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 TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
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 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Assembly/Writer.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/ADT/SmallBitVector.h"
70 #include "llvm/ADT/SetVector.h"
71 #include "llvm/ADT/DenseSet.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Target/TargetLowering.h"
80 static cl::opt<bool> EnableNested(
81 "enable-lsr-nested", cl::Hidden, cl::desc("Enable LSR on nested loops"));
83 static cl::opt<bool> EnableRetry(
84 "enable-lsr-retry", cl::Hidden, cl::desc("Enable LSR retry"));
86 // Temporary flag to cleanup congruent phis after LSR phi expansion.
87 // It's currently disabled until we can determine whether it's truly useful or
88 // not. The flag should be removed after the v3.0 release.
89 // This is now needed for ivchains.
90 static cl::opt<bool> EnablePhiElim(
91 "enable-lsr-phielim", cl::Hidden, cl::init(true),
92 cl::desc("Enable LSR phi elimination"));
95 // Stress test IV chain generation.
96 static cl::opt<bool> StressIVChain(
97 "stress-ivchain", cl::Hidden, cl::init(false),
98 cl::desc("Stress test LSR IV chains"));
100 static bool StressIVChain = false;
105 /// RegSortData - This class holds data which is used to order reuse candidates.
108 /// UsedByIndices - This represents the set of LSRUse indices which reference
109 /// a particular register.
110 SmallBitVector UsedByIndices;
114 void print(raw_ostream &OS) const;
120 void RegSortData::print(raw_ostream &OS) const {
121 OS << "[NumUses=" << UsedByIndices.count() << ']';
124 void RegSortData::dump() const {
125 print(errs()); errs() << '\n';
130 /// RegUseTracker - Map register candidates to information about how they are
132 class RegUseTracker {
133 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
135 RegUsesTy RegUsesMap;
136 SmallVector<const SCEV *, 16> RegSequence;
139 void CountRegister(const SCEV *Reg, size_t LUIdx);
140 void DropRegister(const SCEV *Reg, size_t LUIdx);
141 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
143 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
145 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
149 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
150 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
151 iterator begin() { return RegSequence.begin(); }
152 iterator end() { return RegSequence.end(); }
153 const_iterator begin() const { return RegSequence.begin(); }
154 const_iterator end() const { return RegSequence.end(); }
160 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
161 std::pair<RegUsesTy::iterator, bool> Pair =
162 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
163 RegSortData &RSD = Pair.first->second;
165 RegSequence.push_back(Reg);
166 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
167 RSD.UsedByIndices.set(LUIdx);
171 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
172 RegUsesTy::iterator It = RegUsesMap.find(Reg);
173 assert(It != RegUsesMap.end());
174 RegSortData &RSD = It->second;
175 assert(RSD.UsedByIndices.size() > LUIdx);
176 RSD.UsedByIndices.reset(LUIdx);
180 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
181 assert(LUIdx <= LastLUIdx);
183 // Update RegUses. The data structure is not optimized for this purpose;
184 // we must iterate through it and update each of the bit vectors.
185 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
187 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
188 if (LUIdx < UsedByIndices.size())
189 UsedByIndices[LUIdx] =
190 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
191 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
196 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
197 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
198 if (I == RegUsesMap.end())
200 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
201 int i = UsedByIndices.find_first();
202 if (i == -1) return false;
203 if ((size_t)i != LUIdx) return true;
204 return UsedByIndices.find_next(i) != -1;
207 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
208 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
209 assert(I != RegUsesMap.end() && "Unknown register!");
210 return I->second.UsedByIndices;
213 void RegUseTracker::clear() {
220 /// Formula - This class holds information that describes a formula for
221 /// computing satisfying a use. It may include broken-out immediates and scaled
224 /// AM - This is used to represent complex addressing, as well as other kinds
225 /// of interesting uses.
226 TargetLowering::AddrMode AM;
228 /// BaseRegs - The list of "base" registers for this use. When this is
229 /// non-empty, AM.HasBaseReg should be set to true.
230 SmallVector<const SCEV *, 2> BaseRegs;
232 /// ScaledReg - The 'scaled' register for this use. This should be non-null
233 /// when AM.Scale is not zero.
234 const SCEV *ScaledReg;
236 /// UnfoldedOffset - An additional constant offset which added near the
237 /// use. This requires a temporary register, but the offset itself can
238 /// live in an add immediate field rather than a register.
239 int64_t UnfoldedOffset;
241 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
243 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
245 unsigned getNumRegs() const;
246 Type *getType() const;
248 void DeleteBaseReg(const SCEV *&S);
250 bool referencesReg(const SCEV *S) const;
251 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
252 const RegUseTracker &RegUses) const;
254 void print(raw_ostream &OS) const;
260 /// DoInitialMatch - Recursion helper for InitialMatch.
261 static void DoInitialMatch(const SCEV *S, Loop *L,
262 SmallVectorImpl<const SCEV *> &Good,
263 SmallVectorImpl<const SCEV *> &Bad,
264 ScalarEvolution &SE) {
265 // Collect expressions which properly dominate the loop header.
266 if (SE.properlyDominates(S, L->getHeader())) {
271 // Look at add operands.
272 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
273 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
275 DoInitialMatch(*I, L, Good, Bad, SE);
279 // Look at addrec operands.
280 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
281 if (!AR->getStart()->isZero()) {
282 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
283 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
284 AR->getStepRecurrence(SE),
285 // FIXME: AR->getNoWrapFlags()
286 AR->getLoop(), SCEV::FlagAnyWrap),
291 // Handle a multiplication by -1 (negation) if it didn't fold.
292 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
293 if (Mul->getOperand(0)->isAllOnesValue()) {
294 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
295 const SCEV *NewMul = SE.getMulExpr(Ops);
297 SmallVector<const SCEV *, 4> MyGood;
298 SmallVector<const SCEV *, 4> MyBad;
299 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
300 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
301 SE.getEffectiveSCEVType(NewMul->getType())));
302 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
303 E = MyGood.end(); I != E; ++I)
304 Good.push_back(SE.getMulExpr(NegOne, *I));
305 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
306 E = MyBad.end(); I != E; ++I)
307 Bad.push_back(SE.getMulExpr(NegOne, *I));
311 // Ok, we can't do anything interesting. Just stuff the whole thing into a
312 // register and hope for the best.
316 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
317 /// attempting to keep all loop-invariant and loop-computable values in a
318 /// single base register.
319 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
320 SmallVector<const SCEV *, 4> Good;
321 SmallVector<const SCEV *, 4> Bad;
322 DoInitialMatch(S, L, Good, Bad, SE);
324 const SCEV *Sum = SE.getAddExpr(Good);
326 BaseRegs.push_back(Sum);
327 AM.HasBaseReg = true;
330 const SCEV *Sum = SE.getAddExpr(Bad);
332 BaseRegs.push_back(Sum);
333 AM.HasBaseReg = true;
337 /// getNumRegs - Return the total number of register operands used by this
338 /// formula. This does not include register uses implied by non-constant
340 unsigned Formula::getNumRegs() const {
341 return !!ScaledReg + BaseRegs.size();
344 /// getType - Return the type of this formula, if it has one, or null
345 /// otherwise. This type is meaningless except for the bit size.
346 Type *Formula::getType() const {
347 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
348 ScaledReg ? ScaledReg->getType() :
349 AM.BaseGV ? AM.BaseGV->getType() :
353 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
354 void Formula::DeleteBaseReg(const SCEV *&S) {
355 if (&S != &BaseRegs.back())
356 std::swap(S, BaseRegs.back());
360 /// referencesReg - Test if this formula references the given register.
361 bool Formula::referencesReg(const SCEV *S) const {
362 return S == ScaledReg ||
363 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
366 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
367 /// which are used by uses other than the use with the given index.
368 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
369 const RegUseTracker &RegUses) const {
371 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
373 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
374 E = BaseRegs.end(); I != E; ++I)
375 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
380 void Formula::print(raw_ostream &OS) const {
383 if (!First) OS << " + "; else First = false;
384 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
386 if (AM.BaseOffs != 0) {
387 if (!First) OS << " + "; else First = false;
390 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
391 E = BaseRegs.end(); I != E; ++I) {
392 if (!First) OS << " + "; else First = false;
393 OS << "reg(" << **I << ')';
395 if (AM.HasBaseReg && BaseRegs.empty()) {
396 if (!First) OS << " + "; else First = false;
397 OS << "**error: HasBaseReg**";
398 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
399 if (!First) OS << " + "; else First = false;
400 OS << "**error: !HasBaseReg**";
403 if (!First) OS << " + "; else First = false;
404 OS << AM.Scale << "*reg(";
411 if (UnfoldedOffset != 0) {
412 if (!First) OS << " + "; else First = false;
413 OS << "imm(" << UnfoldedOffset << ')';
417 void Formula::dump() const {
418 print(errs()); errs() << '\n';
421 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
422 /// without changing its value.
423 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
425 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
426 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
429 /// isAddSExtable - Return true if the given add can be sign-extended
430 /// without changing its value.
431 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
433 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
434 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
437 /// isMulSExtable - Return true if the given mul can be sign-extended
438 /// without changing its value.
439 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
441 IntegerType::get(SE.getContext(),
442 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
443 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
446 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
447 /// and if the remainder is known to be zero, or null otherwise. If
448 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
449 /// to Y, ignoring that the multiplication may overflow, which is useful when
450 /// the result will be used in a context where the most significant bits are
452 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
454 bool IgnoreSignificantBits = false) {
455 // Handle the trivial case, which works for any SCEV type.
457 return SE.getConstant(LHS->getType(), 1);
459 // Handle a few RHS special cases.
460 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
462 const APInt &RA = RC->getValue()->getValue();
463 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
465 if (RA.isAllOnesValue())
466 return SE.getMulExpr(LHS, RC);
467 // Handle x /s 1 as x.
472 // Check for a division of a constant by a constant.
473 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
476 const APInt &LA = C->getValue()->getValue();
477 const APInt &RA = RC->getValue()->getValue();
478 if (LA.srem(RA) != 0)
480 return SE.getConstant(LA.sdiv(RA));
483 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
484 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
485 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
486 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
487 IgnoreSignificantBits);
489 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
490 IgnoreSignificantBits);
491 if (!Start) return 0;
492 // FlagNW is independent of the start value, step direction, and is
493 // preserved with smaller magnitude steps.
494 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
495 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
500 // Distribute the sdiv over add operands, if the add doesn't overflow.
501 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
502 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
503 SmallVector<const SCEV *, 8> Ops;
504 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
506 const SCEV *Op = getExactSDiv(*I, RHS, SE,
507 IgnoreSignificantBits);
511 return SE.getAddExpr(Ops);
516 // Check for a multiply operand that we can pull RHS out of.
517 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
518 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
519 SmallVector<const SCEV *, 4> Ops;
521 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
525 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
526 IgnoreSignificantBits)) {
532 return Found ? SE.getMulExpr(Ops) : 0;
537 // Otherwise we don't know.
541 /// ExtractImmediate - If S involves the addition of a constant integer value,
542 /// return that integer value, and mutate S to point to a new SCEV with that
544 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
545 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
546 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
547 S = SE.getConstant(C->getType(), 0);
548 return C->getValue()->getSExtValue();
550 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
551 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
552 int64_t Result = ExtractImmediate(NewOps.front(), SE);
554 S = SE.getAddExpr(NewOps);
556 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
557 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
558 int64_t Result = ExtractImmediate(NewOps.front(), SE);
560 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
561 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
568 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
569 /// return that symbol, and mutate S to point to a new SCEV with that
571 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
572 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
573 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
574 S = SE.getConstant(GV->getType(), 0);
577 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
578 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
579 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
581 S = SE.getAddExpr(NewOps);
583 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
584 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
585 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
587 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
588 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
595 /// isAddressUse - Returns true if the specified instruction is using the
596 /// specified value as an address.
597 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
598 bool isAddress = isa<LoadInst>(Inst);
599 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
600 if (SI->getOperand(1) == OperandVal)
602 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
603 // Addressing modes can also be folded into prefetches and a variety
605 switch (II->getIntrinsicID()) {
607 case Intrinsic::prefetch:
608 case Intrinsic::x86_sse_storeu_ps:
609 case Intrinsic::x86_sse2_storeu_pd:
610 case Intrinsic::x86_sse2_storeu_dq:
611 case Intrinsic::x86_sse2_storel_dq:
612 if (II->getArgOperand(0) == OperandVal)
620 /// getAccessType - Return the type of the memory being accessed.
621 static Type *getAccessType(const Instruction *Inst) {
622 Type *AccessTy = Inst->getType();
623 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
624 AccessTy = SI->getOperand(0)->getType();
625 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
626 // Addressing modes can also be folded into prefetches and a variety
628 switch (II->getIntrinsicID()) {
630 case Intrinsic::x86_sse_storeu_ps:
631 case Intrinsic::x86_sse2_storeu_pd:
632 case Intrinsic::x86_sse2_storeu_dq:
633 case Intrinsic::x86_sse2_storel_dq:
634 AccessTy = II->getArgOperand(0)->getType();
639 // All pointers have the same requirements, so canonicalize them to an
640 // arbitrary pointer type to minimize variation.
641 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
642 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
643 PTy->getAddressSpace());
648 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
649 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
650 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
651 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
652 if (SE.isSCEVable(PN->getType()) &&
653 (SE.getEffectiveSCEVType(PN->getType()) ==
654 SE.getEffectiveSCEVType(AR->getType())) &&
655 SE.getSCEV(PN) == AR)
661 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
662 /// specified set are trivially dead, delete them and see if this makes any of
663 /// their operands subsequently dead.
665 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
666 bool Changed = false;
668 while (!DeadInsts.empty()) {
669 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
671 if (I == 0 || !isInstructionTriviallyDead(I))
674 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
675 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
678 DeadInsts.push_back(U);
681 I->eraseFromParent();
690 /// Cost - This class is used to measure and compare candidate formulae.
692 /// TODO: Some of these could be merged. Also, a lexical ordering
693 /// isn't always optimal.
697 unsigned NumBaseAdds;
703 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
706 bool operator<(const Cost &Other) const;
711 // Once any of the metrics loses, they must all remain losers.
713 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
714 | ImmCost | SetupCost) != ~0u)
715 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
716 & ImmCost & SetupCost) == ~0u);
721 assert(isValid() && "invalid cost");
722 return NumRegs == ~0u;
725 void RateFormula(const Formula &F,
726 SmallPtrSet<const SCEV *, 16> &Regs,
727 const DenseSet<const SCEV *> &VisitedRegs,
729 const SmallVectorImpl<int64_t> &Offsets,
730 ScalarEvolution &SE, DominatorTree &DT,
731 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
733 void print(raw_ostream &OS) const;
737 void RateRegister(const SCEV *Reg,
738 SmallPtrSet<const SCEV *, 16> &Regs,
740 ScalarEvolution &SE, DominatorTree &DT);
741 void RatePrimaryRegister(const SCEV *Reg,
742 SmallPtrSet<const SCEV *, 16> &Regs,
744 ScalarEvolution &SE, DominatorTree &DT,
745 SmallPtrSet<const SCEV *, 16> *LoserRegs);
750 /// RateRegister - Tally up interesting quantities from the given register.
751 void Cost::RateRegister(const SCEV *Reg,
752 SmallPtrSet<const SCEV *, 16> &Regs,
754 ScalarEvolution &SE, DominatorTree &DT) {
755 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
756 if (AR->getLoop() == L)
757 AddRecCost += 1; /// TODO: This should be a function of the stride.
759 // If this is an addrec for another loop, don't second-guess its addrec phi
760 // nodes. LSR isn't currently smart enough to reason about more than one
761 // loop at a time. LSR has either already run on inner loops, will not run
762 // on other loops, and cannot be expected to change sibling loops. If the
763 // AddRec exists, consider it's register free and leave it alone. Otherwise,
764 // do not consider this formula at all.
765 else if (!EnableNested || L->contains(AR->getLoop()) ||
766 (!AR->getLoop()->contains(L) &&
767 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
768 if (isExistingPhi(AR, SE))
771 // For !EnableNested, never rewrite IVs in other loops.
776 // If this isn't one of the addrecs that the loop already has, it
777 // would require a costly new phi and add. TODO: This isn't
778 // precisely modeled right now.
780 if (!Regs.count(AR->getStart())) {
781 RateRegister(AR->getStart(), Regs, L, SE, DT);
787 // Add the step value register, if it needs one.
788 // TODO: The non-affine case isn't precisely modeled here.
789 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
790 if (!Regs.count(AR->getOperand(1))) {
791 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
799 // Rough heuristic; favor registers which don't require extra setup
800 // instructions in the preheader.
801 if (!isa<SCEVUnknown>(Reg) &&
802 !isa<SCEVConstant>(Reg) &&
803 !(isa<SCEVAddRecExpr>(Reg) &&
804 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
805 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
808 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
809 SE.hasComputableLoopEvolution(Reg, L);
812 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
813 /// before, rate it. Optional LoserRegs provides a way to declare any formula
814 /// that refers to one of those regs an instant loser.
815 void Cost::RatePrimaryRegister(const SCEV *Reg,
816 SmallPtrSet<const SCEV *, 16> &Regs,
818 ScalarEvolution &SE, DominatorTree &DT,
819 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
820 if (LoserRegs && LoserRegs->count(Reg)) {
824 if (Regs.insert(Reg)) {
825 RateRegister(Reg, Regs, L, SE, DT);
827 LoserRegs->insert(Reg);
831 void Cost::RateFormula(const Formula &F,
832 SmallPtrSet<const SCEV *, 16> &Regs,
833 const DenseSet<const SCEV *> &VisitedRegs,
835 const SmallVectorImpl<int64_t> &Offsets,
836 ScalarEvolution &SE, DominatorTree &DT,
837 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
838 // Tally up the registers.
839 if (const SCEV *ScaledReg = F.ScaledReg) {
840 if (VisitedRegs.count(ScaledReg)) {
844 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
848 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
849 E = F.BaseRegs.end(); I != E; ++I) {
850 const SCEV *BaseReg = *I;
851 if (VisitedRegs.count(BaseReg)) {
855 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
860 // Determine how many (unfolded) adds we'll need inside the loop.
861 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
862 if (NumBaseParts > 1)
863 NumBaseAdds += NumBaseParts - 1;
865 // Tally up the non-zero immediates.
866 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
867 E = Offsets.end(); I != E; ++I) {
868 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
870 ImmCost += 64; // Handle symbolic values conservatively.
871 // TODO: This should probably be the pointer size.
872 else if (Offset != 0)
873 ImmCost += APInt(64, Offset, true).getMinSignedBits();
875 assert(isValid() && "invalid cost");
878 /// Loose - Set this cost to a losing value.
888 /// operator< - Choose the lower cost.
889 bool Cost::operator<(const Cost &Other) const {
890 if (NumRegs != Other.NumRegs)
891 return NumRegs < Other.NumRegs;
892 if (AddRecCost != Other.AddRecCost)
893 return AddRecCost < Other.AddRecCost;
894 if (NumIVMuls != Other.NumIVMuls)
895 return NumIVMuls < Other.NumIVMuls;
896 if (NumBaseAdds != Other.NumBaseAdds)
897 return NumBaseAdds < Other.NumBaseAdds;
898 if (ImmCost != Other.ImmCost)
899 return ImmCost < Other.ImmCost;
900 if (SetupCost != Other.SetupCost)
901 return SetupCost < Other.SetupCost;
905 void Cost::print(raw_ostream &OS) const {
906 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
908 OS << ", with addrec cost " << AddRecCost;
910 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
911 if (NumBaseAdds != 0)
912 OS << ", plus " << NumBaseAdds << " base add"
913 << (NumBaseAdds == 1 ? "" : "s");
915 OS << ", plus " << ImmCost << " imm cost";
917 OS << ", plus " << SetupCost << " setup cost";
920 void Cost::dump() const {
921 print(errs()); errs() << '\n';
926 /// LSRFixup - An operand value in an instruction which is to be replaced
927 /// with some equivalent, possibly strength-reduced, replacement.
929 /// UserInst - The instruction which will be updated.
930 Instruction *UserInst;
932 /// OperandValToReplace - The operand of the instruction which will
933 /// be replaced. The operand may be used more than once; every instance
934 /// will be replaced.
935 Value *OperandValToReplace;
937 /// PostIncLoops - If this user is to use the post-incremented value of an
938 /// induction variable, this variable is non-null and holds the loop
939 /// associated with the induction variable.
940 PostIncLoopSet PostIncLoops;
942 /// LUIdx - The index of the LSRUse describing the expression which
943 /// this fixup needs, minus an offset (below).
946 /// Offset - A constant offset to be added to the LSRUse expression.
947 /// This allows multiple fixups to share the same LSRUse with different
948 /// offsets, for example in an unrolled loop.
951 bool isUseFullyOutsideLoop(const Loop *L) const;
955 void print(raw_ostream &OS) const;
962 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
964 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
965 /// value outside of the given loop.
966 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
967 // PHI nodes use their value in their incoming blocks.
968 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
969 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
970 if (PN->getIncomingValue(i) == OperandValToReplace &&
971 L->contains(PN->getIncomingBlock(i)))
976 return !L->contains(UserInst);
979 void LSRFixup::print(raw_ostream &OS) const {
981 // Store is common and interesting enough to be worth special-casing.
982 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
984 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
985 } else if (UserInst->getType()->isVoidTy())
986 OS << UserInst->getOpcodeName();
988 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
990 OS << ", OperandValToReplace=";
991 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
993 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
994 E = PostIncLoops.end(); I != E; ++I) {
995 OS << ", PostIncLoop=";
996 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
999 if (LUIdx != ~size_t(0))
1000 OS << ", LUIdx=" << LUIdx;
1003 OS << ", Offset=" << Offset;
1006 void LSRFixup::dump() const {
1007 print(errs()); errs() << '\n';
1012 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1013 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1014 struct UniquifierDenseMapInfo {
1015 static SmallVector<const SCEV *, 2> getEmptyKey() {
1016 SmallVector<const SCEV *, 2> V;
1017 V.push_back(reinterpret_cast<const SCEV *>(-1));
1021 static SmallVector<const SCEV *, 2> getTombstoneKey() {
1022 SmallVector<const SCEV *, 2> V;
1023 V.push_back(reinterpret_cast<const SCEV *>(-2));
1027 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
1028 unsigned Result = 0;
1029 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1030 E = V.end(); I != E; ++I)
1031 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1035 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1036 const SmallVector<const SCEV *, 2> &RHS) {
1041 /// LSRUse - This class holds the state that LSR keeps for each use in
1042 /// IVUsers, as well as uses invented by LSR itself. It includes information
1043 /// about what kinds of things can be folded into the user, information about
1044 /// the user itself, and information about how the use may be satisfied.
1045 /// TODO: Represent multiple users of the same expression in common?
1047 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1050 /// KindType - An enum for a kind of use, indicating what types of
1051 /// scaled and immediate operands it might support.
1053 Basic, ///< A normal use, with no folding.
1054 Special, ///< A special case of basic, allowing -1 scales.
1055 Address, ///< An address use; folding according to TargetLowering
1056 ICmpZero ///< An equality icmp with both operands folded into one.
1057 // TODO: Add a generic icmp too?
1063 SmallVector<int64_t, 8> Offsets;
1067 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1068 /// LSRUse are outside of the loop, in which case some special-case heuristics
1070 bool AllFixupsOutsideLoop;
1072 /// WidestFixupType - This records the widest use type for any fixup using
1073 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1074 /// max fixup widths to be equivalent, because the narrower one may be relying
1075 /// on the implicit truncation to truncate away bogus bits.
1076 Type *WidestFixupType;
1078 /// Formulae - A list of ways to build a value that can satisfy this user.
1079 /// After the list is populated, one of these is selected heuristically and
1080 /// used to formulate a replacement for OperandValToReplace in UserInst.
1081 SmallVector<Formula, 12> Formulae;
1083 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1084 SmallPtrSet<const SCEV *, 4> Regs;
1086 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1087 MinOffset(INT64_MAX),
1088 MaxOffset(INT64_MIN),
1089 AllFixupsOutsideLoop(true),
1090 WidestFixupType(0) {}
1092 bool HasFormulaWithSameRegs(const Formula &F) const;
1093 bool InsertFormula(const Formula &F);
1094 void DeleteFormula(Formula &F);
1095 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1097 void print(raw_ostream &OS) const;
1103 /// HasFormula - Test whether this use as a formula which has the same
1104 /// registers as the given formula.
1105 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1106 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1107 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1108 // Unstable sort by host order ok, because this is only used for uniquifying.
1109 std::sort(Key.begin(), Key.end());
1110 return Uniquifier.count(Key);
1113 /// InsertFormula - If the given formula has not yet been inserted, add it to
1114 /// the list, and return true. Return false otherwise.
1115 bool LSRUse::InsertFormula(const Formula &F) {
1116 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1117 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1118 // Unstable sort by host order ok, because this is only used for uniquifying.
1119 std::sort(Key.begin(), Key.end());
1121 if (!Uniquifier.insert(Key).second)
1124 // Using a register to hold the value of 0 is not profitable.
1125 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1126 "Zero allocated in a scaled register!");
1128 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1129 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1130 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1133 // Add the formula to the list.
1134 Formulae.push_back(F);
1136 // Record registers now being used by this use.
1137 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1142 /// DeleteFormula - Remove the given formula from this use's list.
1143 void LSRUse::DeleteFormula(Formula &F) {
1144 if (&F != &Formulae.back())
1145 std::swap(F, Formulae.back());
1146 Formulae.pop_back();
1149 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1150 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1151 // Now that we've filtered out some formulae, recompute the Regs set.
1152 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1154 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1155 E = Formulae.end(); I != E; ++I) {
1156 const Formula &F = *I;
1157 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1158 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1161 // Update the RegTracker.
1162 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1163 E = OldRegs.end(); I != E; ++I)
1164 if (!Regs.count(*I))
1165 RegUses.DropRegister(*I, LUIdx);
1168 void LSRUse::print(raw_ostream &OS) const {
1169 OS << "LSR Use: Kind=";
1171 case Basic: OS << "Basic"; break;
1172 case Special: OS << "Special"; break;
1173 case ICmpZero: OS << "ICmpZero"; break;
1175 OS << "Address of ";
1176 if (AccessTy->isPointerTy())
1177 OS << "pointer"; // the full pointer type could be really verbose
1182 OS << ", Offsets={";
1183 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1184 E = Offsets.end(); I != E; ++I) {
1186 if (llvm::next(I) != E)
1191 if (AllFixupsOutsideLoop)
1192 OS << ", all-fixups-outside-loop";
1194 if (WidestFixupType)
1195 OS << ", widest fixup type: " << *WidestFixupType;
1198 void LSRUse::dump() const {
1199 print(errs()); errs() << '\n';
1202 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1203 /// be completely folded into the user instruction at isel time. This includes
1204 /// address-mode folding and special icmp tricks.
1205 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1206 LSRUse::KindType Kind, Type *AccessTy,
1207 const TargetLowering *TLI) {
1209 case LSRUse::Address:
1210 // If we have low-level target information, ask the target if it can
1211 // completely fold this address.
1212 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1214 // Otherwise, just guess that reg+reg addressing is legal.
1215 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1217 case LSRUse::ICmpZero:
1218 // There's not even a target hook for querying whether it would be legal to
1219 // fold a GV into an ICmp.
1223 // ICmp only has two operands; don't allow more than two non-trivial parts.
1224 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1227 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1228 // putting the scaled register in the other operand of the icmp.
1229 if (AM.Scale != 0 && AM.Scale != -1)
1232 // If we have low-level target information, ask the target if it can fold an
1233 // integer immediate on an icmp.
1234 if (AM.BaseOffs != 0) {
1235 if (TLI) return TLI->isLegalICmpImmediate(-(uint64_t)AM.BaseOffs);
1242 // Only handle single-register values.
1243 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1245 case LSRUse::Special:
1246 // Only handle -1 scales, or no scale.
1247 return AM.Scale == 0 || AM.Scale == -1;
1253 static bool isLegalUse(TargetLowering::AddrMode AM,
1254 int64_t MinOffset, int64_t MaxOffset,
1255 LSRUse::KindType Kind, Type *AccessTy,
1256 const TargetLowering *TLI) {
1257 // Check for overflow.
1258 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1261 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1262 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1263 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1264 // Check for overflow.
1265 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1268 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1269 return isLegalUse(AM, Kind, AccessTy, TLI);
1274 static bool isAlwaysFoldable(int64_t BaseOffs,
1275 GlobalValue *BaseGV,
1277 LSRUse::KindType Kind, Type *AccessTy,
1278 const TargetLowering *TLI) {
1279 // Fast-path: zero is always foldable.
1280 if (BaseOffs == 0 && !BaseGV) return true;
1282 // Conservatively, create an address with an immediate and a
1283 // base and a scale.
1284 TargetLowering::AddrMode AM;
1285 AM.BaseOffs = BaseOffs;
1287 AM.HasBaseReg = HasBaseReg;
1288 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1290 // Canonicalize a scale of 1 to a base register if the formula doesn't
1291 // already have a base register.
1292 if (!AM.HasBaseReg && AM.Scale == 1) {
1294 AM.HasBaseReg = true;
1297 return isLegalUse(AM, Kind, AccessTy, TLI);
1300 static bool isAlwaysFoldable(const SCEV *S,
1301 int64_t MinOffset, int64_t MaxOffset,
1303 LSRUse::KindType Kind, Type *AccessTy,
1304 const TargetLowering *TLI,
1305 ScalarEvolution &SE) {
1306 // Fast-path: zero is always foldable.
1307 if (S->isZero()) return true;
1309 // Conservatively, create an address with an immediate and a
1310 // base and a scale.
1311 int64_t BaseOffs = ExtractImmediate(S, SE);
1312 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1314 // If there's anything else involved, it's not foldable.
1315 if (!S->isZero()) return false;
1317 // Fast-path: zero is always foldable.
1318 if (BaseOffs == 0 && !BaseGV) return true;
1320 // Conservatively, create an address with an immediate and a
1321 // base and a scale.
1322 TargetLowering::AddrMode AM;
1323 AM.BaseOffs = BaseOffs;
1325 AM.HasBaseReg = HasBaseReg;
1326 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1328 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1333 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1334 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1335 struct UseMapDenseMapInfo {
1336 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1337 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1340 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1341 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1345 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1346 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1347 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1351 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1352 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1357 /// IVInc - An individual increment in a Chain of IV increments.
1358 /// Relate an IV user to an expression that computes the IV it uses from the IV
1359 /// used by the previous link in the Chain.
1361 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1362 /// original IVOperand. The head of the chain's IVOperand is only valid during
1363 /// chain collection, before LSR replaces IV users. During chain generation,
1364 /// IncExpr can be used to find the new IVOperand that computes the same
1367 Instruction *UserInst;
1369 const SCEV *IncExpr;
1371 IVInc(Instruction *U, Value *O, const SCEV *E):
1372 UserInst(U), IVOperand(O), IncExpr(E) {}
1375 // IVChain - The list of IV increments in program order.
1376 // We typically add the head of a chain without finding subsequent links.
1377 typedef SmallVector<IVInc,1> IVChain;
1379 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1380 /// Distinguish between FarUsers that definitely cross IV increments and
1381 /// NearUsers that may be used between IV increments.
1383 SmallPtrSet<Instruction*, 4> FarUsers;
1384 SmallPtrSet<Instruction*, 4> NearUsers;
1387 /// LSRInstance - This class holds state for the main loop strength reduction
1391 ScalarEvolution &SE;
1394 const TargetLowering *const TLI;
1398 /// IVIncInsertPos - This is the insert position that the current loop's
1399 /// induction variable increment should be placed. In simple loops, this is
1400 /// the latch block's terminator. But in more complicated cases, this is a
1401 /// position which will dominate all the in-loop post-increment users.
1402 Instruction *IVIncInsertPos;
1404 /// Factors - Interesting factors between use strides.
1405 SmallSetVector<int64_t, 8> Factors;
1407 /// Types - Interesting use types, to facilitate truncation reuse.
1408 SmallSetVector<Type *, 4> Types;
1410 /// Fixups - The list of operands which are to be replaced.
1411 SmallVector<LSRFixup, 16> Fixups;
1413 /// Uses - The list of interesting uses.
1414 SmallVector<LSRUse, 16> Uses;
1416 /// RegUses - Track which uses use which register candidates.
1417 RegUseTracker RegUses;
1419 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1420 // have more than a few IV increment chains in a loop. Missing a Chain falls
1421 // back to normal LSR behavior for those uses.
1422 static const unsigned MaxChains = 8;
1424 /// IVChainVec - IV users can form a chain of IV increments.
1425 SmallVector<IVChain, MaxChains> IVChainVec;
1427 /// IVIncSet - IV users that belong to profitable IVChains.
1428 SmallPtrSet<Use*, MaxChains> IVIncSet;
1430 void OptimizeShadowIV();
1431 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1432 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1433 void OptimizeLoopTermCond();
1435 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1436 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1437 void FinalizeChain(IVChain &Chain);
1438 void CollectChains();
1439 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1440 SmallVectorImpl<WeakVH> &DeadInsts);
1442 void CollectInterestingTypesAndFactors();
1443 void CollectFixupsAndInitialFormulae();
1445 LSRFixup &getNewFixup() {
1446 Fixups.push_back(LSRFixup());
1447 return Fixups.back();
1450 // Support for sharing of LSRUses between LSRFixups.
1451 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1453 UseMapDenseMapInfo> UseMapTy;
1456 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1457 LSRUse::KindType Kind, Type *AccessTy);
1459 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1460 LSRUse::KindType Kind,
1463 void DeleteUse(LSRUse &LU, size_t LUIdx);
1465 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1467 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1468 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1469 void CountRegisters(const Formula &F, size_t LUIdx);
1470 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1472 void CollectLoopInvariantFixupsAndFormulae();
1474 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1475 unsigned Depth = 0);
1476 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1477 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1478 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1479 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1480 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1481 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1482 void GenerateCrossUseConstantOffsets();
1483 void GenerateAllReuseFormulae();
1485 void FilterOutUndesirableDedicatedRegisters();
1487 size_t EstimateSearchSpaceComplexity() const;
1488 void NarrowSearchSpaceByDetectingSupersets();
1489 void NarrowSearchSpaceByCollapsingUnrolledCode();
1490 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1491 void NarrowSearchSpaceByPickingWinnerRegs();
1492 void NarrowSearchSpaceUsingHeuristics();
1494 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1496 SmallVectorImpl<const Formula *> &Workspace,
1497 const Cost &CurCost,
1498 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1499 DenseSet<const SCEV *> &VisitedRegs) const;
1500 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1502 BasicBlock::iterator
1503 HoistInsertPosition(BasicBlock::iterator IP,
1504 const SmallVectorImpl<Instruction *> &Inputs) const;
1505 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1507 const LSRUse &LU) const;
1509 Value *Expand(const LSRFixup &LF,
1511 BasicBlock::iterator IP,
1512 SCEVExpander &Rewriter,
1513 SmallVectorImpl<WeakVH> &DeadInsts) const;
1514 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1516 SCEVExpander &Rewriter,
1517 SmallVectorImpl<WeakVH> &DeadInsts,
1519 void Rewrite(const LSRFixup &LF,
1521 SCEVExpander &Rewriter,
1522 SmallVectorImpl<WeakVH> &DeadInsts,
1524 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1528 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1530 bool getChanged() const { return Changed; }
1532 void print_factors_and_types(raw_ostream &OS) const;
1533 void print_fixups(raw_ostream &OS) const;
1534 void print_uses(raw_ostream &OS) const;
1535 void print(raw_ostream &OS) const;
1541 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1542 /// inside the loop then try to eliminate the cast operation.
1543 void LSRInstance::OptimizeShadowIV() {
1544 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1545 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1548 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1549 UI != E; /* empty */) {
1550 IVUsers::const_iterator CandidateUI = UI;
1552 Instruction *ShadowUse = CandidateUI->getUser();
1553 Type *DestTy = NULL;
1554 bool IsSigned = false;
1556 /* If shadow use is a int->float cast then insert a second IV
1557 to eliminate this cast.
1559 for (unsigned i = 0; i < n; ++i)
1565 for (unsigned i = 0; i < n; ++i, ++d)
1568 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1570 DestTy = UCast->getDestTy();
1572 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1574 DestTy = SCast->getDestTy();
1576 if (!DestTy) continue;
1579 // If target does not support DestTy natively then do not apply
1580 // this transformation.
1581 EVT DVT = TLI->getValueType(DestTy);
1582 if (!TLI->isTypeLegal(DVT)) continue;
1585 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1587 if (PH->getNumIncomingValues() != 2) continue;
1589 Type *SrcTy = PH->getType();
1590 int Mantissa = DestTy->getFPMantissaWidth();
1591 if (Mantissa == -1) continue;
1592 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1595 unsigned Entry, Latch;
1596 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1604 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1605 if (!Init) continue;
1606 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1607 (double)Init->getSExtValue() :
1608 (double)Init->getZExtValue());
1610 BinaryOperator *Incr =
1611 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1612 if (!Incr) continue;
1613 if (Incr->getOpcode() != Instruction::Add
1614 && Incr->getOpcode() != Instruction::Sub)
1617 /* Initialize new IV, double d = 0.0 in above example. */
1618 ConstantInt *C = NULL;
1619 if (Incr->getOperand(0) == PH)
1620 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1621 else if (Incr->getOperand(1) == PH)
1622 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1628 // Ignore negative constants, as the code below doesn't handle them
1629 // correctly. TODO: Remove this restriction.
1630 if (!C->getValue().isStrictlyPositive()) continue;
1632 /* Add new PHINode. */
1633 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1635 /* create new increment. '++d' in above example. */
1636 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1637 BinaryOperator *NewIncr =
1638 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1639 Instruction::FAdd : Instruction::FSub,
1640 NewPH, CFP, "IV.S.next.", Incr);
1642 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1643 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1645 /* Remove cast operation */
1646 ShadowUse->replaceAllUsesWith(NewPH);
1647 ShadowUse->eraseFromParent();
1653 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1654 /// set the IV user and stride information and return true, otherwise return
1656 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1657 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1658 if (UI->getUser() == Cond) {
1659 // NOTE: we could handle setcc instructions with multiple uses here, but
1660 // InstCombine does it as well for simple uses, it's not clear that it
1661 // occurs enough in real life to handle.
1668 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1669 /// a max computation.
1671 /// This is a narrow solution to a specific, but acute, problem. For loops
1677 /// } while (++i < n);
1679 /// the trip count isn't just 'n', because 'n' might not be positive. And
1680 /// unfortunately this can come up even for loops where the user didn't use
1681 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1682 /// will commonly be lowered like this:
1688 /// } while (++i < n);
1691 /// and then it's possible for subsequent optimization to obscure the if
1692 /// test in such a way that indvars can't find it.
1694 /// When indvars can't find the if test in loops like this, it creates a
1695 /// max expression, which allows it to give the loop a canonical
1696 /// induction variable:
1699 /// max = n < 1 ? 1 : n;
1702 /// } while (++i != max);
1704 /// Canonical induction variables are necessary because the loop passes
1705 /// are designed around them. The most obvious example of this is the
1706 /// LoopInfo analysis, which doesn't remember trip count values. It
1707 /// expects to be able to rediscover the trip count each time it is
1708 /// needed, and it does this using a simple analysis that only succeeds if
1709 /// the loop has a canonical induction variable.
1711 /// However, when it comes time to generate code, the maximum operation
1712 /// can be quite costly, especially if it's inside of an outer loop.
1714 /// This function solves this problem by detecting this type of loop and
1715 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1716 /// the instructions for the maximum computation.
1718 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1719 // Check that the loop matches the pattern we're looking for.
1720 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1721 Cond->getPredicate() != CmpInst::ICMP_NE)
1724 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1725 if (!Sel || !Sel->hasOneUse()) return Cond;
1727 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1728 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1730 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1732 // Add one to the backedge-taken count to get the trip count.
1733 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1734 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1736 // Check for a max calculation that matches the pattern. There's no check
1737 // for ICMP_ULE here because the comparison would be with zero, which
1738 // isn't interesting.
1739 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1740 const SCEVNAryExpr *Max = 0;
1741 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1742 Pred = ICmpInst::ICMP_SLE;
1744 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1745 Pred = ICmpInst::ICMP_SLT;
1747 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1748 Pred = ICmpInst::ICMP_ULT;
1755 // To handle a max with more than two operands, this optimization would
1756 // require additional checking and setup.
1757 if (Max->getNumOperands() != 2)
1760 const SCEV *MaxLHS = Max->getOperand(0);
1761 const SCEV *MaxRHS = Max->getOperand(1);
1763 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1764 // for a comparison with 1. For <= and >=, a comparison with zero.
1766 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1769 // Check the relevant induction variable for conformance to
1771 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1772 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1773 if (!AR || !AR->isAffine() ||
1774 AR->getStart() != One ||
1775 AR->getStepRecurrence(SE) != One)
1778 assert(AR->getLoop() == L &&
1779 "Loop condition operand is an addrec in a different loop!");
1781 // Check the right operand of the select, and remember it, as it will
1782 // be used in the new comparison instruction.
1784 if (ICmpInst::isTrueWhenEqual(Pred)) {
1785 // Look for n+1, and grab n.
1786 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1787 if (isa<ConstantInt>(BO->getOperand(1)) &&
1788 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1789 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1790 NewRHS = BO->getOperand(0);
1791 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1792 if (isa<ConstantInt>(BO->getOperand(1)) &&
1793 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1794 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1795 NewRHS = BO->getOperand(0);
1798 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1799 NewRHS = Sel->getOperand(1);
1800 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1801 NewRHS = Sel->getOperand(2);
1802 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1803 NewRHS = SU->getValue();
1805 // Max doesn't match expected pattern.
1808 // Determine the new comparison opcode. It may be signed or unsigned,
1809 // and the original comparison may be either equality or inequality.
1810 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1811 Pred = CmpInst::getInversePredicate(Pred);
1813 // Ok, everything looks ok to change the condition into an SLT or SGE and
1814 // delete the max calculation.
1816 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1818 // Delete the max calculation instructions.
1819 Cond->replaceAllUsesWith(NewCond);
1820 CondUse->setUser(NewCond);
1821 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1822 Cond->eraseFromParent();
1823 Sel->eraseFromParent();
1824 if (Cmp->use_empty())
1825 Cmp->eraseFromParent();
1829 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1830 /// postinc iv when possible.
1832 LSRInstance::OptimizeLoopTermCond() {
1833 SmallPtrSet<Instruction *, 4> PostIncs;
1835 BasicBlock *LatchBlock = L->getLoopLatch();
1836 SmallVector<BasicBlock*, 8> ExitingBlocks;
1837 L->getExitingBlocks(ExitingBlocks);
1839 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1840 BasicBlock *ExitingBlock = ExitingBlocks[i];
1842 // Get the terminating condition for the loop if possible. If we
1843 // can, we want to change it to use a post-incremented version of its
1844 // induction variable, to allow coalescing the live ranges for the IV into
1845 // one register value.
1847 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1850 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1851 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1854 // Search IVUsesByStride to find Cond's IVUse if there is one.
1855 IVStrideUse *CondUse = 0;
1856 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1857 if (!FindIVUserForCond(Cond, CondUse))
1860 // If the trip count is computed in terms of a max (due to ScalarEvolution
1861 // being unable to find a sufficient guard, for example), change the loop
1862 // comparison to use SLT or ULT instead of NE.
1863 // One consequence of doing this now is that it disrupts the count-down
1864 // optimization. That's not always a bad thing though, because in such
1865 // cases it may still be worthwhile to avoid a max.
1866 Cond = OptimizeMax(Cond, CondUse);
1868 // If this exiting block dominates the latch block, it may also use
1869 // the post-inc value if it won't be shared with other uses.
1870 // Check for dominance.
1871 if (!DT.dominates(ExitingBlock, LatchBlock))
1874 // Conservatively avoid trying to use the post-inc value in non-latch
1875 // exits if there may be pre-inc users in intervening blocks.
1876 if (LatchBlock != ExitingBlock)
1877 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1878 // Test if the use is reachable from the exiting block. This dominator
1879 // query is a conservative approximation of reachability.
1880 if (&*UI != CondUse &&
1881 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1882 // Conservatively assume there may be reuse if the quotient of their
1883 // strides could be a legal scale.
1884 const SCEV *A = IU.getStride(*CondUse, L);
1885 const SCEV *B = IU.getStride(*UI, L);
1886 if (!A || !B) continue;
1887 if (SE.getTypeSizeInBits(A->getType()) !=
1888 SE.getTypeSizeInBits(B->getType())) {
1889 if (SE.getTypeSizeInBits(A->getType()) >
1890 SE.getTypeSizeInBits(B->getType()))
1891 B = SE.getSignExtendExpr(B, A->getType());
1893 A = SE.getSignExtendExpr(A, B->getType());
1895 if (const SCEVConstant *D =
1896 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1897 const ConstantInt *C = D->getValue();
1898 // Stride of one or negative one can have reuse with non-addresses.
1899 if (C->isOne() || C->isAllOnesValue())
1900 goto decline_post_inc;
1901 // Avoid weird situations.
1902 if (C->getValue().getMinSignedBits() >= 64 ||
1903 C->getValue().isMinSignedValue())
1904 goto decline_post_inc;
1905 // Without TLI, assume that any stride might be valid, and so any
1906 // use might be shared.
1908 goto decline_post_inc;
1909 // Check for possible scaled-address reuse.
1910 Type *AccessTy = getAccessType(UI->getUser());
1911 TargetLowering::AddrMode AM;
1912 AM.Scale = C->getSExtValue();
1913 if (TLI->isLegalAddressingMode(AM, AccessTy))
1914 goto decline_post_inc;
1915 AM.Scale = -AM.Scale;
1916 if (TLI->isLegalAddressingMode(AM, AccessTy))
1917 goto decline_post_inc;
1921 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1924 // It's possible for the setcc instruction to be anywhere in the loop, and
1925 // possible for it to have multiple users. If it is not immediately before
1926 // the exiting block branch, move it.
1927 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1928 if (Cond->hasOneUse()) {
1929 Cond->moveBefore(TermBr);
1931 // Clone the terminating condition and insert into the loopend.
1932 ICmpInst *OldCond = Cond;
1933 Cond = cast<ICmpInst>(Cond->clone());
1934 Cond->setName(L->getHeader()->getName() + ".termcond");
1935 ExitingBlock->getInstList().insert(TermBr, Cond);
1937 // Clone the IVUse, as the old use still exists!
1938 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1939 TermBr->replaceUsesOfWith(OldCond, Cond);
1943 // If we get to here, we know that we can transform the setcc instruction to
1944 // use the post-incremented version of the IV, allowing us to coalesce the
1945 // live ranges for the IV correctly.
1946 CondUse->transformToPostInc(L);
1949 PostIncs.insert(Cond);
1953 // Determine an insertion point for the loop induction variable increment. It
1954 // must dominate all the post-inc comparisons we just set up, and it must
1955 // dominate the loop latch edge.
1956 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1957 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1958 E = PostIncs.end(); I != E; ++I) {
1960 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1962 if (BB == (*I)->getParent())
1963 IVIncInsertPos = *I;
1964 else if (BB != IVIncInsertPos->getParent())
1965 IVIncInsertPos = BB->getTerminator();
1969 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
1970 /// at the given offset and other details. If so, update the use and
1973 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1974 LSRUse::KindType Kind, Type *AccessTy) {
1975 int64_t NewMinOffset = LU.MinOffset;
1976 int64_t NewMaxOffset = LU.MaxOffset;
1977 Type *NewAccessTy = AccessTy;
1979 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1980 // something conservative, however this can pessimize in the case that one of
1981 // the uses will have all its uses outside the loop, for example.
1982 if (LU.Kind != Kind)
1984 // Conservatively assume HasBaseReg is true for now.
1985 if (NewOffset < LU.MinOffset) {
1986 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
1987 Kind, AccessTy, TLI))
1989 NewMinOffset = NewOffset;
1990 } else if (NewOffset > LU.MaxOffset) {
1991 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
1992 Kind, AccessTy, TLI))
1994 NewMaxOffset = NewOffset;
1996 // Check for a mismatched access type, and fall back conservatively as needed.
1997 // TODO: Be less conservative when the type is similar and can use the same
1998 // addressing modes.
1999 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2000 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2003 LU.MinOffset = NewMinOffset;
2004 LU.MaxOffset = NewMaxOffset;
2005 LU.AccessTy = NewAccessTy;
2006 if (NewOffset != LU.Offsets.back())
2007 LU.Offsets.push_back(NewOffset);
2011 /// getUse - Return an LSRUse index and an offset value for a fixup which
2012 /// needs the given expression, with the given kind and optional access type.
2013 /// Either reuse an existing use or create a new one, as needed.
2014 std::pair<size_t, int64_t>
2015 LSRInstance::getUse(const SCEV *&Expr,
2016 LSRUse::KindType Kind, Type *AccessTy) {
2017 const SCEV *Copy = Expr;
2018 int64_t Offset = ExtractImmediate(Expr, SE);
2020 // Basic uses can't accept any offset, for example.
2021 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
2026 std::pair<UseMapTy::iterator, bool> P =
2027 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2029 // A use already existed with this base.
2030 size_t LUIdx = P.first->second;
2031 LSRUse &LU = Uses[LUIdx];
2032 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2034 return std::make_pair(LUIdx, Offset);
2037 // Create a new use.
2038 size_t LUIdx = Uses.size();
2039 P.first->second = LUIdx;
2040 Uses.push_back(LSRUse(Kind, AccessTy));
2041 LSRUse &LU = Uses[LUIdx];
2043 // We don't need to track redundant offsets, but we don't need to go out
2044 // of our way here to avoid them.
2045 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2046 LU.Offsets.push_back(Offset);
2048 LU.MinOffset = Offset;
2049 LU.MaxOffset = Offset;
2050 return std::make_pair(LUIdx, Offset);
2053 /// DeleteUse - Delete the given use from the Uses list.
2054 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2055 if (&LU != &Uses.back())
2056 std::swap(LU, Uses.back());
2060 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2063 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2064 /// a formula that has the same registers as the given formula.
2066 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2067 const LSRUse &OrigLU) {
2068 // Search all uses for the formula. This could be more clever.
2069 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2070 LSRUse &LU = Uses[LUIdx];
2071 // Check whether this use is close enough to OrigLU, to see whether it's
2072 // worthwhile looking through its formulae.
2073 // Ignore ICmpZero uses because they may contain formulae generated by
2074 // GenerateICmpZeroScales, in which case adding fixup offsets may
2076 if (&LU != &OrigLU &&
2077 LU.Kind != LSRUse::ICmpZero &&
2078 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2079 LU.WidestFixupType == OrigLU.WidestFixupType &&
2080 LU.HasFormulaWithSameRegs(OrigF)) {
2081 // Scan through this use's formulae.
2082 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2083 E = LU.Formulae.end(); I != E; ++I) {
2084 const Formula &F = *I;
2085 // Check to see if this formula has the same registers and symbols
2087 if (F.BaseRegs == OrigF.BaseRegs &&
2088 F.ScaledReg == OrigF.ScaledReg &&
2089 F.AM.BaseGV == OrigF.AM.BaseGV &&
2090 F.AM.Scale == OrigF.AM.Scale &&
2091 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2092 if (F.AM.BaseOffs == 0)
2094 // This is the formula where all the registers and symbols matched;
2095 // there aren't going to be any others. Since we declined it, we
2096 // can skip the rest of the formulae and procede to the next LSRUse.
2103 // Nothing looked good.
2107 void LSRInstance::CollectInterestingTypesAndFactors() {
2108 SmallSetVector<const SCEV *, 4> Strides;
2110 // Collect interesting types and strides.
2111 SmallVector<const SCEV *, 4> Worklist;
2112 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2113 const SCEV *Expr = IU.getExpr(*UI);
2115 // Collect interesting types.
2116 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2118 // Add strides for mentioned loops.
2119 Worklist.push_back(Expr);
2121 const SCEV *S = Worklist.pop_back_val();
2122 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2123 if (EnableNested || AR->getLoop() == L)
2124 Strides.insert(AR->getStepRecurrence(SE));
2125 Worklist.push_back(AR->getStart());
2126 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2127 Worklist.append(Add->op_begin(), Add->op_end());
2129 } while (!Worklist.empty());
2132 // Compute interesting factors from the set of interesting strides.
2133 for (SmallSetVector<const SCEV *, 4>::const_iterator
2134 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2135 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2136 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2137 const SCEV *OldStride = *I;
2138 const SCEV *NewStride = *NewStrideIter;
2140 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2141 SE.getTypeSizeInBits(NewStride->getType())) {
2142 if (SE.getTypeSizeInBits(OldStride->getType()) >
2143 SE.getTypeSizeInBits(NewStride->getType()))
2144 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2146 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2148 if (const SCEVConstant *Factor =
2149 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2151 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2152 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2153 } else if (const SCEVConstant *Factor =
2154 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2157 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2158 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2162 // If all uses use the same type, don't bother looking for truncation-based
2164 if (Types.size() == 1)
2167 DEBUG(print_factors_and_types(dbgs()));
2170 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2171 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2172 /// Instructions to IVStrideUses, we could partially skip this.
2173 static User::op_iterator
2174 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2175 Loop *L, ScalarEvolution &SE) {
2176 for(; OI != OE; ++OI) {
2177 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2178 if (!SE.isSCEVable(Oper->getType()))
2181 if (const SCEVAddRecExpr *AR =
2182 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2183 if (AR->getLoop() == L)
2191 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2192 /// operands, so wrap it in a convenient helper.
2193 static Value *getWideOperand(Value *Oper) {
2194 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2195 return Trunc->getOperand(0);
2199 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2201 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2202 Type *LType = LVal->getType();
2203 Type *RType = RVal->getType();
2204 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2207 /// Return true if the chain increment is profitable to expand into a loop
2208 /// invariant value, which may require its own register. A profitable chain
2209 /// increment will be an offset relative to the same base. We allow such offsets
2210 /// to potentially be used as chain increment as long as it's not obviously
2211 /// expensive to expand using real instructions.
2213 getProfitableChainIncrement(Value *NextIV, Value *PrevIV,
2214 const IVChain &Chain, Loop *L,
2215 ScalarEvolution &SE, const TargetLowering *TLI) {
2216 const SCEV *IncExpr = SE.getMinusSCEV(SE.getSCEV(NextIV), SE.getSCEV(PrevIV));
2217 if (!SE.isLoopInvariant(IncExpr, L))
2220 // We are not able to expand an increment unless it is loop invariant,
2221 // however, the following checks are purely for profitability.
2229 /// Return true if the number of registers needed for the chain is estimated to
2230 /// be less than the number required for the individual IV users. First prohibit
2231 /// any IV users that keep the IV live across increments (the Users set should
2232 /// be empty). Next count the number and type of increments in the chain.
2234 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2235 /// effectively use postinc addressing modes. Only consider it profitable it the
2236 /// increments can be computed in fewer registers when chained.
2238 /// TODO: Consider IVInc free if it's already used in another chains.
2240 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2241 ScalarEvolution &SE, const TargetLowering *TLI) {
2249 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2251 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2252 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2253 // When IVs are used as types of varying widths, they are generally converted
2254 // to a wider type with some uses remaining narrow under a (free) trunc.
2255 Value *NextIV = getWideOperand(IVOper);
2257 // Visit all existing chains. Check if its IVOper can be computed as a
2258 // profitable loop invariant increment from the last link in the Chain.
2259 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2260 const SCEV *LastIncExpr = 0;
2261 for (; ChainIdx < NChains; ++ChainIdx) {
2262 Value *PrevIV = getWideOperand(IVChainVec[ChainIdx].back().IVOperand);
2263 if (!isCompatibleIVType(PrevIV, NextIV))
2266 // A phi nodes terminates a chain.
2267 if (isa<PHINode>(UserInst)
2268 && isa<PHINode>(IVChainVec[ChainIdx].back().UserInst))
2271 if (const SCEV *IncExpr =
2272 getProfitableChainIncrement(NextIV, PrevIV, IVChainVec[ChainIdx],
2274 LastIncExpr = IncExpr;
2278 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2279 // bother for phi nodes, because they must be last in the chain.
2280 if (ChainIdx == NChains) {
2281 if (isa<PHINode>(UserInst))
2283 if (NChains >= MaxChains && !StressIVChain) {
2284 DEBUG(dbgs() << "IV Chain Limit\n");
2288 IVChainVec.resize(NChains);
2289 ChainUsersVec.resize(NChains);
2290 LastIncExpr = SE.getSCEV(NextIV);
2291 assert(isa<SCEVAddRecExpr>(LastIncExpr) && "expect recurrence at IV user");
2292 DEBUG(dbgs() << "IV Head: (" << *UserInst << ") IV=" << *LastIncExpr
2296 DEBUG(dbgs() << "IV Inc: (" << *UserInst << ") IV+" << *LastIncExpr
2299 // Add this IV user to the end of the chain.
2300 IVChainVec[ChainIdx].push_back(IVInc(UserInst, IVOper, LastIncExpr));
2302 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2303 // This chain's NearUsers become FarUsers.
2304 if (!LastIncExpr->isZero()) {
2305 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2310 // All other uses of IVOperand become near uses of the chain.
2311 // We currently ignore intermediate values within SCEV expressions, assuming
2312 // they will eventually be used be the current chain, or can be computed
2313 // from one of the chain increments. To be more precise we could
2314 // transitively follow its user and only add leaf IV users to the set.
2315 for (Value::use_iterator UseIter = IVOper->use_begin(),
2316 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2317 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2318 if (SE.isSCEVable(OtherUse->getType())
2319 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2320 && IU.isIVUserOrOperand(OtherUse)) {
2323 if (OtherUse && OtherUse != UserInst)
2324 NearUsers.insert(OtherUse);
2327 // Since this user is part of the chain, it's no longer considered a use
2329 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2332 /// CollectChains - Populate the vector of Chains.
2334 /// This decreases ILP at the architecture level. Targets with ample registers,
2335 /// multiple memory ports, and no register renaming probably don't want
2336 /// this. However, such targets should probably disable LSR altogether.
2338 /// The job of LSR is to make a reasonable choice of induction variables across
2339 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2340 /// ILP *within the loop* if the target wants it.
2342 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2343 /// will not reorder memory operations, it will recognize this as a chain, but
2344 /// will generate redundant IV increments. Ideally this would be corrected later
2345 /// by a smart scheduler:
2351 /// TODO: Walk the entire domtree within this loop, not just the path to the
2352 /// loop latch. This will discover chains on side paths, but requires
2353 /// maintaining multiple copies of the Chains state.
2354 void LSRInstance::CollectChains() {
2355 SmallVector<ChainUsers, 8> ChainUsersVec;
2357 SmallVector<BasicBlock *,8> LatchPath;
2358 BasicBlock *LoopHeader = L->getHeader();
2359 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2360 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2361 LatchPath.push_back(Rung->getBlock());
2363 LatchPath.push_back(LoopHeader);
2365 // Walk the instruction stream from the loop header to the loop latch.
2366 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2367 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2368 BBIter != BBEnd; ++BBIter) {
2369 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2371 // Skip instructions that weren't seen by IVUsers analysis.
2372 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2375 // Ignore users that are part of a SCEV expression. This way we only
2376 // consider leaf IV Users. This effectively rediscovers a portion of
2377 // IVUsers analysis but in program order this time.
2378 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2381 // Remove this instruction from any NearUsers set it may be in.
2382 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2383 ChainIdx < NChains; ++ChainIdx) {
2384 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2386 // Search for operands that can be chained.
2387 SmallPtrSet<Instruction*, 4> UniqueOperands;
2388 User::op_iterator IVOpEnd = I->op_end();
2389 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2390 while (IVOpIter != IVOpEnd) {
2391 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2392 if (UniqueOperands.insert(IVOpInst))
2393 ChainInstruction(I, IVOpInst, ChainUsersVec);
2394 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2396 } // Continue walking down the instructions.
2397 } // Continue walking down the domtree.
2398 // Visit phi backedges to determine if the chain can generate the IV postinc.
2399 for (BasicBlock::iterator I = L->getHeader()->begin();
2400 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2401 if (!SE.isSCEVable(PN->getType()))
2405 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2407 ChainInstruction(PN, IncV, ChainUsersVec);
2409 // Remove any unprofitable chains.
2410 unsigned ChainIdx = 0;
2411 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2412 UsersIdx < NChains; ++UsersIdx) {
2413 if (!isProfitableChain(IVChainVec[UsersIdx],
2414 ChainUsersVec[UsersIdx].FarUsers, SE, TLI))
2416 // Preserve the chain at UsesIdx.
2417 if (ChainIdx != UsersIdx)
2418 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2419 FinalizeChain(IVChainVec[ChainIdx]);
2422 IVChainVec.resize(ChainIdx);
2425 void LSRInstance::FinalizeChain(IVChain &Chain) {
2426 assert(!Chain.empty() && "empty IV chains are not allowed");
2427 DEBUG(dbgs() << "Final Chain: " << *Chain[0].UserInst << "\n");
2429 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2431 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2432 User::op_iterator UseI =
2433 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2434 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2435 IVIncSet.insert(UseI);
2439 /// Return true if the IVInc can be folded into an addressing mode.
2440 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2441 Value *Operand, const TargetLowering *TLI) {
2442 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2443 if (!IncConst || !isAddressUse(UserInst, Operand))
2446 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2449 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2450 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false,
2451 LSRUse::Address, getAccessType(UserInst), TLI))
2457 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2458 /// materialize the IV user's operand from the previous IV user's operand.
2459 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2460 SmallVectorImpl<WeakVH> &DeadInsts) {
2461 // Find the new IVOperand for the head of the chain. It may have been replaced
2463 const IVInc &Head = Chain[0];
2464 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2465 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2468 while (IVOpIter != IVOpEnd) {
2469 IVSrc = getWideOperand(*IVOpIter);
2471 // If this operand computes the expression that the chain needs, we may use
2472 // it. (Check this after setting IVSrc which is used below.)
2474 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2475 // narrow for the chain, so we can no longer use it. We do allow using a
2476 // wider phi, assuming the LSR checked for free truncation. In that case we
2477 // should already have a truncate on this operand such that
2478 // getSCEV(IVSrc) == IncExpr.
2479 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2480 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2483 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2485 if (IVOpIter == IVOpEnd) {
2486 // Gracefully give up on this chain.
2487 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2491 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2492 Type *IVTy = IVSrc->getType();
2493 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2494 const SCEV *LeftOverExpr = 0;
2495 for (IVChain::const_iterator IncI = llvm::next(Chain.begin()),
2496 IncE = Chain.end(); IncI != IncE; ++IncI) {
2498 Instruction *InsertPt = IncI->UserInst;
2499 if (isa<PHINode>(InsertPt))
2500 InsertPt = L->getLoopLatch()->getTerminator();
2502 // IVOper will replace the current IV User's operand. IVSrc is the IV
2503 // value currently held in a register.
2504 Value *IVOper = IVSrc;
2505 if (!IncI->IncExpr->isZero()) {
2506 // IncExpr was the result of subtraction of two narrow values, so must
2508 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2509 LeftOverExpr = LeftOverExpr ?
2510 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2512 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2513 // Expand the IV increment.
2514 Rewriter.clearPostInc();
2515 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2516 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2517 SE.getUnknown(IncV));
2518 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2520 // If an IV increment can't be folded, use it as the next IV value.
2521 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2523 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2528 Type *OperTy = IncI->IVOperand->getType();
2529 if (IVTy != OperTy) {
2530 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2531 "cannot extend a chained IV");
2532 IRBuilder<> Builder(InsertPt);
2533 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2535 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2536 DeadInsts.push_back(IncI->IVOperand);
2538 // If LSR created a new, wider phi, we may also replace its postinc. We only
2539 // do this if we also found a wide value for the head of the chain.
2540 if (isa<PHINode>(Chain.back().UserInst)) {
2541 for (BasicBlock::iterator I = L->getHeader()->begin();
2542 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2543 if (!isCompatibleIVType(Phi, IVSrc))
2545 Instruction *PostIncV = dyn_cast<Instruction>(
2546 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2547 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2549 Value *IVOper = IVSrc;
2550 Type *PostIncTy = PostIncV->getType();
2551 if (IVTy != PostIncTy) {
2552 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2553 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2554 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2555 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2557 Phi->replaceUsesOfWith(PostIncV, IVOper);
2558 DeadInsts.push_back(PostIncV);
2563 void LSRInstance::CollectFixupsAndInitialFormulae() {
2564 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2565 Instruction *UserInst = UI->getUser();
2566 // Skip IV users that are part of profitable IV Chains.
2567 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2568 UI->getOperandValToReplace());
2569 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2570 if (IVIncSet.count(UseI))
2574 LSRFixup &LF = getNewFixup();
2575 LF.UserInst = UserInst;
2576 LF.OperandValToReplace = UI->getOperandValToReplace();
2577 LF.PostIncLoops = UI->getPostIncLoops();
2579 LSRUse::KindType Kind = LSRUse::Basic;
2581 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2582 Kind = LSRUse::Address;
2583 AccessTy = getAccessType(LF.UserInst);
2586 const SCEV *S = IU.getExpr(*UI);
2588 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2589 // (N - i == 0), and this allows (N - i) to be the expression that we work
2590 // with rather than just N or i, so we can consider the register
2591 // requirements for both N and i at the same time. Limiting this code to
2592 // equality icmps is not a problem because all interesting loops use
2593 // equality icmps, thanks to IndVarSimplify.
2594 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2595 if (CI->isEquality()) {
2596 // Swap the operands if needed to put the OperandValToReplace on the
2597 // left, for consistency.
2598 Value *NV = CI->getOperand(1);
2599 if (NV == LF.OperandValToReplace) {
2600 CI->setOperand(1, CI->getOperand(0));
2601 CI->setOperand(0, NV);
2602 NV = CI->getOperand(1);
2606 // x == y --> x - y == 0
2607 const SCEV *N = SE.getSCEV(NV);
2608 if (SE.isLoopInvariant(N, L)) {
2609 // S is normalized, so normalize N before folding it into S
2610 // to keep the result normalized.
2611 N = TransformForPostIncUse(Normalize, N, CI, 0,
2612 LF.PostIncLoops, SE, DT);
2613 Kind = LSRUse::ICmpZero;
2614 S = SE.getMinusSCEV(N, S);
2617 // -1 and the negations of all interesting strides (except the negation
2618 // of -1) are now also interesting.
2619 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2620 if (Factors[i] != -1)
2621 Factors.insert(-(uint64_t)Factors[i]);
2625 // Set up the initial formula for this use.
2626 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2628 LF.Offset = P.second;
2629 LSRUse &LU = Uses[LF.LUIdx];
2630 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2631 if (!LU.WidestFixupType ||
2632 SE.getTypeSizeInBits(LU.WidestFixupType) <
2633 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2634 LU.WidestFixupType = LF.OperandValToReplace->getType();
2636 // If this is the first use of this LSRUse, give it a formula.
2637 if (LU.Formulae.empty()) {
2638 InsertInitialFormula(S, LU, LF.LUIdx);
2639 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2643 DEBUG(print_fixups(dbgs()));
2646 /// InsertInitialFormula - Insert a formula for the given expression into
2647 /// the given use, separating out loop-variant portions from loop-invariant
2648 /// and loop-computable portions.
2650 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2652 F.InitialMatch(S, L, SE);
2653 bool Inserted = InsertFormula(LU, LUIdx, F);
2654 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2657 /// InsertSupplementalFormula - Insert a simple single-register formula for
2658 /// the given expression into the given use.
2660 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2661 LSRUse &LU, size_t LUIdx) {
2663 F.BaseRegs.push_back(S);
2664 F.AM.HasBaseReg = true;
2665 bool Inserted = InsertFormula(LU, LUIdx, F);
2666 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2669 /// CountRegisters - Note which registers are used by the given formula,
2670 /// updating RegUses.
2671 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2673 RegUses.CountRegister(F.ScaledReg, LUIdx);
2674 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2675 E = F.BaseRegs.end(); I != E; ++I)
2676 RegUses.CountRegister(*I, LUIdx);
2679 /// InsertFormula - If the given formula has not yet been inserted, add it to
2680 /// the list, and return true. Return false otherwise.
2681 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2682 if (!LU.InsertFormula(F))
2685 CountRegisters(F, LUIdx);
2689 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2690 /// loop-invariant values which we're tracking. These other uses will pin these
2691 /// values in registers, making them less profitable for elimination.
2692 /// TODO: This currently misses non-constant addrec step registers.
2693 /// TODO: Should this give more weight to users inside the loop?
2695 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2696 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2697 SmallPtrSet<const SCEV *, 8> Inserted;
2699 while (!Worklist.empty()) {
2700 const SCEV *S = Worklist.pop_back_val();
2702 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2703 Worklist.append(N->op_begin(), N->op_end());
2704 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2705 Worklist.push_back(C->getOperand());
2706 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2707 Worklist.push_back(D->getLHS());
2708 Worklist.push_back(D->getRHS());
2709 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2710 if (!Inserted.insert(U)) continue;
2711 const Value *V = U->getValue();
2712 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2713 // Look for instructions defined outside the loop.
2714 if (L->contains(Inst)) continue;
2715 } else if (isa<UndefValue>(V))
2716 // Undef doesn't have a live range, so it doesn't matter.
2718 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2720 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2721 // Ignore non-instructions.
2724 // Ignore instructions in other functions (as can happen with
2726 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2728 // Ignore instructions not dominated by the loop.
2729 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2730 UserInst->getParent() :
2731 cast<PHINode>(UserInst)->getIncomingBlock(
2732 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2733 if (!DT.dominates(L->getHeader(), UseBB))
2735 // Ignore uses which are part of other SCEV expressions, to avoid
2736 // analyzing them multiple times.
2737 if (SE.isSCEVable(UserInst->getType())) {
2738 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2739 // If the user is a no-op, look through to its uses.
2740 if (!isa<SCEVUnknown>(UserS))
2744 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2748 // Ignore icmp instructions which are already being analyzed.
2749 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2750 unsigned OtherIdx = !UI.getOperandNo();
2751 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2752 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2756 LSRFixup &LF = getNewFixup();
2757 LF.UserInst = const_cast<Instruction *>(UserInst);
2758 LF.OperandValToReplace = UI.getUse();
2759 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2761 LF.Offset = P.second;
2762 LSRUse &LU = Uses[LF.LUIdx];
2763 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2764 if (!LU.WidestFixupType ||
2765 SE.getTypeSizeInBits(LU.WidestFixupType) <
2766 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2767 LU.WidestFixupType = LF.OperandValToReplace->getType();
2768 InsertSupplementalFormula(U, LU, LF.LUIdx);
2769 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2776 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2777 /// separate registers. If C is non-null, multiply each subexpression by C.
2778 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2779 SmallVectorImpl<const SCEV *> &Ops,
2781 ScalarEvolution &SE) {
2782 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2783 // Break out add operands.
2784 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2786 CollectSubexprs(*I, C, Ops, L, SE);
2788 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2789 // Split a non-zero base out of an addrec.
2790 if (!AR->getStart()->isZero()) {
2791 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2792 AR->getStepRecurrence(SE),
2794 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2797 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2800 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2801 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2802 if (Mul->getNumOperands() == 2)
2803 if (const SCEVConstant *Op0 =
2804 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2805 CollectSubexprs(Mul->getOperand(1),
2806 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2812 // Otherwise use the value itself, optionally with a scale applied.
2813 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2816 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2818 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2821 // Arbitrarily cap recursion to protect compile time.
2822 if (Depth >= 3) return;
2824 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2825 const SCEV *BaseReg = Base.BaseRegs[i];
2827 SmallVector<const SCEV *, 8> AddOps;
2828 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
2830 if (AddOps.size() == 1) continue;
2832 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2833 JE = AddOps.end(); J != JE; ++J) {
2835 // Loop-variant "unknown" values are uninteresting; we won't be able to
2836 // do anything meaningful with them.
2837 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
2840 // Don't pull a constant into a register if the constant could be folded
2841 // into an immediate field.
2842 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2843 Base.getNumRegs() > 1,
2844 LU.Kind, LU.AccessTy, TLI, SE))
2847 // Collect all operands except *J.
2848 SmallVector<const SCEV *, 8> InnerAddOps
2849 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
2851 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
2853 // Don't leave just a constant behind in a register if the constant could
2854 // be folded into an immediate field.
2855 if (InnerAddOps.size() == 1 &&
2856 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2857 Base.getNumRegs() > 1,
2858 LU.Kind, LU.AccessTy, TLI, SE))
2861 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
2862 if (InnerSum->isZero())
2866 // Add the remaining pieces of the add back into the new formula.
2867 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
2868 if (TLI && InnerSumSC &&
2869 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
2870 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2871 InnerSumSC->getValue()->getZExtValue())) {
2872 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2873 InnerSumSC->getValue()->getZExtValue();
2874 F.BaseRegs.erase(F.BaseRegs.begin() + i);
2876 F.BaseRegs[i] = InnerSum;
2878 // Add J as its own register, or an unfolded immediate.
2879 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
2880 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
2881 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
2882 SC->getValue()->getZExtValue()))
2883 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
2884 SC->getValue()->getZExtValue();
2886 F.BaseRegs.push_back(*J);
2888 if (InsertFormula(LU, LUIdx, F))
2889 // If that formula hadn't been seen before, recurse to find more like
2891 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2896 /// GenerateCombinations - Generate a formula consisting of all of the
2897 /// loop-dominating registers added into a single register.
2898 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2900 // This method is only interesting on a plurality of registers.
2901 if (Base.BaseRegs.size() <= 1) return;
2905 SmallVector<const SCEV *, 4> Ops;
2906 for (SmallVectorImpl<const SCEV *>::const_iterator
2907 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2908 const SCEV *BaseReg = *I;
2909 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
2910 !SE.hasComputableLoopEvolution(BaseReg, L))
2911 Ops.push_back(BaseReg);
2913 F.BaseRegs.push_back(BaseReg);
2915 if (Ops.size() > 1) {
2916 const SCEV *Sum = SE.getAddExpr(Ops);
2917 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2918 // opportunity to fold something. For now, just ignore such cases
2919 // rather than proceed with zero in a register.
2920 if (!Sum->isZero()) {
2921 F.BaseRegs.push_back(Sum);
2922 (void)InsertFormula(LU, LUIdx, F);
2927 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2928 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2930 // We can't add a symbolic offset if the address already contains one.
2931 if (Base.AM.BaseGV) return;
2933 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2934 const SCEV *G = Base.BaseRegs[i];
2935 GlobalValue *GV = ExtractSymbol(G, SE);
2936 if (G->isZero() || !GV)
2940 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2941 LU.Kind, LU.AccessTy, TLI))
2944 (void)InsertFormula(LU, LUIdx, F);
2948 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2949 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2951 // TODO: For now, just add the min and max offset, because it usually isn't
2952 // worthwhile looking at everything inbetween.
2953 SmallVector<int64_t, 2> Worklist;
2954 Worklist.push_back(LU.MinOffset);
2955 if (LU.MaxOffset != LU.MinOffset)
2956 Worklist.push_back(LU.MaxOffset);
2958 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2959 const SCEV *G = Base.BaseRegs[i];
2961 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2962 E = Worklist.end(); I != E; ++I) {
2964 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2965 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2966 LU.Kind, LU.AccessTy, TLI)) {
2967 // Add the offset to the base register.
2968 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
2969 // If it cancelled out, drop the base register, otherwise update it.
2970 if (NewG->isZero()) {
2971 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2972 F.BaseRegs.pop_back();
2974 F.BaseRegs[i] = NewG;
2976 (void)InsertFormula(LU, LUIdx, F);
2980 int64_t Imm = ExtractImmediate(G, SE);
2981 if (G->isZero() || Imm == 0)
2984 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2985 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2986 LU.Kind, LU.AccessTy, TLI))
2989 (void)InsertFormula(LU, LUIdx, F);
2993 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2994 /// the comparison. For example, x == y -> x*c == y*c.
2995 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2997 if (LU.Kind != LSRUse::ICmpZero) return;
2999 // Determine the integer type for the base formula.
3000 Type *IntTy = Base.getType();
3002 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3004 // Don't do this if there is more than one offset.
3005 if (LU.MinOffset != LU.MaxOffset) return;
3007 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
3009 // Check each interesting stride.
3010 for (SmallSetVector<int64_t, 8>::const_iterator
3011 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3012 int64_t Factor = *I;
3014 // Check that the multiplication doesn't overflow.
3015 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
3017 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
3018 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
3021 // Check that multiplying with the use offset doesn't overflow.
3022 int64_t Offset = LU.MinOffset;
3023 if (Offset == INT64_MIN && Factor == -1)
3025 Offset = (uint64_t)Offset * Factor;
3026 if (Offset / Factor != LU.MinOffset)
3030 F.AM.BaseOffs = NewBaseOffs;
3032 // Check that this scale is legal.
3033 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
3036 // Compensate for the use having MinOffset built into it.
3037 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
3039 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3041 // Check that multiplying with each base register doesn't overflow.
3042 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3043 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3044 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3048 // Check that multiplying with the scaled register doesn't overflow.
3050 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3051 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3055 // Check that multiplying with the unfolded offset doesn't overflow.
3056 if (F.UnfoldedOffset != 0) {
3057 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3059 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3060 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3064 // If we make it here and it's legal, add it.
3065 (void)InsertFormula(LU, LUIdx, F);
3070 /// GenerateScales - Generate stride factor reuse formulae by making use of
3071 /// scaled-offset address modes, for example.
3072 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3073 // Determine the integer type for the base formula.
3074 Type *IntTy = Base.getType();
3077 // If this Formula already has a scaled register, we can't add another one.
3078 if (Base.AM.Scale != 0) return;
3080 // Check each interesting stride.
3081 for (SmallSetVector<int64_t, 8>::const_iterator
3082 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3083 int64_t Factor = *I;
3085 Base.AM.Scale = Factor;
3086 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
3087 // Check whether this scale is going to be legal.
3088 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3089 LU.Kind, LU.AccessTy, TLI)) {
3090 // As a special-case, handle special out-of-loop Basic users specially.
3091 // TODO: Reconsider this special case.
3092 if (LU.Kind == LSRUse::Basic &&
3093 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3094 LSRUse::Special, LU.AccessTy, TLI) &&
3095 LU.AllFixupsOutsideLoop)
3096 LU.Kind = LSRUse::Special;
3100 // For an ICmpZero, negating a solitary base register won't lead to
3102 if (LU.Kind == LSRUse::ICmpZero &&
3103 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
3105 // For each addrec base reg, apply the scale, if possible.
3106 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3107 if (const SCEVAddRecExpr *AR =
3108 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3109 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3110 if (FactorS->isZero())
3112 // Divide out the factor, ignoring high bits, since we'll be
3113 // scaling the value back up in the end.
3114 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3115 // TODO: This could be optimized to avoid all the copying.
3117 F.ScaledReg = Quotient;
3118 F.DeleteBaseReg(F.BaseRegs[i]);
3119 (void)InsertFormula(LU, LUIdx, F);
3125 /// GenerateTruncates - Generate reuse formulae from different IV types.
3126 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3127 // This requires TargetLowering to tell us which truncates are free.
3130 // Don't bother truncating symbolic values.
3131 if (Base.AM.BaseGV) return;
3133 // Determine the integer type for the base formula.
3134 Type *DstTy = Base.getType();
3136 DstTy = SE.getEffectiveSCEVType(DstTy);
3138 for (SmallSetVector<Type *, 4>::const_iterator
3139 I = Types.begin(), E = Types.end(); I != E; ++I) {
3141 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
3144 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3145 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3146 JE = F.BaseRegs.end(); J != JE; ++J)
3147 *J = SE.getAnyExtendExpr(*J, SrcTy);
3149 // TODO: This assumes we've done basic processing on all uses and
3150 // have an idea what the register usage is.
3151 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3154 (void)InsertFormula(LU, LUIdx, F);
3161 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3162 /// defer modifications so that the search phase doesn't have to worry about
3163 /// the data structures moving underneath it.
3167 const SCEV *OrigReg;
3169 WorkItem(size_t LI, int64_t I, const SCEV *R)
3170 : LUIdx(LI), Imm(I), OrigReg(R) {}
3172 void print(raw_ostream &OS) const;
3178 void WorkItem::print(raw_ostream &OS) const {
3179 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3180 << " , add offset " << Imm;
3183 void WorkItem::dump() const {
3184 print(errs()); errs() << '\n';
3187 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3188 /// distance apart and try to form reuse opportunities between them.
3189 void LSRInstance::GenerateCrossUseConstantOffsets() {
3190 // Group the registers by their value without any added constant offset.
3191 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3192 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3194 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3195 SmallVector<const SCEV *, 8> Sequence;
3196 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3198 const SCEV *Reg = *I;
3199 int64_t Imm = ExtractImmediate(Reg, SE);
3200 std::pair<RegMapTy::iterator, bool> Pair =
3201 Map.insert(std::make_pair(Reg, ImmMapTy()));
3203 Sequence.push_back(Reg);
3204 Pair.first->second.insert(std::make_pair(Imm, *I));
3205 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3208 // Now examine each set of registers with the same base value. Build up
3209 // a list of work to do and do the work in a separate step so that we're
3210 // not adding formulae and register counts while we're searching.
3211 SmallVector<WorkItem, 32> WorkItems;
3212 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3213 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3214 E = Sequence.end(); I != E; ++I) {
3215 const SCEV *Reg = *I;
3216 const ImmMapTy &Imms = Map.find(Reg)->second;
3218 // It's not worthwhile looking for reuse if there's only one offset.
3219 if (Imms.size() == 1)
3222 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3223 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3225 dbgs() << ' ' << J->first;
3228 // Examine each offset.
3229 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3231 const SCEV *OrigReg = J->second;
3233 int64_t JImm = J->first;
3234 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3236 if (!isa<SCEVConstant>(OrigReg) &&
3237 UsedByIndicesMap[Reg].count() == 1) {
3238 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3242 // Conservatively examine offsets between this orig reg a few selected
3244 ImmMapTy::const_iterator OtherImms[] = {
3245 Imms.begin(), prior(Imms.end()),
3246 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3248 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3249 ImmMapTy::const_iterator M = OtherImms[i];
3250 if (M == J || M == JE) continue;
3252 // Compute the difference between the two.
3253 int64_t Imm = (uint64_t)JImm - M->first;
3254 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3255 LUIdx = UsedByIndices.find_next(LUIdx))
3256 // Make a memo of this use, offset, and register tuple.
3257 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3258 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3265 UsedByIndicesMap.clear();
3266 UniqueItems.clear();
3268 // Now iterate through the worklist and add new formulae.
3269 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3270 E = WorkItems.end(); I != E; ++I) {
3271 const WorkItem &WI = *I;
3272 size_t LUIdx = WI.LUIdx;
3273 LSRUse &LU = Uses[LUIdx];
3274 int64_t Imm = WI.Imm;
3275 const SCEV *OrigReg = WI.OrigReg;
3277 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3278 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3279 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3281 // TODO: Use a more targeted data structure.
3282 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3283 const Formula &F = LU.Formulae[L];
3284 // Use the immediate in the scaled register.
3285 if (F.ScaledReg == OrigReg) {
3286 int64_t Offs = (uint64_t)F.AM.BaseOffs +
3287 Imm * (uint64_t)F.AM.Scale;
3288 // Don't create 50 + reg(-50).
3289 if (F.referencesReg(SE.getSCEV(
3290 ConstantInt::get(IntTy, -(uint64_t)Offs))))
3293 NewF.AM.BaseOffs = Offs;
3294 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3295 LU.Kind, LU.AccessTy, TLI))
3297 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3299 // If the new scale is a constant in a register, and adding the constant
3300 // value to the immediate would produce a value closer to zero than the
3301 // immediate itself, then the formula isn't worthwhile.
3302 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3303 if (C->getValue()->isNegative() !=
3304 (NewF.AM.BaseOffs < 0) &&
3305 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
3306 .ule(abs64(NewF.AM.BaseOffs)))
3310 (void)InsertFormula(LU, LUIdx, NewF);
3312 // Use the immediate in a base register.
3313 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3314 const SCEV *BaseReg = F.BaseRegs[N];
3315 if (BaseReg != OrigReg)
3318 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
3319 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3320 LU.Kind, LU.AccessTy, TLI)) {
3322 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3325 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3327 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3329 // If the new formula has a constant in a register, and adding the
3330 // constant value to the immediate would produce a value closer to
3331 // zero than the immediate itself, then the formula isn't worthwhile.
3332 for (SmallVectorImpl<const SCEV *>::const_iterator
3333 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3335 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3336 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
3337 abs64(NewF.AM.BaseOffs)) &&
3338 (C->getValue()->getValue() +
3339 NewF.AM.BaseOffs).countTrailingZeros() >=
3340 CountTrailingZeros_64(NewF.AM.BaseOffs))
3344 (void)InsertFormula(LU, LUIdx, NewF);
3353 /// GenerateAllReuseFormulae - Generate formulae for each use.
3355 LSRInstance::GenerateAllReuseFormulae() {
3356 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3357 // queries are more precise.
3358 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3359 LSRUse &LU = Uses[LUIdx];
3360 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3361 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3362 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3363 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3365 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3366 LSRUse &LU = Uses[LUIdx];
3367 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3368 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3369 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3370 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3371 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3372 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3373 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3374 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3376 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3377 LSRUse &LU = Uses[LUIdx];
3378 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3379 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3382 GenerateCrossUseConstantOffsets();
3384 DEBUG(dbgs() << "\n"
3385 "After generating reuse formulae:\n";
3386 print_uses(dbgs()));
3389 /// If there are multiple formulae with the same set of registers used
3390 /// by other uses, pick the best one and delete the others.
3391 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3392 DenseSet<const SCEV *> VisitedRegs;
3393 SmallPtrSet<const SCEV *, 16> Regs;
3394 SmallPtrSet<const SCEV *, 16> LoserRegs;
3396 bool ChangedFormulae = false;
3399 // Collect the best formula for each unique set of shared registers. This
3400 // is reset for each use.
3401 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
3403 BestFormulaeTy BestFormulae;
3405 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3406 LSRUse &LU = Uses[LUIdx];
3407 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3410 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3411 FIdx != NumForms; ++FIdx) {
3412 Formula &F = LU.Formulae[FIdx];
3414 // Some formulas are instant losers. For example, they may depend on
3415 // nonexistent AddRecs from other loops. These need to be filtered
3416 // immediately, otherwise heuristics could choose them over others leading
3417 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3418 // avoids the need to recompute this information across formulae using the
3419 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3420 // the corresponding bad register from the Regs set.
3423 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3425 if (CostF.isLoser()) {
3426 // During initial formula generation, undesirable formulae are generated
3427 // by uses within other loops that have some non-trivial address mode or
3428 // use the postinc form of the IV. LSR needs to provide these formulae
3429 // as the basis of rediscovering the desired formula that uses an AddRec
3430 // corresponding to the existing phi. Once all formulae have been
3431 // generated, these initial losers may be pruned.
3432 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3436 SmallVector<const SCEV *, 2> Key;
3437 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3438 JE = F.BaseRegs.end(); J != JE; ++J) {
3439 const SCEV *Reg = *J;
3440 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3444 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3445 Key.push_back(F.ScaledReg);
3446 // Unstable sort by host order ok, because this is only used for
3448 std::sort(Key.begin(), Key.end());
3450 std::pair<BestFormulaeTy::const_iterator, bool> P =
3451 BestFormulae.insert(std::make_pair(Key, FIdx));
3455 Formula &Best = LU.Formulae[P.first->second];
3459 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3460 if (CostF < CostBest)
3462 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3464 " in favor of formula "; Best.print(dbgs());
3468 ChangedFormulae = true;
3470 LU.DeleteFormula(F);
3476 // Now that we've filtered out some formulae, recompute the Regs set.
3478 LU.RecomputeRegs(LUIdx, RegUses);
3480 // Reset this to prepare for the next use.
3481 BestFormulae.clear();
3484 DEBUG(if (ChangedFormulae) {
3486 "After filtering out undesirable candidates:\n";
3491 // This is a rough guess that seems to work fairly well.
3492 static const size_t ComplexityLimit = UINT16_MAX;
3494 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3495 /// solutions the solver might have to consider. It almost never considers
3496 /// this many solutions because it prune the search space, but the pruning
3497 /// isn't always sufficient.
3498 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3500 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3501 E = Uses.end(); I != E; ++I) {
3502 size_t FSize = I->Formulae.size();
3503 if (FSize >= ComplexityLimit) {
3504 Power = ComplexityLimit;
3508 if (Power >= ComplexityLimit)
3514 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3515 /// of the registers of another formula, it won't help reduce register
3516 /// pressure (though it may not necessarily hurt register pressure); remove
3517 /// it to simplify the system.
3518 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3519 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3520 DEBUG(dbgs() << "The search space is too complex.\n");
3522 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3523 "which use a superset of registers used by other "
3526 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3527 LSRUse &LU = Uses[LUIdx];
3529 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3530 Formula &F = LU.Formulae[i];
3531 // Look for a formula with a constant or GV in a register. If the use
3532 // also has a formula with that same value in an immediate field,
3533 // delete the one that uses a register.
3534 for (SmallVectorImpl<const SCEV *>::const_iterator
3535 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3536 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3538 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3539 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3540 (I - F.BaseRegs.begin()));
3541 if (LU.HasFormulaWithSameRegs(NewF)) {
3542 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3543 LU.DeleteFormula(F);
3549 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3550 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3553 NewF.AM.BaseGV = GV;
3554 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3555 (I - F.BaseRegs.begin()));
3556 if (LU.HasFormulaWithSameRegs(NewF)) {
3557 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3559 LU.DeleteFormula(F);
3570 LU.RecomputeRegs(LUIdx, RegUses);
3573 DEBUG(dbgs() << "After pre-selection:\n";
3574 print_uses(dbgs()));
3578 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3579 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3581 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3582 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3583 DEBUG(dbgs() << "The search space is too complex.\n");
3585 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3586 "separated by a constant offset will use the same "
3589 // This is especially useful for unrolled loops.
3591 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3592 LSRUse &LU = Uses[LUIdx];
3593 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3594 E = LU.Formulae.end(); I != E; ++I) {
3595 const Formula &F = *I;
3596 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3597 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3598 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3599 /*HasBaseReg=*/false,
3600 LU.Kind, LU.AccessTy)) {
3601 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3604 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3606 // Update the relocs to reference the new use.
3607 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3608 E = Fixups.end(); I != E; ++I) {
3609 LSRFixup &Fixup = *I;
3610 if (Fixup.LUIdx == LUIdx) {
3611 Fixup.LUIdx = LUThatHas - &Uses.front();
3612 Fixup.Offset += F.AM.BaseOffs;
3613 // Add the new offset to LUThatHas' offset list.
3614 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3615 LUThatHas->Offsets.push_back(Fixup.Offset);
3616 if (Fixup.Offset > LUThatHas->MaxOffset)
3617 LUThatHas->MaxOffset = Fixup.Offset;
3618 if (Fixup.Offset < LUThatHas->MinOffset)
3619 LUThatHas->MinOffset = Fixup.Offset;
3621 DEBUG(dbgs() << "New fixup has offset "
3622 << Fixup.Offset << '\n');
3624 if (Fixup.LUIdx == NumUses-1)
3625 Fixup.LUIdx = LUIdx;
3628 // Delete formulae from the new use which are no longer legal.
3630 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3631 Formula &F = LUThatHas->Formulae[i];
3632 if (!isLegalUse(F.AM,
3633 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3634 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3635 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3637 LUThatHas->DeleteFormula(F);
3644 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3646 // Delete the old use.
3647 DeleteUse(LU, LUIdx);
3657 DEBUG(dbgs() << "After pre-selection:\n";
3658 print_uses(dbgs()));
3662 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3663 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3664 /// we've done more filtering, as it may be able to find more formulae to
3666 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3667 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3668 DEBUG(dbgs() << "The search space is too complex.\n");
3670 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3671 "undesirable dedicated registers.\n");
3673 FilterOutUndesirableDedicatedRegisters();
3675 DEBUG(dbgs() << "After pre-selection:\n";
3676 print_uses(dbgs()));
3680 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3681 /// to be profitable, and then in any use which has any reference to that
3682 /// register, delete all formulae which do not reference that register.
3683 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3684 // With all other options exhausted, loop until the system is simple
3685 // enough to handle.
3686 SmallPtrSet<const SCEV *, 4> Taken;
3687 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3688 // Ok, we have too many of formulae on our hands to conveniently handle.
3689 // Use a rough heuristic to thin out the list.
3690 DEBUG(dbgs() << "The search space is too complex.\n");
3692 // Pick the register which is used by the most LSRUses, which is likely
3693 // to be a good reuse register candidate.
3694 const SCEV *Best = 0;
3695 unsigned BestNum = 0;
3696 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3698 const SCEV *Reg = *I;
3699 if (Taken.count(Reg))
3704 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3705 if (Count > BestNum) {
3712 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3713 << " will yield profitable reuse.\n");
3716 // In any use with formulae which references this register, delete formulae
3717 // which don't reference it.
3718 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3719 LSRUse &LU = Uses[LUIdx];
3720 if (!LU.Regs.count(Best)) continue;
3723 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3724 Formula &F = LU.Formulae[i];
3725 if (!F.referencesReg(Best)) {
3726 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3727 LU.DeleteFormula(F);
3731 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3737 LU.RecomputeRegs(LUIdx, RegUses);
3740 DEBUG(dbgs() << "After pre-selection:\n";
3741 print_uses(dbgs()));
3745 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3746 /// formulae to choose from, use some rough heuristics to prune down the number
3747 /// of formulae. This keeps the main solver from taking an extraordinary amount
3748 /// of time in some worst-case scenarios.
3749 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3750 NarrowSearchSpaceByDetectingSupersets();
3751 NarrowSearchSpaceByCollapsingUnrolledCode();
3752 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3753 NarrowSearchSpaceByPickingWinnerRegs();
3756 /// SolveRecurse - This is the recursive solver.
3757 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3759 SmallVectorImpl<const Formula *> &Workspace,
3760 const Cost &CurCost,
3761 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3762 DenseSet<const SCEV *> &VisitedRegs) const {
3765 // - use more aggressive filtering
3766 // - sort the formula so that the most profitable solutions are found first
3767 // - sort the uses too
3769 // - don't compute a cost, and then compare. compare while computing a cost
3771 // - track register sets with SmallBitVector
3773 const LSRUse &LU = Uses[Workspace.size()];
3775 // If this use references any register that's already a part of the
3776 // in-progress solution, consider it a requirement that a formula must
3777 // reference that register in order to be considered. This prunes out
3778 // unprofitable searching.
3779 SmallSetVector<const SCEV *, 4> ReqRegs;
3780 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3781 E = CurRegs.end(); I != E; ++I)
3782 if (LU.Regs.count(*I))
3785 bool AnySatisfiedReqRegs = false;
3786 SmallPtrSet<const SCEV *, 16> NewRegs;
3789 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3790 E = LU.Formulae.end(); I != E; ++I) {
3791 const Formula &F = *I;
3793 // Ignore formulae which do not use any of the required registers.
3794 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3795 JE = ReqRegs.end(); J != JE; ++J) {
3796 const SCEV *Reg = *J;
3797 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3798 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3802 AnySatisfiedReqRegs = true;
3804 // Evaluate the cost of the current formula. If it's already worse than
3805 // the current best, prune the search at that point.
3808 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3809 if (NewCost < SolutionCost) {
3810 Workspace.push_back(&F);
3811 if (Workspace.size() != Uses.size()) {
3812 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3813 NewRegs, VisitedRegs);
3814 if (F.getNumRegs() == 1 && Workspace.size() == 1)
3815 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
3817 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
3818 dbgs() << ".\n Regs:";
3819 for (SmallPtrSet<const SCEV *, 16>::const_iterator
3820 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
3821 dbgs() << ' ' << **I;
3824 SolutionCost = NewCost;
3825 Solution = Workspace;
3827 Workspace.pop_back();
3832 if (!EnableRetry && !AnySatisfiedReqRegs)
3835 // If none of the formulae had all of the required registers, relax the
3836 // constraint so that we don't exclude all formulae.
3837 if (!AnySatisfiedReqRegs) {
3838 assert(!ReqRegs.empty() && "Solver failed even without required registers");
3844 /// Solve - Choose one formula from each use. Return the results in the given
3845 /// Solution vector.
3846 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
3847 SmallVector<const Formula *, 8> Workspace;
3849 SolutionCost.Loose();
3851 SmallPtrSet<const SCEV *, 16> CurRegs;
3852 DenseSet<const SCEV *> VisitedRegs;
3853 Workspace.reserve(Uses.size());
3855 // SolveRecurse does all the work.
3856 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
3857 CurRegs, VisitedRegs);
3858 if (Solution.empty()) {
3859 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
3863 // Ok, we've now made all our decisions.
3864 DEBUG(dbgs() << "\n"
3865 "The chosen solution requires "; SolutionCost.print(dbgs());
3867 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
3869 Uses[i].print(dbgs());
3872 Solution[i]->print(dbgs());
3876 assert(Solution.size() == Uses.size() && "Malformed solution!");
3879 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
3880 /// the dominator tree far as we can go while still being dominated by the
3881 /// input positions. This helps canonicalize the insert position, which
3882 /// encourages sharing.
3883 BasicBlock::iterator
3884 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
3885 const SmallVectorImpl<Instruction *> &Inputs)
3888 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
3889 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
3892 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
3893 if (!Rung) return IP;
3894 Rung = Rung->getIDom();
3895 if (!Rung) return IP;
3896 IDom = Rung->getBlock();
3898 // Don't climb into a loop though.
3899 const Loop *IDomLoop = LI.getLoopFor(IDom);
3900 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
3901 if (IDomDepth <= IPLoopDepth &&
3902 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
3906 bool AllDominate = true;
3907 Instruction *BetterPos = 0;
3908 Instruction *Tentative = IDom->getTerminator();
3909 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
3910 E = Inputs.end(); I != E; ++I) {
3911 Instruction *Inst = *I;
3912 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
3913 AllDominate = false;
3916 // Attempt to find an insert position in the middle of the block,
3917 // instead of at the end, so that it can be used for other expansions.
3918 if (IDom == Inst->getParent() &&
3919 (!BetterPos || DT.dominates(BetterPos, Inst)))
3920 BetterPos = llvm::next(BasicBlock::iterator(Inst));
3933 /// AdjustInsertPositionForExpand - Determine an input position which will be
3934 /// dominated by the operands and which will dominate the result.
3935 BasicBlock::iterator
3936 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
3938 const LSRUse &LU) const {
3939 // Collect some instructions which must be dominated by the
3940 // expanding replacement. These must be dominated by any operands that
3941 // will be required in the expansion.
3942 SmallVector<Instruction *, 4> Inputs;
3943 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
3944 Inputs.push_back(I);
3945 if (LU.Kind == LSRUse::ICmpZero)
3946 if (Instruction *I =
3947 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
3948 Inputs.push_back(I);
3949 if (LF.PostIncLoops.count(L)) {
3950 if (LF.isUseFullyOutsideLoop(L))
3951 Inputs.push_back(L->getLoopLatch()->getTerminator());
3953 Inputs.push_back(IVIncInsertPos);
3955 // The expansion must also be dominated by the increment positions of any
3956 // loops it for which it is using post-inc mode.
3957 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
3958 E = LF.PostIncLoops.end(); I != E; ++I) {
3959 const Loop *PIL = *I;
3960 if (PIL == L) continue;
3962 // Be dominated by the loop exit.
3963 SmallVector<BasicBlock *, 4> ExitingBlocks;
3964 PIL->getExitingBlocks(ExitingBlocks);
3965 if (!ExitingBlocks.empty()) {
3966 BasicBlock *BB = ExitingBlocks[0];
3967 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
3968 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
3969 Inputs.push_back(BB->getTerminator());
3973 // Then, climb up the immediate dominator tree as far as we can go while
3974 // still being dominated by the input positions.
3975 IP = HoistInsertPosition(IP, Inputs);
3977 // Don't insert instructions before PHI nodes.
3978 while (isa<PHINode>(IP)) ++IP;
3980 // Ignore landingpad instructions.
3981 while (isa<LandingPadInst>(IP)) ++IP;
3983 // Ignore debug intrinsics.
3984 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
3989 /// Expand - Emit instructions for the leading candidate expression for this
3990 /// LSRUse (this is called "expanding").
3991 Value *LSRInstance::Expand(const LSRFixup &LF,
3993 BasicBlock::iterator IP,
3994 SCEVExpander &Rewriter,
3995 SmallVectorImpl<WeakVH> &DeadInsts) const {
3996 const LSRUse &LU = Uses[LF.LUIdx];
3998 // Determine an input position which will be dominated by the operands and
3999 // which will dominate the result.
4000 IP = AdjustInsertPositionForExpand(IP, LF, LU);
4002 // Inform the Rewriter if we have a post-increment use, so that it can
4003 // perform an advantageous expansion.
4004 Rewriter.setPostInc(LF.PostIncLoops);
4006 // This is the type that the user actually needs.
4007 Type *OpTy = LF.OperandValToReplace->getType();
4008 // This will be the type that we'll initially expand to.
4009 Type *Ty = F.getType();
4011 // No type known; just expand directly to the ultimate type.
4013 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4014 // Expand directly to the ultimate type if it's the right size.
4016 // This is the type to do integer arithmetic in.
4017 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4019 // Build up a list of operands to add together to form the full base.
4020 SmallVector<const SCEV *, 8> Ops;
4022 // Expand the BaseRegs portion.
4023 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4024 E = F.BaseRegs.end(); I != E; ++I) {
4025 const SCEV *Reg = *I;
4026 assert(!Reg->isZero() && "Zero allocated in a base register!");
4028 // If we're expanding for a post-inc user, make the post-inc adjustment.
4029 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4030 Reg = TransformForPostIncUse(Denormalize, Reg,
4031 LF.UserInst, LF.OperandValToReplace,
4034 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4037 // Flush the operand list to suppress SCEVExpander hoisting.
4039 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4041 Ops.push_back(SE.getUnknown(FullV));
4044 // Expand the ScaledReg portion.
4045 Value *ICmpScaledV = 0;
4046 if (F.AM.Scale != 0) {
4047 const SCEV *ScaledS = F.ScaledReg;
4049 // If we're expanding for a post-inc user, make the post-inc adjustment.
4050 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4051 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4052 LF.UserInst, LF.OperandValToReplace,
4055 if (LU.Kind == LSRUse::ICmpZero) {
4056 // An interesting way of "folding" with an icmp is to use a negated
4057 // scale, which we'll implement by inserting it into the other operand
4059 assert(F.AM.Scale == -1 &&
4060 "The only scale supported by ICmpZero uses is -1!");
4061 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4063 // Otherwise just expand the scaled register and an explicit scale,
4064 // which is expected to be matched as part of the address.
4065 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4066 ScaledS = SE.getMulExpr(ScaledS,
4067 SE.getConstant(ScaledS->getType(), F.AM.Scale));
4068 Ops.push_back(ScaledS);
4070 // Flush the operand list to suppress SCEVExpander hoisting.
4071 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4073 Ops.push_back(SE.getUnknown(FullV));
4077 // Expand the GV portion.
4079 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
4081 // Flush the operand list to suppress SCEVExpander hoisting.
4082 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4084 Ops.push_back(SE.getUnknown(FullV));
4087 // Expand the immediate portion.
4088 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
4090 if (LU.Kind == LSRUse::ICmpZero) {
4091 // The other interesting way of "folding" with an ICmpZero is to use a
4092 // negated immediate.
4094 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4096 Ops.push_back(SE.getUnknown(ICmpScaledV));
4097 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4100 // Just add the immediate values. These again are expected to be matched
4101 // as part of the address.
4102 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4106 // Expand the unfolded offset portion.
4107 int64_t UnfoldedOffset = F.UnfoldedOffset;
4108 if (UnfoldedOffset != 0) {
4109 // Just add the immediate values.
4110 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4114 // Emit instructions summing all the operands.
4115 const SCEV *FullS = Ops.empty() ?
4116 SE.getConstant(IntTy, 0) :
4118 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4120 // We're done expanding now, so reset the rewriter.
4121 Rewriter.clearPostInc();
4123 // An ICmpZero Formula represents an ICmp which we're handling as a
4124 // comparison against zero. Now that we've expanded an expression for that
4125 // form, update the ICmp's other operand.
4126 if (LU.Kind == LSRUse::ICmpZero) {
4127 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4128 DeadInsts.push_back(CI->getOperand(1));
4129 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
4130 "a scale at the same time!");
4131 if (F.AM.Scale == -1) {
4132 if (ICmpScaledV->getType() != OpTy) {
4134 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4136 ICmpScaledV, OpTy, "tmp", CI);
4139 CI->setOperand(1, ICmpScaledV);
4141 assert(F.AM.Scale == 0 &&
4142 "ICmp does not support folding a global value and "
4143 "a scale at the same time!");
4144 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4146 if (C->getType() != OpTy)
4147 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4151 CI->setOperand(1, C);
4158 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4159 /// of their operands effectively happens in their predecessor blocks, so the
4160 /// expression may need to be expanded in multiple places.
4161 void LSRInstance::RewriteForPHI(PHINode *PN,
4164 SCEVExpander &Rewriter,
4165 SmallVectorImpl<WeakVH> &DeadInsts,
4167 DenseMap<BasicBlock *, Value *> Inserted;
4168 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4169 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4170 BasicBlock *BB = PN->getIncomingBlock(i);
4172 // If this is a critical edge, split the edge so that we do not insert
4173 // the code on all predecessor/successor paths. We do this unless this
4174 // is the canonical backedge for this loop, which complicates post-inc
4176 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4177 !isa<IndirectBrInst>(BB->getTerminator())) {
4178 BasicBlock *Parent = PN->getParent();
4179 Loop *PNLoop = LI.getLoopFor(Parent);
4180 if (!PNLoop || Parent != PNLoop->getHeader()) {
4181 // Split the critical edge.
4182 BasicBlock *NewBB = 0;
4183 if (!Parent->isLandingPad()) {
4184 NewBB = SplitCriticalEdge(BB, Parent, P,
4185 /*MergeIdenticalEdges=*/true,
4186 /*DontDeleteUselessPhis=*/true);
4188 SmallVector<BasicBlock*, 2> NewBBs;
4189 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4193 // If PN is outside of the loop and BB is in the loop, we want to
4194 // move the block to be immediately before the PHI block, not
4195 // immediately after BB.
4196 if (L->contains(BB) && !L->contains(PN))
4197 NewBB->moveBefore(PN->getParent());
4199 // Splitting the edge can reduce the number of PHI entries we have.
4200 e = PN->getNumIncomingValues();
4202 i = PN->getBasicBlockIndex(BB);
4206 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4207 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4209 PN->setIncomingValue(i, Pair.first->second);
4211 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4213 // If this is reuse-by-noop-cast, insert the noop cast.
4214 Type *OpTy = LF.OperandValToReplace->getType();
4215 if (FullV->getType() != OpTy)
4217 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4219 FullV, LF.OperandValToReplace->getType(),
4220 "tmp", BB->getTerminator());
4222 PN->setIncomingValue(i, FullV);
4223 Pair.first->second = FullV;
4228 /// Rewrite - Emit instructions for the leading candidate expression for this
4229 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4230 /// the newly expanded value.
4231 void LSRInstance::Rewrite(const LSRFixup &LF,
4233 SCEVExpander &Rewriter,
4234 SmallVectorImpl<WeakVH> &DeadInsts,
4236 // First, find an insertion point that dominates UserInst. For PHI nodes,
4237 // find the nearest block which dominates all the relevant uses.
4238 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4239 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4241 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4243 // If this is reuse-by-noop-cast, insert the noop cast.
4244 Type *OpTy = LF.OperandValToReplace->getType();
4245 if (FullV->getType() != OpTy) {
4247 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4248 FullV, OpTy, "tmp", LF.UserInst);
4252 // Update the user. ICmpZero is handled specially here (for now) because
4253 // Expand may have updated one of the operands of the icmp already, and
4254 // its new value may happen to be equal to LF.OperandValToReplace, in
4255 // which case doing replaceUsesOfWith leads to replacing both operands
4256 // with the same value. TODO: Reorganize this.
4257 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4258 LF.UserInst->setOperand(0, FullV);
4260 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4263 DeadInsts.push_back(LF.OperandValToReplace);
4266 /// ImplementSolution - Rewrite all the fixup locations with new values,
4267 /// following the chosen solution.
4269 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4271 // Keep track of instructions we may have made dead, so that
4272 // we can remove them after we are done working.
4273 SmallVector<WeakVH, 16> DeadInsts;
4275 SCEVExpander Rewriter(SE, "lsr");
4277 Rewriter.setDebugType(DEBUG_TYPE);
4279 Rewriter.disableCanonicalMode();
4280 Rewriter.enableLSRMode();
4281 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4283 // Expand the new value definitions and update the users.
4284 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4285 E = Fixups.end(); I != E; ++I) {
4286 const LSRFixup &Fixup = *I;
4288 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4293 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4294 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4295 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4298 // Clean up after ourselves. This must be done before deleting any
4302 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4305 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
4306 : IU(P->getAnalysis<IVUsers>()),
4307 SE(P->getAnalysis<ScalarEvolution>()),
4308 DT(P->getAnalysis<DominatorTree>()),
4309 LI(P->getAnalysis<LoopInfo>()),
4310 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
4312 // If LoopSimplify form is not available, stay out of trouble.
4313 if (!L->isLoopSimplifyForm())
4316 // All outer loops must have preheaders, or SCEVExpander may not be able to
4317 // materialize an AddRecExpr whose Start is an outer AddRecExpr.
4318 for (const Loop *OuterLoop = L; (OuterLoop = OuterLoop->getParentLoop());) {
4319 if (!OuterLoop->getLoopPreheader())
4322 // If there's no interesting work to be done, bail early.
4323 if (IU.empty()) return;
4325 DEBUG(dbgs() << "\nLSR on loop ";
4326 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4329 // First, perform some low-level loop optimizations.
4331 OptimizeLoopTermCond();
4333 // If loop preparation eliminates all interesting IV users, bail.
4334 if (IU.empty()) return;
4336 // Skip nested loops until we can model them better with formulae.
4337 if (!EnableNested && !L->empty()) {
4338 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4342 // Start collecting data and preparing for the solver.
4344 CollectInterestingTypesAndFactors();
4345 CollectFixupsAndInitialFormulae();
4346 CollectLoopInvariantFixupsAndFormulae();
4348 assert(!Uses.empty() && "IVUsers reported at least one use");
4349 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4350 print_uses(dbgs()));
4352 // Now use the reuse data to generate a bunch of interesting ways
4353 // to formulate the values needed for the uses.
4354 GenerateAllReuseFormulae();
4356 FilterOutUndesirableDedicatedRegisters();
4357 NarrowSearchSpaceUsingHeuristics();
4359 SmallVector<const Formula *, 8> Solution;
4362 // Release memory that is no longer needed.
4367 if (Solution.empty())
4371 // Formulae should be legal.
4372 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4373 E = Uses.end(); I != E; ++I) {
4374 const LSRUse &LU = *I;
4375 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4376 JE = LU.Formulae.end(); J != JE; ++J)
4377 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
4378 LU.Kind, LU.AccessTy, TLI) &&
4379 "Illegal formula generated!");
4383 // Now that we've decided what we want, make it so.
4384 ImplementSolution(Solution, P);
4387 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4388 if (Factors.empty() && Types.empty()) return;
4390 OS << "LSR has identified the following interesting factors and types: ";
4393 for (SmallSetVector<int64_t, 8>::const_iterator
4394 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4395 if (!First) OS << ", ";
4400 for (SmallSetVector<Type *, 4>::const_iterator
4401 I = Types.begin(), E = Types.end(); I != E; ++I) {
4402 if (!First) OS << ", ";
4404 OS << '(' << **I << ')';
4409 void LSRInstance::print_fixups(raw_ostream &OS) const {
4410 OS << "LSR is examining the following fixup sites:\n";
4411 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4412 E = Fixups.end(); I != E; ++I) {
4419 void LSRInstance::print_uses(raw_ostream &OS) const {
4420 OS << "LSR is examining the following uses:\n";
4421 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4422 E = Uses.end(); I != E; ++I) {
4423 const LSRUse &LU = *I;
4427 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4428 JE = LU.Formulae.end(); J != JE; ++J) {
4436 void LSRInstance::print(raw_ostream &OS) const {
4437 print_factors_and_types(OS);
4442 void LSRInstance::dump() const {
4443 print(errs()); errs() << '\n';
4448 class LoopStrengthReduce : public LoopPass {
4449 /// TLI - Keep a pointer of a TargetLowering to consult for determining
4450 /// transformation profitability.
4451 const TargetLowering *const TLI;
4454 static char ID; // Pass ID, replacement for typeid
4455 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
4458 bool runOnLoop(Loop *L, LPPassManager &LPM);
4459 void getAnalysisUsage(AnalysisUsage &AU) const;
4464 char LoopStrengthReduce::ID = 0;
4465 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4466 "Loop Strength Reduction", false, false)
4467 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4468 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4469 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4470 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4471 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4472 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4473 "Loop Strength Reduction", false, false)
4476 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
4477 return new LoopStrengthReduce(TLI);
4480 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
4481 : LoopPass(ID), TLI(tli) {
4482 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4485 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4486 // We split critical edges, so we change the CFG. However, we do update
4487 // many analyses if they are around.
4488 AU.addPreservedID(LoopSimplifyID);
4490 AU.addRequired<LoopInfo>();
4491 AU.addPreserved<LoopInfo>();
4492 AU.addRequiredID(LoopSimplifyID);
4493 AU.addRequired<DominatorTree>();
4494 AU.addPreserved<DominatorTree>();
4495 AU.addRequired<ScalarEvolution>();
4496 AU.addPreserved<ScalarEvolution>();
4497 // Requiring LoopSimplify a second time here prevents IVUsers from running
4498 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4499 AU.addRequiredID(LoopSimplifyID);
4500 AU.addRequired<IVUsers>();
4501 AU.addPreserved<IVUsers>();
4504 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4505 bool Changed = false;
4507 // Run the main LSR transformation.
4508 Changed |= LSRInstance(TLI, L, this).getChanged();
4510 // Remove any extra phis created by processing inner loops.
4511 Changed |= DeleteDeadPHIs(L->getHeader());
4512 if (EnablePhiElim) {
4513 SmallVector<WeakVH, 16> DeadInsts;
4514 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4516 Rewriter.setDebugType(DEBUG_TYPE);
4518 unsigned numFolded = Rewriter.
4519 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI);
4522 DeleteTriviallyDeadInstructions(DeadInsts);
4523 DeleteDeadPHIs(L->getHeader());