsertPosition(BasicBlock::iterator IP,
1675 const SmallVectorImpl<Instruction *> &Inputs) const;
1676 BasicBlock::iterator
1677 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1680 SCEVExpander &Rewriter) const;
1682 Value *Expand(const LSRFixup &LF,
1684 BasicBlock::iterator IP,
1685 SCEVExpander &Rewriter,
1686 SmallVectorImpl<WeakVH> &DeadInsts) const;
1687 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1689 SCEVExpander &Rewriter,
1690 SmallVectorImpl<WeakVH> &DeadInsts,
1692 void Rewrite(const LSRFixup &LF,
1694 SCEVExpander &Rewriter,
1695 SmallVectorImpl<WeakVH> &DeadInsts,
1697 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1701 LSRInstance(Loop *L, Pass *P);
1703 bool getChanged() const { return Changed; }
1705 void print_factors_and_types(raw_ostream &OS) const;
1706 void print_fixups(raw_ostream &OS) const;
1707 void print_uses(raw_ostream &OS) const;
1708 void print(raw_ostream &OS) const;
1714 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1715 /// inside the loop then try to eliminate the cast operation.
1716 void LSRInstance::OptimizeShadowIV() {
1717 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1718 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1721 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1722 UI != E; /* empty */) {
1723 IVUsers::const_iterator CandidateUI = UI;
1725 Instruction *ShadowUse = CandidateUI->getUser();
1726 Type *DestTy = nullptr;
1727 bool IsSigned = false;
1729 /* If shadow use is a int->float cast then insert a second IV
1730 to eliminate this cast.
1732 for (unsigned i = 0; i < n; ++i)
1738 for (unsigned i = 0; i < n; ++i, ++d)
1741 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1743 DestTy = UCast->getDestTy();
1745 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1747 DestTy = SCast->getDestTy();
1749 if (!DestTy) continue;
1751 // If target does not support DestTy natively then do not apply
1752 // this transformation.
1753 if (!TTI.isTypeLegal(DestTy)) continue;
1755 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1757 if (PH->getNumIncomingValues() != 2) continue;
1759 Type *SrcTy = PH->getType();
1760 int Mantissa = DestTy->getFPMantissaWidth();
1761 if (Mantissa == -1) continue;
1762 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1765 unsigned Entry, Latch;
1766 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1774 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1775 if (!Init) continue;
1776 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1777 (double)Init->getSExtValue() :
1778 (double)Init->getZExtValue());
1780 BinaryOperator *Incr =
1781 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1782 if (!Incr) continue;
1783 if (Incr->getOpcode() != Instruction::Add
1784 && Incr->getOpcode() != Instruction::Sub)
1787 /* Initialize new IV, double d = 0.0 in above example. */
1788 ConstantInt *C = nullptr;
1789 if (Incr->getOperand(0) == PH)
1790 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1791 else if (Incr->getOperand(1) == PH)
1792 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1798 // Ignore negative constants, as the code below doesn't handle them
1799 // correctly. TODO: Remove this restriction.
1800 if (!C->getValue().isStrictlyPositive()) continue;
1802 /* Add new PHINode. */
1803 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1805 /* create new increment. '++d' in above example. */
1806 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1807 BinaryOperator *NewIncr =
1808 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1809 Instruction::FAdd : Instruction::FSub,
1810 NewPH, CFP, "IV.S.next.", Incr);
1812 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1813 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1815 /* Remove cast operation */
1816 ShadowUse->replaceAllUsesWith(NewPH);
1817 ShadowUse->eraseFromParent();
1823 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1824 /// set the IV user and stride information and return true, otherwise return
1826 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1827 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1828 if (UI->getUser() == Cond) {
1829 // NOTE: we could handle setcc instructions with multiple uses here, but
1830 // InstCombine does it as well for simple uses, it's not clear that it
1831 // occurs enough in real life to handle.
1838 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1839 /// a max computation.
1841 /// This is a narrow solution to a specific, but acute, problem. For loops
1847 /// } while (++i < n);
1849 /// the trip count isn't just 'n', because 'n' might not be positive. And
1850 /// unfortunately this can come up even for loops where the user didn't use
1851 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1852 /// will commonly be lowered like this:
1858 /// } while (++i < n);
1861 /// and then it's possible for subsequent optimization to obscure the if
1862 /// test in such a way that indvars can't find it.
1864 /// When indvars can't find the if test in loops like this, it creates a
1865 /// max expression, which allows it to give the loop a canonical
1866 /// induction variable:
1869 /// max = n < 1 ? 1 : n;
1872 /// } while (++i != max);
1874 /// Canonical induction variables are necessary because the loop passes
1875 /// are designed around them. The most obvious example of this is the
1876 /// LoopInfo analysis, which doesn't remember trip count values. It
1877 /// expects to be able to rediscover the trip count each time it is
1878 /// needed, and it does this using a simple analysis that only succeeds if
1879 /// the loop has a canonical induction variable.
1881 /// However, when it comes time to generate code, the maximum operation
1882 /// can be quite costly, especially if it's inside of an outer loop.
1884 /// This function solves this problem by detecting this type of loop and
1885 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1886 /// the instructions for the maximum computation.
1888 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1889 // Check that the loop matches the pattern we're looking for.
1890 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1891 Cond->getPredicate() != CmpInst::ICMP_NE)
1894 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1895 if (!Sel || !Sel->hasOneUse()) return Cond;
1897 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1898 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1900 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1902 // Add one to the backedge-taken count to get the trip count.
1903 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1904 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1906 // Check for a max calculation that matches the pattern. There's no check
1907 // for ICMP_ULE here because the comparison would be with zero, which
1908 // isn't interesting.
1909 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1910 const SCEVNAryExpr *Max = nullptr;
1911 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1912 Pred = ICmpInst::ICMP_SLE;
1914 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1915 Pred = ICmpInst::ICMP_SLT;
1917 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1918 Pred = ICmpInst::ICMP_ULT;
1925 // To handle a max with more than two operands, this optimization would
1926 // require additional checking and setup.
1927 if (Max->getNumOperands() != 2)
1930 const SCEV *MaxLHS = Max->getOperand(0);
1931 const SCEV *MaxRHS = Max->getOperand(1);
1933 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1934 // for a comparison with 1. For <= and >=, a comparison with zero.
1936 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1939 // Check the relevant induction variable for conformance to
1941 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1942 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1943 if (!AR || !AR->isAffine() ||
1944 AR->getStart() != One ||
1945 AR->getStepRecurrence(SE) != One)
1948 assert(AR->getLoop() == L &&
1949 "Loop condition operand is an addrec in a different loop!");
1951 // Check the right operand of the select, and remember it, as it will
1952 // be used in the new comparison instruction.
1953 Value *NewRHS = nullptr;
1954 if (ICmpInst::isTrueWhenEqual(Pred)) {
1955 // Look for n+1, and grab n.
1956 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1957 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1958 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1959 NewRHS = BO->getOperand(0);
1960 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1961 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1962 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1963 NewRHS = BO->getOperand(0);
1966 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1967 NewRHS = Sel->getOperand(1);
1968 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1969 NewRHS = Sel->getOperand(2);
1970 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1971 NewRHS = SU->getValue();
1973 // Max doesn't match expected pattern.
1976 // Determine the new comparison opcode. It may be signed or unsigned,
1977 // and the original comparison may be either equality or inequality.
1978 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1979 Pred = CmpInst::getInversePredicate(Pred);
1981 // Ok, everything looks ok to change the condition into an SLT or SGE and
1982 // delete the max calculation.
1984 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1986 // Delete the max calculation instructions.
1987 Cond->replaceAllUsesWith(NewCond);
1988 CondUse->setUser(NewCond);
1989 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1990 Cond->eraseFromParent();
1991 Sel->eraseFromParent();
1992 if (Cmp->use_empty())
1993 Cmp->eraseFromParent();
1997 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1998 /// postinc iv when possible.
2000 LSRInstance::OptimizeLoopTermCond() {
2001 SmallPtrSet<Instruction *, 4> PostIncs;
2003 BasicBlock *LatchBlock = L->getLoopLatch();
2004 SmallVector<BasicBlock*, 8> ExitingBlocks;
2005 L->getExitingBlocks(ExitingBlocks);
2007 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2008 BasicBlock *ExitingBlock = ExitingBlocks[i];
2010 // Get the terminating condition for the loop if possible. If we
2011 // can, we want to change it to use a post-incremented version of its
2012 // induction variable, to allow coalescing the live ranges for the IV into
2013 // one register value.
2015 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2018 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2019 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2022 // Search IVUsesByStride to find Cond's IVUse if there is one.
2023 IVStrideUse *CondUse = nullptr;
2024 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2025 if (!FindIVUserForCond(Cond, CondUse))
2028 // If the trip count is computed in terms of a max (due to ScalarEvolution
2029 // being unable to find a sufficient guard, for example), change the loop
2030 // comparison to use SLT or ULT instead of NE.
2031 // One consequence of doing this now is that it disrupts the count-down
2032 // optimization. That's not always a bad thing though, because in such
2033 // cases it may still be worthwhile to avoid a max.
2034 Cond = OptimizeMax(Cond, CondUse);
2036 // If this exiting block dominates the latch block, it may also use
2037 // the post-inc value if it won't be shared with other uses.
2038 // Check for dominance.
2039 if (!DT.dominates(ExitingBlock, LatchBlock))
2042 // Conservatively avoid trying to use the post-inc value in non-latch
2043 // exits if there may be pre-inc users in intervening blocks.
2044 if (LatchBlock != ExitingBlock)
2045 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2046 // Test if the use is reachable from the exiting block. This dominator
2047 // query is a conservative approximation of reachability.
2048 if (&*UI != CondUse &&
2049 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2050 // Conservatively assume there may be reuse if the quotient of their
2051 // strides could be a legal scale.
2052 const SCEV *A = IU.getStride(*CondUse, L);
2053 const SCEV *B = IU.getStride(*UI, L);
2054 if (!A || !B) continue;
2055 if (SE.getTypeSizeInBits(A->getType()) !=
2056 SE.getTypeSizeInBits(B->getType())) {
2057 if (SE.getTypeSizeInBits(A->getType()) >
2058 SE.getTypeSizeInBits(B->getType()))
2059 B = SE.getSignExtendExpr(B, A->getType());
2061 A = SE.getSignExtendExpr(A, B->getType());
2063 if (const SCEVConstant *D =
2064 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2065 const ConstantInt *C = D->getValue();
2066 // Stride of one or negative one can have reuse with non-addresses.
2067 if (C->isOne() || C->isAllOnesValue())
2068 goto decline_post_inc;
2069 // Avoid weird situations.
2070 if (C->getValue().getMinSignedBits() >= 64 ||
2071 C->getValue().isMinSignedValue())
2072 goto decline_post_inc;
2073 // Check for possible scaled-address reuse.
2074 Type *AccessTy = getAccessType(UI->getUser());
2075 int64_t Scale = C->getSExtValue();
2076 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ nullptr,
2078 /*HasBaseReg=*/ false, Scale))
2079 goto decline_post_inc;
2081 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ nullptr,
2083 /*HasBaseReg=*/ false, Scale))
2084 goto decline_post_inc;
2088 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2091 // It's possible for the setcc instruction to be anywhere in the loop, and
2092 // possible for it to have multiple users. If it is not immediately before
2093 // the exiting block branch, move it.
2094 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2095 if (Cond->hasOneUse()) {
2096 Cond->moveBefore(TermBr);
2098 // Clone the terminating condition and insert into the loopend.
2099 ICmpInst *OldCond = Cond;
2100 Cond = cast<ICmpInst>(Cond->clone());
2101 Cond->setName(L->getHeader()->getName() + ".termcond");
2102 ExitingBlock->getInstList().insert(TermBr, Cond);
2104 // Clone the IVUse, as the old use still exists!
2105 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2106 TermBr->replaceUsesOfWith(OldCond, Cond);
2110 // If we get to here, we know that we can transform the setcc instruction to
2111 // use the post-incremented version of the IV, allowing us to coalesce the
2112 // live ranges for the IV correctly.
2113 CondUse->transformToPostInc(L);
2116 PostIncs.insert(Cond);
2120 // Determine an insertion point for the loop induction variable increment. It
2121 // must dominate all the post-inc comparisons we just set up, and it must
2122 // dominate the loop latch edge.
2123 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2124 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2125 E = PostIncs.end(); I != E; ++I) {
2127 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2129 if (BB == (*I)->getParent())
2130 IVIncInsertPos = *I;
2131 else if (BB != IVIncInsertPos->getParent())
2132 IVIncInsertPos = BB->getTerminator();
2136 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2137 /// at the given offset and other details. If so, update the use and
2140 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2141 LSRUse::KindType Kind, Type *AccessTy) {
2142 int64_t NewMinOffset = LU.MinOffset;
2143 int64_t NewMaxOffset = LU.MaxOffset;
2144 Type *NewAccessTy = AccessTy;
2146 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2147 // something conservative, however this can pessimize in the case that one of
2148 // the uses will have all its uses outside the loop, for example.
2149 if (LU.Kind != Kind)
2151 // Conservatively assume HasBaseReg is true for now.
2152 if (NewOffset < LU.MinOffset) {
2153 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2154 LU.MaxOffset - NewOffset, HasBaseReg))
2156 NewMinOffset = NewOffset;
2157 } else if (NewOffset > LU.MaxOffset) {
2158 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2159 NewOffset - LU.MinOffset, HasBaseReg))
2161 NewMaxOffset = NewOffset;
2163 // Check for a mismatched access type, and fall back conservatively as needed.
2164 // TODO: Be less conservative when the type is similar and can use the same
2165 // addressing modes.
2166 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2167 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2170 LU.MinOffset = NewMinOffset;
2171 LU.MaxOffset = NewMaxOffset;
2172 LU.AccessTy = NewAccessTy;
2173 if (NewOffset != LU.Offsets.back())
2174 LU.Offsets.push_back(NewOffset);
2178 /// getUse - Return an LSRUse index and an offset value for a fixup which
2179 /// needs the given expression, with the given kind and optional access type.
2180 /// Either reuse an existing use or create a new one, as needed.
2181 std::pair<size_t, int64_t>
2182 LSRInstance::getUse(const SCEV *&Expr,
2183 LSRUse::KindType Kind, Type *AccessTy) {
2184 const SCEV *Copy = Expr;
2185 int64_t Offset = ExtractImmediate(Expr, SE);
2187 // Basic uses can't accept any offset, for example.
2188 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2189 Offset, /*HasBaseReg=*/ true)) {
2194 std::pair<UseMapTy::iterator, bool> P =
2195 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2197 // A use already existed with this base.
2198 size_t LUIdx = P.first->second;
2199 LSRUse &LU = Uses[LUIdx];
2200 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2202 return std::make_pair(LUIdx, Offset);
2205 // Create a new use.
2206 size_t LUIdx = Uses.size();
2207 P.first->second = LUIdx;
2208 Uses.push_back(LSRUse(Kind, AccessTy));
2209 LSRUse &LU = Uses[LUIdx];
2211 // We don't need to track redundant offsets, but we don't need to go out
2212 // of our way here to avoid them.
2213 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2214 LU.Offsets.push_back(Offset);
2216 LU.MinOffset = Offset;
2217 LU.MaxOffset = Offset;
2218 return std::make_pair(LUIdx, Offset);
2221 /// DeleteUse - Delete the given use from the Uses list.
2222 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2223 if (&LU != &Uses.back())
2224 std::swap(LU, Uses.back());
2228 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2231 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2232 /// a formula that has the same registers as the given formula.
2234 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2235 const LSRUse &OrigLU) {
2236 // Search all uses for the formula. This could be more clever.
2237 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2238 LSRUse &LU = Uses[LUIdx];
2239 // Check whether this use is close enough to OrigLU, to see whether it's
2240 // worthwhile looking through its formulae.
2241 // Ignore ICmpZero uses because they may contain formulae generated by
2242 // GenerateICmpZeroScales, in which case adding fixup offsets may
2244 if (&LU != &OrigLU &&
2245 LU.Kind != LSRUse::ICmpZero &&
2246 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2247 LU.WidestFixupType == OrigLU.WidestFixupType &&
2248 LU.HasFormulaWithSameRegs(OrigF)) {
2249 // Scan through this use's formulae.
2250 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2251 E = LU.Formulae.end(); I != E; ++I) {
2252 const Formula &F = *I;
2253 // Check to see if this formula has the same registers and symbols
2255 if (F.BaseRegs == OrigF.BaseRegs &&
2256 F.ScaledReg == OrigF.ScaledReg &&
2257 F.BaseGV == OrigF.BaseGV &&
2258 F.Scale == OrigF.Scale &&
2259 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2260 if (F.BaseOffset == 0)
2262 // This is the formula where all the registers and symbols matched;
2263 // there aren't going to be any others. Since we declined it, we
2264 // can skip the rest of the formulae and proceed to the next LSRUse.
2271 // Nothing looked good.
2275 void LSRInstance::CollectInterestingTypesAndFactors() {
2276 SmallSetVector<const SCEV *, 4> Strides;
2278 // Collect interesting types and strides.
2279 SmallVector<const SCEV *, 4> Worklist;
2280 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2281 const SCEV *Expr = IU.getExpr(*UI);
2283 // Collect interesting types.
2284 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2286 // Add strides for mentioned loops.
2287 Worklist.push_back(Expr);
2289 const SCEV *S = Worklist.pop_back_val();
2290 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2291 if (AR->getLoop() == L)
2292 Strides.insert(AR->getStepRecurrence(SE));
2293 Worklist.push_back(AR->getStart());
2294 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2295 Worklist.append(Add->op_begin(), Add->op_end());
2297 } while (!Worklist.empty());
2300 // Compute interesting factors from the set of interesting strides.
2301 for (SmallSetVector<const SCEV *, 4>::const_iterator
2302 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2303 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2304 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2305 const SCEV *OldStride = *I;
2306 const SCEV *NewStride = *NewStrideIter;
2308 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2309 SE.getTypeSizeInBits(NewStride->getType())) {
2310 if (SE.getTypeSizeInBits(OldStride->getType()) >
2311 SE.getTypeSizeInBits(NewStride->getType()))
2312 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2314 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2316 if (const SCEVConstant *Factor =
2317 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2319 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2320 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2321 } else if (const SCEVConstant *Factor =
2322 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2325 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2326 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2330 // If all uses use the same type, don't bother looking for truncation-based
2332 if (Types.size() == 1)
2335 DEBUG(print_factors_and_types(dbgs()));
2338 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2339 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2340 /// Instructions to IVStrideUses, we could partially skip this.
2341 static User::op_iterator
2342 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2343 Loop *L, ScalarEvolution &SE) {
2344 for(; OI != OE; ++OI) {
2345 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2346 if (!SE.isSCEVable(Oper->getType()))
2349 if (const SCEVAddRecExpr *AR =
2350 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2351 if (AR->getLoop() == L)
2359 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2360 /// operands, so wrap it in a convenient helper.
2361 static Value *getWideOperand(Value *Oper) {
2362 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2363 return Trunc->getOperand(0);
2367 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2369 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2370 Type *LType = LVal->getType();
2371 Type *RType = RVal->getType();
2372 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2375 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2376 /// NULL for any constant. Returning the expression itself is
2377 /// conservative. Returning a deeper subexpression is more precise and valid as
2378 /// long as it isn't less complex than another subexpression. For expressions
2379 /// involving multiple unscaled values, we need to return the pointer-type
2380 /// SCEVUnknown. This avoids forming chains across objects, such as:
2381 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2383 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2384 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2385 static const SCEV *getExprBase(const SCEV *S) {
2386 switch (S->getSCEVType()) {
2387 default: // uncluding scUnknown.
2392 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2394 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2396 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2398 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2399 // there's nothing more complex.
2400 // FIXME: not sure if we want to recognize negation.
2401 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2402 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2403 E(Add->op_begin()); I != E; ++I) {
2404 const SCEV *SubExpr = *I;
2405 if (SubExpr->getSCEVType() == scAddExpr)
2406 return getExprBase(SubExpr);
2408 if (SubExpr->getSCEVType() != scMulExpr)
2411 return S; // all operands are scaled, be conservative.
2414 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2418 /// Return true if the chain increment is profitable to expand into a loop
2419 /// invariant value, which may require its own register. A profitable chain
2420 /// increment will be an offset relative to the same base. We allow such offsets
2421 /// to potentially be used as chain increment as long as it's not obviously
2422 /// expensive to expand using real instructions.
2423 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2424 const SCEV *IncExpr,
2425 ScalarEvolution &SE) {
2426 // Aggressively form chains when -stress-ivchain.
2430 // Do not replace a constant offset from IV head with a nonconstant IV
2432 if (!isa<SCEVConstant>(IncExpr)) {
2433 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2434 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2438 SmallPtrSet<const SCEV*, 8> Processed;
2439 return !isHighCostExpansion(IncExpr, Processed, SE);
2442 /// Return true if the number of registers needed for the chain is estimated to
2443 /// be less than the number required for the individual IV users. First prohibit
2444 /// any IV users that keep the IV live across increments (the Users set should
2445 /// be empty). Next count the number and type of increments in the chain.
2447 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2448 /// effectively use postinc addressing modes. Only consider it profitable it the
2449 /// increments can be computed in fewer registers when chained.
2451 /// TODO: Consider IVInc free if it's already used in another chains.
2453 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2454 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2458 if (!Chain.hasIncs())
2461 if (!Users.empty()) {
2462 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2463 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2464 E = Users.end(); I != E; ++I) {
2465 dbgs() << " " << **I << "\n";
2469 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2471 // The chain itself may require a register, so intialize cost to 1.
2474 // A complete chain likely eliminates the need for keeping the original IV in
2475 // a register. LSR does not currently know how to form a complete chain unless
2476 // the header phi already exists.
2477 if (isa<PHINode>(Chain.tailUserInst())
2478 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2481 const SCEV *LastIncExpr = nullptr;
2482 unsigned NumConstIncrements = 0;
2483 unsigned NumVarIncrements = 0;
2484 unsigned NumReusedIncrements = 0;
2485 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2488 if (I->IncExpr->isZero())
2491 // Incrementing by zero or some constant is neutral. We assume constants can
2492 // be folded into an addressing mode or an add's immediate operand.
2493 if (isa<SCEVConstant>(I->IncExpr)) {
2494 ++NumConstIncrements;
2498 if (I->IncExpr == LastIncExpr)
2499 ++NumReusedIncrements;
2503 LastIncExpr = I->IncExpr;
2505 // An IV chain with a single increment is handled by LSR's postinc
2506 // uses. However, a chain with multiple increments requires keeping the IV's
2507 // value live longer than it needs to be if chained.
2508 if (NumConstIncrements > 1)
2511 // Materializing increment expressions in the preheader that didn't exist in
2512 // the original code may cost a register. For example, sign-extended array
2513 // indices can produce ridiculous increments like this:
2514 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2515 cost += NumVarIncrements;
2517 // Reusing variable increments likely saves a register to hold the multiple of
2519 cost -= NumReusedIncrements;
2521 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2527 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2529 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2530 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2531 // When IVs are used as types of varying widths, they are generally converted
2532 // to a wider type with some uses remaining narrow under a (free) trunc.
2533 Value *const NextIV = getWideOperand(IVOper);
2534 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2535 const SCEV *const OperExprBase = getExprBase(OperExpr);
2537 // Visit all existing chains. Check if its IVOper can be computed as a
2538 // profitable loop invariant increment from the last link in the Chain.
2539 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2540 const SCEV *LastIncExpr = nullptr;
2541 for (; ChainIdx < NChains; ++ChainIdx) {
2542 IVChain &Chain = IVChainVec[ChainIdx];
2544 // Prune the solution space aggressively by checking that both IV operands
2545 // are expressions that operate on the same unscaled SCEVUnknown. This
2546 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2547 // first avoids creating extra SCEV expressions.
2548 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2551 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2552 if (!isCompatibleIVType(PrevIV, NextIV))
2555 // A phi node terminates a chain.
2556 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2559 // The increment must be loop-invariant so it can be kept in a register.
2560 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2561 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2562 if (!SE.isLoopInvariant(IncExpr, L))
2565 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2566 LastIncExpr = IncExpr;
2570 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2571 // bother for phi nodes, because they must be last in the chain.
2572 if (ChainIdx == NChains) {
2573 if (isa<PHINode>(UserInst))
2575 if (NChains >= MaxChains && !StressIVChain) {
2576 DEBUG(dbgs() << "IV Chain Limit\n");
2579 LastIncExpr = OperExpr;
2580 // IVUsers may have skipped over sign/zero extensions. We don't currently
2581 // attempt to form chains involving extensions unless they can be hoisted
2582 // into this loop's AddRec.
2583 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2586 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2588 ChainUsersVec.resize(NChains);
2589 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2590 << ") IV=" << *LastIncExpr << "\n");
2592 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2593 << ") IV+" << *LastIncExpr << "\n");
2594 // Add this IV user to the end of the chain.
2595 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2597 IVChain &Chain = IVChainVec[ChainIdx];
2599 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2600 // This chain's NearUsers become FarUsers.
2601 if (!LastIncExpr->isZero()) {
2602 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2607 // All other uses of IVOperand become near uses of the chain.
2608 // We currently ignore intermediate values within SCEV expressions, assuming
2609 // they will eventually be used be the current chain, or can be computed
2610 // from one of the chain increments. To be more precise we could
2611 // transitively follow its user and only add leaf IV users to the set.
2612 for (User *U : IVOper->users()) {
2613 Instruction *OtherUse = dyn_cast<Instruction>(U);
2616 // Uses in the chain will no longer be uses if the chain is formed.
2617 // Include the head of the chain in this iteration (not Chain.begin()).
2618 IVChain::const_iterator IncIter = Chain.Incs.begin();
2619 IVChain::const_iterator IncEnd = Chain.Incs.end();
2620 for( ; IncIter != IncEnd; ++IncIter) {
2621 if (IncIter->UserInst == OtherUse)
2624 if (IncIter != IncEnd)
2627 if (SE.isSCEVable(OtherUse->getType())
2628 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2629 && IU.isIVUserOrOperand(OtherUse)) {
2632 NearUsers.insert(OtherUse);
2635 // Since this user is part of the chain, it's no longer considered a use
2637 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2640 /// CollectChains - Populate the vector of Chains.
2642 /// This decreases ILP at the architecture level. Targets with ample registers,
2643 /// multiple memory ports, and no register renaming probably don't want
2644 /// this. However, such targets should probably disable LSR altogether.
2646 /// The job of LSR is to make a reasonable choice of induction variables across
2647 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2648 /// ILP *within the loop* if the target wants it.
2650 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2651 /// will not reorder memory operations, it will recognize this as a chain, but
2652 /// will generate redundant IV increments. Ideally this would be corrected later
2653 /// by a smart scheduler:
2659 /// TODO: Walk the entire domtree within this loop, not just the path to the
2660 /// loop latch. This will discover chains on side paths, but requires
2661 /// maintaining multiple copies of the Chains state.
2662 void LSRInstance::CollectChains() {
2663 DEBUG(dbgs() << "Collecting IV Chains.\n");
2664 SmallVector<ChainUsers, 8> ChainUsersVec;
2666 SmallVector<BasicBlock *,8> LatchPath;
2667 BasicBlock *LoopHeader = L->getHeader();
2668 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2669 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2670 LatchPath.push_back(Rung->getBlock());
2672 LatchPath.push_back(LoopHeader);
2674 // Walk the instruction stream from the loop header to the loop latch.
2675 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2676 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2677 BBIter != BBEnd; ++BBIter) {
2678 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2680 // Skip instructions that weren't seen by IVUsers analysis.
2681 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2684 // Ignore users that are part of a SCEV expression. This way we only
2685 // consider leaf IV Users. This effectively rediscovers a portion of
2686 // IVUsers analysis but in program order this time.
2687 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2690 // Remove this instruction from any NearUsers set it may be in.
2691 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2692 ChainIdx < NChains; ++ChainIdx) {
2693 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2695 // Search for operands that can be chained.
2696 SmallPtrSet<Instruction*, 4> UniqueOperands;
2697 User::op_iterator IVOpEnd = I->op_end();
2698 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2699 while (IVOpIter != IVOpEnd) {
2700 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2701 if (UniqueOperands.insert(IVOpInst))
2702 ChainInstruction(I, IVOpInst, ChainUsersVec);
2703 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2705 } // Continue walking down the instructions.
2706 } // Continue walking down the domtree.
2707 // Visit phi backedges to determine if the chain can generate the IV postinc.
2708 for (BasicBlock::iterator I = L->getHeader()->begin();
2709 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2710 if (!SE.isSCEVable(PN->getType()))
2714 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2716 ChainInstruction(PN, IncV, ChainUsersVec);
2718 // Remove any unprofitable chains.
2719 unsigned ChainIdx = 0;
2720 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2721 UsersIdx < NChains; ++UsersIdx) {
2722 if (!isProfitableChain(IVChainVec[UsersIdx],
2723 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2725 // Preserve the chain at UsesIdx.
2726 if (ChainIdx != UsersIdx)
2727 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2728 FinalizeChain(IVChainVec[ChainIdx]);
2731 IVChainVec.resize(ChainIdx);
2734 void LSRInstance::FinalizeChain(IVChain &Chain) {
2735 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2736 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2738 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2740 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2741 User::op_iterator UseI =
2742 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2743 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2744 IVIncSet.insert(UseI);
2748 /// Return true if the IVInc can be folded into an addressing mode.
2749 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2750 Value *Operand, const TargetTransformInfo &TTI) {
2751 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2752 if (!IncConst || !isAddressUse(UserInst, Operand))
2755 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2758 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2759 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2760 getAccessType(UserInst), /*BaseGV=*/ nullptr,
2761 IncOffset, /*HaseBaseReg=*/ false))
2767 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2768 /// materialize the IV user's operand from the previous IV user's operand.
2769 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2770 SmallVectorImpl<WeakVH> &DeadInsts) {
2771 // Find the new IVOperand for the head of the chain. It may have been replaced
2773 const IVInc &Head = Chain.Incs[0];
2774 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2775 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2776 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2778 Value *IVSrc = nullptr;
2779 while (IVOpIter != IVOpEnd) {
2780 IVSrc = getWideOperand(*IVOpIter);
2782 // If this operand computes the expression that the chain needs, we may use
2783 // it. (Check this after setting IVSrc which is used below.)
2785 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2786 // narrow for the chain, so we can no longer use it. We do allow using a
2787 // wider phi, assuming the LSR checked for free truncation. In that case we
2788 // should already have a truncate on this operand such that
2789 // getSCEV(IVSrc) == IncExpr.
2790 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2791 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2794 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2796 if (IVOpIter == IVOpEnd) {
2797 // Gracefully give up on this chain.
2798 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2802 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2803 Type *IVTy = IVSrc->getType();
2804 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2805 const SCEV *LeftOverExpr = nullptr;
2806 for (IVChain::const_iterator IncI = Chain.begin(),
2807 IncE = Chain.end(); IncI != IncE; ++IncI) {
2809 Instruction *InsertPt = IncI->UserInst;
2810 if (isa<PHINode>(InsertPt))
2811 InsertPt = L->getLoopLatch()->getTerminator();
2813 // IVOper will replace the current IV User's operand. IVSrc is the IV
2814 // value currently held in a register.
2815 Value *IVOper = IVSrc;
2816 if (!IncI->IncExpr->isZero()) {
2817 // IncExpr was the result of subtraction of two narrow values, so must
2819 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2820 LeftOverExpr = LeftOverExpr ?
2821 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2823 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2824 // Expand the IV increment.
2825 Rewriter.clearPostInc();
2826 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2827 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2828 SE.getUnknown(IncV));
2829 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2831 // If an IV increment can't be folded, use it as the next IV value.
2832 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2834 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2836 LeftOverExpr = nullptr;
2839 Type *OperTy = IncI->IVOperand->getType();
2840 if (IVTy != OperTy) {
2841 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2842 "cannot extend a chained IV");
2843 IRBuilder<> Builder(InsertPt);
2844 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2846 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2847 DeadInsts.push_back(IncI->IVOperand);
2849 // If LSR created a new, wider phi, we may also replace its postinc. We only
2850 // do this if we also found a wide value for the head of the chain.
2851 if (isa<PHINode>(Chain.tailUserInst())) {
2852 for (BasicBlock::iterator I = L->getHeader()->begin();
2853 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2854 if (!isCompatibleIVType(Phi, IVSrc))
2856 Instruction *PostIncV = dyn_cast<Instruction>(
2857 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2858 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2860 Value *IVOper = IVSrc;
2861 Type *PostIncTy = PostIncV->getType();
2862 if (IVTy != PostIncTy) {
2863 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2864 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2865 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2866 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2868 Phi->replaceUsesOfWith(PostIncV, IVOper);
2869 DeadInsts.push_back(PostIncV);
2874 void LSRInstance::CollectFixupsAndInitialFormulae() {
2875 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2876 Instruction *UserInst = UI->getUser();
2877 // Skip IV users that are part of profitable IV Chains.
2878 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2879 UI->getOperandValToReplace());
2880 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2881 if (IVIncSet.count(UseI))
2885 LSRFixup &LF = getNewFixup();
2886 LF.UserInst = UserInst;
2887 LF.OperandValToReplace = UI->getOperandValToReplace();
2888 LF.PostIncLoops = UI->getPostIncLoops();
2890 LSRUse::KindType Kind = LSRUse::Basic;
2891 Type *AccessTy = nullptr;
2892 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2893 Kind = LSRUse::Address;
2894 AccessTy = getAccessType(LF.UserInst);
2897 const SCEV *S = IU.getExpr(*UI);
2899 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2900 // (N - i == 0), and this allows (N - i) to be the expression that we work
2901 // with rather than just N or i, so we can consider the register
2902 // requirements for both N and i at the same time. Limiting this code to
2903 // equality icmps is not a problem because all interesting loops use
2904 // equality icmps, thanks to IndVarSimplify.
2905 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2906 if (CI->isEquality()) {
2907 // Swap the operands if needed to put the OperandValToReplace on the
2908 // left, for consistency.
2909 Value *NV = CI->getOperand(1);
2910 if (NV == LF.OperandValToReplace) {
2911 CI->setOperand(1, CI->getOperand(0));
2912 CI->setOperand(0, NV);
2913 NV = CI->getOperand(1);
2917 // x == y --> x - y == 0
2918 const SCEV *N = SE.getSCEV(NV);
2919 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2920 // S is normalized, so normalize N before folding it into S
2921 // to keep the result normalized.
2922 N = TransformForPostIncUse(Normalize, N, CI, nullptr,
2923 LF.PostIncLoops, SE, DT);
2924 Kind = LSRUse::ICmpZero;
2925 S = SE.getMinusSCEV(N, S);
2928 // -1 and the negations of all interesting strides (except the negation
2929 // of -1) are now also interesting.
2930 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2931 if (Factors[i] != -1)
2932 Factors.insert(-(uint64_t)Factors[i]);
2936 // Set up the initial formula for this use.
2937 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2939 LF.Offset = P.second;
2940 LSRUse &LU = Uses[LF.LUIdx];
2941 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2942 if (!LU.WidestFixupType ||
2943 SE.getTypeSizeInBits(LU.WidestFixupType) <
2944 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2945 LU.WidestFixupType = LF.OperandValToReplace->getType();
2947 // If this is the first use of this LSRUse, give it a formula.
2948 if (LU.Formulae.empty()) {
2949 InsertInitialFormula(S, LU, LF.LUIdx);
2950 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2954 DEBUG(print_fixups(dbgs()));
2957 /// InsertInitialFormula - Insert a formula for the given expression into
2958 /// the given use, separating out loop-variant portions from loop-invariant
2959 /// and loop-computable portions.
2961 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2962 // Mark uses whose expressions cannot be expanded.
2963 if (!isSafeToExpand(S, SE))
2964 LU.RigidFormula = true;
2967 F.InitialMatch(S, L, SE);
2968 bool Inserted = InsertFormula(LU, LUIdx, F);
2969 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2972 /// InsertSupplementalFormula - Insert a simple single-register formula for
2973 /// the given expression into the given use.
2975 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2976 LSRUse &LU, size_t LUIdx) {
2978 F.BaseRegs.push_back(S);
2979 F.HasBaseReg = true;
2980 bool Inserted = InsertFormula(LU, LUIdx, F);
2981 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2984 /// CountRegisters - Note which registers are used by the given formula,
2985 /// updating RegUses.
2986 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2988 RegUses.CountRegister(F.ScaledReg, LUIdx);
2989 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2990 E = F.BaseRegs.end(); I != E; ++I)
2991 RegUses.CountRegister(*I, LUIdx);
2994 /// InsertFormula - If the given formula has not yet been inserted, add it to
2995 /// the list, and return true. Return false otherwise.
2996 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2997 if (!LU.InsertFormula(F))
3000 CountRegisters(F, LUIdx);
3004 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3005 /// loop-invariant values which we're tracking. These other uses will pin these
3006 /// values in registers, making them less profitable for elimination.
3007 /// TODO: This currently misses non-constant addrec step registers.
3008 /// TODO: Should this give more weight to users inside the loop?
3010 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3011 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3012 SmallPtrSet<const SCEV *, 8> Inserted;
3014 while (!Worklist.empty()) {
3015 const SCEV *S = Worklist.pop_back_val();
3017 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3018 Worklist.append(N->op_begin(), N->op_end());
3019 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3020 Worklist.push_back(C->getOperand());
3021 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3022 Worklist.push_back(D->getLHS());
3023 Worklist.push_back(D->getRHS());
3024 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3025 if (!Inserted.insert(US)) continue;
3026 const Value *V = US->getValue();
3027 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3028 // Look for instructions defined outside the loop.
3029 if (L->contains(Inst)) continue;
3030 } else if (isa<UndefValue>(V))
3031 // Undef doesn't have a live range, so it doesn't matter.
3033 for (const Use &U : V->uses()) {
3034 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3035 // Ignore non-instructions.
3038 // Ignore instructions in other functions (as can happen with
3040 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3042 // Ignore instructions not dominated by the loop.
3043 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3044 UserInst->getParent() :
3045 cast<PHINode>(UserInst)->getIncomingBlock(
3046 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3047 if (!DT.dominates(L->getHeader(), UseBB))
3049 // Ignore uses which are part of other SCEV expressions, to avoid
3050 // analyzing them multiple times.
3051 if (SE.isSCEVable(UserInst->getType())) {
3052 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3053 // If the user is a no-op, look through to its uses.
3054 if (!isa<SCEVUnknown>(UserS))
3058 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3062 // Ignore icmp instructions which are already being analyzed.
3063 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3064 unsigned OtherIdx = !U.getOperandNo();
3065 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3066 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3070 LSRFixup &LF = getNewFixup();
3071 LF.UserInst = const_cast<Instruction *>(UserInst);
3072 LF.OperandValToReplace = U;
3073 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, nullptr);
3075 LF.Offset = P.second;
3076 LSRUse &LU = Uses[LF.LUIdx];
3077 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3078 if (!LU.WidestFixupType ||
3079 SE.getTypeSizeInBits(LU.WidestFixupType) <
3080 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3081 LU.WidestFixupType = LF.OperandValToReplace->getType();
3082 InsertSupplementalFormula(US, LU, LF.LUIdx);
3083 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3090 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3091 /// separate registers. If C is non-null, multiply each subexpression by C.
3093 /// Return remainder expression after factoring the subexpressions captured by
3094 /// Ops. If Ops is complete, return NULL.
3095 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3096 SmallVectorImpl<const SCEV *> &Ops,
3098 ScalarEvolution &SE,
3099 unsigned Depth = 0) {
3100 // Arbitrarily cap recursion to protect compile time.
3104 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3105 // Break out add operands.
3106 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3108 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3110 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3113 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3114 // Split a non-zero base out of an addrec.
3115 if (AR->getStart()->isZero())
3118 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3119 C, Ops, L, SE, Depth+1);
3120 // Split the non-zero AddRec unless it is part of a nested recurrence that
3121 // does not pertain to this loop.
3122 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3123 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3124 Remainder = nullptr;
3126 if (Remainder != AR->getStart()) {
3128 Remainder = SE.getConstant(AR->getType(), 0);
3129 return SE.getAddRecExpr(Remainder,
3130 AR->getStepRecurrence(SE),
3132 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3135 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3136 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3137 if (Mul->getNumOperands() != 2)
3139 if (const SCEVConstant *Op0 =
3140 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3141 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3142 const SCEV *Remainder =
3143 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3145 Ops.push_back(SE.getMulExpr(C, Remainder));
3152 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3154 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3157 // Arbitrarily cap recursion to protect compile time.
3158 if (Depth >= 3) return;
3160 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3161 const SCEV *BaseReg = Base.BaseRegs[i];
3163 SmallVector<const SCEV *, 8> AddOps;
3164 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3166 AddOps.push_back(Remainder);
3168 if (AddOps.size() == 1) continue;
3170 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3171 JE = AddOps.end(); J != JE; ++J) {
3173 // Loop-variant "unknown" values are uninteresting; we won't be able to
3174 // do anything meaningful with them.
3175 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3178 // Don't pull a constant into a register if the constant could be folded
3179 // into an immediate field.
3180 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3181 LU.AccessTy, *J, Base.getNumRegs() > 1))
3184 // Collect all operands except *J.
3185 SmallVector<const SCEV *, 8> InnerAddOps(
3186 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3187 InnerAddOps.append(std::next(J),
3188 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3190 // Don't leave just a constant behind in a register if the constant could
3191 // be folded into an immediate field.
3192 if (InnerAddOps.size() == 1 &&
3193 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3194 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3197 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3198 if (InnerSum->isZero())
3202 // Add the remaining pieces of the add back into the new formula.
3203 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3205 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3206 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3207 InnerSumSC->getValue()->getZExtValue())) {
3208 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3209 InnerSumSC->getValue()->getZExtValue();
3210 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3212 F.BaseRegs[i] = InnerSum;
3214 // Add J as its own register, or an unfolded immediate.
3215 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3216 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3217 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3218 SC->getValue()->getZExtValue()))
3219 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3220 SC->getValue()->getZExtValue();
3222 F.BaseRegs.push_back(*J);
3224 if (InsertFormula(LU, LUIdx, F))
3225 // If that formula hadn't been seen before, recurse to find more like
3227 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3232 /// GenerateCombinations - Generate a formula consisting of all of the
3233 /// loop-dominating registers added into a single register.
3234 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3236 // This method is only interesting on a plurality of registers.
3237 if (Base.BaseRegs.size() <= 1) return;
3241 SmallVector<const SCEV *, 4> Ops;
3242 for (SmallVectorImpl<const SCEV *>::const_iterator
3243 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3244 const SCEV *BaseReg = *I;
3245 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3246 !SE.hasComputableLoopEvolution(BaseReg, L))
3247 Ops.push_back(BaseReg);
3249 F.BaseRegs.push_back(BaseReg);
3251 if (Ops.size() > 1) {
3252 const SCEV *Sum = SE.getAddExpr(Ops);
3253 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3254 // opportunity to fold something. For now, just ignore such cases
3255 // rather than proceed with zero in a register.
3256 if (!Sum->isZero()) {
3257 F.BaseRegs.push_back(Sum);
3258 (void)InsertFormula(LU, LUIdx, F);
3263 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3264 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3266 // We can't add a symbolic offset if the address already contains one.
3267 if (Base.BaseGV) return;
3269 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3270 const SCEV *G = Base.BaseRegs[i];
3271 GlobalValue *GV = ExtractSymbol(G, SE);
3272 if (G->isZero() || !GV)
3276 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3279 (void)InsertFormula(LU, LUIdx, F);
3283 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3284 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3286 // TODO: For now, just add the min and max offset, because it usually isn't
3287 // worthwhile looking at everything inbetween.
3288 SmallVector<int64_t, 2> Worklist;
3289 Worklist.push_back(LU.MinOffset);
3290 if (LU.MaxOffset != LU.MinOffset)
3291 Worklist.push_back(LU.MaxOffset);
3293 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3294 const SCEV *G = Base.BaseRegs[i];
3296 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3297 E = Worklist.end(); I != E; ++I) {
3299 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3300 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3302 // Add the offset to the base register.
3303 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3304 // If it cancelled out, drop the base register, otherwise update it.
3305 if (NewG->isZero()) {
3306 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3307 F.BaseRegs.pop_back();
3309 F.BaseRegs[i] = NewG;
3311 (void)InsertFormula(LU, LUIdx, F);
3315 int64_t Imm = ExtractImmediate(G, SE);
3316 if (G->isZero() || Imm == 0)
3319 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3320 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3323 (void)InsertFormula(LU, LUIdx, F);
3327 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3328 /// the comparison. For example, x == y -> x*c == y*c.
3329 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3331 if (LU.Kind != LSRUse::ICmpZero) return;
3333 // Determine the integer type for the base formula.
3334 Type *IntTy = Base.getType();
3336 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3338 // Don't do this if there is more than one offset.
3339 if (LU.MinOffset != LU.MaxOffset) return;
3341 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3343 // Check each interesting stride.
3344 for (SmallSetVector<int64_t, 8>::const_iterator
3345 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3346 int64_t Factor = *I;
3348 // Check that the multiplication doesn't overflow.
3349 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3351 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3352 if (NewBaseOffset / Factor != Base.BaseOffset)
3354 // If the offset will be truncated at this use, check that it is in bounds.
3355 if (!IntTy->isPointerTy() &&
3356 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3359 // Check that multiplying with the use offset doesn't overflow.
3360 int64_t Offset = LU.MinOffset;
3361 if (Offset == INT64_MIN && Factor == -1)
3363 Offset = (uint64_t)Offset * Factor;
3364 if (Offset / Factor != LU.MinOffset)
3366 // If the offset will be truncated at this use, check that it is in bounds.
3367 if (!IntTy->isPointerTy() &&
3368 !ConstantInt::isValueValidForType(IntTy, Offset))
3372 F.BaseOffset = NewBaseOffset;
3374 // Check that this scale is legal.
3375 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3378 // Compensate for the use having MinOffset built into it.
3379 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3381 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3383 // Check that multiplying with each base register doesn't overflow.
3384 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3385 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3386 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3390 // Check that multiplying with the scaled register doesn't overflow.
3392 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3393 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3397 // Check that multiplying with the unfolded offset doesn't overflow.
3398 if (F.UnfoldedOffset != 0) {
3399 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3401 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3402 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3404 // If the offset will be truncated, check that it is in bounds.
3405 if (!IntTy->isPointerTy() &&
3406 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3410 // If we make it here and it's legal, add it.
3411 (void)InsertFormula(LU, LUIdx, F);
3416 /// GenerateScales - Generate stride factor reuse formulae by making use of
3417 /// scaled-offset address modes, for example.
3418 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3419 // Determine the integer type for the base formula.
3420 Type *IntTy = Base.getType();
3423 // If this Formula already has a scaled register, we can't add another one.
3424 if (Base.Scale != 0) return;
3426 // Check each interesting stride.
3427 for (SmallSetVector<int64_t, 8>::const_iterator
3428 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3429 int64_t Factor = *I;
3431 Base.Scale = Factor;
3432 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3433 // Check whether this scale is going to be legal.
3434 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3436 // As a special-case, handle special out-of-loop Basic users specially.
3437 // TODO: Reconsider this special case.
3438 if (LU.Kind == LSRUse::Basic &&
3439 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3440 LU.AccessTy, Base) &&
3441 LU.AllFixupsOutsideLoop)
3442 LU.Kind = LSRUse::Special;
3446 // For an ICmpZero, negating a solitary base register won't lead to
3448 if (LU.Kind == LSRUse::ICmpZero &&
3449 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3451 // For each addrec base reg, apply the scale, if possible.
3452 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3453 if (const SCEVAddRecExpr *AR =
3454 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3455 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3456 if (FactorS->isZero())
3458 // Divide out the factor, ignoring high bits, since we'll be
3459 // scaling the value back up in the end.
3460 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3461 // TODO: This could be optimized to avoid all the copying.
3463 F.ScaledReg = Quotient;
3464 F.DeleteBaseReg(F.BaseRegs[i]);
3465 (void)InsertFormula(LU, LUIdx, F);
3471 /// GenerateTruncates - Generate reuse formulae from different IV types.
3472 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3473 // Don't bother truncating symbolic values.
3474 if (Base.BaseGV) return;
3476 // Determine the integer type for the base formula.
3477 Type *DstTy = Base.getType();
3479 DstTy = SE.getEffectiveSCEVType(DstTy);
3481 for (SmallSetVector<Type *, 4>::const_iterator
3482 I = Types.begin(), E = Types.end(); I != E; ++I) {
3484 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3487 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3488 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3489 JE = F.BaseRegs.end(); J != JE; ++J)
3490 *J = SE.getAnyExtendExpr(*J, SrcTy);
3492 // TODO: This assumes we've done basic processing on all uses and
3493 // have an idea what the register usage is.
3494 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3497 (void)InsertFormula(LU, LUIdx, F);
3504 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3505 /// defer modifications so that the search phase doesn't have to worry about
3506 /// the data structures moving underneath it.
3510 const SCEV *OrigReg;
3512 WorkItem(size_t LI, int64_t I, const SCEV *R)
3513 : LUIdx(LI), Imm(I), OrigReg(R) {}
3515 void print(raw_ostream &OS) const;
3521 void WorkItem::print(raw_ostream &OS) const {
3522 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3523 << " , add offset " << Imm;
3526 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3527 void WorkItem::dump() const {
3528 print(errs()); errs() << '\n';
3532 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3533 /// distance apart and try to form reuse opportunities between them.
3534 void LSRInstance::GenerateCrossUseConstantOffsets() {
3535 // Group the registers by their value without any added constant offset.
3536 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3537 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3539 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3540 SmallVector<const SCEV *, 8> Sequence;
3541 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3543 const SCEV *Reg = *I;
3544 int64_t Imm = ExtractImmediate(Reg, SE);
3545 std::pair<RegMapTy::iterator, bool> Pair =
3546 Map.insert(std::make_pair(Reg, ImmMapTy()));
3548 Sequence.push_back(Reg);
3549 Pair.first->second.insert(std::make_pair(Imm, *I));
3550 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3553 // Now examine each set of registers with the same base value. Build up
3554 // a list of work to do and do the work in a separate step so that we're
3555 // not adding formulae and register counts while we're searching.
3556 SmallVector<WorkItem, 32> WorkItems;
3557 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3558 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3559 E = Sequence.end(); I != E; ++I) {
3560 const SCEV *Reg = *I;
3561 const ImmMapTy &Imms = Map.find(Reg)->second;
3563 // It's not worthwhile looking for reuse if there's only one offset.
3564 if (Imms.size() == 1)
3567 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3568 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3570 dbgs() << ' ' << J->first;
3573 // Examine each offset.
3574 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3576 const SCEV *OrigReg = J->second;
3578 int64_t JImm = J->first;
3579 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3581 if (!isa<SCEVConstant>(OrigReg) &&
3582 UsedByIndicesMap[Reg].count() == 1) {
3583 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3587 // Conservatively examine offsets between this orig reg a few selected
3589 ImmMapTy::const_iterator OtherImms[] = {
3590 Imms.begin(), std::prev(Imms.end()),
3591 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3594 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3595 ImmMapTy::const_iterator M = OtherImms[i];
3596 if (M == J || M == JE) continue;
3598 // Compute the difference between the two.
3599 int64_t Imm = (uint64_t)JImm - M->first;
3600 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3601 LUIdx = UsedByIndices.find_next(LUIdx))
3602 // Make a memo of this use, offset, and register tuple.
3603 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3604 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3611 UsedByIndicesMap.clear();
3612 UniqueItems.clear();
3614 // Now iterate through the worklist and add new formulae.
3615 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3616 E = WorkItems.end(); I != E; ++I) {
3617 const WorkItem &WI = *I;
3618 size_t LUIdx = WI.LUIdx;
3619 LSRUse &LU = Uses[LUIdx];
3620 int64_t Imm = WI.Imm;
3621 const SCEV *OrigReg = WI.OrigReg;
3623 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3624 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3625 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3627 // TODO: Use a more targeted data structure.
3628 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3629 const Formula &F = LU.Formulae[L];
3630 // Use the immediate in the scaled register.
3631 if (F.ScaledReg == OrigReg) {
3632 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3633 // Don't create 50 + reg(-50).
3634 if (F.referencesReg(SE.getSCEV(
3635 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3638 NewF.BaseOffset = Offset;
3639 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3642 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3644 // If the new scale is a constant in a register, and adding the constant
3645 // value to the immediate would produce a value closer to zero than the
3646 // immediate itself, then the formula isn't worthwhile.
3647 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3648 if (C->getValue()->isNegative() !=
3649 (NewF.BaseOffset < 0) &&
3650 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3651 .ule(abs64(NewF.BaseOffset)))
3655 (void)InsertFormula(LU, LUIdx, NewF);
3657 // Use the immediate in a base register.
3658 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3659 const SCEV *BaseReg = F.BaseRegs[N];
3660 if (BaseReg != OrigReg)
3663 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3664 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3665 LU.Kind, LU.AccessTy, NewF)) {
3666 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3669 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3671 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3673 // If the new formula has a constant in a register, and adding the
3674 // constant value to the immediate would produce a value closer to
3675 // zero than the immediate itself, then the formula isn't worthwhile.
3676 for (SmallVectorImpl<const SCEV *>::const_iterator
3677 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3679 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3680 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3681 abs64(NewF.BaseOffset)) &&
3682 (C->getValue()->getValue() +
3683 NewF.BaseOffset).countTrailingZeros() >=
3684 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3688 (void)InsertFormula(LU, LUIdx, NewF);
3697 /// GenerateAllReuseFormulae - Generate formulae for each use.
3699 LSRInstance::GenerateAllReuseFormulae() {
3700 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3701 // queries are more precise.
3702 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3703 LSRUse &LU = Uses[LUIdx];
3704 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3705 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3706 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3707 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3709 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3710 LSRUse &LU = Uses[LUIdx];
3711 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3712 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3713 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3714 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3715 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3716 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3717 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3718 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3720 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3721 LSRUse &LU = Uses[LUIdx];
3722 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3723 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3726 GenerateCrossUseConstantOffsets();
3728 DEBUG(dbgs() << "\n"
3729 "After generating reuse formulae:\n";
3730 print_uses(dbgs()));
3733 /// If there are multiple formulae with the same set of registers used
3734 /// by other uses, pick the best one and delete the others.
3735 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3736 DenseSet<const SCEV *> VisitedRegs;
3737 SmallPtrSet<const SCEV *, 16> Regs;
3738 SmallPtrSet<const SCEV *, 16> LoserRegs;
3740 bool ChangedFormulae = false;
3743 // Collect the best formula for each unique set of shared registers. This
3744 // is reset for each use.
3745 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3747 BestFormulaeTy BestFormulae;
3749 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3750 LSRUse &LU = Uses[LUIdx];
3751 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3754 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3755 FIdx != NumForms; ++FIdx) {
3756 Formula &F = LU.Formulae[FIdx];
3758 // Some formulas are instant losers. For example, they may depend on
3759 // nonexistent AddRecs from other loops. These need to be filtered
3760 // immediately, otherwise heuristics could choose them over others leading
3761 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3762 // avoids the need to recompute this information across formulae using the
3763 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3764 // the corresponding bad register from the Regs set.
3767 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3769 if (CostF.isLoser()) {
3770 // During initial formula generation, undesirable formulae are generated
3771 // by uses within other loops that have some non-trivial address mode or
3772 // use the postinc form of the IV. LSR needs to provide these formulae
3773 // as the basis of rediscovering the desired formula that uses an AddRec
3774 // corresponding to the existing phi. Once all formulae have been
3775 // generated, these initial losers may be pruned.
3776 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3780 SmallVector<const SCEV *, 4> Key;
3781 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3782 JE = F.BaseRegs.end(); J != JE; ++J) {
3783 const SCEV *Reg = *J;
3784 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3788 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3789 Key.push_back(F.ScaledReg);
3790 // Unstable sort by host order ok, because this is only used for
3792 std::sort(Key.begin(), Key.end());
3794 std::pair<BestFormulaeTy::const_iterator, bool> P =
3795 BestFormulae.insert(std::make_pair(Key, FIdx));
3799 Formula &Best = LU.Formulae[P.first->second];
3803 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3805 if (CostF < CostBest)
3807 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3809 " in favor of formula "; Best.print(dbgs());
3813 ChangedFormulae = true;
3815 LU.DeleteFormula(F);
3821 // Now that we've filtered out some formulae, recompute the Regs set.
3823 LU.RecomputeRegs(LUIdx, RegUses);
3825 // Reset this to prepare for the next use.
3826 BestFormulae.clear();
3829 DEBUG(if (ChangedFormulae) {
3831 "After filtering out undesirable candidates:\n";
3836 // This is a rough guess that seems to work fairly well.
3837 static const size_t ComplexityLimit = UINT16_MAX;
3839 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3840 /// solutions the solver might have to consider. It almost never considers
3841 /// this many solutions because it prune the search space, but the pruning
3842 /// isn't always sufficient.
3843 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3845 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3846 E = Uses.end(); I != E; ++I) {
3847 size_t FSize = I->Formulae.size();
3848 if (FSize >= ComplexityLimit) {
3849 Power = ComplexityLimit;
3853 if (Power >= ComplexityLimit)
3859 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3860 /// of the registers of another formula, it won't help reduce register
3861 /// pressure (though it may not necessarily hurt register pressure); remove
3862 /// it to simplify the system.
3863 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3864 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3865 DEBUG(dbgs() << "The search space is too complex.\n");
3867 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3868 "which use a superset of registers used by other "
3871 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3872 LSRUse &LU = Uses[LUIdx];
3874 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3875 Formula &F = LU.Formulae[i];
3876 // Look for a formula with a constant or GV in a register. If the use
3877 // also has a formula with that same value in an immediate field,
3878 // delete the one that uses a register.
3879 for (SmallVectorImpl<const SCEV *>::const_iterator
3880 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3881 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3883 NewF.BaseOffset += C->getValue()->getSExtValue();
3884 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3885 (I - F.BaseRegs.begin()));
3886 if (LU.HasFormulaWithSameRegs(NewF)) {
3887 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3888 LU.DeleteFormula(F);
3894 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3895 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3899 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3900 (I - F.BaseRegs.begin()));
3901 if (LU.HasFormulaWithSameRegs(NewF)) {
3902 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3904 LU.DeleteFormula(F);
3915 LU.RecomputeRegs(LUIdx, RegUses);
3918 DEBUG(dbgs() << "After pre-selection:\n";
3919 print_uses(dbgs()));
3923 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3924 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3926 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3927 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
3930 DEBUG(dbgs() << "The search space is too complex.\n"
3931 "Narrowing the search space by assuming that uses separated "
3932 "by a constant offset will use the same registers.\n");
3934 // This is especially useful for unrolled loops.
3936 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3937 LSRUse &LU = Uses[LUIdx];
3938 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3939 E = LU.Formulae.end(); I != E; ++I) {
3940 const Formula &F = *I;
3941 if (F.BaseOffset == 0 || F.Scale != 0)
3944 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
3948 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
3949 LU.Kind, LU.AccessTy))
3952 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
3954 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3956 // Update the relocs to reference the new use.
3957 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3958 E = Fixups.end(); I != E; ++I) {
3959 LSRFixup &Fixup = *I;
3960 if (Fixup.LUIdx == LUIdx) {
3961 Fixup.LUIdx = LUThatHas - &Uses.front();
3962 Fixup.Offset += F.BaseOffset;
3963 // Add the new offset to LUThatHas' offset list.
3964 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3965 LUThatHas->Offsets.push_back(Fixup.Offset);
3966 if (Fixup.Offset > LUThatHas->MaxOffset)
3967 LUThatHas->MaxOffset = Fixup.Offset;
3968 if (Fixup.Offset < LUThatHas->MinOffset)
3969 LUThatHas->MinOffset = Fixup.Offset;
3971 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
3973 if (Fixup.LUIdx == NumUses-1)
3974 Fixup.LUIdx = LUIdx;
3977 // Delete formulae from the new use which are no longer legal.
3979 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3980 Formula &F = LUThatHas->Formulae[i];
3981 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
3982 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
3983 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3985 LUThatHas->DeleteFormula(F);
3993 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3995 // Delete the old use.
3996 DeleteUse(LU, LUIdx);
4003 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4006 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4007 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4008 /// we've done more filtering, as it may be able to find more formulae to
4010 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4011 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4012 DEBUG(dbgs() << "The search space is too complex.\n");
4014 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4015 "undesirable dedicated registers.\n");
4017 FilterOutUndesirableDedicatedRegisters();
4019 DEBUG(dbgs() << "After pre-selection:\n";
4020 print_uses(dbgs()));
4024 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4025 /// to be profitable, and then in any use which has any reference to that
4026 /// register, delete all formulae which do not reference that register.
4027 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4028 // With all other options exhausted, loop until the system is simple
4029 // enough to handle.
4030 SmallPtrSet<const SCEV *, 4> Taken;
4031 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4032 // Ok, we have too many of formulae on our hands to conveniently handle.
4033 // Use a rough heuristic to thin out the list.
4034 DEBUG(dbgs() << "The search space is too complex.\n");
4036 // Pick the register which is used by the most LSRUses, which is likely
4037 // to be a good reuse register candidate.
4038 const SCEV *Best = nullptr;
4039 unsigned BestNum = 0;
4040 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4042 const SCEV *Reg = *I;
4043 if (Taken.count(Reg))
4048 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4049 if (Count > BestNum) {
4056 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4057 << " will yield profitable reuse.\n");
4060 // In any use with formulae which references this register, delete formulae
4061 // which don't reference it.
4062 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4063 LSRUse &LU = Uses[LUIdx];
4064 if (!LU.Regs.count(Best)) continue;
4067 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4068 Formula &F = LU.Formulae[i];
4069 if (!F.referencesReg(Best)) {
4070 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4071 LU.DeleteFormula(F);
4075 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4081 LU.RecomputeRegs(LUIdx, RegUses);
4084 DEBUG(dbgs() << "After pre-selection:\n";
4085 print_uses(dbgs()));
4089 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4090 /// formulae to choose from, use some rough heuristics to prune down the number
4091 /// of formulae. This keeps the main solver from taking an extraordinary amount
4092 /// of time in some worst-case scenarios.
4093 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4094 NarrowSearchSpaceByDetectingSupersets();
4095 NarrowSearchSpaceByCollapsingUnrolledCode();
4096 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4097 NarrowSearchSpaceByPickingWinnerRegs();
4100 /// SolveRecurse - This is the recursive solver.
4101 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4103 SmallVectorImpl<const Formula *> &Workspace,
4104 const Cost &CurCost,
4105 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4106 DenseSet<const SCEV *> &VisitedRegs) const {
4109 // - use more aggressive filtering
4110 // - sort the formula so that the most profitable solutions are found first
4111 // - sort the uses too
4113 // - don't compute a cost, and then compare. compare while computing a cost
4115 // - track register sets with SmallBitVector
4117 const LSRUse &LU = Uses[Workspace.size()];
4119 // If this use references any register that's already a part of the
4120 // in-progress solution, consider it a requirement that a formula must
4121 // reference that register in order to be considered. This prunes out
4122 // unprofitable searching.
4123 SmallSetVector<const SCEV *, 4> ReqRegs;
4124 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4125 E = CurRegs.end(); I != E; ++I)
4126 if (LU.Regs.count(*I))
4129 SmallPtrSet<const SCEV *, 16> NewRegs;
4131 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4132 E = LU.Formulae.end(); I != E; ++I) {
4133 const Formula &F = *I;
4135 // Ignore formulae which may not be ideal in terms of register reuse of
4136 // ReqRegs. The formula should use all required registers before
4137 // introducing new ones.
4138 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4139 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4140 JE = ReqRegs.end(); J != JE; ++J) {
4141 const SCEV *Reg = *J;
4142 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4143 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) !=
4146 if (NumReqRegsToFind == 0)
4150 if (NumReqRegsToFind != 0) {
4151 // If none of the formulae satisfied the required registers, then we could
4152 // clear ReqRegs and try again. Currently, we simply give up in this case.
4156 // Evaluate the cost of the current formula. If it's already worse than
4157 // the current best, prune the search at that point.
4160 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4162 if (NewCost < SolutionCost) {
4163 Workspace.push_back(&F);
4164 if (Workspace.size() != Uses.size()) {
4165 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4166 NewRegs, VisitedRegs);
4167 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4168 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4170 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4171 dbgs() << ".\n Regs:";
4172 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4173 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4174 dbgs() << ' ' << **I;
4177 SolutionCost = NewCost;
4178 Solution = Workspace;
4180 Workspace.pop_back();
4185 /// Solve - Choose one formula from each use. Return the results in the given
4186 /// Solution vector.
4187 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4188 SmallVector<const Formula *, 8> Workspace;
4190 SolutionCost.Lose();
4192 SmallPtrSet<const SCEV *, 16> CurRegs;
4193 DenseSet<const SCEV *> VisitedRegs;
4194 Workspace.reserve(Uses.size());
4196 // SolveRecurse does all the work.
4197 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4198 CurRegs, VisitedRegs);
4199 if (Solution.empty()) {
4200 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4204 // Ok, we've now made all our decisions.
4205 DEBUG(dbgs() << "\n"
4206 "The chosen solution requires "; SolutionCost.print(dbgs());
4208 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4210 Uses[i].print(dbgs());
4213 Solution[i]->print(dbgs());
4217 assert(Solution.size() == Uses.size() && "Malformed solution!");
4220 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4221 /// the dominator tree far as we can go while still being dominated by the
4222 /// input positions. This helps canonicalize the insert position, which
4223 /// encourages sharing.
4224 BasicBlock::iterator
4225 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4226 const SmallVectorImpl<Instruction *> &Inputs)
4229 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4230 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4233 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4234 if (!Rung) return IP;
4235 Rung = Rung->getIDom();
4236 if (!Rung) return IP;
4237 IDom = Rung->getBlock();
4239 // Don't climb into a loop though.
4240 const Loop *IDomLoop = LI.getLoopFor(IDom);
4241 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4242 if (IDomDepth <= IPLoopDepth &&
4243 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4247 bool AllDominate = true;
4248 Instruction *BetterPos = nullptr;
4249 Instruction *Tentative = IDom->getTerminator();
4250 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4251 E = Inputs.end(); I != E; ++I) {
4252 Instruction *Inst = *I;
4253 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4254 AllDominate = false;
4257 // Attempt to find an insert position in the middle of the block,
4258 // instead of at the end, so that it can be used for other expansions.
4259 if (IDom == Inst->getParent() &&
4260 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4261 BetterPos = std::next(BasicBlock::iterator(Inst));
4274 /// AdjustInsertPositionForExpand - Determine an input position which will be
4275 /// dominated by the operands and which will dominate the result.
4276 BasicBlock::iterator
4277 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4280 SCEVExpander &Rewriter) const {
4281 // Collect some instructions which must be dominated by the
4282 // expanding replacement. These must be dominated by any operands that
4283 // will be required in the expansion.
4284 SmallVector<Instruction *, 4> Inputs;
4285 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4286 Inputs.push_back(I);
4287 if (LU.Kind == LSRUse::ICmpZero)
4288 if (Instruction *I =
4289 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4290 Inputs.push_back(I);
4291 if (LF.PostIncLoops.count(L)) {
4292 if (LF.isUseFullyOutsideLoop(L))
4293 Inputs.push_back(L->getLoopLatch()->getTerminator());
4295 Inputs.push_back(IVIncInsertPos);
4297 // The expansion must also be dominated by the increment positions of any
4298 // loops it for which it is using post-inc mode.
4299 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4300 E = LF.PostIncLoops.end(); I != E; ++I) {
4301 const Loop *PIL = *I;
4302 if (PIL == L) continue;
4304 // Be dominated by the loop exit.
4305 SmallVector<BasicBlock *, 4> ExitingBlocks;
4306 PIL->getExitingBlocks(ExitingBlocks);
4307 if (!ExitingBlocks.empty()) {
4308 BasicBlock *BB = ExitingBlocks[0];
4309 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4310 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4311 Inputs.push_back(BB->getTerminator());
4315 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4316 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4317 "Insertion point must be a normal instruction");
4319 // Then, climb up the immediate dominator tree as far as we can go while
4320 // still being dominated by the input positions.
4321 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4323 // Don't insert instructions before PHI nodes.
4324 while (isa<PHINode>(IP)) ++IP;
4326 // Ignore landingpad instructions.
4327 while (isa<LandingPadInst>(IP)) ++IP;
4329 // Ignore debug intrinsics.
4330 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4332 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4333 // IP consistent across expansions and allows the previously inserted
4334 // instructions to be reused by subsequent expansion.
4335 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4340 /// Expand - Emit instructions for the leading candidate expression for this
4341 /// LSRUse (this is called "expanding").
4342 Value *LSRInstance::Expand(const LSRFixup &LF,
4344 BasicBlock::iterator IP,
4345 SCEVExpander &Rewriter,
4346 SmallVectorImpl<WeakVH> &DeadInsts) const {
4347 const LSRUse &LU = Uses[LF.LUIdx];
4348 if (LU.RigidFormula)
4349 return LF.OperandValToReplace;
4351 // Determine an input position which will be dominated by the operands and
4352 // which will dominate the result.
4353 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4355 // Inform the Rewriter if we have a post-increment use, so that it can
4356 // perform an advantageous expansion.
4357 Rewriter.setPostInc(LF.PostIncLoops);
4359 // This is the type that the user actually needs.
4360 Type *OpTy = LF.OperandValToReplace->getType();
4361 // This will be the type that we'll initially expand to.
4362 Type *Ty = F.getType();
4364 // No type known; just expand directly to the ultimate type.
4366 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4367 // Expand directly to the ultimate type if it's the right size.
4369 // This is the type to do integer arithmetic in.
4370 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4372 // Build up a list of operands to add together to form the full base.
4373 SmallVector<const SCEV *, 8> Ops;
4375 // Expand the BaseRegs portion.
4376 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4377 E = F.BaseRegs.end(); I != E; ++I) {
4378 const SCEV *Reg = *I;
4379 assert(!Reg->isZero() && "Zero allocated in a base register!");
4381 // If we're expanding for a post-inc user, make the post-inc adjustment.
4382 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4383 Reg = TransformForPostIncUse(Denormalize, Reg,
4384 LF.UserInst, LF.OperandValToReplace,
4387 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr, IP)));
4390 // Expand the ScaledReg portion.
4391 Value *ICmpScaledV = nullptr;
4393 const SCEV *ScaledS = F.ScaledReg;
4395 // If we're expanding for a post-inc user, make the post-inc adjustment.
4396 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4397 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4398 LF.UserInst, LF.OperandValToReplace,
4401 if (LU.Kind == LSRUse::ICmpZero) {
4402 // An interesting way of "folding" with an icmp is to use a negated
4403 // scale, which we'll implement by inserting it into the other operand
4405 assert(F.Scale == -1 &&
4406 "The only scale supported by ICmpZero uses is -1!");
4407 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr, IP);
4409 // Otherwise just expand the scaled register and an explicit scale,
4410 // which is expected to be matched as part of the address.
4412 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4413 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4414 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4416 Ops.push_back(SE.getUnknown(FullV));
4418 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, IP));
4419 ScaledS = SE.getMulExpr(ScaledS,
4420 SE.getConstant(ScaledS->getType(), F.Scale));
4421 Ops.push_back(ScaledS);
4425 // Expand the GV portion.
4427 // Flush the operand list to suppress SCEVExpander hoisting.
4429 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4431 Ops.push_back(SE.getUnknown(FullV));
4433 Ops.push_back(SE.getUnknown(F.BaseGV));
4436 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4437 // unfolded offsets. LSR assumes they both live next to their uses.
4439 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4441 Ops.push_back(SE.getUnknown(FullV));
4444 // Expand the immediate portion.
4445 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4447 if (LU.Kind == LSRUse::ICmpZero) {
4448 // The other interesting way of "folding" with an ICmpZero is to use a
4449 // negated immediate.
4451 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4453 Ops.push_back(SE.getUnknown(ICmpScaledV));
4454 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4457 // Just add the immediate values. These again are expected to be matched
4458 // as part of the address.
4459 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4463 // Expand the unfolded offset portion.
4464 int64_t UnfoldedOffset = F.UnfoldedOffset;
4465 if (UnfoldedOffset != 0) {
4466 // Just add the immediate values.
4467 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4471 // Emit instructions summing all the operands.
4472 const SCEV *FullS = Ops.empty() ?
4473 SE.getConstant(IntTy, 0) :
4475 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4477 // We're done expanding now, so reset the rewriter.
4478 Rewriter.clearPostInc();
4480 // An ICmpZero Formula represents an ICmp which we're handling as a
4481 // comparison against zero. Now that we've expanded an expression for that
4482 // form, update the ICmp's other operand.
4483 if (LU.Kind == LSRUse::ICmpZero) {
4484 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4485 DeadInsts.push_back(CI->getOperand(1));
4486 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4487 "a scale at the same time!");
4488 if (F.Scale == -1) {
4489 if (ICmpScaledV->getType() != OpTy) {
4491 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4493 ICmpScaledV, OpTy, "tmp", CI);
4496 CI->setOperand(1, ICmpScaledV);
4498 assert(F.Scale == 0 &&
4499 "ICmp does not support folding a global value and "
4500 "a scale at the same time!");
4501 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4503 if (C->getType() != OpTy)
4504 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4508 CI->setOperand(1, C);
4515 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4516 /// of their operands effectively happens in their predecessor blocks, so the
4517 /// expression may need to be expanded in multiple places.
4518 void LSRInstance::RewriteForPHI(PHINode *PN,
4521 SCEVExpander &Rewriter,
4522 SmallVectorImpl<WeakVH> &DeadInsts,
4524 DenseMap<BasicBlock *, Value *> Inserted;
4525 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4526 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4527 BasicBlock *BB = PN->getIncomingBlock(i);
4529 // If this is a critical edge, split the edge so that we do not insert
4530 // the code on all predecessor/successor paths. We do this unless this
4531 // is the canonical backedge for this loop, which complicates post-inc
4533 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4534 !isa<IndirectBrInst>(BB->getTerminator())) {
4535 BasicBlock *Parent = PN->getParent();
4536 Loop *PNLoop = LI.getLoopFor(Parent);
4537 if (!PNLoop || Parent != PNLoop->getHeader()) {
4538 // Split the critical edge.
4539 BasicBlock *NewBB = nullptr;
4540 if (!Parent->isLandingPad()) {
4541 NewBB = SplitCriticalEdge(BB, Parent, P,
4542 /*MergeIdenticalEdges=*/true,
4543 /*DontDeleteUselessPhis=*/true);
4545 SmallVector<BasicBlock*, 2> NewBBs;
4546 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4549 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4550 // phi predecessors are identical. The simple thing to do is skip
4551 // splitting in this case rather than complicate the API.
4553 // If PN is outside of the loop and BB is in the loop, we want to
4554 // move the block to be immediately before the PHI block, not
4555 // immediately after BB.
4556 if (L->contains(BB) && !L->contains(PN))
4557 NewBB->moveBefore(PN->getParent());
4559 // Splitting the edge can reduce the number of PHI entries we have.
4560 e = PN->getNumIncomingValues();
4562 i = PN->getBasicBlockIndex(BB);
4567 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4568 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
4570 PN->setIncomingValue(i, Pair.first->second);
4572 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4574 // If this is reuse-by-noop-cast, insert the noop cast.
4575 Type *OpTy = LF.OperandValToReplace->getType();
4576 if (FullV->getType() != OpTy)
4578 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4580 FullV, LF.OperandValToReplace->getType(),
4581 "tmp", BB->getTerminator());
4583 PN->setIncomingValue(i, FullV);
4584 Pair.first->second = FullV;
4589 /// Rewrite - Emit instructions for the leading candidate expression for this
4590 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4591 /// the newly expanded value.
4592 void LSRInstance::Rewrite(const LSRFixup &LF,
4594 SCEVExpander &Rewriter,
4595 SmallVectorImpl<WeakVH> &DeadInsts,
4597 // First, find an insertion point that dominates UserInst. For PHI nodes,
4598 // find the nearest block which dominates all the relevant uses.
4599 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4600 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4602 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4604 // If this is reuse-by-noop-cast, insert the noop cast.
4605 Type *OpTy = LF.OperandValToReplace->getType();
4606 if (FullV->getType() != OpTy) {
4608 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4609 FullV, OpTy, "tmp", LF.UserInst);
4613 // Update the user. ICmpZero is handled specially here (for now) because
4614 // Expand may have updated one of the operands of the icmp already, and
4615 // its new value may happen to be equal to LF.OperandValToReplace, in
4616 // which case doing replaceUsesOfWith leads to replacing both operands
4617 // with the same value. TODO: Reorganize this.
4618 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4619 LF.UserInst->setOperand(0, FullV);
4621 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4624 DeadInsts.push_back(LF.OperandValToReplace);
4627 /// ImplementSolution - Rewrite all the fixup locations with new values,
4628 /// following the chosen solution.
4630 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4632 // Keep track of instructions we may have made dead, so that
4633 // we can remove them after we are done working.
4634 SmallVector<WeakVH, 16> DeadInsts;
4636 SCEVExpander Rewriter(SE, "lsr");
4638 Rewriter.setDebugType(DEBUG_TYPE);
4640 Rewriter.disableCanonicalMode();
4641 Rewriter.enableLSRMode();
4642 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4644 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4645 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4646 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4647 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4648 Rewriter.setChainedPhi(PN);
4651 // Expand the new value definitions and update the users.
4652 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4653 E = Fixups.end(); I != E; ++I) {
4654 const LSRFixup &Fixup = *I;
4656 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4661 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4662 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4663 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4666 // Clean up after ourselves. This must be done before deleting any
4670 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4673 LSRInstance::LSRInstance(Loop *L, Pass *P)
4674 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4675 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4676 LI(P->getAnalysis<LoopInfo>()),
4677 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4678 IVIncInsertPos(nullptr) {
4679 // If LoopSimplify form is not available, stay out of trouble.
4680 if (!L->isLoopSimplifyForm())
4683 // If there's no interesting work to be done, bail early.
4684 if (IU.empty()) return;
4686 // If there's too much analysis to be done, bail early. We won't be able to
4687 // model the problem anyway.
4688 unsigned NumUsers = 0;
4689 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4690 if (++NumUsers > MaxIVUsers) {
4691 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4698 // All dominating loops must have preheaders, or SCEVExpander may not be able
4699 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4701 // IVUsers analysis should only create users that are dominated by simple loop
4702 // headers. Since this loop should dominate all of its users, its user list
4703 // should be empty if this loop itself is not within a simple loop nest.
4704 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4705 Rung; Rung = Rung->getIDom()) {
4706 BasicBlock *BB = Rung->getBlock();
4707 const Loop *DomLoop = LI.getLoopFor(BB);
4708 if (DomLoop && DomLoop->getHeader() == BB) {
4709 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4714 DEBUG(dbgs() << "\nLSR on loop ";
4715 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4718 // First, perform some low-level loop optimizations.
4720 OptimizeLoopTermCond();
4722 // If loop preparation eliminates all interesting IV users, bail.
4723 if (IU.empty()) return;
4725 // Skip nested loops until we can model them better with formulae.
4727 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4731 // Start collecting data and preparing for the solver.
4733 CollectInterestingTypesAndFactors();
4734 CollectFixupsAndInitialFormulae();
4735 CollectLoopInvariantFixupsAndFormulae();
4737 assert(!Uses.empty() && "IVUsers reported at least one use");
4738 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4739 print_uses(dbgs()));
4741 // Now use the reuse data to generate a bunch of interesting ways
4742 // to formulate the values needed for the uses.
4743 GenerateAllReuseFormulae();
4745 FilterOutUndesirableDedicatedRegisters();
4746 NarrowSearchSpaceUsingHeuristics();
4748 SmallVector<const Formula *, 8> Solution;
4751 // Release memory that is no longer needed.
4756 if (Solution.empty())
4760 // Formulae should be legal.
4761 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4763 const LSRUse &LU = *I;
4764 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4765 JE = LU.Formulae.end();
4767 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4768 *J) && "Illegal formula generated!");
4772 // Now that we've decided what we want, make it so.
4773 ImplementSolution(Solution, P);
4776 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4777 if (Factors.empty() && Types.empty()) return;
4779 OS << "LSR has identified the following interesting factors and types: ";
4782 for (SmallSetVector<int64_t, 8>::const_iterator
4783 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4784 if (!First) OS << ", ";
4789 for (SmallSetVector<Type *, 4>::const_iterator
4790 I = Types.begin(), E = Types.end(); I != E; ++I) {
4791 if (!First) OS << ", ";
4793 OS << '(' << **I << ')';
4798 void LSRInstance::print_fixups(raw_ostream &OS) const {
4799 OS << "LSR is examining the following fixup sites:\n";
4800 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4801 E = Fixups.end(); I != E; ++I) {
4808 void LSRInstance::print_uses(raw_ostream &OS) const {
4809 OS << "LSR is examining the following uses:\n";
4810 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4811 E = Uses.end(); I != E; ++I) {
4812 const LSRUse &LU = *I;
4816 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4817 JE = LU.Formulae.end(); J != JE; ++J) {
4825 void LSRInstance::print(raw_ostream &OS) const {
4826 print_factors_and_types(OS);
4831 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4832 void LSRInstance::dump() const {
4833 print(errs()); errs() << '\n';
4839 class LoopStrengthReduce : public LoopPass {
4841 static char ID; // Pass ID, replacement for typeid
4842 LoopStrengthReduce();
4845 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
4846 void getAnalysisUsage(AnalysisUsage &AU) const override;
4851 char LoopStrengthReduce::ID = 0;
4852 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4853 "Loop Strength Reduction", false, false)
4854 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4855 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4856 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4857 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4858 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4859 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4860 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4861 "Loop Strength Reduction", false, false)
4864 Pass *llvm::createLoopStrengthReducePass() {
4865 return new LoopStrengthReduce();
4868 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4869 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4872 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4873 // We split critical edges, so we change the CFG. However, we do update
4874 // many analyses if they are around.
4875 AU.addPreservedID(LoopSimplifyID);
4877 AU.addRequired<LoopInfo>();
4878 AU.addPreserved<LoopInfo>();
4879 AU.addRequiredID(LoopSimplifyID);
4880 AU.addRequired<DominatorTreeWrapperPass>();
4881 AU.addPreserved<DominatorTreeWrapperPass>();
4882 AU.addRequired<ScalarEvolution>();
4883 AU.addPreserved<ScalarEvolution>();
4884 // Requiring LoopSimplify a second time here prevents IVUsers from running
4885 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4886 AU.addRequiredID(LoopSimplifyID);
4887 AU.addRequired<IVUsers>();
4888 AU.addPreserved<IVUsers>();
4889 AU.addRequired<TargetTransformInfo>();
4892 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4893 if (skipOptnoneFunction(L))
4896 bool Changed = false;
4898 // Run the main LSR transformation.
4899 Changed |= LSRInstance(L, this).getChanged();
4901 // Remove any extra phis created by processing inner loops.
4902 Changed |= DeleteDeadPHIs(L->getHeader());
4903 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4904 SmallVector<WeakVH, 16> DeadInsts;
4905 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4907 Rewriter.setDebugType(DEBUG_TYPE);
4909 unsigned numFolded = Rewriter.replaceCongruentIVs(
4910 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
4911 &getAnalysis<TargetTransformInfo>());
4914 DeleteTriviallyDeadInstructions(DeadInsts);
4915 DeleteDeadPHIs(L->getHeader());