1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/ADT/DenseSet.h"
59 #include "llvm/ADT/STLExtras.h"
60 #include "llvm/ADT/SetVector.h"
61 #include "llvm/ADT/SmallBitVector.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/LoopPass.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/IR/Constants.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/Instructions.h"
70 #include "llvm/IR/IntrinsicInst.h"
71 #include "llvm/IR/ValueHandle.h"
72 #include "llvm/Support/CommandLine.h"
73 #include "llvm/Support/Debug.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
76 #include "llvm/Transforms/Utils/Local.h"
80 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
81 /// bail out. This threshold is far beyond the number of users that LSR can
82 /// conceivably solve, so it should not affect generated code, but catches the
83 /// worst cases before LSR burns too much compile time and stack space.
84 static const unsigned MaxIVUsers = 200;
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 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
125 void RegSortData::dump() const {
126 print(errs()); errs() << '\n';
132 /// RegUseTracker - Map register candidates to information about how they are
134 class RegUseTracker {
135 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
137 RegUsesTy RegUsesMap;
138 SmallVector<const SCEV *, 16> RegSequence;
141 void CountRegister(const SCEV *Reg, size_t LUIdx);
142 void DropRegister(const SCEV *Reg, size_t LUIdx);
143 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
145 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
147 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
151 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
152 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
153 iterator begin() { return RegSequence.begin(); }
154 iterator end() { return RegSequence.end(); }
155 const_iterator begin() const { return RegSequence.begin(); }
156 const_iterator end() const { return RegSequence.end(); }
162 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
163 std::pair<RegUsesTy::iterator, bool> Pair =
164 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
165 RegSortData &RSD = Pair.first->second;
167 RegSequence.push_back(Reg);
168 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
169 RSD.UsedByIndices.set(LUIdx);
173 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
174 RegUsesTy::iterator It = RegUsesMap.find(Reg);
175 assert(It != RegUsesMap.end());
176 RegSortData &RSD = It->second;
177 assert(RSD.UsedByIndices.size() > LUIdx);
178 RSD.UsedByIndices.reset(LUIdx);
182 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
183 assert(LUIdx <= LastLUIdx);
185 // Update RegUses. The data structure is not optimized for this purpose;
186 // we must iterate through it and update each of the bit vectors.
187 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
189 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
190 if (LUIdx < UsedByIndices.size())
191 UsedByIndices[LUIdx] =
192 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
193 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
198 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
199 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
200 if (I == RegUsesMap.end())
202 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
203 int i = UsedByIndices.find_first();
204 if (i == -1) return false;
205 if ((size_t)i != LUIdx) return true;
206 return UsedByIndices.find_next(i) != -1;
209 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
210 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
211 assert(I != RegUsesMap.end() && "Unknown register!");
212 return I->second.UsedByIndices;
215 void RegUseTracker::clear() {
222 /// Formula - This class holds information that describes a formula for
223 /// computing satisfying a use. It may include broken-out immediates and scaled
226 /// Global base address used for complex addressing.
229 /// Base offset for complex addressing.
232 /// Whether any complex addressing has a base register.
235 /// The scale of any complex addressing.
238 /// BaseRegs - The list of "base" registers for this use. When this is
240 SmallVector<const SCEV *, 4> BaseRegs;
242 /// ScaledReg - The 'scaled' register for this use. This should be non-null
243 /// when Scale is not zero.
244 const SCEV *ScaledReg;
246 /// UnfoldedOffset - An additional constant offset which added near the
247 /// use. This requires a temporary register, but the offset itself can
248 /// live in an add immediate field rather than a register.
249 int64_t UnfoldedOffset;
252 : BaseGV(0), BaseOffset(0), HasBaseReg(false), Scale(0), ScaledReg(0),
255 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
257 unsigned getNumRegs() const;
258 Type *getType() const;
260 void DeleteBaseReg(const SCEV *&S);
262 bool referencesReg(const SCEV *S) const;
263 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
264 const RegUseTracker &RegUses) const;
266 void print(raw_ostream &OS) const;
272 /// DoInitialMatch - Recursion helper for InitialMatch.
273 static void DoInitialMatch(const SCEV *S, Loop *L,
274 SmallVectorImpl<const SCEV *> &Good,
275 SmallVectorImpl<const SCEV *> &Bad,
276 ScalarEvolution &SE) {
277 // Collect expressions which properly dominate the loop header.
278 if (SE.properlyDominates(S, L->getHeader())) {
283 // Look at add operands.
284 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
285 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
287 DoInitialMatch(*I, L, Good, Bad, SE);
291 // Look at addrec operands.
292 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
293 if (!AR->getStart()->isZero()) {
294 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
295 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
296 AR->getStepRecurrence(SE),
297 // FIXME: AR->getNoWrapFlags()
298 AR->getLoop(), SCEV::FlagAnyWrap),
303 // Handle a multiplication by -1 (negation) if it didn't fold.
304 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
305 if (Mul->getOperand(0)->isAllOnesValue()) {
306 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
307 const SCEV *NewMul = SE.getMulExpr(Ops);
309 SmallVector<const SCEV *, 4> MyGood;
310 SmallVector<const SCEV *, 4> MyBad;
311 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
312 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
313 SE.getEffectiveSCEVType(NewMul->getType())));
314 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
315 E = MyGood.end(); I != E; ++I)
316 Good.push_back(SE.getMulExpr(NegOne, *I));
317 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
318 E = MyBad.end(); I != E; ++I)
319 Bad.push_back(SE.getMulExpr(NegOne, *I));
323 // Ok, we can't do anything interesting. Just stuff the whole thing into a
324 // register and hope for the best.
328 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
329 /// attempting to keep all loop-invariant and loop-computable values in a
330 /// single base register.
331 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
332 SmallVector<const SCEV *, 4> Good;
333 SmallVector<const SCEV *, 4> Bad;
334 DoInitialMatch(S, L, Good, Bad, SE);
336 const SCEV *Sum = SE.getAddExpr(Good);
338 BaseRegs.push_back(Sum);
342 const SCEV *Sum = SE.getAddExpr(Bad);
344 BaseRegs.push_back(Sum);
349 /// getNumRegs - Return the total number of register operands used by this
350 /// formula. This does not include register uses implied by non-constant
352 unsigned Formula::getNumRegs() const {
353 return !!ScaledReg + BaseRegs.size();
356 /// getType - Return the type of this formula, if it has one, or null
357 /// otherwise. This type is meaningless except for the bit size.
358 Type *Formula::getType() const {
359 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
360 ScaledReg ? ScaledReg->getType() :
361 BaseGV ? BaseGV->getType() :
365 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
366 void Formula::DeleteBaseReg(const SCEV *&S) {
367 if (&S != &BaseRegs.back())
368 std::swap(S, BaseRegs.back());
372 /// referencesReg - Test if this formula references the given register.
373 bool Formula::referencesReg(const SCEV *S) const {
374 return S == ScaledReg ||
375 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
378 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
379 /// which are used by uses other than the use with the given index.
380 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
381 const RegUseTracker &RegUses) const {
383 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
385 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
386 E = BaseRegs.end(); I != E; ++I)
387 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
392 void Formula::print(raw_ostream &OS) const {
395 if (!First) OS << " + "; else First = false;
396 BaseGV->printAsOperand(OS, /*PrintType=*/false);
398 if (BaseOffset != 0) {
399 if (!First) OS << " + "; else First = false;
402 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
403 E = BaseRegs.end(); I != E; ++I) {
404 if (!First) OS << " + "; else First = false;
405 OS << "reg(" << **I << ')';
407 if (HasBaseReg && BaseRegs.empty()) {
408 if (!First) OS << " + "; else First = false;
409 OS << "**error: HasBaseReg**";
410 } else if (!HasBaseReg && !BaseRegs.empty()) {
411 if (!First) OS << " + "; else First = false;
412 OS << "**error: !HasBaseReg**";
415 if (!First) OS << " + "; else First = false;
416 OS << Scale << "*reg(";
423 if (UnfoldedOffset != 0) {
424 if (!First) OS << " + "; else First = false;
425 OS << "imm(" << UnfoldedOffset << ')';
429 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
430 void Formula::dump() const {
431 print(errs()); errs() << '\n';
435 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
436 /// without changing its value.
437 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
439 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
440 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
443 /// isAddSExtable - Return true if the given add can be sign-extended
444 /// without changing its value.
445 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
447 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
448 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
451 /// isMulSExtable - Return true if the given mul can be sign-extended
452 /// without changing its value.
453 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
455 IntegerType::get(SE.getContext(),
456 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
457 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
460 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
461 /// and if the remainder is known to be zero, or null otherwise. If
462 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
463 /// to Y, ignoring that the multiplication may overflow, which is useful when
464 /// the result will be used in a context where the most significant bits are
466 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
468 bool IgnoreSignificantBits = false) {
469 // Handle the trivial case, which works for any SCEV type.
471 return SE.getConstant(LHS->getType(), 1);
473 // Handle a few RHS special cases.
474 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
476 const APInt &RA = RC->getValue()->getValue();
477 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
479 if (RA.isAllOnesValue())
480 return SE.getMulExpr(LHS, RC);
481 // Handle x /s 1 as x.
486 // Check for a division of a constant by a constant.
487 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
490 const APInt &LA = C->getValue()->getValue();
491 const APInt &RA = RC->getValue()->getValue();
492 if (LA.srem(RA) != 0)
494 return SE.getConstant(LA.sdiv(RA));
497 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
498 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
499 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
500 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
501 IgnoreSignificantBits);
503 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
504 IgnoreSignificantBits);
505 if (!Start) return 0;
506 // FlagNW is independent of the start value, step direction, and is
507 // preserved with smaller magnitude steps.
508 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
509 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
514 // Distribute the sdiv over add operands, if the add doesn't overflow.
515 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
516 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
517 SmallVector<const SCEV *, 8> Ops;
518 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
520 const SCEV *Op = getExactSDiv(*I, RHS, SE,
521 IgnoreSignificantBits);
525 return SE.getAddExpr(Ops);
530 // Check for a multiply operand that we can pull RHS out of.
531 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
532 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
533 SmallVector<const SCEV *, 4> Ops;
535 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
539 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
540 IgnoreSignificantBits)) {
546 return Found ? SE.getMulExpr(Ops) : 0;
551 // Otherwise we don't know.
555 /// ExtractImmediate - If S involves the addition of a constant integer value,
556 /// return that integer value, and mutate S to point to a new SCEV with that
558 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
559 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
560 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
561 S = SE.getConstant(C->getType(), 0);
562 return C->getValue()->getSExtValue();
564 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
565 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
566 int64_t Result = ExtractImmediate(NewOps.front(), SE);
568 S = SE.getAddExpr(NewOps);
570 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
571 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
572 int64_t Result = ExtractImmediate(NewOps.front(), SE);
574 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
575 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
582 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
583 /// return that symbol, and mutate S to point to a new SCEV with that
585 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
586 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
587 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
588 S = SE.getConstant(GV->getType(), 0);
591 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
592 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
593 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
595 S = SE.getAddExpr(NewOps);
597 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
598 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
599 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
601 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
602 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
609 /// isAddressUse - Returns true if the specified instruction is using the
610 /// specified value as an address.
611 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
612 bool isAddress = isa<LoadInst>(Inst);
613 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
614 if (SI->getOperand(1) == OperandVal)
616 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
617 // Addressing modes can also be folded into prefetches and a variety
619 switch (II->getIntrinsicID()) {
621 case Intrinsic::prefetch:
622 case Intrinsic::x86_sse_storeu_ps:
623 case Intrinsic::x86_sse2_storeu_pd:
624 case Intrinsic::x86_sse2_storeu_dq:
625 case Intrinsic::x86_sse2_storel_dq:
626 if (II->getArgOperand(0) == OperandVal)
634 /// getAccessType - Return the type of the memory being accessed.
635 static Type *getAccessType(const Instruction *Inst) {
636 Type *AccessTy = Inst->getType();
637 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
638 AccessTy = SI->getOperand(0)->getType();
639 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
640 // Addressing modes can also be folded into prefetches and a variety
642 switch (II->getIntrinsicID()) {
644 case Intrinsic::x86_sse_storeu_ps:
645 case Intrinsic::x86_sse2_storeu_pd:
646 case Intrinsic::x86_sse2_storeu_dq:
647 case Intrinsic::x86_sse2_storel_dq:
648 AccessTy = II->getArgOperand(0)->getType();
653 // All pointers have the same requirements, so canonicalize them to an
654 // arbitrary pointer type to minimize variation.
655 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
656 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
657 PTy->getAddressSpace());
662 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
663 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
664 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
665 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
666 if (SE.isSCEVable(PN->getType()) &&
667 (SE.getEffectiveSCEVType(PN->getType()) ==
668 SE.getEffectiveSCEVType(AR->getType())) &&
669 SE.getSCEV(PN) == AR)
675 /// Check if expanding this expression is likely to incur significant cost. This
676 /// is tricky because SCEV doesn't track which expressions are actually computed
677 /// by the current IR.
679 /// We currently allow expansion of IV increments that involve adds,
680 /// multiplication by constants, and AddRecs from existing phis.
682 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
683 /// obvious multiple of the UDivExpr.
684 static bool isHighCostExpansion(const SCEV *S,
685 SmallPtrSet<const SCEV*, 8> &Processed,
686 ScalarEvolution &SE) {
687 // Zero/One operand expressions
688 switch (S->getSCEVType()) {
693 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
696 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
699 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
703 if (!Processed.insert(S))
706 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
707 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
709 if (isHighCostExpansion(*I, Processed, SE))
715 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
716 if (Mul->getNumOperands() == 2) {
717 // Multiplication by a constant is ok
718 if (isa<SCEVConstant>(Mul->getOperand(0)))
719 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
721 // If we have the value of one operand, check if an existing
722 // multiplication already generates this expression.
723 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
724 Value *UVal = U->getValue();
725 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
727 // If U is a constant, it may be used by a ConstantExpr.
728 Instruction *User = dyn_cast<Instruction>(*UI);
729 if (User && User->getOpcode() == Instruction::Mul
730 && SE.isSCEVable(User->getType())) {
731 return SE.getSCEV(User) == Mul;
738 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
739 if (isExistingPhi(AR, SE))
743 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
747 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
748 /// specified set are trivially dead, delete them and see if this makes any of
749 /// their operands subsequently dead.
751 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
752 bool Changed = false;
754 while (!DeadInsts.empty()) {
755 Value *V = DeadInsts.pop_back_val();
756 Instruction *I = dyn_cast_or_null<Instruction>(V);
758 if (I == 0 || !isInstructionTriviallyDead(I))
761 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
762 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
765 DeadInsts.push_back(U);
768 I->eraseFromParent();
778 // Check if it is legal to fold 2 base registers.
779 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
781 // Get the cost of the scaling factor used in F for LU.
782 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
783 const LSRUse &LU, const Formula &F);
787 /// Cost - This class is used to measure and compare candidate formulae.
789 /// TODO: Some of these could be merged. Also, a lexical ordering
790 /// isn't always optimal.
794 unsigned NumBaseAdds;
801 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
802 SetupCost(0), ScaleCost(0) {}
804 bool operator<(const Cost &Other) const;
809 // Once any of the metrics loses, they must all remain losers.
811 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
812 | ImmCost | SetupCost | ScaleCost) != ~0u)
813 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
814 & ImmCost & SetupCost & ScaleCost) == ~0u);
819 assert(isValid() && "invalid cost");
820 return NumRegs == ~0u;
823 void RateFormula(const TargetTransformInfo &TTI,
825 SmallPtrSet<const SCEV *, 16> &Regs,
826 const DenseSet<const SCEV *> &VisitedRegs,
828 const SmallVectorImpl<int64_t> &Offsets,
829 ScalarEvolution &SE, DominatorTree &DT,
831 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
833 void print(raw_ostream &OS) const;
837 void RateRegister(const SCEV *Reg,
838 SmallPtrSet<const SCEV *, 16> &Regs,
840 ScalarEvolution &SE, DominatorTree &DT);
841 void RatePrimaryRegister(const SCEV *Reg,
842 SmallPtrSet<const SCEV *, 16> &Regs,
844 ScalarEvolution &SE, DominatorTree &DT,
845 SmallPtrSet<const SCEV *, 16> *LoserRegs);
850 /// RateRegister - Tally up interesting quantities from the given register.
851 void Cost::RateRegister(const SCEV *Reg,
852 SmallPtrSet<const SCEV *, 16> &Regs,
854 ScalarEvolution &SE, DominatorTree &DT) {
855 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
856 // If this is an addrec for another loop, don't second-guess its addrec phi
857 // nodes. LSR isn't currently smart enough to reason about more than one
858 // loop at a time. LSR has already run on inner loops, will not run on outer
859 // loops, and cannot be expected to change sibling loops.
860 if (AR->getLoop() != L) {
861 // If the AddRec exists, consider it's register free and leave it alone.
862 if (isExistingPhi(AR, SE))
865 // Otherwise, do not consider this formula at all.
869 AddRecCost += 1; /// TODO: This should be a function of the stride.
871 // Add the step value register, if it needs one.
872 // TODO: The non-affine case isn't precisely modeled here.
873 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
874 if (!Regs.count(AR->getOperand(1))) {
875 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
883 // Rough heuristic; favor registers which don't require extra setup
884 // instructions in the preheader.
885 if (!isa<SCEVUnknown>(Reg) &&
886 !isa<SCEVConstant>(Reg) &&
887 !(isa<SCEVAddRecExpr>(Reg) &&
888 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
889 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
892 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
893 SE.hasComputableLoopEvolution(Reg, L);
896 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
897 /// before, rate it. Optional LoserRegs provides a way to declare any formula
898 /// that refers to one of those regs an instant loser.
899 void Cost::RatePrimaryRegister(const SCEV *Reg,
900 SmallPtrSet<const SCEV *, 16> &Regs,
902 ScalarEvolution &SE, DominatorTree &DT,
903 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
904 if (LoserRegs && LoserRegs->count(Reg)) {
908 if (Regs.insert(Reg)) {
909 RateRegister(Reg, Regs, L, SE, DT);
910 if (LoserRegs && isLoser())
911 LoserRegs->insert(Reg);
915 void Cost::RateFormula(const TargetTransformInfo &TTI,
917 SmallPtrSet<const SCEV *, 16> &Regs,
918 const DenseSet<const SCEV *> &VisitedRegs,
920 const SmallVectorImpl<int64_t> &Offsets,
921 ScalarEvolution &SE, DominatorTree &DT,
923 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
924 // Tally up the registers.
925 if (const SCEV *ScaledReg = F.ScaledReg) {
926 if (VisitedRegs.count(ScaledReg)) {
930 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
934 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
935 E = F.BaseRegs.end(); I != E; ++I) {
936 const SCEV *BaseReg = *I;
937 if (VisitedRegs.count(BaseReg)) {
941 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
946 // Determine how many (unfolded) adds we'll need inside the loop.
947 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
948 if (NumBaseParts > 1)
949 // Do not count the base and a possible second register if the target
950 // allows to fold 2 registers.
951 NumBaseAdds += NumBaseParts - (1 + isLegal2RegAMUse(TTI, LU, F));
953 // Accumulate non-free scaling amounts.
954 ScaleCost += getScalingFactorCost(TTI, LU, F);
956 // Tally up the non-zero immediates.
957 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
958 E = Offsets.end(); I != E; ++I) {
959 int64_t Offset = (uint64_t)*I + F.BaseOffset;
961 ImmCost += 64; // Handle symbolic values conservatively.
962 // TODO: This should probably be the pointer size.
963 else if (Offset != 0)
964 ImmCost += APInt(64, Offset, true).getMinSignedBits();
966 assert(isValid() && "invalid cost");
969 /// Lose - Set this cost to a losing value.
980 /// operator< - Choose the lower cost.
981 bool Cost::operator<(const Cost &Other) const {
982 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
983 ImmCost, SetupCost) <
984 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
985 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
989 void Cost::print(raw_ostream &OS) const {
990 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
992 OS << ", with addrec cost " << AddRecCost;
994 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
995 if (NumBaseAdds != 0)
996 OS << ", plus " << NumBaseAdds << " base add"
997 << (NumBaseAdds == 1 ? "" : "s");
999 OS << ", plus " << ScaleCost << " scale cost";
1001 OS << ", plus " << ImmCost << " imm cost";
1003 OS << ", plus " << SetupCost << " setup cost";
1006 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1007 void Cost::dump() const {
1008 print(errs()); errs() << '\n';
1014 /// LSRFixup - An operand value in an instruction which is to be replaced
1015 /// with some equivalent, possibly strength-reduced, replacement.
1017 /// UserInst - The instruction which will be updated.
1018 Instruction *UserInst;
1020 /// OperandValToReplace - The operand of the instruction which will
1021 /// be replaced. The operand may be used more than once; every instance
1022 /// will be replaced.
1023 Value *OperandValToReplace;
1025 /// PostIncLoops - If this user is to use the post-incremented value of an
1026 /// induction variable, this variable is non-null and holds the loop
1027 /// associated with the induction variable.
1028 PostIncLoopSet PostIncLoops;
1030 /// LUIdx - The index of the LSRUse describing the expression which
1031 /// this fixup needs, minus an offset (below).
1034 /// Offset - A constant offset to be added to the LSRUse expression.
1035 /// This allows multiple fixups to share the same LSRUse with different
1036 /// offsets, for example in an unrolled loop.
1039 bool isUseFullyOutsideLoop(const Loop *L) const;
1043 void print(raw_ostream &OS) const;
1049 LSRFixup::LSRFixup()
1050 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1052 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1053 /// value outside of the given loop.
1054 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1055 // PHI nodes use their value in their incoming blocks.
1056 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1057 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1058 if (PN->getIncomingValue(i) == OperandValToReplace &&
1059 L->contains(PN->getIncomingBlock(i)))
1064 return !L->contains(UserInst);
1067 void LSRFixup::print(raw_ostream &OS) const {
1069 // Store is common and interesting enough to be worth special-casing.
1070 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1072 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1073 } else if (UserInst->getType()->isVoidTy())
1074 OS << UserInst->getOpcodeName();
1076 UserInst->printAsOperand(OS, /*PrintType=*/false);
1078 OS << ", OperandValToReplace=";
1079 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1081 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1082 E = PostIncLoops.end(); I != E; ++I) {
1083 OS << ", PostIncLoop=";
1084 (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1087 if (LUIdx != ~size_t(0))
1088 OS << ", LUIdx=" << LUIdx;
1091 OS << ", Offset=" << Offset;
1094 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1095 void LSRFixup::dump() const {
1096 print(errs()); errs() << '\n';
1102 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1103 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1104 struct UniquifierDenseMapInfo {
1105 static SmallVector<const SCEV *, 4> getEmptyKey() {
1106 SmallVector<const SCEV *, 4> V;
1107 V.push_back(reinterpret_cast<const SCEV *>(-1));
1111 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1112 SmallVector<const SCEV *, 4> V;
1113 V.push_back(reinterpret_cast<const SCEV *>(-2));
1117 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1118 unsigned Result = 0;
1119 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1120 E = V.end(); I != E; ++I)
1121 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1125 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1126 const SmallVector<const SCEV *, 4> &RHS) {
1131 /// LSRUse - This class holds the state that LSR keeps for each use in
1132 /// IVUsers, as well as uses invented by LSR itself. It includes information
1133 /// about what kinds of things can be folded into the user, information about
1134 /// the user itself, and information about how the use may be satisfied.
1135 /// TODO: Represent multiple users of the same expression in common?
1137 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1140 /// KindType - An enum for a kind of use, indicating what types of
1141 /// scaled and immediate operands it might support.
1143 Basic, ///< A normal use, with no folding.
1144 Special, ///< A special case of basic, allowing -1 scales.
1145 Address, ///< An address use; folding according to TargetLowering
1146 ICmpZero ///< An equality icmp with both operands folded into one.
1147 // TODO: Add a generic icmp too?
1153 SmallVector<int64_t, 8> Offsets;
1157 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1158 /// LSRUse are outside of the loop, in which case some special-case heuristics
1160 bool AllFixupsOutsideLoop;
1162 /// RigidFormula is set to true to guarantee that this use will be associated
1163 /// with a single formula--the one that initially matched. Some SCEV
1164 /// expressions cannot be expanded. This allows LSR to consider the registers
1165 /// used by those expressions without the need to expand them later after
1166 /// changing the formula.
1169 /// WidestFixupType - This records the widest use type for any fixup using
1170 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1171 /// max fixup widths to be equivalent, because the narrower one may be relying
1172 /// on the implicit truncation to truncate away bogus bits.
1173 Type *WidestFixupType;
1175 /// Formulae - A list of ways to build a value that can satisfy this user.
1176 /// After the list is populated, one of these is selected heuristically and
1177 /// used to formulate a replacement for OperandValToReplace in UserInst.
1178 SmallVector<Formula, 12> Formulae;
1180 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1181 SmallPtrSet<const SCEV *, 4> Regs;
1183 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1184 MinOffset(INT64_MAX),
1185 MaxOffset(INT64_MIN),
1186 AllFixupsOutsideLoop(true),
1187 RigidFormula(false),
1188 WidestFixupType(0) {}
1190 bool HasFormulaWithSameRegs(const Formula &F) const;
1191 bool InsertFormula(const Formula &F);
1192 void DeleteFormula(Formula &F);
1193 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1195 void print(raw_ostream &OS) const;
1201 /// HasFormula - Test whether this use as a formula which has the same
1202 /// registers as the given formula.
1203 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1204 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1205 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1206 // Unstable sort by host order ok, because this is only used for uniquifying.
1207 std::sort(Key.begin(), Key.end());
1208 return Uniquifier.count(Key);
1211 /// InsertFormula - If the given formula has not yet been inserted, add it to
1212 /// the list, and return true. Return false otherwise.
1213 bool LSRUse::InsertFormula(const Formula &F) {
1214 if (!Formulae.empty() && RigidFormula)
1217 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1218 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1219 // Unstable sort by host order ok, because this is only used for uniquifying.
1220 std::sort(Key.begin(), Key.end());
1222 if (!Uniquifier.insert(Key).second)
1225 // Using a register to hold the value of 0 is not profitable.
1226 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1227 "Zero allocated in a scaled register!");
1229 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1230 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1231 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1234 // Add the formula to the list.
1235 Formulae.push_back(F);
1237 // Record registers now being used by this use.
1238 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1243 /// DeleteFormula - Remove the given formula from this use's list.
1244 void LSRUse::DeleteFormula(Formula &F) {
1245 if (&F != &Formulae.back())
1246 std::swap(F, Formulae.back());
1247 Formulae.pop_back();
1250 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1251 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1252 // Now that we've filtered out some formulae, recompute the Regs set.
1253 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1255 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1256 E = Formulae.end(); I != E; ++I) {
1257 const Formula &F = *I;
1258 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1259 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1262 // Update the RegTracker.
1263 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1264 E = OldRegs.end(); I != E; ++I)
1265 if (!Regs.count(*I))
1266 RegUses.DropRegister(*I, LUIdx);
1269 void LSRUse::print(raw_ostream &OS) const {
1270 OS << "LSR Use: Kind=";
1272 case Basic: OS << "Basic"; break;
1273 case Special: OS << "Special"; break;
1274 case ICmpZero: OS << "ICmpZero"; break;
1276 OS << "Address of ";
1277 if (AccessTy->isPointerTy())
1278 OS << "pointer"; // the full pointer type could be really verbose
1283 OS << ", Offsets={";
1284 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1285 E = Offsets.end(); I != E; ++I) {
1287 if (std::next(I) != E)
1292 if (AllFixupsOutsideLoop)
1293 OS << ", all-fixups-outside-loop";
1295 if (WidestFixupType)
1296 OS << ", widest fixup type: " << *WidestFixupType;
1299 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1300 void LSRUse::dump() const {
1301 print(errs()); errs() << '\n';
1305 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1306 /// be completely folded into the user instruction at isel time. This includes
1307 /// address-mode folding and special icmp tricks.
1308 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1309 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1310 bool HasBaseReg, int64_t Scale) {
1312 case LSRUse::Address:
1313 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1315 // Otherwise, just guess that reg+reg addressing is legal.
1318 case LSRUse::ICmpZero:
1319 // There's not even a target hook for querying whether it would be legal to
1320 // fold a GV into an ICmp.
1324 // ICmp only has two operands; don't allow more than two non-trivial parts.
1325 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1328 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1329 // putting the scaled register in the other operand of the icmp.
1330 if (Scale != 0 && Scale != -1)
1333 // If we have low-level target information, ask the target if it can fold an
1334 // integer immediate on an icmp.
1335 if (BaseOffset != 0) {
1337 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1338 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1339 // Offs is the ICmp immediate.
1341 // The cast does the right thing with INT64_MIN.
1342 BaseOffset = -(uint64_t)BaseOffset;
1343 return TTI.isLegalICmpImmediate(BaseOffset);
1346 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1350 // Only handle single-register values.
1351 return !BaseGV && Scale == 0 && BaseOffset == 0;
1353 case LSRUse::Special:
1354 // Special case Basic to handle -1 scales.
1355 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1358 llvm_unreachable("Invalid LSRUse Kind!");
1361 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1362 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1363 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1365 // Check for overflow.
1366 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1369 MinOffset = (uint64_t)BaseOffset + MinOffset;
1370 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1373 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1375 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1377 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1380 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1381 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1383 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1384 F.BaseOffset, F.HasBaseReg, F.Scale);
1387 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
1389 // If F is used as an Addressing Mode, it may fold one Base plus one
1390 // scaled register. If the scaled register is nil, do as if another
1391 // element of the base regs is a 1-scaled register.
1392 // This is possible if BaseRegs has at least 2 registers.
1394 // If this is not an address calculation, this is not an addressing mode
1396 if (LU.Kind != LSRUse::Address)
1399 // F is already scaled.
1403 // We need to keep one register for the base and one to scale.
1404 if (F.BaseRegs.size() < 2)
1407 return isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
1408 F.BaseGV, F.BaseOffset, F.HasBaseReg, 1);
1411 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1412 const LSRUse &LU, const Formula &F) {
1415 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1416 LU.AccessTy, F) && "Illegal formula in use.");
1419 case LSRUse::Address: {
1420 // Check the scaling factor cost with both the min and max offsets.
1421 int ScaleCostMinOffset =
1422 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1423 F.BaseOffset + LU.MinOffset,
1424 F.HasBaseReg, F.Scale);
1425 int ScaleCostMaxOffset =
1426 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1427 F.BaseOffset + LU.MaxOffset,
1428 F.HasBaseReg, F.Scale);
1430 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1431 "Legal addressing mode has an illegal cost!");
1432 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1434 case LSRUse::ICmpZero:
1435 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
1436 // Therefore, return 0 in case F.Scale == -1.
1437 return F.Scale != -1;
1440 case LSRUse::Special:
1444 llvm_unreachable("Invalid LSRUse Kind!");
1447 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1448 LSRUse::KindType Kind, Type *AccessTy,
1449 GlobalValue *BaseGV, int64_t BaseOffset,
1451 // Fast-path: zero is always foldable.
1452 if (BaseOffset == 0 && !BaseGV) return true;
1454 // Conservatively, create an address with an immediate and a
1455 // base and a scale.
1456 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1458 // Canonicalize a scale of 1 to a base register if the formula doesn't
1459 // already have a base register.
1460 if (!HasBaseReg && Scale == 1) {
1465 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1468 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1469 ScalarEvolution &SE, int64_t MinOffset,
1470 int64_t MaxOffset, LSRUse::KindType Kind,
1471 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1472 // Fast-path: zero is always foldable.
1473 if (S->isZero()) return true;
1475 // Conservatively, create an address with an immediate and a
1476 // base and a scale.
1477 int64_t BaseOffset = ExtractImmediate(S, SE);
1478 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1480 // If there's anything else involved, it's not foldable.
1481 if (!S->isZero()) return false;
1483 // Fast-path: zero is always foldable.
1484 if (BaseOffset == 0 && !BaseGV) return true;
1486 // Conservatively, create an address with an immediate and a
1487 // base and a scale.
1488 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1490 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1491 BaseOffset, HasBaseReg, Scale);
1496 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1497 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1498 struct UseMapDenseMapInfo {
1499 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1500 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1503 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1504 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1508 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1509 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1510 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1514 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1515 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1520 /// IVInc - An individual increment in a Chain of IV increments.
1521 /// Relate an IV user to an expression that computes the IV it uses from the IV
1522 /// used by the previous link in the Chain.
1524 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1525 /// original IVOperand. The head of the chain's IVOperand is only valid during
1526 /// chain collection, before LSR replaces IV users. During chain generation,
1527 /// IncExpr can be used to find the new IVOperand that computes the same
1530 Instruction *UserInst;
1532 const SCEV *IncExpr;
1534 IVInc(Instruction *U, Value *O, const SCEV *E):
1535 UserInst(U), IVOperand(O), IncExpr(E) {}
1538 // IVChain - The list of IV increments in program order.
1539 // We typically add the head of a chain without finding subsequent links.
1541 SmallVector<IVInc,1> Incs;
1542 const SCEV *ExprBase;
1544 IVChain() : ExprBase(0) {}
1546 IVChain(const IVInc &Head, const SCEV *Base)
1547 : Incs(1, Head), ExprBase(Base) {}
1549 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1551 // begin - return the first increment in the chain.
1552 const_iterator begin() const {
1553 assert(!Incs.empty());
1554 return std::next(Incs.begin());
1556 const_iterator end() const {
1560 // hasIncs - Returns true if this chain contains any increments.
1561 bool hasIncs() const { return Incs.size() >= 2; }
1563 // add - Add an IVInc to the end of this chain.
1564 void add(const IVInc &X) { Incs.push_back(X); }
1566 // tailUserInst - Returns the last UserInst in the chain.
1567 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1569 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1571 bool isProfitableIncrement(const SCEV *OperExpr,
1572 const SCEV *IncExpr,
1576 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1577 /// Distinguish between FarUsers that definitely cross IV increments and
1578 /// NearUsers that may be used between IV increments.
1580 SmallPtrSet<Instruction*, 4> FarUsers;
1581 SmallPtrSet<Instruction*, 4> NearUsers;
1584 /// LSRInstance - This class holds state for the main loop strength reduction
1588 ScalarEvolution &SE;
1591 const TargetTransformInfo &TTI;
1595 /// IVIncInsertPos - This is the insert position that the current loop's
1596 /// induction variable increment should be placed. In simple loops, this is
1597 /// the latch block's terminator. But in more complicated cases, this is a
1598 /// position which will dominate all the in-loop post-increment users.
1599 Instruction *IVIncInsertPos;
1601 /// Factors - Interesting factors between use strides.
1602 SmallSetVector<int64_t, 8> Factors;
1604 /// Types - Interesting use types, to facilitate truncation reuse.
1605 SmallSetVector<Type *, 4> Types;
1607 /// Fixups - The list of operands which are to be replaced.
1608 SmallVector<LSRFixup, 16> Fixups;
1610 /// Uses - The list of interesting uses.
1611 SmallVector<LSRUse, 16> Uses;
1613 /// RegUses - Track which uses use which register candidates.
1614 RegUseTracker RegUses;
1616 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1617 // have more than a few IV increment chains in a loop. Missing a Chain falls
1618 // back to normal LSR behavior for those uses.
1619 static const unsigned MaxChains = 8;
1621 /// IVChainVec - IV users can form a chain of IV increments.
1622 SmallVector<IVChain, MaxChains> IVChainVec;
1624 /// IVIncSet - IV users that belong to profitable IVChains.
1625 SmallPtrSet<Use*, MaxChains> IVIncSet;
1627 void OptimizeShadowIV();
1628 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1629 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1630 void OptimizeLoopTermCond();
1632 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1633 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1634 void FinalizeChain(IVChain &Chain);
1635 void CollectChains();
1636 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1637 SmallVectorImpl<WeakVH> &DeadInsts);
1639 void CollectInterestingTypesAndFactors();
1640 void CollectFixupsAndInitialFormulae();
1642 LSRFixup &getNewFixup() {
1643 Fixups.push_back(LSRFixup());
1644 return Fixups.back();
1647 // Support for sharing of LSRUses between LSRFixups.
1648 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1650 UseMapDenseMapInfo> UseMapTy;
1653 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1654 LSRUse::KindType Kind, Type *AccessTy);
1656 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1657 LSRUse::KindType Kind,
1660 void DeleteUse(LSRUse &LU, size_t LUIdx);
1662 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1664 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1665 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1666 void CountRegisters(const Formula &F, size_t LUIdx);
1667 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1669 void CollectLoopInvariantFixupsAndFormulae();
1671 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1672 unsigned Depth = 0);
1673 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1674 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1675 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1676 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1677 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1678 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1679 void GenerateCrossUseConstantOffsets();
1680 void GenerateAllReuseFormulae();
1682 void FilterOutUndesirableDedicatedRegisters();
1684 size_t EstimateSearchSpaceComplexity() const;
1685 void NarrowSearchSpaceByDetectingSupersets();
1686 void NarrowSearchSpaceByCollapsingUnrolledCode();
1687 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1688 void NarrowSearchSpaceByPickingWinnerRegs();
1689 void NarrowSearchSpaceUsingHeuristics();
1691 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1693 SmallVectorImpl<const Formula *> &Workspace,
1694 const Cost &CurCost,
1695 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1696 DenseSet<const SCEV *> &VisitedRegs) const;
1697 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1699 BasicBlock::iterator
1700 HoistInsertPosition(BasicBlock::iterator IP,
1701 const SmallVectorImpl<Instruction *> &Inputs) const;
1702 BasicBlock::iterator
1703 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1706 SCEVExpander &Rewriter) const;
1708 Value *Expand(const LSRFixup &LF,
1710 BasicBlock::iterator IP,
1711 SCEVExpander &Rewriter,
1712 SmallVectorImpl<WeakVH> &DeadInsts) const;
1713 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1715 SCEVExpander &Rewriter,
1716 SmallVectorImpl<WeakVH> &DeadInsts,
1718 void Rewrite(const LSRFixup &LF,
1720 SCEVExpander &Rewriter,
1721 SmallVectorImpl<WeakVH> &DeadInsts,
1723 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1727 LSRInstance(Loop *L, Pass *P);
1729 bool getChanged() const { return Changed; }
1731 void print_factors_and_types(raw_ostream &OS) const;
1732 void print_fixups(raw_ostream &OS) const;
1733 void print_uses(raw_ostream &OS) const;
1734 void print(raw_ostream &OS) const;
1740 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1741 /// inside the loop then try to eliminate the cast operation.
1742 void LSRInstance::OptimizeShadowIV() {
1743 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1744 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1747 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1748 UI != E; /* empty */) {
1749 IVUsers::const_iterator CandidateUI = UI;
1751 Instruction *ShadowUse = CandidateUI->getUser();
1753 bool IsSigned = false;
1755 /* If shadow use is a int->float cast then insert a second IV
1756 to eliminate this cast.
1758 for (unsigned i = 0; i < n; ++i)
1764 for (unsigned i = 0; i < n; ++i, ++d)
1767 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1769 DestTy = UCast->getDestTy();
1771 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1773 DestTy = SCast->getDestTy();
1775 if (!DestTy) continue;
1777 // If target does not support DestTy natively then do not apply
1778 // this transformation.
1779 if (!TTI.isTypeLegal(DestTy)) continue;
1781 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1783 if (PH->getNumIncomingValues() != 2) continue;
1785 Type *SrcTy = PH->getType();
1786 int Mantissa = DestTy->getFPMantissaWidth();
1787 if (Mantissa == -1) continue;
1788 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1791 unsigned Entry, Latch;
1792 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1800 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1801 if (!Init) continue;
1802 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1803 (double)Init->getSExtValue() :
1804 (double)Init->getZExtValue());
1806 BinaryOperator *Incr =
1807 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1808 if (!Incr) continue;
1809 if (Incr->getOpcode() != Instruction::Add
1810 && Incr->getOpcode() != Instruction::Sub)
1813 /* Initialize new IV, double d = 0.0 in above example. */
1815 if (Incr->getOperand(0) == PH)
1816 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1817 else if (Incr->getOperand(1) == PH)
1818 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1824 // Ignore negative constants, as the code below doesn't handle them
1825 // correctly. TODO: Remove this restriction.
1826 if (!C->getValue().isStrictlyPositive()) continue;
1828 /* Add new PHINode. */
1829 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1831 /* create new increment. '++d' in above example. */
1832 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1833 BinaryOperator *NewIncr =
1834 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1835 Instruction::FAdd : Instruction::FSub,
1836 NewPH, CFP, "IV.S.next.", Incr);
1838 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1839 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1841 /* Remove cast operation */
1842 ShadowUse->replaceAllUsesWith(NewPH);
1843 ShadowUse->eraseFromParent();
1849 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1850 /// set the IV user and stride information and return true, otherwise return
1852 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1853 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1854 if (UI->getUser() == Cond) {
1855 // NOTE: we could handle setcc instructions with multiple uses here, but
1856 // InstCombine does it as well for simple uses, it's not clear that it
1857 // occurs enough in real life to handle.
1864 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1865 /// a max computation.
1867 /// This is a narrow solution to a specific, but acute, problem. For loops
1873 /// } while (++i < n);
1875 /// the trip count isn't just 'n', because 'n' might not be positive. And
1876 /// unfortunately this can come up even for loops where the user didn't use
1877 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1878 /// will commonly be lowered like this:
1884 /// } while (++i < n);
1887 /// and then it's possible for subsequent optimization to obscure the if
1888 /// test in such a way that indvars can't find it.
1890 /// When indvars can't find the if test in loops like this, it creates a
1891 /// max expression, which allows it to give the loop a canonical
1892 /// induction variable:
1895 /// max = n < 1 ? 1 : n;
1898 /// } while (++i != max);
1900 /// Canonical induction variables are necessary because the loop passes
1901 /// are designed around them. The most obvious example of this is the
1902 /// LoopInfo analysis, which doesn't remember trip count values. It
1903 /// expects to be able to rediscover the trip count each time it is
1904 /// needed, and it does this using a simple analysis that only succeeds if
1905 /// the loop has a canonical induction variable.
1907 /// However, when it comes time to generate code, the maximum operation
1908 /// can be quite costly, especially if it's inside of an outer loop.
1910 /// This function solves this problem by detecting this type of loop and
1911 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1912 /// the instructions for the maximum computation.
1914 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1915 // Check that the loop matches the pattern we're looking for.
1916 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1917 Cond->getPredicate() != CmpInst::ICMP_NE)
1920 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1921 if (!Sel || !Sel->hasOneUse()) return Cond;
1923 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1924 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1926 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1928 // Add one to the backedge-taken count to get the trip count.
1929 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1930 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1932 // Check for a max calculation that matches the pattern. There's no check
1933 // for ICMP_ULE here because the comparison would be with zero, which
1934 // isn't interesting.
1935 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1936 const SCEVNAryExpr *Max = 0;
1937 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1938 Pred = ICmpInst::ICMP_SLE;
1940 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1941 Pred = ICmpInst::ICMP_SLT;
1943 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1944 Pred = ICmpInst::ICMP_ULT;
1951 // To handle a max with more than two operands, this optimization would
1952 // require additional checking and setup.
1953 if (Max->getNumOperands() != 2)
1956 const SCEV *MaxLHS = Max->getOperand(0);
1957 const SCEV *MaxRHS = Max->getOperand(1);
1959 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1960 // for a comparison with 1. For <= and >=, a comparison with zero.
1962 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1965 // Check the relevant induction variable for conformance to
1967 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1968 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1969 if (!AR || !AR->isAffine() ||
1970 AR->getStart() != One ||
1971 AR->getStepRecurrence(SE) != One)
1974 assert(AR->getLoop() == L &&
1975 "Loop condition operand is an addrec in a different loop!");
1977 // Check the right operand of the select, and remember it, as it will
1978 // be used in the new comparison instruction.
1980 if (ICmpInst::isTrueWhenEqual(Pred)) {
1981 // Look for n+1, and grab n.
1982 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1983 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1984 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1985 NewRHS = BO->getOperand(0);
1986 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1987 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1988 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1989 NewRHS = BO->getOperand(0);
1992 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1993 NewRHS = Sel->getOperand(1);
1994 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1995 NewRHS = Sel->getOperand(2);
1996 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1997 NewRHS = SU->getValue();
1999 // Max doesn't match expected pattern.
2002 // Determine the new comparison opcode. It may be signed or unsigned,
2003 // and the original comparison may be either equality or inequality.
2004 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2005 Pred = CmpInst::getInversePredicate(Pred);
2007 // Ok, everything looks ok to change the condition into an SLT or SGE and
2008 // delete the max calculation.
2010 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2012 // Delete the max calculation instructions.
2013 Cond->replaceAllUsesWith(NewCond);
2014 CondUse->setUser(NewCond);
2015 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2016 Cond->eraseFromParent();
2017 Sel->eraseFromParent();
2018 if (Cmp->use_empty())
2019 Cmp->eraseFromParent();
2023 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2024 /// postinc iv when possible.
2026 LSRInstance::OptimizeLoopTermCond() {
2027 SmallPtrSet<Instruction *, 4> PostIncs;
2029 BasicBlock *LatchBlock = L->getLoopLatch();
2030 SmallVector<BasicBlock*, 8> ExitingBlocks;
2031 L->getExitingBlocks(ExitingBlocks);
2033 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2034 BasicBlock *ExitingBlock = ExitingBlocks[i];
2036 // Get the terminating condition for the loop if possible. If we
2037 // can, we want to change it to use a post-incremented version of its
2038 // induction variable, to allow coalescing the live ranges for the IV into
2039 // one register value.
2041 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2044 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2045 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2048 // Search IVUsesByStride to find Cond's IVUse if there is one.
2049 IVStrideUse *CondUse = 0;
2050 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2051 if (!FindIVUserForCond(Cond, CondUse))
2054 // If the trip count is computed in terms of a max (due to ScalarEvolution
2055 // being unable to find a sufficient guard, for example), change the loop
2056 // comparison to use SLT or ULT instead of NE.
2057 // One consequence of doing this now is that it disrupts the count-down
2058 // optimization. That's not always a bad thing though, because in such
2059 // cases it may still be worthwhile to avoid a max.
2060 Cond = OptimizeMax(Cond, CondUse);
2062 // If this exiting block dominates the latch block, it may also use
2063 // the post-inc value if it won't be shared with other uses.
2064 // Check for dominance.
2065 if (!DT.dominates(ExitingBlock, LatchBlock))
2068 // Conservatively avoid trying to use the post-inc value in non-latch
2069 // exits if there may be pre-inc users in intervening blocks.
2070 if (LatchBlock != ExitingBlock)
2071 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2072 // Test if the use is reachable from the exiting block. This dominator
2073 // query is a conservative approximation of reachability.
2074 if (&*UI != CondUse &&
2075 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2076 // Conservatively assume there may be reuse if the quotient of their
2077 // strides could be a legal scale.
2078 const SCEV *A = IU.getStride(*CondUse, L);
2079 const SCEV *B = IU.getStride(*UI, L);
2080 if (!A || !B) continue;
2081 if (SE.getTypeSizeInBits(A->getType()) !=
2082 SE.getTypeSizeInBits(B->getType())) {
2083 if (SE.getTypeSizeInBits(A->getType()) >
2084 SE.getTypeSizeInBits(B->getType()))
2085 B = SE.getSignExtendExpr(B, A->getType());
2087 A = SE.getSignExtendExpr(A, B->getType());
2089 if (const SCEVConstant *D =
2090 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2091 const ConstantInt *C = D->getValue();
2092 // Stride of one or negative one can have reuse with non-addresses.
2093 if (C->isOne() || C->isAllOnesValue())
2094 goto decline_post_inc;
2095 // Avoid weird situations.
2096 if (C->getValue().getMinSignedBits() >= 64 ||
2097 C->getValue().isMinSignedValue())
2098 goto decline_post_inc;
2099 // Check for possible scaled-address reuse.
2100 Type *AccessTy = getAccessType(UI->getUser());
2101 int64_t Scale = C->getSExtValue();
2102 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2104 /*HasBaseReg=*/ false, Scale))
2105 goto decline_post_inc;
2107 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2109 /*HasBaseReg=*/ false, Scale))
2110 goto decline_post_inc;
2114 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2117 // It's possible for the setcc instruction to be anywhere in the loop, and
2118 // possible for it to have multiple users. If it is not immediately before
2119 // the exiting block branch, move it.
2120 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2121 if (Cond->hasOneUse()) {
2122 Cond->moveBefore(TermBr);
2124 // Clone the terminating condition and insert into the loopend.
2125 ICmpInst *OldCond = Cond;
2126 Cond = cast<ICmpInst>(Cond->clone());
2127 Cond->setName(L->getHeader()->getName() + ".termcond");
2128 ExitingBlock->getInstList().insert(TermBr, Cond);
2130 // Clone the IVUse, as the old use still exists!
2131 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2132 TermBr->replaceUsesOfWith(OldCond, Cond);
2136 // If we get to here, we know that we can transform the setcc instruction to
2137 // use the post-incremented version of the IV, allowing us to coalesce the
2138 // live ranges for the IV correctly.
2139 CondUse->transformToPostInc(L);
2142 PostIncs.insert(Cond);
2146 // Determine an insertion point for the loop induction variable increment. It
2147 // must dominate all the post-inc comparisons we just set up, and it must
2148 // dominate the loop latch edge.
2149 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2150 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2151 E = PostIncs.end(); I != E; ++I) {
2153 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2155 if (BB == (*I)->getParent())
2156 IVIncInsertPos = *I;
2157 else if (BB != IVIncInsertPos->getParent())
2158 IVIncInsertPos = BB->getTerminator();
2162 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2163 /// at the given offset and other details. If so, update the use and
2166 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2167 LSRUse::KindType Kind, Type *AccessTy) {
2168 int64_t NewMinOffset = LU.MinOffset;
2169 int64_t NewMaxOffset = LU.MaxOffset;
2170 Type *NewAccessTy = AccessTy;
2172 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2173 // something conservative, however this can pessimize in the case that one of
2174 // the uses will have all its uses outside the loop, for example.
2175 if (LU.Kind != Kind)
2177 // Conservatively assume HasBaseReg is true for now.
2178 if (NewOffset < LU.MinOffset) {
2179 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2180 LU.MaxOffset - NewOffset, HasBaseReg))
2182 NewMinOffset = NewOffset;
2183 } else if (NewOffset > LU.MaxOffset) {
2184 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2185 NewOffset - LU.MinOffset, HasBaseReg))
2187 NewMaxOffset = NewOffset;
2189 // Check for a mismatched access type, and fall back conservatively as needed.
2190 // TODO: Be less conservative when the type is similar and can use the same
2191 // addressing modes.
2192 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2193 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2196 LU.MinOffset = NewMinOffset;
2197 LU.MaxOffset = NewMaxOffset;
2198 LU.AccessTy = NewAccessTy;
2199 if (NewOffset != LU.Offsets.back())
2200 LU.Offsets.push_back(NewOffset);
2204 /// getUse - Return an LSRUse index and an offset value for a fixup which
2205 /// needs the given expression, with the given kind and optional access type.
2206 /// Either reuse an existing use or create a new one, as needed.
2207 std::pair<size_t, int64_t>
2208 LSRInstance::getUse(const SCEV *&Expr,
2209 LSRUse::KindType Kind, Type *AccessTy) {
2210 const SCEV *Copy = Expr;
2211 int64_t Offset = ExtractImmediate(Expr, SE);
2213 // Basic uses can't accept any offset, for example.
2214 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2215 Offset, /*HasBaseReg=*/ true)) {
2220 std::pair<UseMapTy::iterator, bool> P =
2221 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2223 // A use already existed with this base.
2224 size_t LUIdx = P.first->second;
2225 LSRUse &LU = Uses[LUIdx];
2226 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2228 return std::make_pair(LUIdx, Offset);
2231 // Create a new use.
2232 size_t LUIdx = Uses.size();
2233 P.first->second = LUIdx;
2234 Uses.push_back(LSRUse(Kind, AccessTy));
2235 LSRUse &LU = Uses[LUIdx];
2237 // We don't need to track redundant offsets, but we don't need to go out
2238 // of our way here to avoid them.
2239 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2240 LU.Offsets.push_back(Offset);
2242 LU.MinOffset = Offset;
2243 LU.MaxOffset = Offset;
2244 return std::make_pair(LUIdx, Offset);
2247 /// DeleteUse - Delete the given use from the Uses list.
2248 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2249 if (&LU != &Uses.back())
2250 std::swap(LU, Uses.back());
2254 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2257 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2258 /// a formula that has the same registers as the given formula.
2260 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2261 const LSRUse &OrigLU) {
2262 // Search all uses for the formula. This could be more clever.
2263 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2264 LSRUse &LU = Uses[LUIdx];
2265 // Check whether this use is close enough to OrigLU, to see whether it's
2266 // worthwhile looking through its formulae.
2267 // Ignore ICmpZero uses because they may contain formulae generated by
2268 // GenerateICmpZeroScales, in which case adding fixup offsets may
2270 if (&LU != &OrigLU &&
2271 LU.Kind != LSRUse::ICmpZero &&
2272 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2273 LU.WidestFixupType == OrigLU.WidestFixupType &&
2274 LU.HasFormulaWithSameRegs(OrigF)) {
2275 // Scan through this use's formulae.
2276 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2277 E = LU.Formulae.end(); I != E; ++I) {
2278 const Formula &F = *I;
2279 // Check to see if this formula has the same registers and symbols
2281 if (F.BaseRegs == OrigF.BaseRegs &&
2282 F.ScaledReg == OrigF.ScaledReg &&
2283 F.BaseGV == OrigF.BaseGV &&
2284 F.Scale == OrigF.Scale &&
2285 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2286 if (F.BaseOffset == 0)
2288 // This is the formula where all the registers and symbols matched;
2289 // there aren't going to be any others. Since we declined it, we
2290 // can skip the rest of the formulae and proceed to the next LSRUse.
2297 // Nothing looked good.
2301 void LSRInstance::CollectInterestingTypesAndFactors() {
2302 SmallSetVector<const SCEV *, 4> Strides;
2304 // Collect interesting types and strides.
2305 SmallVector<const SCEV *, 4> Worklist;
2306 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2307 const SCEV *Expr = IU.getExpr(*UI);
2309 // Collect interesting types.
2310 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2312 // Add strides for mentioned loops.
2313 Worklist.push_back(Expr);
2315 const SCEV *S = Worklist.pop_back_val();
2316 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2317 if (AR->getLoop() == L)
2318 Strides.insert(AR->getStepRecurrence(SE));
2319 Worklist.push_back(AR->getStart());
2320 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2321 Worklist.append(Add->op_begin(), Add->op_end());
2323 } while (!Worklist.empty());
2326 // Compute interesting factors from the set of interesting strides.
2327 for (SmallSetVector<const SCEV *, 4>::const_iterator
2328 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2329 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2330 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2331 const SCEV *OldStride = *I;
2332 const SCEV *NewStride = *NewStrideIter;
2334 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2335 SE.getTypeSizeInBits(NewStride->getType())) {
2336 if (SE.getTypeSizeInBits(OldStride->getType()) >
2337 SE.getTypeSizeInBits(NewStride->getType()))
2338 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2340 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2342 if (const SCEVConstant *Factor =
2343 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2345 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2346 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2347 } else if (const SCEVConstant *Factor =
2348 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2351 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2352 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2356 // If all uses use the same type, don't bother looking for truncation-based
2358 if (Types.size() == 1)
2361 DEBUG(print_factors_and_types(dbgs()));
2364 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2365 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2366 /// Instructions to IVStrideUses, we could partially skip this.
2367 static User::op_iterator
2368 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2369 Loop *L, ScalarEvolution &SE) {
2370 for(; OI != OE; ++OI) {
2371 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2372 if (!SE.isSCEVable(Oper->getType()))
2375 if (const SCEVAddRecExpr *AR =
2376 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2377 if (AR->getLoop() == L)
2385 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2386 /// operands, so wrap it in a convenient helper.
2387 static Value *getWideOperand(Value *Oper) {
2388 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2389 return Trunc->getOperand(0);
2393 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2395 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2396 Type *LType = LVal->getType();
2397 Type *RType = RVal->getType();
2398 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2401 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2402 /// NULL for any constant. Returning the expression itself is
2403 /// conservative. Returning a deeper subexpression is more precise and valid as
2404 /// long as it isn't less complex than another subexpression. For expressions
2405 /// involving multiple unscaled values, we need to return the pointer-type
2406 /// SCEVUnknown. This avoids forming chains across objects, such as:
2407 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2409 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2410 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2411 static const SCEV *getExprBase(const SCEV *S) {
2412 switch (S->getSCEVType()) {
2413 default: // uncluding scUnknown.
2418 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2420 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2422 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2424 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2425 // there's nothing more complex.
2426 // FIXME: not sure if we want to recognize negation.
2427 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2428 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2429 E(Add->op_begin()); I != E; ++I) {
2430 const SCEV *SubExpr = *I;
2431 if (SubExpr->getSCEVType() == scAddExpr)
2432 return getExprBase(SubExpr);
2434 if (SubExpr->getSCEVType() != scMulExpr)
2437 return S; // all operands are scaled, be conservative.
2440 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2444 /// Return true if the chain increment is profitable to expand into a loop
2445 /// invariant value, which may require its own register. A profitable chain
2446 /// increment will be an offset relative to the same base. We allow such offsets
2447 /// to potentially be used as chain increment as long as it's not obviously
2448 /// expensive to expand using real instructions.
2449 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2450 const SCEV *IncExpr,
2451 ScalarEvolution &SE) {
2452 // Aggressively form chains when -stress-ivchain.
2456 // Do not replace a constant offset from IV head with a nonconstant IV
2458 if (!isa<SCEVConstant>(IncExpr)) {
2459 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2460 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2464 SmallPtrSet<const SCEV*, 8> Processed;
2465 return !isHighCostExpansion(IncExpr, Processed, SE);
2468 /// Return true if the number of registers needed for the chain is estimated to
2469 /// be less than the number required for the individual IV users. First prohibit
2470 /// any IV users that keep the IV live across increments (the Users set should
2471 /// be empty). Next count the number and type of increments in the chain.
2473 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2474 /// effectively use postinc addressing modes. Only consider it profitable it the
2475 /// increments can be computed in fewer registers when chained.
2477 /// TODO: Consider IVInc free if it's already used in another chains.
2479 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2480 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2484 if (!Chain.hasIncs())
2487 if (!Users.empty()) {
2488 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2489 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2490 E = Users.end(); I != E; ++I) {
2491 dbgs() << " " << **I << "\n";
2495 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2497 // The chain itself may require a register, so intialize cost to 1.
2500 // A complete chain likely eliminates the need for keeping the original IV in
2501 // a register. LSR does not currently know how to form a complete chain unless
2502 // the header phi already exists.
2503 if (isa<PHINode>(Chain.tailUserInst())
2504 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2507 const SCEV *LastIncExpr = 0;
2508 unsigned NumConstIncrements = 0;
2509 unsigned NumVarIncrements = 0;
2510 unsigned NumReusedIncrements = 0;
2511 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2514 if (I->IncExpr->isZero())
2517 // Incrementing by zero or some constant is neutral. We assume constants can
2518 // be folded into an addressing mode or an add's immediate operand.
2519 if (isa<SCEVConstant>(I->IncExpr)) {
2520 ++NumConstIncrements;
2524 if (I->IncExpr == LastIncExpr)
2525 ++NumReusedIncrements;
2529 LastIncExpr = I->IncExpr;
2531 // An IV chain with a single increment is handled by LSR's postinc
2532 // uses. However, a chain with multiple increments requires keeping the IV's
2533 // value live longer than it needs to be if chained.
2534 if (NumConstIncrements > 1)
2537 // Materializing increment expressions in the preheader that didn't exist in
2538 // the original code may cost a register. For example, sign-extended array
2539 // indices can produce ridiculous increments like this:
2540 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2541 cost += NumVarIncrements;
2543 // Reusing variable increments likely saves a register to hold the multiple of
2545 cost -= NumReusedIncrements;
2547 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2553 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2555 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2556 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2557 // When IVs are used as types of varying widths, they are generally converted
2558 // to a wider type with some uses remaining narrow under a (free) trunc.
2559 Value *const NextIV = getWideOperand(IVOper);
2560 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2561 const SCEV *const OperExprBase = getExprBase(OperExpr);
2563 // Visit all existing chains. Check if its IVOper can be computed as a
2564 // profitable loop invariant increment from the last link in the Chain.
2565 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2566 const SCEV *LastIncExpr = 0;
2567 for (; ChainIdx < NChains; ++ChainIdx) {
2568 IVChain &Chain = IVChainVec[ChainIdx];
2570 // Prune the solution space aggressively by checking that both IV operands
2571 // are expressions that operate on the same unscaled SCEVUnknown. This
2572 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2573 // first avoids creating extra SCEV expressions.
2574 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2577 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2578 if (!isCompatibleIVType(PrevIV, NextIV))
2581 // A phi node terminates a chain.
2582 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2585 // The increment must be loop-invariant so it can be kept in a register.
2586 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2587 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2588 if (!SE.isLoopInvariant(IncExpr, L))
2591 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2592 LastIncExpr = IncExpr;
2596 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2597 // bother for phi nodes, because they must be last in the chain.
2598 if (ChainIdx == NChains) {
2599 if (isa<PHINode>(UserInst))
2601 if (NChains >= MaxChains && !StressIVChain) {
2602 DEBUG(dbgs() << "IV Chain Limit\n");
2605 LastIncExpr = OperExpr;
2606 // IVUsers may have skipped over sign/zero extensions. We don't currently
2607 // attempt to form chains involving extensions unless they can be hoisted
2608 // into this loop's AddRec.
2609 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2612 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2614 ChainUsersVec.resize(NChains);
2615 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2616 << ") IV=" << *LastIncExpr << "\n");
2618 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2619 << ") IV+" << *LastIncExpr << "\n");
2620 // Add this IV user to the end of the chain.
2621 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2623 IVChain &Chain = IVChainVec[ChainIdx];
2625 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2626 // This chain's NearUsers become FarUsers.
2627 if (!LastIncExpr->isZero()) {
2628 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2633 // All other uses of IVOperand become near uses of the chain.
2634 // We currently ignore intermediate values within SCEV expressions, assuming
2635 // they will eventually be used be the current chain, or can be computed
2636 // from one of the chain increments. To be more precise we could
2637 // transitively follow its user and only add leaf IV users to the set.
2638 for (Value::use_iterator UseIter = IVOper->use_begin(),
2639 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2640 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2643 // Uses in the chain will no longer be uses if the chain is formed.
2644 // Include the head of the chain in this iteration (not Chain.begin()).
2645 IVChain::const_iterator IncIter = Chain.Incs.begin();
2646 IVChain::const_iterator IncEnd = Chain.Incs.end();
2647 for( ; IncIter != IncEnd; ++IncIter) {
2648 if (IncIter->UserInst == OtherUse)
2651 if (IncIter != IncEnd)
2654 if (SE.isSCEVable(OtherUse->getType())
2655 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2656 && IU.isIVUserOrOperand(OtherUse)) {
2659 NearUsers.insert(OtherUse);
2662 // Since this user is part of the chain, it's no longer considered a use
2664 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2667 /// CollectChains - Populate the vector of Chains.
2669 /// This decreases ILP at the architecture level. Targets with ample registers,
2670 /// multiple memory ports, and no register renaming probably don't want
2671 /// this. However, such targets should probably disable LSR altogether.
2673 /// The job of LSR is to make a reasonable choice of induction variables across
2674 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2675 /// ILP *within the loop* if the target wants it.
2677 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2678 /// will not reorder memory operations, it will recognize this as a chain, but
2679 /// will generate redundant IV increments. Ideally this would be corrected later
2680 /// by a smart scheduler:
2686 /// TODO: Walk the entire domtree within this loop, not just the path to the
2687 /// loop latch. This will discover chains on side paths, but requires
2688 /// maintaining multiple copies of the Chains state.
2689 void LSRInstance::CollectChains() {
2690 DEBUG(dbgs() << "Collecting IV Chains.\n");
2691 SmallVector<ChainUsers, 8> ChainUsersVec;
2693 SmallVector<BasicBlock *,8> LatchPath;
2694 BasicBlock *LoopHeader = L->getHeader();
2695 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2696 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2697 LatchPath.push_back(Rung->getBlock());
2699 LatchPath.push_back(LoopHeader);
2701 // Walk the instruction stream from the loop header to the loop latch.
2702 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2703 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2704 BBIter != BBEnd; ++BBIter) {
2705 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2707 // Skip instructions that weren't seen by IVUsers analysis.
2708 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2711 // Ignore users that are part of a SCEV expression. This way we only
2712 // consider leaf IV Users. This effectively rediscovers a portion of
2713 // IVUsers analysis but in program order this time.
2714 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2717 // Remove this instruction from any NearUsers set it may be in.
2718 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2719 ChainIdx < NChains; ++ChainIdx) {
2720 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2722 // Search for operands that can be chained.
2723 SmallPtrSet<Instruction*, 4> UniqueOperands;
2724 User::op_iterator IVOpEnd = I->op_end();
2725 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2726 while (IVOpIter != IVOpEnd) {
2727 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2728 if (UniqueOperands.insert(IVOpInst))
2729 ChainInstruction(I, IVOpInst, ChainUsersVec);
2730 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2732 } // Continue walking down the instructions.
2733 } // Continue walking down the domtree.
2734 // Visit phi backedges to determine if the chain can generate the IV postinc.
2735 for (BasicBlock::iterator I = L->getHeader()->begin();
2736 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2737 if (!SE.isSCEVable(PN->getType()))
2741 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2743 ChainInstruction(PN, IncV, ChainUsersVec);
2745 // Remove any unprofitable chains.
2746 unsigned ChainIdx = 0;
2747 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2748 UsersIdx < NChains; ++UsersIdx) {
2749 if (!isProfitableChain(IVChainVec[UsersIdx],
2750 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2752 // Preserve the chain at UsesIdx.
2753 if (ChainIdx != UsersIdx)
2754 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2755 FinalizeChain(IVChainVec[ChainIdx]);
2758 IVChainVec.resize(ChainIdx);
2761 void LSRInstance::FinalizeChain(IVChain &Chain) {
2762 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2763 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2765 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2767 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2768 User::op_iterator UseI =
2769 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2770 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2771 IVIncSet.insert(UseI);
2775 /// Return true if the IVInc can be folded into an addressing mode.
2776 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2777 Value *Operand, const TargetTransformInfo &TTI) {
2778 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2779 if (!IncConst || !isAddressUse(UserInst, Operand))
2782 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2785 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2786 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2787 getAccessType(UserInst), /*BaseGV=*/ 0,
2788 IncOffset, /*HaseBaseReg=*/ false))
2794 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2795 /// materialize the IV user's operand from the previous IV user's operand.
2796 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2797 SmallVectorImpl<WeakVH> &DeadInsts) {
2798 // Find the new IVOperand for the head of the chain. It may have been replaced
2800 const IVInc &Head = Chain.Incs[0];
2801 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2802 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2803 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2806 while (IVOpIter != IVOpEnd) {
2807 IVSrc = getWideOperand(*IVOpIter);
2809 // If this operand computes the expression that the chain needs, we may use
2810 // it. (Check this after setting IVSrc which is used below.)
2812 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2813 // narrow for the chain, so we can no longer use it. We do allow using a
2814 // wider phi, assuming the LSR checked for free truncation. In that case we
2815 // should already have a truncate on this operand such that
2816 // getSCEV(IVSrc) == IncExpr.
2817 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2818 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2821 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2823 if (IVOpIter == IVOpEnd) {
2824 // Gracefully give up on this chain.
2825 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2829 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2830 Type *IVTy = IVSrc->getType();
2831 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2832 const SCEV *LeftOverExpr = 0;
2833 for (IVChain::const_iterator IncI = Chain.begin(),
2834 IncE = Chain.end(); IncI != IncE; ++IncI) {
2836 Instruction *InsertPt = IncI->UserInst;
2837 if (isa<PHINode>(InsertPt))
2838 InsertPt = L->getLoopLatch()->getTerminator();
2840 // IVOper will replace the current IV User's operand. IVSrc is the IV
2841 // value currently held in a register.
2842 Value *IVOper = IVSrc;
2843 if (!IncI->IncExpr->isZero()) {
2844 // IncExpr was the result of subtraction of two narrow values, so must
2846 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2847 LeftOverExpr = LeftOverExpr ?
2848 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2850 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2851 // Expand the IV increment.
2852 Rewriter.clearPostInc();
2853 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2854 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2855 SE.getUnknown(IncV));
2856 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2858 // If an IV increment can't be folded, use it as the next IV value.
2859 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2861 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2866 Type *OperTy = IncI->IVOperand->getType();
2867 if (IVTy != OperTy) {
2868 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2869 "cannot extend a chained IV");
2870 IRBuilder<> Builder(InsertPt);
2871 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2873 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2874 DeadInsts.push_back(IncI->IVOperand);
2876 // If LSR created a new, wider phi, we may also replace its postinc. We only
2877 // do this if we also found a wide value for the head of the chain.
2878 if (isa<PHINode>(Chain.tailUserInst())) {
2879 for (BasicBlock::iterator I = L->getHeader()->begin();
2880 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2881 if (!isCompatibleIVType(Phi, IVSrc))
2883 Instruction *PostIncV = dyn_cast<Instruction>(
2884 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2885 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2887 Value *IVOper = IVSrc;
2888 Type *PostIncTy = PostIncV->getType();
2889 if (IVTy != PostIncTy) {
2890 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2891 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2892 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2893 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2895 Phi->replaceUsesOfWith(PostIncV, IVOper);
2896 DeadInsts.push_back(PostIncV);
2901 void LSRInstance::CollectFixupsAndInitialFormulae() {
2902 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2903 Instruction *UserInst = UI->getUser();
2904 // Skip IV users that are part of profitable IV Chains.
2905 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2906 UI->getOperandValToReplace());
2907 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2908 if (IVIncSet.count(UseI))
2912 LSRFixup &LF = getNewFixup();
2913 LF.UserInst = UserInst;
2914 LF.OperandValToReplace = UI->getOperandValToReplace();
2915 LF.PostIncLoops = UI->getPostIncLoops();
2917 LSRUse::KindType Kind = LSRUse::Basic;
2919 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2920 Kind = LSRUse::Address;
2921 AccessTy = getAccessType(LF.UserInst);
2924 const SCEV *S = IU.getExpr(*UI);
2926 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2927 // (N - i == 0), and this allows (N - i) to be the expression that we work
2928 // with rather than just N or i, so we can consider the register
2929 // requirements for both N and i at the same time. Limiting this code to
2930 // equality icmps is not a problem because all interesting loops use
2931 // equality icmps, thanks to IndVarSimplify.
2932 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2933 if (CI->isEquality()) {
2934 // Swap the operands if needed to put the OperandValToReplace on the
2935 // left, for consistency.
2936 Value *NV = CI->getOperand(1);
2937 if (NV == LF.OperandValToReplace) {
2938 CI->setOperand(1, CI->getOperand(0));
2939 CI->setOperand(0, NV);
2940 NV = CI->getOperand(1);
2944 // x == y --> x - y == 0
2945 const SCEV *N = SE.getSCEV(NV);
2946 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2947 // S is normalized, so normalize N before folding it into S
2948 // to keep the result normalized.
2949 N = TransformForPostIncUse(Normalize, N, CI, 0,
2950 LF.PostIncLoops, SE, DT);
2951 Kind = LSRUse::ICmpZero;
2952 S = SE.getMinusSCEV(N, S);
2955 // -1 and the negations of all interesting strides (except the negation
2956 // of -1) are now also interesting.
2957 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2958 if (Factors[i] != -1)
2959 Factors.insert(-(uint64_t)Factors[i]);
2963 // Set up the initial formula for this use.
2964 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2966 LF.Offset = P.second;
2967 LSRUse &LU = Uses[LF.LUIdx];
2968 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2969 if (!LU.WidestFixupType ||
2970 SE.getTypeSizeInBits(LU.WidestFixupType) <
2971 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2972 LU.WidestFixupType = LF.OperandValToReplace->getType();
2974 // If this is the first use of this LSRUse, give it a formula.
2975 if (LU.Formulae.empty()) {
2976 InsertInitialFormula(S, LU, LF.LUIdx);
2977 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2981 DEBUG(print_fixups(dbgs()));
2984 /// InsertInitialFormula - Insert a formula for the given expression into
2985 /// the given use, separating out loop-variant portions from loop-invariant
2986 /// and loop-computable portions.
2988 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2989 // Mark uses whose expressions cannot be expanded.
2990 if (!isSafeToExpand(S, SE))
2991 LU.RigidFormula = true;
2994 F.InitialMatch(S, L, SE);
2995 bool Inserted = InsertFormula(LU, LUIdx, F);
2996 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2999 /// InsertSupplementalFormula - Insert a simple single-register formula for
3000 /// the given expression into the given use.
3002 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3003 LSRUse &LU, size_t LUIdx) {
3005 F.BaseRegs.push_back(S);
3006 F.HasBaseReg = true;
3007 bool Inserted = InsertFormula(LU, LUIdx, F);
3008 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3011 /// CountRegisters - Note which registers are used by the given formula,
3012 /// updating RegUses.
3013 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3015 RegUses.CountRegister(F.ScaledReg, LUIdx);
3016 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3017 E = F.BaseRegs.end(); I != E; ++I)
3018 RegUses.CountRegister(*I, LUIdx);
3021 /// InsertFormula - If the given formula has not yet been inserted, add it to
3022 /// the list, and return true. Return false otherwise.
3023 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3024 if (!LU.InsertFormula(F))
3027 CountRegisters(F, LUIdx);
3031 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3032 /// loop-invariant values which we're tracking. These other uses will pin these
3033 /// values in registers, making them less profitable for elimination.
3034 /// TODO: This currently misses non-constant addrec step registers.
3035 /// TODO: Should this give more weight to users inside the loop?
3037 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3038 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3039 SmallPtrSet<const SCEV *, 8> Inserted;
3041 while (!Worklist.empty()) {
3042 const SCEV *S = Worklist.pop_back_val();
3044 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3045 Worklist.append(N->op_begin(), N->op_end());
3046 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3047 Worklist.push_back(C->getOperand());
3048 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3049 Worklist.push_back(D->getLHS());
3050 Worklist.push_back(D->getRHS());
3051 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3052 if (!Inserted.insert(U)) continue;
3053 const Value *V = U->getValue();
3054 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3055 // Look for instructions defined outside the loop.
3056 if (L->contains(Inst)) continue;
3057 } else if (isa<UndefValue>(V))
3058 // Undef doesn't have a live range, so it doesn't matter.
3060 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
3062 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
3063 // Ignore non-instructions.
3066 // Ignore instructions in other functions (as can happen with
3068 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3070 // Ignore instructions not dominated by the loop.
3071 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3072 UserInst->getParent() :
3073 cast<PHINode>(UserInst)->getIncomingBlock(
3074 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
3075 if (!DT.dominates(L->getHeader(), UseBB))
3077 // Ignore uses which are part of other SCEV expressions, to avoid
3078 // analyzing them multiple times.
3079 if (SE.isSCEVable(UserInst->getType())) {
3080 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3081 // If the user is a no-op, look through to its uses.
3082 if (!isa<SCEVUnknown>(UserS))
3086 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3090 // Ignore icmp instructions which are already being analyzed.
3091 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3092 unsigned OtherIdx = !UI.getOperandNo();
3093 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3094 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3098 LSRFixup &LF = getNewFixup();
3099 LF.UserInst = const_cast<Instruction *>(UserInst);
3100 LF.OperandValToReplace = UI.getUse();
3101 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
3103 LF.Offset = P.second;
3104 LSRUse &LU = Uses[LF.LUIdx];
3105 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3106 if (!LU.WidestFixupType ||
3107 SE.getTypeSizeInBits(LU.WidestFixupType) <
3108 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3109 LU.WidestFixupType = LF.OperandValToReplace->getType();
3110 InsertSupplementalFormula(U, LU, LF.LUIdx);
3111 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3118 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3119 /// separate registers. If C is non-null, multiply each subexpression by C.
3121 /// Return remainder expression after factoring the subexpressions captured by
3122 /// Ops. If Ops is complete, return NULL.
3123 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3124 SmallVectorImpl<const SCEV *> &Ops,
3126 ScalarEvolution &SE,
3127 unsigned Depth = 0) {
3128 // Arbitrarily cap recursion to protect compile time.
3132 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3133 // Break out add operands.
3134 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3136 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3138 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3141 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3142 // Split a non-zero base out of an addrec.
3143 if (AR->getStart()->isZero())
3146 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3147 C, Ops, L, SE, Depth+1);
3148 // Split the non-zero AddRec unless it is part of a nested recurrence that
3149 // does not pertain to this loop.
3150 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3151 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3154 if (Remainder != AR->getStart()) {
3156 Remainder = SE.getConstant(AR->getType(), 0);
3157 return SE.getAddRecExpr(Remainder,
3158 AR->getStepRecurrence(SE),
3160 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3163 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3164 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3165 if (Mul->getNumOperands() != 2)
3167 if (const SCEVConstant *Op0 =
3168 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3169 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3170 const SCEV *Remainder =
3171 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3173 Ops.push_back(SE.getMulExpr(C, Remainder));
3180 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3182 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3185 // Arbitrarily cap recursion to protect compile time.
3186 if (Depth >= 3) return;
3188 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3189 const SCEV *BaseReg = Base.BaseRegs[i];
3191 SmallVector<const SCEV *, 8> AddOps;
3192 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3194 AddOps.push_back(Remainder);
3196 if (AddOps.size() == 1) continue;
3198 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3199 JE = AddOps.end(); J != JE; ++J) {
3201 // Loop-variant "unknown" values are uninteresting; we won't be able to
3202 // do anything meaningful with them.
3203 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3206 // Don't pull a constant into a register if the constant could be folded
3207 // into an immediate field.
3208 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3209 LU.AccessTy, *J, Base.getNumRegs() > 1))
3212 // Collect all operands except *J.
3213 SmallVector<const SCEV *, 8> InnerAddOps(
3214 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3215 InnerAddOps.append(std::next(J),
3216 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3218 // Don't leave just a constant behind in a register if the constant could
3219 // be folded into an immediate field.
3220 if (InnerAddOps.size() == 1 &&
3221 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3222 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3225 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3226 if (InnerSum->isZero())
3230 // Add the remaining pieces of the add back into the new formula.
3231 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3233 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3234 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3235 InnerSumSC->getValue()->getZExtValue())) {
3236 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3237 InnerSumSC->getValue()->getZExtValue();
3238 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3240 F.BaseRegs[i] = InnerSum;
3242 // Add J as its own register, or an unfolded immediate.
3243 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3244 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3245 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3246 SC->getValue()->getZExtValue()))
3247 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3248 SC->getValue()->getZExtValue();
3250 F.BaseRegs.push_back(*J);
3252 if (InsertFormula(LU, LUIdx, F))
3253 // If that formula hadn't been seen before, recurse to find more like
3255 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3260 /// GenerateCombinations - Generate a formula consisting of all of the
3261 /// loop-dominating registers added into a single register.
3262 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3264 // This method is only interesting on a plurality of registers.
3265 if (Base.BaseRegs.size() <= 1) return;
3269 SmallVector<const SCEV *, 4> Ops;
3270 for (SmallVectorImpl<const SCEV *>::const_iterator
3271 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3272 const SCEV *BaseReg = *I;
3273 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3274 !SE.hasComputableLoopEvolution(BaseReg, L))
3275 Ops.push_back(BaseReg);
3277 F.BaseRegs.push_back(BaseReg);
3279 if (Ops.size() > 1) {
3280 const SCEV *Sum = SE.getAddExpr(Ops);
3281 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3282 // opportunity to fold something. For now, just ignore such cases
3283 // rather than proceed with zero in a register.
3284 if (!Sum->isZero()) {
3285 F.BaseRegs.push_back(Sum);
3286 (void)InsertFormula(LU, LUIdx, F);
3291 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3292 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3294 // We can't add a symbolic offset if the address already contains one.
3295 if (Base.BaseGV) return;
3297 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3298 const SCEV *G = Base.BaseRegs[i];
3299 GlobalValue *GV = ExtractSymbol(G, SE);
3300 if (G->isZero() || !GV)
3304 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3307 (void)InsertFormula(LU, LUIdx, F);
3311 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3312 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3314 // TODO: For now, just add the min and max offset, because it usually isn't
3315 // worthwhile looking at everything inbetween.
3316 SmallVector<int64_t, 2> Worklist;
3317 Worklist.push_back(LU.MinOffset);
3318 if (LU.MaxOffset != LU.MinOffset)
3319 Worklist.push_back(LU.MaxOffset);
3321 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3322 const SCEV *G = Base.BaseRegs[i];
3324 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3325 E = Worklist.end(); I != E; ++I) {
3327 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3328 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3330 // Add the offset to the base register.
3331 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3332 // If it cancelled out, drop the base register, otherwise update it.
3333 if (NewG->isZero()) {
3334 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3335 F.BaseRegs.pop_back();
3337 F.BaseRegs[i] = NewG;
3339 (void)InsertFormula(LU, LUIdx, F);
3343 int64_t Imm = ExtractImmediate(G, SE);
3344 if (G->isZero() || Imm == 0)
3347 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3348 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3351 (void)InsertFormula(LU, LUIdx, F);
3355 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3356 /// the comparison. For example, x == y -> x*c == y*c.
3357 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3359 if (LU.Kind != LSRUse::ICmpZero) return;
3361 // Determine the integer type for the base formula.
3362 Type *IntTy = Base.getType();
3364 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3366 // Don't do this if there is more than one offset.
3367 if (LU.MinOffset != LU.MaxOffset) return;
3369 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3371 // Check each interesting stride.
3372 for (SmallSetVector<int64_t, 8>::const_iterator
3373 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3374 int64_t Factor = *I;
3376 // Check that the multiplication doesn't overflow.
3377 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3379 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3380 if (NewBaseOffset / Factor != Base.BaseOffset)
3382 // If the offset will be truncated at this use, check that it is in bounds.
3383 if (!IntTy->isPointerTy() &&
3384 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3387 // Check that multiplying with the use offset doesn't overflow.
3388 int64_t Offset = LU.MinOffset;
3389 if (Offset == INT64_MIN && Factor == -1)
3391 Offset = (uint64_t)Offset * Factor;
3392 if (Offset / Factor != LU.MinOffset)
3394 // If the offset will be truncated at this use, check that it is in bounds.
3395 if (!IntTy->isPointerTy() &&
3396 !ConstantInt::isValueValidForType(IntTy, Offset))
3400 F.BaseOffset = NewBaseOffset;
3402 // Check that this scale is legal.
3403 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3406 // Compensate for the use having MinOffset built into it.
3407 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3409 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3411 // Check that multiplying with each base register doesn't overflow.
3412 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3413 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3414 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3418 // Check that multiplying with the scaled register doesn't overflow.
3420 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3421 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3425 // Check that multiplying with the unfolded offset doesn't overflow.
3426 if (F.UnfoldedOffset != 0) {
3427 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3429 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3430 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3432 // If the offset will be truncated, check that it is in bounds.
3433 if (!IntTy->isPointerTy() &&
3434 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3438 // If we make it here and it's legal, add it.
3439 (void)InsertFormula(LU, LUIdx, F);
3444 /// GenerateScales - Generate stride factor reuse formulae by making use of
3445 /// scaled-offset address modes, for example.
3446 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3447 // Determine the integer type for the base formula.
3448 Type *IntTy = Base.getType();
3451 // If this Formula already has a scaled register, we can't add another one.
3452 if (Base.Scale != 0) return;
3454 // Check each interesting stride.
3455 for (SmallSetVector<int64_t, 8>::const_iterator
3456 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3457 int64_t Factor = *I;
3459 Base.Scale = Factor;
3460 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3461 // Check whether this scale is going to be legal.
3462 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3464 // As a special-case, handle special out-of-loop Basic users specially.
3465 // TODO: Reconsider this special case.
3466 if (LU.Kind == LSRUse::Basic &&
3467 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3468 LU.AccessTy, Base) &&
3469 LU.AllFixupsOutsideLoop)
3470 LU.Kind = LSRUse::Special;
3474 // For an ICmpZero, negating a solitary base register won't lead to
3476 if (LU.Kind == LSRUse::ICmpZero &&
3477 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3479 // For each addrec base reg, apply the scale, if possible.
3480 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3481 if (const SCEVAddRecExpr *AR =
3482 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3483 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3484 if (FactorS->isZero())
3486 // Divide out the factor, ignoring high bits, since we'll be
3487 // scaling the value back up in the end.
3488 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3489 // TODO: This could be optimized to avoid all the copying.
3491 F.ScaledReg = Quotient;
3492 F.DeleteBaseReg(F.BaseRegs[i]);
3493 (void)InsertFormula(LU, LUIdx, F);
3499 /// GenerateTruncates - Generate reuse formulae from different IV types.
3500 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3501 // Don't bother truncating symbolic values.
3502 if (Base.BaseGV) return;
3504 // Determine the integer type for the base formula.
3505 Type *DstTy = Base.getType();
3507 DstTy = SE.getEffectiveSCEVType(DstTy);
3509 for (SmallSetVector<Type *, 4>::const_iterator
3510 I = Types.begin(), E = Types.end(); I != E; ++I) {
3512 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3515 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3516 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3517 JE = F.BaseRegs.end(); J != JE; ++J)
3518 *J = SE.getAnyExtendExpr(*J, SrcTy);
3520 // TODO: This assumes we've done basic processing on all uses and
3521 // have an idea what the register usage is.
3522 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3525 (void)InsertFormula(LU, LUIdx, F);
3532 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3533 /// defer modifications so that the search phase doesn't have to worry about
3534 /// the data structures moving underneath it.
3538 const SCEV *OrigReg;
3540 WorkItem(size_t LI, int64_t I, const SCEV *R)
3541 : LUIdx(LI), Imm(I), OrigReg(R) {}
3543 void print(raw_ostream &OS) const;
3549 void WorkItem::print(raw_ostream &OS) const {
3550 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3551 << " , add offset " << Imm;
3554 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3555 void WorkItem::dump() const {
3556 print(errs()); errs() << '\n';
3560 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3561 /// distance apart and try to form reuse opportunities between them.
3562 void LSRInstance::GenerateCrossUseConstantOffsets() {
3563 // Group the registers by their value without any added constant offset.
3564 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3565 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3567 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3568 SmallVector<const SCEV *, 8> Sequence;
3569 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3571 const SCEV *Reg = *I;
3572 int64_t Imm = ExtractImmediate(Reg, SE);
3573 std::pair<RegMapTy::iterator, bool> Pair =
3574 Map.insert(std::make_pair(Reg, ImmMapTy()));
3576 Sequence.push_back(Reg);
3577 Pair.first->second.insert(std::make_pair(Imm, *I));
3578 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3581 // Now examine each set of registers with the same base value. Build up
3582 // a list of work to do and do the work in a separate step so that we're
3583 // not adding formulae and register counts while we're searching.
3584 SmallVector<WorkItem, 32> WorkItems;
3585 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3586 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3587 E = Sequence.end(); I != E; ++I) {
3588 const SCEV *Reg = *I;
3589 const ImmMapTy &Imms = Map.find(Reg)->second;
3591 // It's not worthwhile looking for reuse if there's only one offset.
3592 if (Imms.size() == 1)
3595 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3596 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3598 dbgs() << ' ' << J->first;
3601 // Examine each offset.
3602 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3604 const SCEV *OrigReg = J->second;
3606 int64_t JImm = J->first;
3607 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3609 if (!isa<SCEVConstant>(OrigReg) &&
3610 UsedByIndicesMap[Reg].count() == 1) {
3611 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3615 // Conservatively examine offsets between this orig reg a few selected
3617 ImmMapTy::const_iterator OtherImms[] = {
3618 Imms.begin(), std::prev(Imms.end()),
3619 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3622 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3623 ImmMapTy::const_iterator M = OtherImms[i];
3624 if (M == J || M == JE) continue;
3626 // Compute the difference between the two.
3627 int64_t Imm = (uint64_t)JImm - M->first;
3628 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3629 LUIdx = UsedByIndices.find_next(LUIdx))
3630 // Make a memo of this use, offset, and register tuple.
3631 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3632 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3639 UsedByIndicesMap.clear();
3640 UniqueItems.clear();
3642 // Now iterate through the worklist and add new formulae.
3643 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3644 E = WorkItems.end(); I != E; ++I) {
3645 const WorkItem &WI = *I;
3646 size_t LUIdx = WI.LUIdx;
3647 LSRUse &LU = Uses[LUIdx];
3648 int64_t Imm = WI.Imm;
3649 const SCEV *OrigReg = WI.OrigReg;
3651 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3652 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3653 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3655 // TODO: Use a more targeted data structure.
3656 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3657 const Formula &F = LU.Formulae[L];
3658 // Use the immediate in the scaled register.
3659 if (F.ScaledReg == OrigReg) {
3660 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3661 // Don't create 50 + reg(-50).
3662 if (F.referencesReg(SE.getSCEV(
3663 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3666 NewF.BaseOffset = Offset;
3667 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3670 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3672 // If the new scale is a constant in a register, and adding the constant
3673 // value to the immediate would produce a value closer to zero than the
3674 // immediate itself, then the formula isn't worthwhile.
3675 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3676 if (C->getValue()->isNegative() !=
3677 (NewF.BaseOffset < 0) &&
3678 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3679 .ule(abs64(NewF.BaseOffset)))
3683 (void)InsertFormula(LU, LUIdx, NewF);
3685 // Use the immediate in a base register.
3686 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3687 const SCEV *BaseReg = F.BaseRegs[N];
3688 if (BaseReg != OrigReg)
3691 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3692 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3693 LU.Kind, LU.AccessTy, NewF)) {
3694 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3697 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3699 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3701 // If the new formula has a constant in a register, and adding the
3702 // constant value to the immediate would produce a value closer to
3703 // zero than the immediate itself, then the formula isn't worthwhile.
3704 for (SmallVectorImpl<const SCEV *>::const_iterator
3705 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3707 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3708 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3709 abs64(NewF.BaseOffset)) &&
3710 (C->getValue()->getValue() +
3711 NewF.BaseOffset).countTrailingZeros() >=
3712 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3716 (void)InsertFormula(LU, LUIdx, NewF);
3725 /// GenerateAllReuseFormulae - Generate formulae for each use.
3727 LSRInstance::GenerateAllReuseFormulae() {
3728 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3729 // queries are more precise.
3730 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3731 LSRUse &LU = Uses[LUIdx];
3732 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3733 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3734 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3735 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3737 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3738 LSRUse &LU = Uses[LUIdx];
3739 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3740 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3741 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3742 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3743 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3744 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3745 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3746 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3748 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3749 LSRUse &LU = Uses[LUIdx];
3750 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3751 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3754 GenerateCrossUseConstantOffsets();
3756 DEBUG(dbgs() << "\n"
3757 "After generating reuse formulae:\n";
3758 print_uses(dbgs()));
3761 /// If there are multiple formulae with the same set of registers used
3762 /// by other uses, pick the best one and delete the others.
3763 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3764 DenseSet<const SCEV *> VisitedRegs;
3765 SmallPtrSet<const SCEV *, 16> Regs;
3766 SmallPtrSet<const SCEV *, 16> LoserRegs;
3768 bool ChangedFormulae = false;
3771 // Collect the best formula for each unique set of shared registers. This
3772 // is reset for each use.
3773 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3775 BestFormulaeTy BestFormulae;
3777 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3778 LSRUse &LU = Uses[LUIdx];
3779 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3782 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3783 FIdx != NumForms; ++FIdx) {
3784 Formula &F = LU.Formulae[FIdx];
3786 // Some formulas are instant losers. For example, they may depend on
3787 // nonexistent AddRecs from other loops. These need to be filtered
3788 // immediately, otherwise heuristics could choose them over others leading
3789 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3790 // avoids the need to recompute this information across formulae using the
3791 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3792 // the corresponding bad register from the Regs set.
3795 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3797 if (CostF.isLoser()) {
3798 // During initial formula generation, undesirable formulae are generated
3799 // by uses within other loops that have some non-trivial address mode or
3800 // use the postinc form of the IV. LSR needs to provide these formulae
3801 // as the basis of rediscovering the desired formula that uses an AddRec
3802 // corresponding to the existing phi. Once all formulae have been
3803 // generated, these initial losers may be pruned.
3804 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3808 SmallVector<const SCEV *, 4> Key;
3809 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3810 JE = F.BaseRegs.end(); J != JE; ++J) {
3811 const SCEV *Reg = *J;
3812 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3816 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3817 Key.push_back(F.ScaledReg);
3818 // Unstable sort by host order ok, because this is only used for
3820 std::sort(Key.begin(), Key.end());
3822 std::pair<BestFormulaeTy::const_iterator, bool> P =
3823 BestFormulae.insert(std::make_pair(Key, FIdx));
3827 Formula &Best = LU.Formulae[P.first->second];
3831 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3833 if (CostF < CostBest)
3835 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3837 " in favor of formula "; Best.print(dbgs());
3841 ChangedFormulae = true;
3843 LU.DeleteFormula(F);
3849 // Now that we've filtered out some formulae, recompute the Regs set.
3851 LU.RecomputeRegs(LUIdx, RegUses);
3853 // Reset this to prepare for the next use.
3854 BestFormulae.clear();
3857 DEBUG(if (ChangedFormulae) {
3859 "After filtering out undesirable candidates:\n";
3864 // This is a rough guess that seems to work fairly well.
3865 static const size_t ComplexityLimit = UINT16_MAX;
3867 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3868 /// solutions the solver might have to consider. It almost never considers
3869 /// this many solutions because it prune the search space, but the pruning
3870 /// isn't always sufficient.
3871 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3873 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3874 E = Uses.end(); I != E; ++I) {
3875 size_t FSize = I->Formulae.size();
3876 if (FSize >= ComplexityLimit) {
3877 Power = ComplexityLimit;
3881 if (Power >= ComplexityLimit)
3887 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3888 /// of the registers of another formula, it won't help reduce register
3889 /// pressure (though it may not necessarily hurt register pressure); remove
3890 /// it to simplify the system.
3891 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3892 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3893 DEBUG(dbgs() << "The search space is too complex.\n");
3895 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3896 "which use a superset of registers used by other "
3899 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3900 LSRUse &LU = Uses[LUIdx];
3902 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3903 Formula &F = LU.Formulae[i];
3904 // Look for a formula with a constant or GV in a register. If the use
3905 // also has a formula with that same value in an immediate field,
3906 // delete the one that uses a register.
3907 for (SmallVectorImpl<const SCEV *>::const_iterator
3908 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3909 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3911 NewF.BaseOffset += C->getValue()->getSExtValue();
3912 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3913 (I - F.BaseRegs.begin()));
3914 if (LU.HasFormulaWithSameRegs(NewF)) {
3915 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3916 LU.DeleteFormula(F);
3922 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3923 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3927 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3928 (I - F.BaseRegs.begin()));
3929 if (LU.HasFormulaWithSameRegs(NewF)) {
3930 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3932 LU.DeleteFormula(F);
3943 LU.RecomputeRegs(LUIdx, RegUses);
3946 DEBUG(dbgs() << "After pre-selection:\n";
3947 print_uses(dbgs()));
3951 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3952 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3954 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3955 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
3958 DEBUG(dbgs() << "The search space is too complex.\n"
3959 "Narrowing the search space by assuming that uses separated "
3960 "by a constant offset will use the same registers.\n");
3962 // This is especially useful for unrolled loops.
3964 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3965 LSRUse &LU = Uses[LUIdx];
3966 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3967 E = LU.Formulae.end(); I != E; ++I) {
3968 const Formula &F = *I;
3969 if (F.BaseOffset == 0 || F.Scale != 0)
3972 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
3976 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
3977 LU.Kind, LU.AccessTy))
3980 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
3982 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3984 // Update the relocs to reference the new use.
3985 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3986 E = Fixups.end(); I != E; ++I) {
3987 LSRFixup &Fixup = *I;
3988 if (Fixup.LUIdx == LUIdx) {
3989 Fixup.LUIdx = LUThatHas - &Uses.front();
3990 Fixup.Offset += F.BaseOffset;
3991 // Add the new offset to LUThatHas' offset list.
3992 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3993 LUThatHas->Offsets.push_back(Fixup.Offset);
3994 if (Fixup.Offset > LUThatHas->MaxOffset)
3995 LUThatHas->MaxOffset = Fixup.Offset;
3996 if (Fixup.Offset < LUThatHas->MinOffset)
3997 LUThatHas->MinOffset = Fixup.Offset;
3999 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4001 if (Fixup.LUIdx == NumUses-1)
4002 Fixup.LUIdx = LUIdx;
4005 // Delete formulae from the new use which are no longer legal.
4007 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4008 Formula &F = LUThatHas->Formulae[i];
4009 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4010 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4011 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4013 LUThatHas->DeleteFormula(F);
4021 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4023 // Delete the old use.
4024 DeleteUse(LU, LUIdx);
4031 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4034 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4035 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4036 /// we've done more filtering, as it may be able to find more formulae to
4038 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4039 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4040 DEBUG(dbgs() << "The search space is too complex.\n");
4042 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4043 "undesirable dedicated registers.\n");
4045 FilterOutUndesirableDedicatedRegisters();
4047 DEBUG(dbgs() << "After pre-selection:\n";
4048 print_uses(dbgs()));
4052 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4053 /// to be profitable, and then in any use which has any reference to that
4054 /// register, delete all formulae which do not reference that register.
4055 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4056 // With all other options exhausted, loop until the system is simple
4057 // enough to handle.
4058 SmallPtrSet<const SCEV *, 4> Taken;
4059 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4060 // Ok, we have too many of formulae on our hands to conveniently handle.
4061 // Use a rough heuristic to thin out the list.
4062 DEBUG(dbgs() << "The search space is too complex.\n");
4064 // Pick the register which is used by the most LSRUses, which is likely
4065 // to be a good reuse register candidate.
4066 const SCEV *Best = 0;
4067 unsigned BestNum = 0;
4068 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4070 const SCEV *Reg = *I;
4071 if (Taken.count(Reg))
4076 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4077 if (Count > BestNum) {
4084 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4085 << " will yield profitable reuse.\n");
4088 // In any use with formulae which references this register, delete formulae
4089 // which don't reference it.
4090 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4091 LSRUse &LU = Uses[LUIdx];
4092 if (!LU.Regs.count(Best)) continue;
4095 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4096 Formula &F = LU.Formulae[i];
4097 if (!F.referencesReg(Best)) {
4098 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4099 LU.DeleteFormula(F);
4103 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4109 LU.RecomputeRegs(LUIdx, RegUses);
4112 DEBUG(dbgs() << "After pre-selection:\n";
4113 print_uses(dbgs()));
4117 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4118 /// formulae to choose from, use some rough heuristics to prune down the number
4119 /// of formulae. This keeps the main solver from taking an extraordinary amount
4120 /// of time in some worst-case scenarios.
4121 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4122 NarrowSearchSpaceByDetectingSupersets();
4123 NarrowSearchSpaceByCollapsingUnrolledCode();
4124 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4125 NarrowSearchSpaceByPickingWinnerRegs();
4128 /// SolveRecurse - This is the recursive solver.
4129 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4131 SmallVectorImpl<const Formula *> &Workspace,
4132 const Cost &CurCost,
4133 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4134 DenseSet<const SCEV *> &VisitedRegs) const {
4137 // - use more aggressive filtering
4138 // - sort the formula so that the most profitable solutions are found first
4139 // - sort the uses too
4141 // - don't compute a cost, and then compare. compare while computing a cost
4143 // - track register sets with SmallBitVector
4145 const LSRUse &LU = Uses[Workspace.size()];
4147 // If this use references any register that's already a part of the
4148 // in-progress solution, consider it a requirement that a formula must
4149 // reference that register in order to be considered. This prunes out
4150 // unprofitable searching.
4151 SmallSetVector<const SCEV *, 4> ReqRegs;
4152 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4153 E = CurRegs.end(); I != E; ++I)
4154 if (LU.Regs.count(*I))
4157 SmallPtrSet<const SCEV *, 16> NewRegs;
4159 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4160 E = LU.Formulae.end(); I != E; ++I) {
4161 const Formula &F = *I;
4163 // Ignore formulae which do not use any of the required registers.
4164 bool SatisfiedReqReg = true;
4165 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4166 JE = ReqRegs.end(); J != JE; ++J) {
4167 const SCEV *Reg = *J;
4168 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4169 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4171 SatisfiedReqReg = false;
4175 if (!SatisfiedReqReg) {
4176 // If none of the formulae satisfied the required registers, then we could
4177 // clear ReqRegs and try again. Currently, we simply give up in this case.
4181 // Evaluate the cost of the current formula. If it's already worse than
4182 // the current best, prune the search at that point.
4185 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4187 if (NewCost < SolutionCost) {
4188 Workspace.push_back(&F);
4189 if (Workspace.size() != Uses.size()) {
4190 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4191 NewRegs, VisitedRegs);
4192 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4193 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4195 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4196 dbgs() << ".\n Regs:";
4197 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4198 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4199 dbgs() << ' ' << **I;
4202 SolutionCost = NewCost;
4203 Solution = Workspace;
4205 Workspace.pop_back();
4210 /// Solve - Choose one formula from each use. Return the results in the given
4211 /// Solution vector.
4212 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4213 SmallVector<const Formula *, 8> Workspace;
4215 SolutionCost.Lose();
4217 SmallPtrSet<const SCEV *, 16> CurRegs;
4218 DenseSet<const SCEV *> VisitedRegs;
4219 Workspace.reserve(Uses.size());
4221 // SolveRecurse does all the work.
4222 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4223 CurRegs, VisitedRegs);
4224 if (Solution.empty()) {
4225 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4229 // Ok, we've now made all our decisions.
4230 DEBUG(dbgs() << "\n"
4231 "The chosen solution requires "; SolutionCost.print(dbgs());
4233 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4235 Uses[i].print(dbgs());
4238 Solution[i]->print(dbgs());
4242 assert(Solution.size() == Uses.size() && "Malformed solution!");
4245 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4246 /// the dominator tree far as we can go while still being dominated by the
4247 /// input positions. This helps canonicalize the insert position, which
4248 /// encourages sharing.
4249 BasicBlock::iterator
4250 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4251 const SmallVectorImpl<Instruction *> &Inputs)
4254 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4255 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4258 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4259 if (!Rung) return IP;
4260 Rung = Rung->getIDom();
4261 if (!Rung) return IP;
4262 IDom = Rung->getBlock();
4264 // Don't climb into a loop though.
4265 const Loop *IDomLoop = LI.getLoopFor(IDom);
4266 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4267 if (IDomDepth <= IPLoopDepth &&
4268 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4272 bool AllDominate = true;
4273 Instruction *BetterPos = 0;
4274 Instruction *Tentative = IDom->getTerminator();
4275 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4276 E = Inputs.end(); I != E; ++I) {
4277 Instruction *Inst = *I;
4278 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4279 AllDominate = false;
4282 // Attempt to find an insert position in the middle of the block,
4283 // instead of at the end, so that it can be used for other expansions.
4284 if (IDom == Inst->getParent() &&
4285 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4286 BetterPos = std::next(BasicBlock::iterator(Inst));
4299 /// AdjustInsertPositionForExpand - Determine an input position which will be
4300 /// dominated by the operands and which will dominate the result.
4301 BasicBlock::iterator
4302 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4305 SCEVExpander &Rewriter) const {
4306 // Collect some instructions which must be dominated by the
4307 // expanding replacement. These must be dominated by any operands that
4308 // will be required in the expansion.
4309 SmallVector<Instruction *, 4> Inputs;
4310 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4311 Inputs.push_back(I);
4312 if (LU.Kind == LSRUse::ICmpZero)
4313 if (Instruction *I =
4314 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4315 Inputs.push_back(I);
4316 if (LF.PostIncLoops.count(L)) {
4317 if (LF.isUseFullyOutsideLoop(L))
4318 Inputs.push_back(L->getLoopLatch()->getTerminator());
4320 Inputs.push_back(IVIncInsertPos);
4322 // The expansion must also be dominated by the increment positions of any
4323 // loops it for which it is using post-inc mode.
4324 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4325 E = LF.PostIncLoops.end(); I != E; ++I) {
4326 const Loop *PIL = *I;
4327 if (PIL == L) continue;
4329 // Be dominated by the loop exit.
4330 SmallVector<BasicBlock *, 4> ExitingBlocks;
4331 PIL->getExitingBlocks(ExitingBlocks);
4332 if (!ExitingBlocks.empty()) {
4333 BasicBlock *BB = ExitingBlocks[0];
4334 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4335 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4336 Inputs.push_back(BB->getTerminator());
4340 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4341 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4342 "Insertion point must be a normal instruction");
4344 // Then, climb up the immediate dominator tree as far as we can go while
4345 // still being dominated by the input positions.
4346 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4348 // Don't insert instructions before PHI nodes.
4349 while (isa<PHINode>(IP)) ++IP;
4351 // Ignore landingpad instructions.
4352 while (isa<LandingPadInst>(IP)) ++IP;
4354 // Ignore debug intrinsics.
4355 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4357 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4358 // IP consistent across expansions and allows the previously inserted
4359 // instructions to be reused by subsequent expansion.
4360 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4365 /// Expand - Emit instructions for the leading candidate expression for this
4366 /// LSRUse (this is called "expanding").
4367 Value *LSRInstance::Expand(const LSRFixup &LF,
4369 BasicBlock::iterator IP,
4370 SCEVExpander &Rewriter,
4371 SmallVectorImpl<WeakVH> &DeadInsts) const {
4372 const LSRUse &LU = Uses[LF.LUIdx];
4373 if (LU.RigidFormula)
4374 return LF.OperandValToReplace;
4376 // Determine an input position which will be dominated by the operands and
4377 // which will dominate the result.
4378 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4380 // Inform the Rewriter if we have a post-increment use, so that it can
4381 // perform an advantageous expansion.
4382 Rewriter.setPostInc(LF.PostIncLoops);
4384 // This is the type that the user actually needs.
4385 Type *OpTy = LF.OperandValToReplace->getType();
4386 // This will be the type that we'll initially expand to.
4387 Type *Ty = F.getType();
4389 // No type known; just expand directly to the ultimate type.
4391 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4392 // Expand directly to the ultimate type if it's the right size.
4394 // This is the type to do integer arithmetic in.
4395 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4397 // Build up a list of operands to add together to form the full base.
4398 SmallVector<const SCEV *, 8> Ops;
4400 // Expand the BaseRegs portion.
4401 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4402 E = F.BaseRegs.end(); I != E; ++I) {
4403 const SCEV *Reg = *I;
4404 assert(!Reg->isZero() && "Zero allocated in a base register!");
4406 // If we're expanding for a post-inc user, make the post-inc adjustment.
4407 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4408 Reg = TransformForPostIncUse(Denormalize, Reg,
4409 LF.UserInst, LF.OperandValToReplace,
4412 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4415 // Expand the ScaledReg portion.
4416 Value *ICmpScaledV = 0;
4418 const SCEV *ScaledS = F.ScaledReg;
4420 // If we're expanding for a post-inc user, make the post-inc adjustment.
4421 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4422 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4423 LF.UserInst, LF.OperandValToReplace,
4426 if (LU.Kind == LSRUse::ICmpZero) {
4427 // An interesting way of "folding" with an icmp is to use a negated
4428 // scale, which we'll implement by inserting it into the other operand
4430 assert(F.Scale == -1 &&
4431 "The only scale supported by ICmpZero uses is -1!");
4432 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4434 // Otherwise just expand the scaled register and an explicit scale,
4435 // which is expected to be matched as part of the address.
4437 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4438 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4439 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4441 Ops.push_back(SE.getUnknown(FullV));
4443 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4444 ScaledS = SE.getMulExpr(ScaledS,
4445 SE.getConstant(ScaledS->getType(), F.Scale));
4446 Ops.push_back(ScaledS);
4450 // Expand the GV portion.
4452 // Flush the operand list to suppress SCEVExpander hoisting.
4454 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4456 Ops.push_back(SE.getUnknown(FullV));
4458 Ops.push_back(SE.getUnknown(F.BaseGV));
4461 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4462 // unfolded offsets. LSR assumes they both live next to their uses.
4464 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4466 Ops.push_back(SE.getUnknown(FullV));
4469 // Expand the immediate portion.
4470 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4472 if (LU.Kind == LSRUse::ICmpZero) {
4473 // The other interesting way of "folding" with an ICmpZero is to use a
4474 // negated immediate.
4476 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4478 Ops.push_back(SE.getUnknown(ICmpScaledV));
4479 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4482 // Just add the immediate values. These again are expected to be matched
4483 // as part of the address.
4484 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4488 // Expand the unfolded offset portion.
4489 int64_t UnfoldedOffset = F.UnfoldedOffset;
4490 if (UnfoldedOffset != 0) {
4491 // Just add the immediate values.
4492 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4496 // Emit instructions summing all the operands.
4497 const SCEV *FullS = Ops.empty() ?
4498 SE.getConstant(IntTy, 0) :
4500 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4502 // We're done expanding now, so reset the rewriter.
4503 Rewriter.clearPostInc();
4505 // An ICmpZero Formula represents an ICmp which we're handling as a
4506 // comparison against zero. Now that we've expanded an expression for that
4507 // form, update the ICmp's other operand.
4508 if (LU.Kind == LSRUse::ICmpZero) {
4509 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4510 DeadInsts.push_back(CI->getOperand(1));
4511 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4512 "a scale at the same time!");
4513 if (F.Scale == -1) {
4514 if (ICmpScaledV->getType() != OpTy) {
4516 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4518 ICmpScaledV, OpTy, "tmp", CI);
4521 CI->setOperand(1, ICmpScaledV);
4523 assert(F.Scale == 0 &&
4524 "ICmp does not support folding a global value and "
4525 "a scale at the same time!");
4526 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4528 if (C->getType() != OpTy)
4529 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4533 CI->setOperand(1, C);
4540 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4541 /// of their operands effectively happens in their predecessor blocks, so the
4542 /// expression may need to be expanded in multiple places.
4543 void LSRInstance::RewriteForPHI(PHINode *PN,
4546 SCEVExpander &Rewriter,
4547 SmallVectorImpl<WeakVH> &DeadInsts,
4549 DenseMap<BasicBlock *, Value *> Inserted;
4550 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4551 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4552 BasicBlock *BB = PN->getIncomingBlock(i);
4554 // If this is a critical edge, split the edge so that we do not insert
4555 // the code on all predecessor/successor paths. We do this unless this
4556 // is the canonical backedge for this loop, which complicates post-inc
4558 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4559 !isa<IndirectBrInst>(BB->getTerminator())) {
4560 BasicBlock *Parent = PN->getParent();
4561 Loop *PNLoop = LI.getLoopFor(Parent);
4562 if (!PNLoop || Parent != PNLoop->getHeader()) {
4563 // Split the critical edge.
4564 BasicBlock *NewBB = 0;
4565 if (!Parent->isLandingPad()) {
4566 NewBB = SplitCriticalEdge(BB, Parent, P,
4567 /*MergeIdenticalEdges=*/true,
4568 /*DontDeleteUselessPhis=*/true);
4570 SmallVector<BasicBlock*, 2> NewBBs;
4571 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4574 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4575 // phi predecessors are identical. The simple thing to do is skip
4576 // splitting in this case rather than complicate the API.
4578 // If PN is outside of the loop and BB is in the loop, we want to
4579 // move the block to be immediately before the PHI block, not
4580 // immediately after BB.
4581 if (L->contains(BB) && !L->contains(PN))
4582 NewBB->moveBefore(PN->getParent());
4584 // Splitting the edge can reduce the number of PHI entries we have.
4585 e = PN->getNumIncomingValues();
4587 i = PN->getBasicBlockIndex(BB);
4592 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4593 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4595 PN->setIncomingValue(i, Pair.first->second);
4597 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4599 // If this is reuse-by-noop-cast, insert the noop cast.
4600 Type *OpTy = LF.OperandValToReplace->getType();
4601 if (FullV->getType() != OpTy)
4603 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4605 FullV, LF.OperandValToReplace->getType(),
4606 "tmp", BB->getTerminator());
4608 PN->setIncomingValue(i, FullV);
4609 Pair.first->second = FullV;
4614 /// Rewrite - Emit instructions for the leading candidate expression for this
4615 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4616 /// the newly expanded value.
4617 void LSRInstance::Rewrite(const LSRFixup &LF,
4619 SCEVExpander &Rewriter,
4620 SmallVectorImpl<WeakVH> &DeadInsts,
4622 // First, find an insertion point that dominates UserInst. For PHI nodes,
4623 // find the nearest block which dominates all the relevant uses.
4624 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4625 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4627 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4629 // If this is reuse-by-noop-cast, insert the noop cast.
4630 Type *OpTy = LF.OperandValToReplace->getType();
4631 if (FullV->getType() != OpTy) {
4633 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4634 FullV, OpTy, "tmp", LF.UserInst);
4638 // Update the user. ICmpZero is handled specially here (for now) because
4639 // Expand may have updated one of the operands of the icmp already, and
4640 // its new value may happen to be equal to LF.OperandValToReplace, in
4641 // which case doing replaceUsesOfWith leads to replacing both operands
4642 // with the same value. TODO: Reorganize this.
4643 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4644 LF.UserInst->setOperand(0, FullV);
4646 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4649 DeadInsts.push_back(LF.OperandValToReplace);
4652 /// ImplementSolution - Rewrite all the fixup locations with new values,
4653 /// following the chosen solution.
4655 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4657 // Keep track of instructions we may have made dead, so that
4658 // we can remove them after we are done working.
4659 SmallVector<WeakVH, 16> DeadInsts;
4661 SCEVExpander Rewriter(SE, "lsr");
4663 Rewriter.setDebugType(DEBUG_TYPE);
4665 Rewriter.disableCanonicalMode();
4666 Rewriter.enableLSRMode();
4667 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4669 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4670 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4671 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4672 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4673 Rewriter.setChainedPhi(PN);
4676 // Expand the new value definitions and update the users.
4677 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4678 E = Fixups.end(); I != E; ++I) {
4679 const LSRFixup &Fixup = *I;
4681 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4686 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4687 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4688 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4691 // Clean up after ourselves. This must be done before deleting any
4695 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4698 LSRInstance::LSRInstance(Loop *L, Pass *P)
4699 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4700 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4701 LI(P->getAnalysis<LoopInfo>()),
4702 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4704 // If LoopSimplify form is not available, stay out of trouble.
4705 if (!L->isLoopSimplifyForm())
4708 // If there's no interesting work to be done, bail early.
4709 if (IU.empty()) return;
4711 // If there's too much analysis to be done, bail early. We won't be able to
4712 // model the problem anyway.
4713 unsigned NumUsers = 0;
4714 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4715 if (++NumUsers > MaxIVUsers) {
4716 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4723 // All dominating loops must have preheaders, or SCEVExpander may not be able
4724 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4726 // IVUsers analysis should only create users that are dominated by simple loop
4727 // headers. Since this loop should dominate all of its users, its user list
4728 // should be empty if this loop itself is not within a simple loop nest.
4729 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4730 Rung; Rung = Rung->getIDom()) {
4731 BasicBlock *BB = Rung->getBlock();
4732 const Loop *DomLoop = LI.getLoopFor(BB);
4733 if (DomLoop && DomLoop->getHeader() == BB) {
4734 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4739 DEBUG(dbgs() << "\nLSR on loop ";
4740 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4743 // First, perform some low-level loop optimizations.
4745 OptimizeLoopTermCond();
4747 // If loop preparation eliminates all interesting IV users, bail.
4748 if (IU.empty()) return;
4750 // Skip nested loops until we can model them better with formulae.
4752 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4756 // Start collecting data and preparing for the solver.
4758 CollectInterestingTypesAndFactors();
4759 CollectFixupsAndInitialFormulae();
4760 CollectLoopInvariantFixupsAndFormulae();
4762 assert(!Uses.empty() && "IVUsers reported at least one use");
4763 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4764 print_uses(dbgs()));
4766 // Now use the reuse data to generate a bunch of interesting ways
4767 // to formulate the values needed for the uses.
4768 GenerateAllReuseFormulae();
4770 FilterOutUndesirableDedicatedRegisters();
4771 NarrowSearchSpaceUsingHeuristics();
4773 SmallVector<const Formula *, 8> Solution;
4776 // Release memory that is no longer needed.
4781 if (Solution.empty())
4785 // Formulae should be legal.
4786 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4788 const LSRUse &LU = *I;
4789 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4790 JE = LU.Formulae.end();
4792 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4793 *J) && "Illegal formula generated!");
4797 // Now that we've decided what we want, make it so.
4798 ImplementSolution(Solution, P);
4801 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4802 if (Factors.empty() && Types.empty()) return;
4804 OS << "LSR has identified the following interesting factors and types: ";
4807 for (SmallSetVector<int64_t, 8>::const_iterator
4808 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4809 if (!First) OS << ", ";
4814 for (SmallSetVector<Type *, 4>::const_iterator
4815 I = Types.begin(), E = Types.end(); I != E; ++I) {
4816 if (!First) OS << ", ";
4818 OS << '(' << **I << ')';
4823 void LSRInstance::print_fixups(raw_ostream &OS) const {
4824 OS << "LSR is examining the following fixup sites:\n";
4825 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4826 E = Fixups.end(); I != E; ++I) {
4833 void LSRInstance::print_uses(raw_ostream &OS) const {
4834 OS << "LSR is examining the following uses:\n";
4835 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4836 E = Uses.end(); I != E; ++I) {
4837 const LSRUse &LU = *I;
4841 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4842 JE = LU.Formulae.end(); J != JE; ++J) {
4850 void LSRInstance::print(raw_ostream &OS) const {
4851 print_factors_and_types(OS);
4856 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4857 void LSRInstance::dump() const {
4858 print(errs()); errs() << '\n';
4864 class LoopStrengthReduce : public LoopPass {
4866 static char ID; // Pass ID, replacement for typeid
4867 LoopStrengthReduce();
4870 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
4871 void getAnalysisUsage(AnalysisUsage &AU) const override;
4876 char LoopStrengthReduce::ID = 0;
4877 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4878 "Loop Strength Reduction", false, false)
4879 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4880 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4881 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4882 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4883 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4884 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4885 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4886 "Loop Strength Reduction", false, false)
4889 Pass *llvm::createLoopStrengthReducePass() {
4890 return new LoopStrengthReduce();
4893 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4894 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4897 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4898 // We split critical edges, so we change the CFG. However, we do update
4899 // many analyses if they are around.
4900 AU.addPreservedID(LoopSimplifyID);
4902 AU.addRequired<LoopInfo>();
4903 AU.addPreserved<LoopInfo>();
4904 AU.addRequiredID(LoopSimplifyID);
4905 AU.addRequired<DominatorTreeWrapperPass>();
4906 AU.addPreserved<DominatorTreeWrapperPass>();
4907 AU.addRequired<ScalarEvolution>();
4908 AU.addPreserved<ScalarEvolution>();
4909 // Requiring LoopSimplify a second time here prevents IVUsers from running
4910 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4911 AU.addRequiredID(LoopSimplifyID);
4912 AU.addRequired<IVUsers>();
4913 AU.addPreserved<IVUsers>();
4914 AU.addRequired<TargetTransformInfo>();
4917 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4918 if (skipOptnoneFunction(L))
4921 bool Changed = false;
4923 // Run the main LSR transformation.
4924 Changed |= LSRInstance(L, this).getChanged();
4926 // Remove any extra phis created by processing inner loops.
4927 Changed |= DeleteDeadPHIs(L->getHeader());
4928 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4929 SmallVector<WeakVH, 16> DeadInsts;
4930 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4932 Rewriter.setDebugType(DEBUG_TYPE);
4934 unsigned numFolded = Rewriter.replaceCongruentIVs(
4935 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
4936 &getAnalysis<TargetTransformInfo>());
4939 DeleteTriviallyDeadInstructions(DeadInsts);
4940 DeleteDeadPHIs(L->getHeader());