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 (User *UR : UVal->users()) {
726 // If U is a constant, it may be used by a ConstantExpr.
727 Instruction *UI = dyn_cast<Instruction>(UR);
728 if (UI && UI->getOpcode() == Instruction::Mul &&
729 SE.isSCEVable(UI->getType())) {
730 return SE.getSCEV(UI) == Mul;
737 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
738 if (isExistingPhi(AR, SE))
742 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
746 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
747 /// specified set are trivially dead, delete them and see if this makes any of
748 /// their operands subsequently dead.
750 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
751 bool Changed = false;
753 while (!DeadInsts.empty()) {
754 Value *V = DeadInsts.pop_back_val();
755 Instruction *I = dyn_cast_or_null<Instruction>(V);
757 if (I == 0 || !isInstructionTriviallyDead(I))
760 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
761 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
764 DeadInsts.push_back(U);
767 I->eraseFromParent();
777 // Check if it is legal to fold 2 base registers.
778 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
780 // Get the cost of the scaling factor used in F for LU.
781 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
782 const LSRUse &LU, const Formula &F);
786 /// Cost - This class is used to measure and compare candidate formulae.
788 /// TODO: Some of these could be merged. Also, a lexical ordering
789 /// isn't always optimal.
793 unsigned NumBaseAdds;
800 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
801 SetupCost(0), ScaleCost(0) {}
803 bool operator<(const Cost &Other) const;
808 // Once any of the metrics loses, they must all remain losers.
810 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
811 | ImmCost | SetupCost | ScaleCost) != ~0u)
812 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
813 & ImmCost & SetupCost & ScaleCost) == ~0u);
818 assert(isValid() && "invalid cost");
819 return NumRegs == ~0u;
822 void RateFormula(const TargetTransformInfo &TTI,
824 SmallPtrSet<const SCEV *, 16> &Regs,
825 const DenseSet<const SCEV *> &VisitedRegs,
827 const SmallVectorImpl<int64_t> &Offsets,
828 ScalarEvolution &SE, DominatorTree &DT,
830 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
832 void print(raw_ostream &OS) const;
836 void RateRegister(const SCEV *Reg,
837 SmallPtrSet<const SCEV *, 16> &Regs,
839 ScalarEvolution &SE, DominatorTree &DT);
840 void RatePrimaryRegister(const SCEV *Reg,
841 SmallPtrSet<const SCEV *, 16> &Regs,
843 ScalarEvolution &SE, DominatorTree &DT,
844 SmallPtrSet<const SCEV *, 16> *LoserRegs);
849 /// RateRegister - Tally up interesting quantities from the given register.
850 void Cost::RateRegister(const SCEV *Reg,
851 SmallPtrSet<const SCEV *, 16> &Regs,
853 ScalarEvolution &SE, DominatorTree &DT) {
854 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
855 // If this is an addrec for another loop, don't second-guess its addrec phi
856 // nodes. LSR isn't currently smart enough to reason about more than one
857 // loop at a time. LSR has already run on inner loops, will not run on outer
858 // loops, and cannot be expected to change sibling loops.
859 if (AR->getLoop() != L) {
860 // If the AddRec exists, consider it's register free and leave it alone.
861 if (isExistingPhi(AR, SE))
864 // Otherwise, do not consider this formula at all.
868 AddRecCost += 1; /// TODO: This should be a function of the stride.
870 // Add the step value register, if it needs one.
871 // TODO: The non-affine case isn't precisely modeled here.
872 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
873 if (!Regs.count(AR->getOperand(1))) {
874 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
882 // Rough heuristic; favor registers which don't require extra setup
883 // instructions in the preheader.
884 if (!isa<SCEVUnknown>(Reg) &&
885 !isa<SCEVConstant>(Reg) &&
886 !(isa<SCEVAddRecExpr>(Reg) &&
887 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
888 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
891 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
892 SE.hasComputableLoopEvolution(Reg, L);
895 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
896 /// before, rate it. Optional LoserRegs provides a way to declare any formula
897 /// that refers to one of those regs an instant loser.
898 void Cost::RatePrimaryRegister(const SCEV *Reg,
899 SmallPtrSet<const SCEV *, 16> &Regs,
901 ScalarEvolution &SE, DominatorTree &DT,
902 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
903 if (LoserRegs && LoserRegs->count(Reg)) {
907 if (Regs.insert(Reg)) {
908 RateRegister(Reg, Regs, L, SE, DT);
909 if (LoserRegs && isLoser())
910 LoserRegs->insert(Reg);
914 void Cost::RateFormula(const TargetTransformInfo &TTI,
916 SmallPtrSet<const SCEV *, 16> &Regs,
917 const DenseSet<const SCEV *> &VisitedRegs,
919 const SmallVectorImpl<int64_t> &Offsets,
920 ScalarEvolution &SE, DominatorTree &DT,
922 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
923 // Tally up the registers.
924 if (const SCEV *ScaledReg = F.ScaledReg) {
925 if (VisitedRegs.count(ScaledReg)) {
929 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
933 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
934 E = F.BaseRegs.end(); I != E; ++I) {
935 const SCEV *BaseReg = *I;
936 if (VisitedRegs.count(BaseReg)) {
940 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
945 // Determine how many (unfolded) adds we'll need inside the loop.
946 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
947 if (NumBaseParts > 1)
948 // Do not count the base and a possible second register if the target
949 // allows to fold 2 registers.
950 NumBaseAdds += NumBaseParts - (1 + isLegal2RegAMUse(TTI, LU, F));
952 // Accumulate non-free scaling amounts.
953 ScaleCost += getScalingFactorCost(TTI, LU, F);
955 // Tally up the non-zero immediates.
956 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
957 E = Offsets.end(); I != E; ++I) {
958 int64_t Offset = (uint64_t)*I + F.BaseOffset;
960 ImmCost += 64; // Handle symbolic values conservatively.
961 // TODO: This should probably be the pointer size.
962 else if (Offset != 0)
963 ImmCost += APInt(64, Offset, true).getMinSignedBits();
965 assert(isValid() && "invalid cost");
968 /// Lose - Set this cost to a losing value.
979 /// operator< - Choose the lower cost.
980 bool Cost::operator<(const Cost &Other) const {
981 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
982 ImmCost, SetupCost) <
983 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
984 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
988 void Cost::print(raw_ostream &OS) const {
989 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
991 OS << ", with addrec cost " << AddRecCost;
993 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
994 if (NumBaseAdds != 0)
995 OS << ", plus " << NumBaseAdds << " base add"
996 << (NumBaseAdds == 1 ? "" : "s");
998 OS << ", plus " << ScaleCost << " scale cost";
1000 OS << ", plus " << ImmCost << " imm cost";
1002 OS << ", plus " << SetupCost << " setup cost";
1005 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1006 void Cost::dump() const {
1007 print(errs()); errs() << '\n';
1013 /// LSRFixup - An operand value in an instruction which is to be replaced
1014 /// with some equivalent, possibly strength-reduced, replacement.
1016 /// UserInst - The instruction which will be updated.
1017 Instruction *UserInst;
1019 /// OperandValToReplace - The operand of the instruction which will
1020 /// be replaced. The operand may be used more than once; every instance
1021 /// will be replaced.
1022 Value *OperandValToReplace;
1024 /// PostIncLoops - If this user is to use the post-incremented value of an
1025 /// induction variable, this variable is non-null and holds the loop
1026 /// associated with the induction variable.
1027 PostIncLoopSet PostIncLoops;
1029 /// LUIdx - The index of the LSRUse describing the expression which
1030 /// this fixup needs, minus an offset (below).
1033 /// Offset - A constant offset to be added to the LSRUse expression.
1034 /// This allows multiple fixups to share the same LSRUse with different
1035 /// offsets, for example in an unrolled loop.
1038 bool isUseFullyOutsideLoop(const Loop *L) const;
1042 void print(raw_ostream &OS) const;
1048 LSRFixup::LSRFixup()
1049 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1051 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1052 /// value outside of the given loop.
1053 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1054 // PHI nodes use their value in their incoming blocks.
1055 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1056 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1057 if (PN->getIncomingValue(i) == OperandValToReplace &&
1058 L->contains(PN->getIncomingBlock(i)))
1063 return !L->contains(UserInst);
1066 void LSRFixup::print(raw_ostream &OS) const {
1068 // Store is common and interesting enough to be worth special-casing.
1069 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1071 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1072 } else if (UserInst->getType()->isVoidTy())
1073 OS << UserInst->getOpcodeName();
1075 UserInst->printAsOperand(OS, /*PrintType=*/false);
1077 OS << ", OperandValToReplace=";
1078 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1080 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1081 E = PostIncLoops.end(); I != E; ++I) {
1082 OS << ", PostIncLoop=";
1083 (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1086 if (LUIdx != ~size_t(0))
1087 OS << ", LUIdx=" << LUIdx;
1090 OS << ", Offset=" << Offset;
1093 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1094 void LSRFixup::dump() const {
1095 print(errs()); errs() << '\n';
1101 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1102 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1103 struct UniquifierDenseMapInfo {
1104 static SmallVector<const SCEV *, 4> getEmptyKey() {
1105 SmallVector<const SCEV *, 4> V;
1106 V.push_back(reinterpret_cast<const SCEV *>(-1));
1110 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1111 SmallVector<const SCEV *, 4> V;
1112 V.push_back(reinterpret_cast<const SCEV *>(-2));
1116 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1117 unsigned Result = 0;
1118 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1119 E = V.end(); I != E; ++I)
1120 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1124 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1125 const SmallVector<const SCEV *, 4> &RHS) {
1130 /// LSRUse - This class holds the state that LSR keeps for each use in
1131 /// IVUsers, as well as uses invented by LSR itself. It includes information
1132 /// about what kinds of things can be folded into the user, information about
1133 /// the user itself, and information about how the use may be satisfied.
1134 /// TODO: Represent multiple users of the same expression in common?
1136 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1139 /// KindType - An enum for a kind of use, indicating what types of
1140 /// scaled and immediate operands it might support.
1142 Basic, ///< A normal use, with no folding.
1143 Special, ///< A special case of basic, allowing -1 scales.
1144 Address, ///< An address use; folding according to TargetLowering
1145 ICmpZero ///< An equality icmp with both operands folded into one.
1146 // TODO: Add a generic icmp too?
1152 SmallVector<int64_t, 8> Offsets;
1156 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1157 /// LSRUse are outside of the loop, in which case some special-case heuristics
1159 bool AllFixupsOutsideLoop;
1161 /// RigidFormula is set to true to guarantee that this use will be associated
1162 /// with a single formula--the one that initially matched. Some SCEV
1163 /// expressions cannot be expanded. This allows LSR to consider the registers
1164 /// used by those expressions without the need to expand them later after
1165 /// changing the formula.
1168 /// WidestFixupType - This records the widest use type for any fixup using
1169 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1170 /// max fixup widths to be equivalent, because the narrower one may be relying
1171 /// on the implicit truncation to truncate away bogus bits.
1172 Type *WidestFixupType;
1174 /// Formulae - A list of ways to build a value that can satisfy this user.
1175 /// After the list is populated, one of these is selected heuristically and
1176 /// used to formulate a replacement for OperandValToReplace in UserInst.
1177 SmallVector<Formula, 12> Formulae;
1179 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1180 SmallPtrSet<const SCEV *, 4> Regs;
1182 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1183 MinOffset(INT64_MAX),
1184 MaxOffset(INT64_MIN),
1185 AllFixupsOutsideLoop(true),
1186 RigidFormula(false),
1187 WidestFixupType(0) {}
1189 bool HasFormulaWithSameRegs(const Formula &F) const;
1190 bool InsertFormula(const Formula &F);
1191 void DeleteFormula(Formula &F);
1192 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1194 void print(raw_ostream &OS) const;
1200 /// HasFormula - Test whether this use as a formula which has the same
1201 /// registers as the given formula.
1202 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1203 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1204 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1205 // Unstable sort by host order ok, because this is only used for uniquifying.
1206 std::sort(Key.begin(), Key.end());
1207 return Uniquifier.count(Key);
1210 /// InsertFormula - If the given formula has not yet been inserted, add it to
1211 /// the list, and return true. Return false otherwise.
1212 bool LSRUse::InsertFormula(const Formula &F) {
1213 if (!Formulae.empty() && RigidFormula)
1216 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1217 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1218 // Unstable sort by host order ok, because this is only used for uniquifying.
1219 std::sort(Key.begin(), Key.end());
1221 if (!Uniquifier.insert(Key).second)
1224 // Using a register to hold the value of 0 is not profitable.
1225 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1226 "Zero allocated in a scaled register!");
1228 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1229 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1230 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1233 // Add the formula to the list.
1234 Formulae.push_back(F);
1236 // Record registers now being used by this use.
1237 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1242 /// DeleteFormula - Remove the given formula from this use's list.
1243 void LSRUse::DeleteFormula(Formula &F) {
1244 if (&F != &Formulae.back())
1245 std::swap(F, Formulae.back());
1246 Formulae.pop_back();
1249 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1250 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1251 // Now that we've filtered out some formulae, recompute the Regs set.
1252 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1254 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1255 E = Formulae.end(); I != E; ++I) {
1256 const Formula &F = *I;
1257 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1258 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1261 // Update the RegTracker.
1262 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1263 E = OldRegs.end(); I != E; ++I)
1264 if (!Regs.count(*I))
1265 RegUses.DropRegister(*I, LUIdx);
1268 void LSRUse::print(raw_ostream &OS) const {
1269 OS << "LSR Use: Kind=";
1271 case Basic: OS << "Basic"; break;
1272 case Special: OS << "Special"; break;
1273 case ICmpZero: OS << "ICmpZero"; break;
1275 OS << "Address of ";
1276 if (AccessTy->isPointerTy())
1277 OS << "pointer"; // the full pointer type could be really verbose
1282 OS << ", Offsets={";
1283 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1284 E = Offsets.end(); I != E; ++I) {
1286 if (std::next(I) != E)
1291 if (AllFixupsOutsideLoop)
1292 OS << ", all-fixups-outside-loop";
1294 if (WidestFixupType)
1295 OS << ", widest fixup type: " << *WidestFixupType;
1298 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1299 void LSRUse::dump() const {
1300 print(errs()); errs() << '\n';
1304 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1305 /// be completely folded into the user instruction at isel time. This includes
1306 /// address-mode folding and special icmp tricks.
1307 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1308 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1309 bool HasBaseReg, int64_t Scale) {
1311 case LSRUse::Address:
1312 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1314 // Otherwise, just guess that reg+reg addressing is legal.
1317 case LSRUse::ICmpZero:
1318 // There's not even a target hook for querying whether it would be legal to
1319 // fold a GV into an ICmp.
1323 // ICmp only has two operands; don't allow more than two non-trivial parts.
1324 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1327 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1328 // putting the scaled register in the other operand of the icmp.
1329 if (Scale != 0 && Scale != -1)
1332 // If we have low-level target information, ask the target if it can fold an
1333 // integer immediate on an icmp.
1334 if (BaseOffset != 0) {
1336 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1337 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1338 // Offs is the ICmp immediate.
1340 // The cast does the right thing with INT64_MIN.
1341 BaseOffset = -(uint64_t)BaseOffset;
1342 return TTI.isLegalICmpImmediate(BaseOffset);
1345 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1349 // Only handle single-register values.
1350 return !BaseGV && Scale == 0 && BaseOffset == 0;
1352 case LSRUse::Special:
1353 // Special case Basic to handle -1 scales.
1354 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1357 llvm_unreachable("Invalid LSRUse Kind!");
1360 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1361 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1362 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1364 // Check for overflow.
1365 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1368 MinOffset = (uint64_t)BaseOffset + MinOffset;
1369 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1372 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1374 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1376 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1379 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1380 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1382 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1383 F.BaseOffset, F.HasBaseReg, F.Scale);
1386 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
1388 // If F is used as an Addressing Mode, it may fold one Base plus one
1389 // scaled register. If the scaled register is nil, do as if another
1390 // element of the base regs is a 1-scaled register.
1391 // This is possible if BaseRegs has at least 2 registers.
1393 // If this is not an address calculation, this is not an addressing mode
1395 if (LU.Kind != LSRUse::Address)
1398 // F is already scaled.
1402 // We need to keep one register for the base and one to scale.
1403 if (F.BaseRegs.size() < 2)
1406 return isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
1407 F.BaseGV, F.BaseOffset, F.HasBaseReg, 1);
1410 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1411 const LSRUse &LU, const Formula &F) {
1414 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1415 LU.AccessTy, F) && "Illegal formula in use.");
1418 case LSRUse::Address: {
1419 // Check the scaling factor cost with both the min and max offsets.
1420 int ScaleCostMinOffset =
1421 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1422 F.BaseOffset + LU.MinOffset,
1423 F.HasBaseReg, F.Scale);
1424 int ScaleCostMaxOffset =
1425 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1426 F.BaseOffset + LU.MaxOffset,
1427 F.HasBaseReg, F.Scale);
1429 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1430 "Legal addressing mode has an illegal cost!");
1431 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1433 case LSRUse::ICmpZero:
1434 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
1435 // Therefore, return 0 in case F.Scale == -1.
1436 return F.Scale != -1;
1439 case LSRUse::Special:
1443 llvm_unreachable("Invalid LSRUse Kind!");
1446 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1447 LSRUse::KindType Kind, Type *AccessTy,
1448 GlobalValue *BaseGV, int64_t BaseOffset,
1450 // Fast-path: zero is always foldable.
1451 if (BaseOffset == 0 && !BaseGV) return true;
1453 // Conservatively, create an address with an immediate and a
1454 // base and a scale.
1455 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1457 // Canonicalize a scale of 1 to a base register if the formula doesn't
1458 // already have a base register.
1459 if (!HasBaseReg && Scale == 1) {
1464 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1467 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1468 ScalarEvolution &SE, int64_t MinOffset,
1469 int64_t MaxOffset, LSRUse::KindType Kind,
1470 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1471 // Fast-path: zero is always foldable.
1472 if (S->isZero()) return true;
1474 // Conservatively, create an address with an immediate and a
1475 // base and a scale.
1476 int64_t BaseOffset = ExtractImmediate(S, SE);
1477 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1479 // If there's anything else involved, it's not foldable.
1480 if (!S->isZero()) return false;
1482 // Fast-path: zero is always foldable.
1483 if (BaseOffset == 0 && !BaseGV) return true;
1485 // Conservatively, create an address with an immediate and a
1486 // base and a scale.
1487 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1489 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1490 BaseOffset, HasBaseReg, Scale);
1495 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1496 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1497 struct UseMapDenseMapInfo {
1498 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1499 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1502 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1503 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1507 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1508 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1509 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1513 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1514 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1519 /// IVInc - An individual increment in a Chain of IV increments.
1520 /// Relate an IV user to an expression that computes the IV it uses from the IV
1521 /// used by the previous link in the Chain.
1523 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1524 /// original IVOperand. The head of the chain's IVOperand is only valid during
1525 /// chain collection, before LSR replaces IV users. During chain generation,
1526 /// IncExpr can be used to find the new IVOperand that computes the same
1529 Instruction *UserInst;
1531 const SCEV *IncExpr;
1533 IVInc(Instruction *U, Value *O, const SCEV *E):
1534 UserInst(U), IVOperand(O), IncExpr(E) {}
1537 // IVChain - The list of IV increments in program order.
1538 // We typically add the head of a chain without finding subsequent links.
1540 SmallVector<IVInc,1> Incs;
1541 const SCEV *ExprBase;
1543 IVChain() : ExprBase(0) {}
1545 IVChain(const IVInc &Head, const SCEV *Base)
1546 : Incs(1, Head), ExprBase(Base) {}
1548 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1550 // begin - return the first increment in the chain.
1551 const_iterator begin() const {
1552 assert(!Incs.empty());
1553 return std::next(Incs.begin());
1555 const_iterator end() const {
1559 // hasIncs - Returns true if this chain contains any increments.
1560 bool hasIncs() const { return Incs.size() >= 2; }
1562 // add - Add an IVInc to the end of this chain.
1563 void add(const IVInc &X) { Incs.push_back(X); }
1565 // tailUserInst - Returns the last UserInst in the chain.
1566 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1568 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1570 bool isProfitableIncrement(const SCEV *OperExpr,
1571 const SCEV *IncExpr,
1575 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1576 /// Distinguish between FarUsers that definitely cross IV increments and
1577 /// NearUsers that may be used between IV increments.
1579 SmallPtrSet<Instruction*, 4> FarUsers;
1580 SmallPtrSet<Instruction*, 4> NearUsers;
1583 /// LSRInstance - This class holds state for the main loop strength reduction
1587 ScalarEvolution &SE;
1590 const TargetTransformInfo &TTI;
1594 /// IVIncInsertPos - This is the insert position that the current loop's
1595 /// induction variable increment should be placed. In simple loops, this is
1596 /// the latch block's terminator. But in more complicated cases, this is a
1597 /// position which will dominate all the in-loop post-increment users.
1598 Instruction *IVIncInsertPos;
1600 /// Factors - Interesting factors between use strides.
1601 SmallSetVector<int64_t, 8> Factors;
1603 /// Types - Interesting use types, to facilitate truncation reuse.
1604 SmallSetVector<Type *, 4> Types;
1606 /// Fixups - The list of operands which are to be replaced.
1607 SmallVector<LSRFixup, 16> Fixups;
1609 /// Uses - The list of interesting uses.
1610 SmallVector<LSRUse, 16> Uses;
1612 /// RegUses - Track which uses use which register candidates.
1613 RegUseTracker RegUses;
1615 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1616 // have more than a few IV increment chains in a loop. Missing a Chain falls
1617 // back to normal LSR behavior for those uses.
1618 static const unsigned MaxChains = 8;
1620 /// IVChainVec - IV users can form a chain of IV increments.
1621 SmallVector<IVChain, MaxChains> IVChainVec;
1623 /// IVIncSet - IV users that belong to profitable IVChains.
1624 SmallPtrSet<Use*, MaxChains> IVIncSet;
1626 void OptimizeShadowIV();
1627 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1628 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1629 void OptimizeLoopTermCond();
1631 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1632 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1633 void FinalizeChain(IVChain &Chain);
1634 void CollectChains();
1635 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1636 SmallVectorImpl<WeakVH> &DeadInsts);
1638 void CollectInterestingTypesAndFactors();
1639 void CollectFixupsAndInitialFormulae();
1641 LSRFixup &getNewFixup() {
1642 Fixups.push_back(LSRFixup());
1643 return Fixups.back();
1646 // Support for sharing of LSRUses between LSRFixups.
1647 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1649 UseMapDenseMapInfo> UseMapTy;
1652 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1653 LSRUse::KindType Kind, Type *AccessTy);
1655 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1656 LSRUse::KindType Kind,
1659 void DeleteUse(LSRUse &LU, size_t LUIdx);
1661 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1663 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1664 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1665 void CountRegisters(const Formula &F, size_t LUIdx);
1666 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1668 void CollectLoopInvariantFixupsAndFormulae();
1670 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1671 unsigned Depth = 0);
1672 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1673 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1674 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1675 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1676 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1677 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1678 void GenerateCrossUseConstantOffsets();
1679 void GenerateAllReuseFormulae();
1681 void FilterOutUndesirableDedicatedRegisters();
1683 size_t EstimateSearchSpaceComplexity() const;
1684 void NarrowSearchSpaceByDetectingSupersets();
1685 void NarrowSearchSpaceByCollapsingUnrolledCode();
1686 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1687 void NarrowSearchSpaceByPickingWinnerRegs();
1688 void NarrowSearchSpaceUsingHeuristics();
1690 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1692 SmallVectorImpl<const Formula *> &Workspace,
1693 const Cost &CurCost,
1694 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1695 DenseSet<const SCEV *> &VisitedRegs) const;
1696 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1698 BasicBlock::iterator
1699 HoistInsertPosition(BasicBlock::iterator IP,
1700 const SmallVectorImpl<Instruction *> &Inputs) const;
1701 BasicBlock::iterator
1702 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1705 SCEVExpander &Rewriter) const;
1707 Value *Expand(const LSRFixup &LF,
1709 BasicBlock::iterator IP,
1710 SCEVExpander &Rewriter,
1711 SmallVectorImpl<WeakVH> &DeadInsts) const;
1712 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1714 SCEVExpander &Rewriter,
1715 SmallVectorImpl<WeakVH> &DeadInsts,
1717 void Rewrite(const LSRFixup &LF,
1719 SCEVExpander &Rewriter,
1720 SmallVectorImpl<WeakVH> &DeadInsts,
1722 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1726 LSRInstance(Loop *L, Pass *P);
1728 bool getChanged() const { return Changed; }
1730 void print_factors_and_types(raw_ostream &OS) const;
1731 void print_fixups(raw_ostream &OS) const;
1732 void print_uses(raw_ostream &OS) const;
1733 void print(raw_ostream &OS) const;
1739 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1740 /// inside the loop then try to eliminate the cast operation.
1741 void LSRInstance::OptimizeShadowIV() {
1742 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1743 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1746 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1747 UI != E; /* empty */) {
1748 IVUsers::const_iterator CandidateUI = UI;
1750 Instruction *ShadowUse = CandidateUI->getUser();
1752 bool IsSigned = false;
1754 /* If shadow use is a int->float cast then insert a second IV
1755 to eliminate this cast.
1757 for (unsigned i = 0; i < n; ++i)
1763 for (unsigned i = 0; i < n; ++i, ++d)
1766 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1768 DestTy = UCast->getDestTy();
1770 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1772 DestTy = SCast->getDestTy();
1774 if (!DestTy) continue;
1776 // If target does not support DestTy natively then do not apply
1777 // this transformation.
1778 if (!TTI.isTypeLegal(DestTy)) continue;
1780 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1782 if (PH->getNumIncomingValues() != 2) continue;
1784 Type *SrcTy = PH->getType();
1785 int Mantissa = DestTy->getFPMantissaWidth();
1786 if (Mantissa == -1) continue;
1787 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1790 unsigned Entry, Latch;
1791 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1799 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1800 if (!Init) continue;
1801 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1802 (double)Init->getSExtValue() :
1803 (double)Init->getZExtValue());
1805 BinaryOperator *Incr =
1806 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1807 if (!Incr) continue;
1808 if (Incr->getOpcode() != Instruction::Add
1809 && Incr->getOpcode() != Instruction::Sub)
1812 /* Initialize new IV, double d = 0.0 in above example. */
1814 if (Incr->getOperand(0) == PH)
1815 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1816 else if (Incr->getOperand(1) == PH)
1817 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1823 // Ignore negative constants, as the code below doesn't handle them
1824 // correctly. TODO: Remove this restriction.
1825 if (!C->getValue().isStrictlyPositive()) continue;
1827 /* Add new PHINode. */
1828 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1830 /* create new increment. '++d' in above example. */
1831 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1832 BinaryOperator *NewIncr =
1833 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1834 Instruction::FAdd : Instruction::FSub,
1835 NewPH, CFP, "IV.S.next.", Incr);
1837 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1838 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1840 /* Remove cast operation */
1841 ShadowUse->replaceAllUsesWith(NewPH);
1842 ShadowUse->eraseFromParent();
1848 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1849 /// set the IV user and stride information and return true, otherwise return
1851 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1852 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1853 if (UI->getUser() == Cond) {
1854 // NOTE: we could handle setcc instructions with multiple uses here, but
1855 // InstCombine does it as well for simple uses, it's not clear that it
1856 // occurs enough in real life to handle.
1863 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1864 /// a max computation.
1866 /// This is a narrow solution to a specific, but acute, problem. For loops
1872 /// } while (++i < n);
1874 /// the trip count isn't just 'n', because 'n' might not be positive. And
1875 /// unfortunately this can come up even for loops where the user didn't use
1876 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1877 /// will commonly be lowered like this:
1883 /// } while (++i < n);
1886 /// and then it's possible for subsequent optimization to obscure the if
1887 /// test in such a way that indvars can't find it.
1889 /// When indvars can't find the if test in loops like this, it creates a
1890 /// max expression, which allows it to give the loop a canonical
1891 /// induction variable:
1894 /// max = n < 1 ? 1 : n;
1897 /// } while (++i != max);
1899 /// Canonical induction variables are necessary because the loop passes
1900 /// are designed around them. The most obvious example of this is the
1901 /// LoopInfo analysis, which doesn't remember trip count values. It
1902 /// expects to be able to rediscover the trip count each time it is
1903 /// needed, and it does this using a simple analysis that only succeeds if
1904 /// the loop has a canonical induction variable.
1906 /// However, when it comes time to generate code, the maximum operation
1907 /// can be quite costly, especially if it's inside of an outer loop.
1909 /// This function solves this problem by detecting this type of loop and
1910 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1911 /// the instructions for the maximum computation.
1913 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1914 // Check that the loop matches the pattern we're looking for.
1915 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1916 Cond->getPredicate() != CmpInst::ICMP_NE)
1919 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1920 if (!Sel || !Sel->hasOneUse()) return Cond;
1922 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1923 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1925 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1927 // Add one to the backedge-taken count to get the trip count.
1928 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1929 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1931 // Check for a max calculation that matches the pattern. There's no check
1932 // for ICMP_ULE here because the comparison would be with zero, which
1933 // isn't interesting.
1934 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1935 const SCEVNAryExpr *Max = 0;
1936 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1937 Pred = ICmpInst::ICMP_SLE;
1939 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1940 Pred = ICmpInst::ICMP_SLT;
1942 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1943 Pred = ICmpInst::ICMP_ULT;
1950 // To handle a max with more than two operands, this optimization would
1951 // require additional checking and setup.
1952 if (Max->getNumOperands() != 2)
1955 const SCEV *MaxLHS = Max->getOperand(0);
1956 const SCEV *MaxRHS = Max->getOperand(1);
1958 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1959 // for a comparison with 1. For <= and >=, a comparison with zero.
1961 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1964 // Check the relevant induction variable for conformance to
1966 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1967 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1968 if (!AR || !AR->isAffine() ||
1969 AR->getStart() != One ||
1970 AR->getStepRecurrence(SE) != One)
1973 assert(AR->getLoop() == L &&
1974 "Loop condition operand is an addrec in a different loop!");
1976 // Check the right operand of the select, and remember it, as it will
1977 // be used in the new comparison instruction.
1979 if (ICmpInst::isTrueWhenEqual(Pred)) {
1980 // Look for n+1, and grab n.
1981 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1982 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1983 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1984 NewRHS = BO->getOperand(0);
1985 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1986 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1987 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1988 NewRHS = BO->getOperand(0);
1991 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1992 NewRHS = Sel->getOperand(1);
1993 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1994 NewRHS = Sel->getOperand(2);
1995 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1996 NewRHS = SU->getValue();
1998 // Max doesn't match expected pattern.
2001 // Determine the new comparison opcode. It may be signed or unsigned,
2002 // and the original comparison may be either equality or inequality.
2003 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2004 Pred = CmpInst::getInversePredicate(Pred);
2006 // Ok, everything looks ok to change the condition into an SLT or SGE and
2007 // delete the max calculation.
2009 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2011 // Delete the max calculation instructions.
2012 Cond->replaceAllUsesWith(NewCond);
2013 CondUse->setUser(NewCond);
2014 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2015 Cond->eraseFromParent();
2016 Sel->eraseFromParent();
2017 if (Cmp->use_empty())
2018 Cmp->eraseFromParent();
2022 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2023 /// postinc iv when possible.
2025 LSRInstance::OptimizeLoopTermCond() {
2026 SmallPtrSet<Instruction *, 4> PostIncs;
2028 BasicBlock *LatchBlock = L->getLoopLatch();
2029 SmallVector<BasicBlock*, 8> ExitingBlocks;
2030 L->getExitingBlocks(ExitingBlocks);
2032 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2033 BasicBlock *ExitingBlock = ExitingBlocks[i];
2035 // Get the terminating condition for the loop if possible. If we
2036 // can, we want to change it to use a post-incremented version of its
2037 // induction variable, to allow coalescing the live ranges for the IV into
2038 // one register value.
2040 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2043 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2044 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2047 // Search IVUsesByStride to find Cond's IVUse if there is one.
2048 IVStrideUse *CondUse = 0;
2049 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2050 if (!FindIVUserForCond(Cond, CondUse))
2053 // If the trip count is computed in terms of a max (due to ScalarEvolution
2054 // being unable to find a sufficient guard, for example), change the loop
2055 // comparison to use SLT or ULT instead of NE.
2056 // One consequence of doing this now is that it disrupts the count-down
2057 // optimization. That's not always a bad thing though, because in such
2058 // cases it may still be worthwhile to avoid a max.
2059 Cond = OptimizeMax(Cond, CondUse);
2061 // If this exiting block dominates the latch block, it may also use
2062 // the post-inc value if it won't be shared with other uses.
2063 // Check for dominance.
2064 if (!DT.dominates(ExitingBlock, LatchBlock))
2067 // Conservatively avoid trying to use the post-inc value in non-latch
2068 // exits if there may be pre-inc users in intervening blocks.
2069 if (LatchBlock != ExitingBlock)
2070 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2071 // Test if the use is reachable from the exiting block. This dominator
2072 // query is a conservative approximation of reachability.
2073 if (&*UI != CondUse &&
2074 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2075 // Conservatively assume there may be reuse if the quotient of their
2076 // strides could be a legal scale.
2077 const SCEV *A = IU.getStride(*CondUse, L);
2078 const SCEV *B = IU.getStride(*UI, L);
2079 if (!A || !B) continue;
2080 if (SE.getTypeSizeInBits(A->getType()) !=
2081 SE.getTypeSizeInBits(B->getType())) {
2082 if (SE.getTypeSizeInBits(A->getType()) >
2083 SE.getTypeSizeInBits(B->getType()))
2084 B = SE.getSignExtendExpr(B, A->getType());
2086 A = SE.getSignExtendExpr(A, B->getType());
2088 if (const SCEVConstant *D =
2089 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2090 const ConstantInt *C = D->getValue();
2091 // Stride of one or negative one can have reuse with non-addresses.
2092 if (C->isOne() || C->isAllOnesValue())
2093 goto decline_post_inc;
2094 // Avoid weird situations.
2095 if (C->getValue().getMinSignedBits() >= 64 ||
2096 C->getValue().isMinSignedValue())
2097 goto decline_post_inc;
2098 // Check for possible scaled-address reuse.
2099 Type *AccessTy = getAccessType(UI->getUser());
2100 int64_t Scale = C->getSExtValue();
2101 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2103 /*HasBaseReg=*/ false, Scale))
2104 goto decline_post_inc;
2106 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2108 /*HasBaseReg=*/ false, Scale))
2109 goto decline_post_inc;
2113 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2116 // It's possible for the setcc instruction to be anywhere in the loop, and
2117 // possible for it to have multiple users. If it is not immediately before
2118 // the exiting block branch, move it.
2119 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2120 if (Cond->hasOneUse()) {
2121 Cond->moveBefore(TermBr);
2123 // Clone the terminating condition and insert into the loopend.
2124 ICmpInst *OldCond = Cond;
2125 Cond = cast<ICmpInst>(Cond->clone());
2126 Cond->setName(L->getHeader()->getName() + ".termcond");
2127 ExitingBlock->getInstList().insert(TermBr, Cond);
2129 // Clone the IVUse, as the old use still exists!
2130 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2131 TermBr->replaceUsesOfWith(OldCond, Cond);
2135 // If we get to here, we know that we can transform the setcc instruction to
2136 // use the post-incremented version of the IV, allowing us to coalesce the
2137 // live ranges for the IV correctly.
2138 CondUse->transformToPostInc(L);
2141 PostIncs.insert(Cond);
2145 // Determine an insertion point for the loop induction variable increment. It
2146 // must dominate all the post-inc comparisons we just set up, and it must
2147 // dominate the loop latch edge.
2148 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2149 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2150 E = PostIncs.end(); I != E; ++I) {
2152 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2154 if (BB == (*I)->getParent())
2155 IVIncInsertPos = *I;
2156 else if (BB != IVIncInsertPos->getParent())
2157 IVIncInsertPos = BB->getTerminator();
2161 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2162 /// at the given offset and other details. If so, update the use and
2165 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2166 LSRUse::KindType Kind, Type *AccessTy) {
2167 int64_t NewMinOffset = LU.MinOffset;
2168 int64_t NewMaxOffset = LU.MaxOffset;
2169 Type *NewAccessTy = AccessTy;
2171 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2172 // something conservative, however this can pessimize in the case that one of
2173 // the uses will have all its uses outside the loop, for example.
2174 if (LU.Kind != Kind)
2176 // Conservatively assume HasBaseReg is true for now.
2177 if (NewOffset < LU.MinOffset) {
2178 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2179 LU.MaxOffset - NewOffset, HasBaseReg))
2181 NewMinOffset = NewOffset;
2182 } else if (NewOffset > LU.MaxOffset) {
2183 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2184 NewOffset - LU.MinOffset, HasBaseReg))
2186 NewMaxOffset = NewOffset;
2188 // Check for a mismatched access type, and fall back conservatively as needed.
2189 // TODO: Be less conservative when the type is similar and can use the same
2190 // addressing modes.
2191 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2192 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2195 LU.MinOffset = NewMinOffset;
2196 LU.MaxOffset = NewMaxOffset;
2197 LU.AccessTy = NewAccessTy;
2198 if (NewOffset != LU.Offsets.back())
2199 LU.Offsets.push_back(NewOffset);
2203 /// getUse - Return an LSRUse index and an offset value for a fixup which
2204 /// needs the given expression, with the given kind and optional access type.
2205 /// Either reuse an existing use or create a new one, as needed.
2206 std::pair<size_t, int64_t>
2207 LSRInstance::getUse(const SCEV *&Expr,
2208 LSRUse::KindType Kind, Type *AccessTy) {
2209 const SCEV *Copy = Expr;
2210 int64_t Offset = ExtractImmediate(Expr, SE);
2212 // Basic uses can't accept any offset, for example.
2213 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2214 Offset, /*HasBaseReg=*/ true)) {
2219 std::pair<UseMapTy::iterator, bool> P =
2220 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2222 // A use already existed with this base.
2223 size_t LUIdx = P.first->second;
2224 LSRUse &LU = Uses[LUIdx];
2225 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2227 return std::make_pair(LUIdx, Offset);
2230 // Create a new use.
2231 size_t LUIdx = Uses.size();
2232 P.first->second = LUIdx;
2233 Uses.push_back(LSRUse(Kind, AccessTy));
2234 LSRUse &LU = Uses[LUIdx];
2236 // We don't need to track redundant offsets, but we don't need to go out
2237 // of our way here to avoid them.
2238 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2239 LU.Offsets.push_back(Offset);
2241 LU.MinOffset = Offset;
2242 LU.MaxOffset = Offset;
2243 return std::make_pair(LUIdx, Offset);
2246 /// DeleteUse - Delete the given use from the Uses list.
2247 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2248 if (&LU != &Uses.back())
2249 std::swap(LU, Uses.back());
2253 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2256 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2257 /// a formula that has the same registers as the given formula.
2259 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2260 const LSRUse &OrigLU) {
2261 // Search all uses for the formula. This could be more clever.
2262 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2263 LSRUse &LU = Uses[LUIdx];
2264 // Check whether this use is close enough to OrigLU, to see whether it's
2265 // worthwhile looking through its formulae.
2266 // Ignore ICmpZero uses because they may contain formulae generated by
2267 // GenerateICmpZeroScales, in which case adding fixup offsets may
2269 if (&LU != &OrigLU &&
2270 LU.Kind != LSRUse::ICmpZero &&
2271 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2272 LU.WidestFixupType == OrigLU.WidestFixupType &&
2273 LU.HasFormulaWithSameRegs(OrigF)) {
2274 // Scan through this use's formulae.
2275 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2276 E = LU.Formulae.end(); I != E; ++I) {
2277 const Formula &F = *I;
2278 // Check to see if this formula has the same registers and symbols
2280 if (F.BaseRegs == OrigF.BaseRegs &&
2281 F.ScaledReg == OrigF.ScaledReg &&
2282 F.BaseGV == OrigF.BaseGV &&
2283 F.Scale == OrigF.Scale &&
2284 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2285 if (F.BaseOffset == 0)
2287 // This is the formula where all the registers and symbols matched;
2288 // there aren't going to be any others. Since we declined it, we
2289 // can skip the rest of the formulae and proceed to the next LSRUse.
2296 // Nothing looked good.
2300 void LSRInstance::CollectInterestingTypesAndFactors() {
2301 SmallSetVector<const SCEV *, 4> Strides;
2303 // Collect interesting types and strides.
2304 SmallVector<const SCEV *, 4> Worklist;
2305 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2306 const SCEV *Expr = IU.getExpr(*UI);
2308 // Collect interesting types.
2309 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2311 // Add strides for mentioned loops.
2312 Worklist.push_back(Expr);
2314 const SCEV *S = Worklist.pop_back_val();
2315 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2316 if (AR->getLoop() == L)
2317 Strides.insert(AR->getStepRecurrence(SE));
2318 Worklist.push_back(AR->getStart());
2319 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2320 Worklist.append(Add->op_begin(), Add->op_end());
2322 } while (!Worklist.empty());
2325 // Compute interesting factors from the set of interesting strides.
2326 for (SmallSetVector<const SCEV *, 4>::const_iterator
2327 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2328 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2329 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2330 const SCEV *OldStride = *I;
2331 const SCEV *NewStride = *NewStrideIter;
2333 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2334 SE.getTypeSizeInBits(NewStride->getType())) {
2335 if (SE.getTypeSizeInBits(OldStride->getType()) >
2336 SE.getTypeSizeInBits(NewStride->getType()))
2337 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2339 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2341 if (const SCEVConstant *Factor =
2342 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2344 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2345 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2346 } else if (const SCEVConstant *Factor =
2347 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2350 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2351 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2355 // If all uses use the same type, don't bother looking for truncation-based
2357 if (Types.size() == 1)
2360 DEBUG(print_factors_and_types(dbgs()));
2363 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2364 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2365 /// Instructions to IVStrideUses, we could partially skip this.
2366 static User::op_iterator
2367 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2368 Loop *L, ScalarEvolution &SE) {
2369 for(; OI != OE; ++OI) {
2370 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2371 if (!SE.isSCEVable(Oper->getType()))
2374 if (const SCEVAddRecExpr *AR =
2375 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2376 if (AR->getLoop() == L)
2384 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2385 /// operands, so wrap it in a convenient helper.
2386 static Value *getWideOperand(Value *Oper) {
2387 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2388 return Trunc->getOperand(0);
2392 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2394 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2395 Type *LType = LVal->getType();
2396 Type *RType = RVal->getType();
2397 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2400 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2401 /// NULL for any constant. Returning the expression itself is
2402 /// conservative. Returning a deeper subexpression is more precise and valid as
2403 /// long as it isn't less complex than another subexpression. For expressions
2404 /// involving multiple unscaled values, we need to return the pointer-type
2405 /// SCEVUnknown. This avoids forming chains across objects, such as:
2406 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2408 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2409 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2410 static const SCEV *getExprBase(const SCEV *S) {
2411 switch (S->getSCEVType()) {
2412 default: // uncluding scUnknown.
2417 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2419 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2421 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2423 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2424 // there's nothing more complex.
2425 // FIXME: not sure if we want to recognize negation.
2426 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2427 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2428 E(Add->op_begin()); I != E; ++I) {
2429 const SCEV *SubExpr = *I;
2430 if (SubExpr->getSCEVType() == scAddExpr)
2431 return getExprBase(SubExpr);
2433 if (SubExpr->getSCEVType() != scMulExpr)
2436 return S; // all operands are scaled, be conservative.
2439 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2443 /// Return true if the chain increment is profitable to expand into a loop
2444 /// invariant value, which may require its own register. A profitable chain
2445 /// increment will be an offset relative to the same base. We allow such offsets
2446 /// to potentially be used as chain increment as long as it's not obviously
2447 /// expensive to expand using real instructions.
2448 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2449 const SCEV *IncExpr,
2450 ScalarEvolution &SE) {
2451 // Aggressively form chains when -stress-ivchain.
2455 // Do not replace a constant offset from IV head with a nonconstant IV
2457 if (!isa<SCEVConstant>(IncExpr)) {
2458 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2459 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2463 SmallPtrSet<const SCEV*, 8> Processed;
2464 return !isHighCostExpansion(IncExpr, Processed, SE);
2467 /// Return true if the number of registers needed for the chain is estimated to
2468 /// be less than the number required for the individual IV users. First prohibit
2469 /// any IV users that keep the IV live across increments (the Users set should
2470 /// be empty). Next count the number and type of increments in the chain.
2472 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2473 /// effectively use postinc addressing modes. Only consider it profitable it the
2474 /// increments can be computed in fewer registers when chained.
2476 /// TODO: Consider IVInc free if it's already used in another chains.
2478 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2479 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2483 if (!Chain.hasIncs())
2486 if (!Users.empty()) {
2487 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2488 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2489 E = Users.end(); I != E; ++I) {
2490 dbgs() << " " << **I << "\n";
2494 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2496 // The chain itself may require a register, so intialize cost to 1.
2499 // A complete chain likely eliminates the need for keeping the original IV in
2500 // a register. LSR does not currently know how to form a complete chain unless
2501 // the header phi already exists.
2502 if (isa<PHINode>(Chain.tailUserInst())
2503 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2506 const SCEV *LastIncExpr = 0;
2507 unsigned NumConstIncrements = 0;
2508 unsigned NumVarIncrements = 0;
2509 unsigned NumReusedIncrements = 0;
2510 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2513 if (I->IncExpr->isZero())
2516 // Incrementing by zero or some constant is neutral. We assume constants can
2517 // be folded into an addressing mode or an add's immediate operand.
2518 if (isa<SCEVConstant>(I->IncExpr)) {
2519 ++NumConstIncrements;
2523 if (I->IncExpr == LastIncExpr)
2524 ++NumReusedIncrements;
2528 LastIncExpr = I->IncExpr;
2530 // An IV chain with a single increment is handled by LSR's postinc
2531 // uses. However, a chain with multiple increments requires keeping the IV's
2532 // value live longer than it needs to be if chained.
2533 if (NumConstIncrements > 1)
2536 // Materializing increment expressions in the preheader that didn't exist in
2537 // the original code may cost a register. For example, sign-extended array
2538 // indices can produce ridiculous increments like this:
2539 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2540 cost += NumVarIncrements;
2542 // Reusing variable increments likely saves a register to hold the multiple of
2544 cost -= NumReusedIncrements;
2546 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2552 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2554 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2555 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2556 // When IVs are used as types of varying widths, they are generally converted
2557 // to a wider type with some uses remaining narrow under a (free) trunc.
2558 Value *const NextIV = getWideOperand(IVOper);
2559 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2560 const SCEV *const OperExprBase = getExprBase(OperExpr);
2562 // Visit all existing chains. Check if its IVOper can be computed as a
2563 // profitable loop invariant increment from the last link in the Chain.
2564 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2565 const SCEV *LastIncExpr = 0;
2566 for (; ChainIdx < NChains; ++ChainIdx) {
2567 IVChain &Chain = IVChainVec[ChainIdx];
2569 // Prune the solution space aggressively by checking that both IV operands
2570 // are expressions that operate on the same unscaled SCEVUnknown. This
2571 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2572 // first avoids creating extra SCEV expressions.
2573 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2576 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2577 if (!isCompatibleIVType(PrevIV, NextIV))
2580 // A phi node terminates a chain.
2581 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2584 // The increment must be loop-invariant so it can be kept in a register.
2585 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2586 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2587 if (!SE.isLoopInvariant(IncExpr, L))
2590 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2591 LastIncExpr = IncExpr;
2595 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2596 // bother for phi nodes, because they must be last in the chain.
2597 if (ChainIdx == NChains) {
2598 if (isa<PHINode>(UserInst))
2600 if (NChains >= MaxChains && !StressIVChain) {
2601 DEBUG(dbgs() << "IV Chain Limit\n");
2604 LastIncExpr = OperExpr;
2605 // IVUsers may have skipped over sign/zero extensions. We don't currently
2606 // attempt to form chains involving extensions unless they can be hoisted
2607 // into this loop's AddRec.
2608 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2611 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2613 ChainUsersVec.resize(NChains);
2614 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2615 << ") IV=" << *LastIncExpr << "\n");
2617 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2618 << ") IV+" << *LastIncExpr << "\n");
2619 // Add this IV user to the end of the chain.
2620 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2622 IVChain &Chain = IVChainVec[ChainIdx];
2624 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2625 // This chain's NearUsers become FarUsers.
2626 if (!LastIncExpr->isZero()) {
2627 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2632 // All other uses of IVOperand become near uses of the chain.
2633 // We currently ignore intermediate values within SCEV expressions, assuming
2634 // they will eventually be used be the current chain, or can be computed
2635 // from one of the chain increments. To be more precise we could
2636 // transitively follow its user and only add leaf IV users to the set.
2637 for (User *U : IVOper->users()) {
2638 Instruction *OtherUse = dyn_cast<Instruction>(U);
2641 // Uses in the chain will no longer be uses if the chain is formed.
2642 // Include the head of the chain in this iteration (not Chain.begin()).
2643 IVChain::const_iterator IncIter = Chain.Incs.begin();
2644 IVChain::const_iterator IncEnd = Chain.Incs.end();
2645 for( ; IncIter != IncEnd; ++IncIter) {
2646 if (IncIter->UserInst == OtherUse)
2649 if (IncIter != IncEnd)
2652 if (SE.isSCEVable(OtherUse->getType())
2653 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2654 && IU.isIVUserOrOperand(OtherUse)) {
2657 NearUsers.insert(OtherUse);
2660 // Since this user is part of the chain, it's no longer considered a use
2662 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2665 /// CollectChains - Populate the vector of Chains.
2667 /// This decreases ILP at the architecture level. Targets with ample registers,
2668 /// multiple memory ports, and no register renaming probably don't want
2669 /// this. However, such targets should probably disable LSR altogether.
2671 /// The job of LSR is to make a reasonable choice of induction variables across
2672 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2673 /// ILP *within the loop* if the target wants it.
2675 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2676 /// will not reorder memory operations, it will recognize this as a chain, but
2677 /// will generate redundant IV increments. Ideally this would be corrected later
2678 /// by a smart scheduler:
2684 /// TODO: Walk the entire domtree within this loop, not just the path to the
2685 /// loop latch. This will discover chains on side paths, but requires
2686 /// maintaining multiple copies of the Chains state.
2687 void LSRInstance::CollectChains() {
2688 DEBUG(dbgs() << "Collecting IV Chains.\n");
2689 SmallVector<ChainUsers, 8> ChainUsersVec;
2691 SmallVector<BasicBlock *,8> LatchPath;
2692 BasicBlock *LoopHeader = L->getHeader();
2693 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2694 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2695 LatchPath.push_back(Rung->getBlock());
2697 LatchPath.push_back(LoopHeader);
2699 // Walk the instruction stream from the loop header to the loop latch.
2700 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2701 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2702 BBIter != BBEnd; ++BBIter) {
2703 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2705 // Skip instructions that weren't seen by IVUsers analysis.
2706 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2709 // Ignore users that are part of a SCEV expression. This way we only
2710 // consider leaf IV Users. This effectively rediscovers a portion of
2711 // IVUsers analysis but in program order this time.
2712 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2715 // Remove this instruction from any NearUsers set it may be in.
2716 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2717 ChainIdx < NChains; ++ChainIdx) {
2718 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2720 // Search for operands that can be chained.
2721 SmallPtrSet<Instruction*, 4> UniqueOperands;
2722 User::op_iterator IVOpEnd = I->op_end();
2723 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2724 while (IVOpIter != IVOpEnd) {
2725 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2726 if (UniqueOperands.insert(IVOpInst))
2727 ChainInstruction(I, IVOpInst, ChainUsersVec);
2728 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2730 } // Continue walking down the instructions.
2731 } // Continue walking down the domtree.
2732 // Visit phi backedges to determine if the chain can generate the IV postinc.
2733 for (BasicBlock::iterator I = L->getHeader()->begin();
2734 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2735 if (!SE.isSCEVable(PN->getType()))
2739 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2741 ChainInstruction(PN, IncV, ChainUsersVec);
2743 // Remove any unprofitable chains.
2744 unsigned ChainIdx = 0;
2745 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2746 UsersIdx < NChains; ++UsersIdx) {
2747 if (!isProfitableChain(IVChainVec[UsersIdx],
2748 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2750 // Preserve the chain at UsesIdx.
2751 if (ChainIdx != UsersIdx)
2752 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2753 FinalizeChain(IVChainVec[ChainIdx]);
2756 IVChainVec.resize(ChainIdx);
2759 void LSRInstance::FinalizeChain(IVChain &Chain) {
2760 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2761 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2763 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2765 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2766 User::op_iterator UseI =
2767 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2768 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2769 IVIncSet.insert(UseI);
2773 /// Return true if the IVInc can be folded into an addressing mode.
2774 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2775 Value *Operand, const TargetTransformInfo &TTI) {
2776 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2777 if (!IncConst || !isAddressUse(UserInst, Operand))
2780 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2783 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2784 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2785 getAccessType(UserInst), /*BaseGV=*/ 0,
2786 IncOffset, /*HaseBaseReg=*/ false))
2792 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2793 /// materialize the IV user's operand from the previous IV user's operand.
2794 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2795 SmallVectorImpl<WeakVH> &DeadInsts) {
2796 // Find the new IVOperand for the head of the chain. It may have been replaced
2798 const IVInc &Head = Chain.Incs[0];
2799 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2800 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2801 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2804 while (IVOpIter != IVOpEnd) {
2805 IVSrc = getWideOperand(*IVOpIter);
2807 // If this operand computes the expression that the chain needs, we may use
2808 // it. (Check this after setting IVSrc which is used below.)
2810 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2811 // narrow for the chain, so we can no longer use it. We do allow using a
2812 // wider phi, assuming the LSR checked for free truncation. In that case we
2813 // should already have a truncate on this operand such that
2814 // getSCEV(IVSrc) == IncExpr.
2815 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2816 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2819 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2821 if (IVOpIter == IVOpEnd) {
2822 // Gracefully give up on this chain.
2823 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2827 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2828 Type *IVTy = IVSrc->getType();
2829 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2830 const SCEV *LeftOverExpr = 0;
2831 for (IVChain::const_iterator IncI = Chain.begin(),
2832 IncE = Chain.end(); IncI != IncE; ++IncI) {
2834 Instruction *InsertPt = IncI->UserInst;
2835 if (isa<PHINode>(InsertPt))
2836 InsertPt = L->getLoopLatch()->getTerminator();
2838 // IVOper will replace the current IV User's operand. IVSrc is the IV
2839 // value currently held in a register.
2840 Value *IVOper = IVSrc;
2841 if (!IncI->IncExpr->isZero()) {
2842 // IncExpr was the result of subtraction of two narrow values, so must
2844 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2845 LeftOverExpr = LeftOverExpr ?
2846 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2848 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2849 // Expand the IV increment.
2850 Rewriter.clearPostInc();
2851 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2852 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2853 SE.getUnknown(IncV));
2854 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2856 // If an IV increment can't be folded, use it as the next IV value.
2857 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2859 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2864 Type *OperTy = IncI->IVOperand->getType();
2865 if (IVTy != OperTy) {
2866 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2867 "cannot extend a chained IV");
2868 IRBuilder<> Builder(InsertPt);
2869 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2871 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2872 DeadInsts.push_back(IncI->IVOperand);
2874 // If LSR created a new, wider phi, we may also replace its postinc. We only
2875 // do this if we also found a wide value for the head of the chain.
2876 if (isa<PHINode>(Chain.tailUserInst())) {
2877 for (BasicBlock::iterator I = L->getHeader()->begin();
2878 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2879 if (!isCompatibleIVType(Phi, IVSrc))
2881 Instruction *PostIncV = dyn_cast<Instruction>(
2882 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2883 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2885 Value *IVOper = IVSrc;
2886 Type *PostIncTy = PostIncV->getType();
2887 if (IVTy != PostIncTy) {
2888 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2889 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2890 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2891 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2893 Phi->replaceUsesOfWith(PostIncV, IVOper);
2894 DeadInsts.push_back(PostIncV);
2899 void LSRInstance::CollectFixupsAndInitialFormulae() {
2900 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2901 Instruction *UserInst = UI->getUser();
2902 // Skip IV users that are part of profitable IV Chains.
2903 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2904 UI->getOperandValToReplace());
2905 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2906 if (IVIncSet.count(UseI))
2910 LSRFixup &LF = getNewFixup();
2911 LF.UserInst = UserInst;
2912 LF.OperandValToReplace = UI->getOperandValToReplace();
2913 LF.PostIncLoops = UI->getPostIncLoops();
2915 LSRUse::KindType Kind = LSRUse::Basic;
2917 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2918 Kind = LSRUse::Address;
2919 AccessTy = getAccessType(LF.UserInst);
2922 const SCEV *S = IU.getExpr(*UI);
2924 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2925 // (N - i == 0), and this allows (N - i) to be the expression that we work
2926 // with rather than just N or i, so we can consider the register
2927 // requirements for both N and i at the same time. Limiting this code to
2928 // equality icmps is not a problem because all interesting loops use
2929 // equality icmps, thanks to IndVarSimplify.
2930 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2931 if (CI->isEquality()) {
2932 // Swap the operands if needed to put the OperandValToReplace on the
2933 // left, for consistency.
2934 Value *NV = CI->getOperand(1);
2935 if (NV == LF.OperandValToReplace) {
2936 CI->setOperand(1, CI->getOperand(0));
2937 CI->setOperand(0, NV);
2938 NV = CI->getOperand(1);
2942 // x == y --> x - y == 0
2943 const SCEV *N = SE.getSCEV(NV);
2944 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2945 // S is normalized, so normalize N before folding it into S
2946 // to keep the result normalized.
2947 N = TransformForPostIncUse(Normalize, N, CI, 0,
2948 LF.PostIncLoops, SE, DT);
2949 Kind = LSRUse::ICmpZero;
2950 S = SE.getMinusSCEV(N, S);
2953 // -1 and the negations of all interesting strides (except the negation
2954 // of -1) are now also interesting.
2955 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2956 if (Factors[i] != -1)
2957 Factors.insert(-(uint64_t)Factors[i]);
2961 // Set up the initial formula for this use.
2962 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2964 LF.Offset = P.second;
2965 LSRUse &LU = Uses[LF.LUIdx];
2966 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2967 if (!LU.WidestFixupType ||
2968 SE.getTypeSizeInBits(LU.WidestFixupType) <
2969 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2970 LU.WidestFixupType = LF.OperandValToReplace->getType();
2972 // If this is the first use of this LSRUse, give it a formula.
2973 if (LU.Formulae.empty()) {
2974 InsertInitialFormula(S, LU, LF.LUIdx);
2975 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2979 DEBUG(print_fixups(dbgs()));
2982 /// InsertInitialFormula - Insert a formula for the given expression into
2983 /// the given use, separating out loop-variant portions from loop-invariant
2984 /// and loop-computable portions.
2986 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2987 // Mark uses whose expressions cannot be expanded.
2988 if (!isSafeToExpand(S, SE))
2989 LU.RigidFormula = true;
2992 F.InitialMatch(S, L, SE);
2993 bool Inserted = InsertFormula(LU, LUIdx, F);
2994 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2997 /// InsertSupplementalFormula - Insert a simple single-register formula for
2998 /// the given expression into the given use.
3000 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3001 LSRUse &LU, size_t LUIdx) {
3003 F.BaseRegs.push_back(S);
3004 F.HasBaseReg = true;
3005 bool Inserted = InsertFormula(LU, LUIdx, F);
3006 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3009 /// CountRegisters - Note which registers are used by the given formula,
3010 /// updating RegUses.
3011 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3013 RegUses.CountRegister(F.ScaledReg, LUIdx);
3014 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3015 E = F.BaseRegs.end(); I != E; ++I)
3016 RegUses.CountRegister(*I, LUIdx);
3019 /// InsertFormula - If the given formula has not yet been inserted, add it to
3020 /// the list, and return true. Return false otherwise.
3021 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3022 if (!LU.InsertFormula(F))
3025 CountRegisters(F, LUIdx);
3029 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3030 /// loop-invariant values which we're tracking. These other uses will pin these
3031 /// values in registers, making them less profitable for elimination.
3032 /// TODO: This currently misses non-constant addrec step registers.
3033 /// TODO: Should this give more weight to users inside the loop?
3035 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3036 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3037 SmallPtrSet<const SCEV *, 8> Inserted;
3039 while (!Worklist.empty()) {
3040 const SCEV *S = Worklist.pop_back_val();
3042 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3043 Worklist.append(N->op_begin(), N->op_end());
3044 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3045 Worklist.push_back(C->getOperand());
3046 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3047 Worklist.push_back(D->getLHS());
3048 Worklist.push_back(D->getRHS());
3049 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3050 if (!Inserted.insert(US)) continue;
3051 const Value *V = US->getValue();
3052 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3053 // Look for instructions defined outside the loop.
3054 if (L->contains(Inst)) continue;
3055 } else if (isa<UndefValue>(V))
3056 // Undef doesn't have a live range, so it doesn't matter.
3058 for (const Use &U : V->uses()) {
3059 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3060 // Ignore non-instructions.
3063 // Ignore instructions in other functions (as can happen with
3065 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3067 // Ignore instructions not dominated by the loop.
3068 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3069 UserInst->getParent() :
3070 cast<PHINode>(UserInst)->getIncomingBlock(
3071 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3072 if (!DT.dominates(L->getHeader(), UseBB))
3074 // Ignore uses which are part of other SCEV expressions, to avoid
3075 // analyzing them multiple times.
3076 if (SE.isSCEVable(UserInst->getType())) {
3077 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3078 // If the user is a no-op, look through to its uses.
3079 if (!isa<SCEVUnknown>(UserS))
3083 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3087 // Ignore icmp instructions which are already being analyzed.
3088 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3089 unsigned OtherIdx = !U.getOperandNo();
3090 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3091 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3095 LSRFixup &LF = getNewFixup();
3096 LF.UserInst = const_cast<Instruction *>(UserInst);
3097 LF.OperandValToReplace = U;
3098 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
3100 LF.Offset = P.second;
3101 LSRUse &LU = Uses[LF.LUIdx];
3102 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3103 if (!LU.WidestFixupType ||
3104 SE.getTypeSizeInBits(LU.WidestFixupType) <
3105 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3106 LU.WidestFixupType = LF.OperandValToReplace->getType();
3107 InsertSupplementalFormula(US, LU, LF.LUIdx);
3108 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3115 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3116 /// separate registers. If C is non-null, multiply each subexpression by C.
3118 /// Return remainder expression after factoring the subexpressions captured by
3119 /// Ops. If Ops is complete, return NULL.
3120 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3121 SmallVectorImpl<const SCEV *> &Ops,
3123 ScalarEvolution &SE,
3124 unsigned Depth = 0) {
3125 // Arbitrarily cap recursion to protect compile time.
3129 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3130 // Break out add operands.
3131 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3133 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3135 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3138 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3139 // Split a non-zero base out of an addrec.
3140 if (AR->getStart()->isZero())
3143 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3144 C, Ops, L, SE, Depth+1);
3145 // Split the non-zero AddRec unless it is part of a nested recurrence that
3146 // does not pertain to this loop.
3147 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3148 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3151 if (Remainder != AR->getStart()) {
3153 Remainder = SE.getConstant(AR->getType(), 0);
3154 return SE.getAddRecExpr(Remainder,
3155 AR->getStepRecurrence(SE),
3157 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3160 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3161 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3162 if (Mul->getNumOperands() != 2)
3164 if (const SCEVConstant *Op0 =
3165 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3166 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3167 const SCEV *Remainder =
3168 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3170 Ops.push_back(SE.getMulExpr(C, Remainder));
3177 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3179 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3182 // Arbitrarily cap recursion to protect compile time.
3183 if (Depth >= 3) return;
3185 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3186 const SCEV *BaseReg = Base.BaseRegs[i];
3188 SmallVector<const SCEV *, 8> AddOps;
3189 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3191 AddOps.push_back(Remainder);
3193 if (AddOps.size() == 1) continue;
3195 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3196 JE = AddOps.end(); J != JE; ++J) {
3198 // Loop-variant "unknown" values are uninteresting; we won't be able to
3199 // do anything meaningful with them.
3200 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3203 // Don't pull a constant into a register if the constant could be folded
3204 // into an immediate field.
3205 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3206 LU.AccessTy, *J, Base.getNumRegs() > 1))
3209 // Collect all operands except *J.
3210 SmallVector<const SCEV *, 8> InnerAddOps(
3211 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3212 InnerAddOps.append(std::next(J),
3213 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3215 // Don't leave just a constant behind in a register if the constant could
3216 // be folded into an immediate field.
3217 if (InnerAddOps.size() == 1 &&
3218 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3219 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3222 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3223 if (InnerSum->isZero())
3227 // Add the remaining pieces of the add back into the new formula.
3228 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3230 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3231 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3232 InnerSumSC->getValue()->getZExtValue())) {
3233 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3234 InnerSumSC->getValue()->getZExtValue();
3235 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3237 F.BaseRegs[i] = InnerSum;
3239 // Add J as its own register, or an unfolded immediate.
3240 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3241 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3242 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3243 SC->getValue()->getZExtValue()))
3244 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3245 SC->getValue()->getZExtValue();
3247 F.BaseRegs.push_back(*J);
3249 if (InsertFormula(LU, LUIdx, F))
3250 // If that formula hadn't been seen before, recurse to find more like
3252 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3257 /// GenerateCombinations - Generate a formula consisting of all of the
3258 /// loop-dominating registers added into a single register.
3259 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3261 // This method is only interesting on a plurality of registers.
3262 if (Base.BaseRegs.size() <= 1) return;
3266 SmallVector<const SCEV *, 4> Ops;
3267 for (SmallVectorImpl<const SCEV *>::const_iterator
3268 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3269 const SCEV *BaseReg = *I;
3270 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3271 !SE.hasComputableLoopEvolution(BaseReg, L))
3272 Ops.push_back(BaseReg);
3274 F.BaseRegs.push_back(BaseReg);
3276 if (Ops.size() > 1) {
3277 const SCEV *Sum = SE.getAddExpr(Ops);
3278 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3279 // opportunity to fold something. For now, just ignore such cases
3280 // rather than proceed with zero in a register.
3281 if (!Sum->isZero()) {
3282 F.BaseRegs.push_back(Sum);
3283 (void)InsertFormula(LU, LUIdx, F);
3288 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3289 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3291 // We can't add a symbolic offset if the address already contains one.
3292 if (Base.BaseGV) return;
3294 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3295 const SCEV *G = Base.BaseRegs[i];
3296 GlobalValue *GV = ExtractSymbol(G, SE);
3297 if (G->isZero() || !GV)
3301 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3304 (void)InsertFormula(LU, LUIdx, F);
3308 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3309 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3311 // TODO: For now, just add the min and max offset, because it usually isn't
3312 // worthwhile looking at everything inbetween.
3313 SmallVector<int64_t, 2> Worklist;
3314 Worklist.push_back(LU.MinOffset);
3315 if (LU.MaxOffset != LU.MinOffset)
3316 Worklist.push_back(LU.MaxOffset);
3318 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3319 const SCEV *G = Base.BaseRegs[i];
3321 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3322 E = Worklist.end(); I != E; ++I) {
3324 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3325 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3327 // Add the offset to the base register.
3328 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3329 // If it cancelled out, drop the base register, otherwise update it.
3330 if (NewG->isZero()) {
3331 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3332 F.BaseRegs.pop_back();
3334 F.BaseRegs[i] = NewG;
3336 (void)InsertFormula(LU, LUIdx, F);
3340 int64_t Imm = ExtractImmediate(G, SE);
3341 if (G->isZero() || Imm == 0)
3344 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3345 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3348 (void)InsertFormula(LU, LUIdx, F);
3352 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3353 /// the comparison. For example, x == y -> x*c == y*c.
3354 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3356 if (LU.Kind != LSRUse::ICmpZero) return;
3358 // Determine the integer type for the base formula.
3359 Type *IntTy = Base.getType();
3361 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3363 // Don't do this if there is more than one offset.
3364 if (LU.MinOffset != LU.MaxOffset) return;
3366 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3368 // Check each interesting stride.
3369 for (SmallSetVector<int64_t, 8>::const_iterator
3370 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3371 int64_t Factor = *I;
3373 // Check that the multiplication doesn't overflow.
3374 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3376 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3377 if (NewBaseOffset / Factor != Base.BaseOffset)
3379 // If the offset will be truncated at this use, check that it is in bounds.
3380 if (!IntTy->isPointerTy() &&
3381 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3384 // Check that multiplying with the use offset doesn't overflow.
3385 int64_t Offset = LU.MinOffset;
3386 if (Offset == INT64_MIN && Factor == -1)
3388 Offset = (uint64_t)Offset * Factor;
3389 if (Offset / Factor != LU.MinOffset)
3391 // If the offset will be truncated at this use, check that it is in bounds.
3392 if (!IntTy->isPointerTy() &&
3393 !ConstantInt::isValueValidForType(IntTy, Offset))
3397 F.BaseOffset = NewBaseOffset;
3399 // Check that this scale is legal.
3400 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3403 // Compensate for the use having MinOffset built into it.
3404 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3406 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3408 // Check that multiplying with each base register doesn't overflow.
3409 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3410 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3411 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3415 // Check that multiplying with the scaled register doesn't overflow.
3417 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3418 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3422 // Check that multiplying with the unfolded offset doesn't overflow.
3423 if (F.UnfoldedOffset != 0) {
3424 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3426 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3427 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3429 // If the offset will be truncated, check that it is in bounds.
3430 if (!IntTy->isPointerTy() &&
3431 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3435 // If we make it here and it's legal, add it.
3436 (void)InsertFormula(LU, LUIdx, F);
3441 /// GenerateScales - Generate stride factor reuse formulae by making use of
3442 /// scaled-offset address modes, for example.
3443 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3444 // Determine the integer type for the base formula.
3445 Type *IntTy = Base.getType();
3448 // If this Formula already has a scaled register, we can't add another one.
3449 if (Base.Scale != 0) return;
3451 // Check each interesting stride.
3452 for (SmallSetVector<int64_t, 8>::const_iterator
3453 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3454 int64_t Factor = *I;
3456 Base.Scale = Factor;
3457 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3458 // Check whether this scale is going to be legal.
3459 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3461 // As a special-case, handle special out-of-loop Basic users specially.
3462 // TODO: Reconsider this special case.
3463 if (LU.Kind == LSRUse::Basic &&
3464 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3465 LU.AccessTy, Base) &&
3466 LU.AllFixupsOutsideLoop)
3467 LU.Kind = LSRUse::Special;
3471 // For an ICmpZero, negating a solitary base register won't lead to
3473 if (LU.Kind == LSRUse::ICmpZero &&
3474 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3476 // For each addrec base reg, apply the scale, if possible.
3477 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3478 if (const SCEVAddRecExpr *AR =
3479 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3480 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3481 if (FactorS->isZero())
3483 // Divide out the factor, ignoring high bits, since we'll be
3484 // scaling the value back up in the end.
3485 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3486 // TODO: This could be optimized to avoid all the copying.
3488 F.ScaledReg = Quotient;
3489 F.DeleteBaseReg(F.BaseRegs[i]);
3490 (void)InsertFormula(LU, LUIdx, F);
3496 /// GenerateTruncates - Generate reuse formulae from different IV types.
3497 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3498 // Don't bother truncating symbolic values.
3499 if (Base.BaseGV) return;
3501 // Determine the integer type for the base formula.
3502 Type *DstTy = Base.getType();
3504 DstTy = SE.getEffectiveSCEVType(DstTy);
3506 for (SmallSetVector<Type *, 4>::const_iterator
3507 I = Types.begin(), E = Types.end(); I != E; ++I) {
3509 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3512 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3513 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3514 JE = F.BaseRegs.end(); J != JE; ++J)
3515 *J = SE.getAnyExtendExpr(*J, SrcTy);
3517 // TODO: This assumes we've done basic processing on all uses and
3518 // have an idea what the register usage is.
3519 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3522 (void)InsertFormula(LU, LUIdx, F);
3529 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3530 /// defer modifications so that the search phase doesn't have to worry about
3531 /// the data structures moving underneath it.
3535 const SCEV *OrigReg;
3537 WorkItem(size_t LI, int64_t I, const SCEV *R)
3538 : LUIdx(LI), Imm(I), OrigReg(R) {}
3540 void print(raw_ostream &OS) const;
3546 void WorkItem::print(raw_ostream &OS) const {
3547 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3548 << " , add offset " << Imm;
3551 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3552 void WorkItem::dump() const {
3553 print(errs()); errs() << '\n';
3557 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3558 /// distance apart and try to form reuse opportunities between them.
3559 void LSRInstance::GenerateCrossUseConstantOffsets() {
3560 // Group the registers by their value without any added constant offset.
3561 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3562 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3564 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3565 SmallVector<const SCEV *, 8> Sequence;
3566 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3568 const SCEV *Reg = *I;
3569 int64_t Imm = ExtractImmediate(Reg, SE);
3570 std::pair<RegMapTy::iterator, bool> Pair =
3571 Map.insert(std::make_pair(Reg, ImmMapTy()));
3573 Sequence.push_back(Reg);
3574 Pair.first->second.insert(std::make_pair(Imm, *I));
3575 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3578 // Now examine each set of registers with the same base value. Build up
3579 // a list of work to do and do the work in a separate step so that we're
3580 // not adding formulae and register counts while we're searching.
3581 SmallVector<WorkItem, 32> WorkItems;
3582 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3583 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3584 E = Sequence.end(); I != E; ++I) {
3585 const SCEV *Reg = *I;
3586 const ImmMapTy &Imms = Map.find(Reg)->second;
3588 // It's not worthwhile looking for reuse if there's only one offset.
3589 if (Imms.size() == 1)
3592 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3593 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3595 dbgs() << ' ' << J->first;
3598 // Examine each offset.
3599 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3601 const SCEV *OrigReg = J->second;
3603 int64_t JImm = J->first;
3604 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3606 if (!isa<SCEVConstant>(OrigReg) &&
3607 UsedByIndicesMap[Reg].count() == 1) {
3608 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3612 // Conservatively examine offsets between this orig reg a few selected
3614 ImmMapTy::const_iterator OtherImms[] = {
3615 Imms.begin(), std::prev(Imms.end()),
3616 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3619 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3620 ImmMapTy::const_iterator M = OtherImms[i];
3621 if (M == J || M == JE) continue;
3623 // Compute the difference between the two.
3624 int64_t Imm = (uint64_t)JImm - M->first;
3625 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3626 LUIdx = UsedByIndices.find_next(LUIdx))
3627 // Make a memo of this use, offset, and register tuple.
3628 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3629 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3636 UsedByIndicesMap.clear();
3637 UniqueItems.clear();
3639 // Now iterate through the worklist and add new formulae.
3640 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3641 E = WorkItems.end(); I != E; ++I) {
3642 const WorkItem &WI = *I;
3643 size_t LUIdx = WI.LUIdx;
3644 LSRUse &LU = Uses[LUIdx];
3645 int64_t Imm = WI.Imm;
3646 const SCEV *OrigReg = WI.OrigReg;
3648 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3649 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3650 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3652 // TODO: Use a more targeted data structure.
3653 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3654 const Formula &F = LU.Formulae[L];
3655 // Use the immediate in the scaled register.
3656 if (F.ScaledReg == OrigReg) {
3657 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3658 // Don't create 50 + reg(-50).
3659 if (F.referencesReg(SE.getSCEV(
3660 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3663 NewF.BaseOffset = Offset;
3664 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3667 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3669 // If the new scale is a constant in a register, and adding the constant
3670 // value to the immediate would produce a value closer to zero than the
3671 // immediate itself, then the formula isn't worthwhile.
3672 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3673 if (C->getValue()->isNegative() !=
3674 (NewF.BaseOffset < 0) &&
3675 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3676 .ule(abs64(NewF.BaseOffset)))
3680 (void)InsertFormula(LU, LUIdx, NewF);
3682 // Use the immediate in a base register.
3683 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3684 const SCEV *BaseReg = F.BaseRegs[N];
3685 if (BaseReg != OrigReg)
3688 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3689 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3690 LU.Kind, LU.AccessTy, NewF)) {
3691 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3694 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3696 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3698 // If the new formula has a constant in a register, and adding the
3699 // constant value to the immediate would produce a value closer to
3700 // zero than the immediate itself, then the formula isn't worthwhile.
3701 for (SmallVectorImpl<const SCEV *>::const_iterator
3702 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3704 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3705 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3706 abs64(NewF.BaseOffset)) &&
3707 (C->getValue()->getValue() +
3708 NewF.BaseOffset).countTrailingZeros() >=
3709 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3713 (void)InsertFormula(LU, LUIdx, NewF);
3722 /// GenerateAllReuseFormulae - Generate formulae for each use.
3724 LSRInstance::GenerateAllReuseFormulae() {
3725 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3726 // queries are more precise.
3727 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3728 LSRUse &LU = Uses[LUIdx];
3729 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3730 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3731 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3732 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3734 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3735 LSRUse &LU = Uses[LUIdx];
3736 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3737 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3738 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3739 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3740 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3741 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3742 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3743 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3745 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3746 LSRUse &LU = Uses[LUIdx];
3747 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3748 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3751 GenerateCrossUseConstantOffsets();
3753 DEBUG(dbgs() << "\n"
3754 "After generating reuse formulae:\n";
3755 print_uses(dbgs()));
3758 /// If there are multiple formulae with the same set of registers used
3759 /// by other uses, pick the best one and delete the others.
3760 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3761 DenseSet<const SCEV *> VisitedRegs;
3762 SmallPtrSet<const SCEV *, 16> Regs;
3763 SmallPtrSet<const SCEV *, 16> LoserRegs;
3765 bool ChangedFormulae = false;
3768 // Collect the best formula for each unique set of shared registers. This
3769 // is reset for each use.
3770 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3772 BestFormulaeTy BestFormulae;
3774 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3775 LSRUse &LU = Uses[LUIdx];
3776 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3779 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3780 FIdx != NumForms; ++FIdx) {
3781 Formula &F = LU.Formulae[FIdx];
3783 // Some formulas are instant losers. For example, they may depend on
3784 // nonexistent AddRecs from other loops. These need to be filtered
3785 // immediately, otherwise heuristics could choose them over others leading
3786 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3787 // avoids the need to recompute this information across formulae using the
3788 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3789 // the corresponding bad register from the Regs set.
3792 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3794 if (CostF.isLoser()) {
3795 // During initial formula generation, undesirable formulae are generated
3796 // by uses within other loops that have some non-trivial address mode or
3797 // use the postinc form of the IV. LSR needs to provide these formulae
3798 // as the basis of rediscovering the desired formula that uses an AddRec
3799 // corresponding to the existing phi. Once all formulae have been
3800 // generated, these initial losers may be pruned.
3801 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3805 SmallVector<const SCEV *, 4> Key;
3806 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3807 JE = F.BaseRegs.end(); J != JE; ++J) {
3808 const SCEV *Reg = *J;
3809 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3813 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3814 Key.push_back(F.ScaledReg);
3815 // Unstable sort by host order ok, because this is only used for
3817 std::sort(Key.begin(), Key.end());
3819 std::pair<BestFormulaeTy::const_iterator, bool> P =
3820 BestFormulae.insert(std::make_pair(Key, FIdx));
3824 Formula &Best = LU.Formulae[P.first->second];
3828 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3830 if (CostF < CostBest)
3832 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3834 " in favor of formula "; Best.print(dbgs());
3838 ChangedFormulae = true;
3840 LU.DeleteFormula(F);
3846 // Now that we've filtered out some formulae, recompute the Regs set.
3848 LU.RecomputeRegs(LUIdx, RegUses);
3850 // Reset this to prepare for the next use.
3851 BestFormulae.clear();
3854 DEBUG(if (ChangedFormulae) {
3856 "After filtering out undesirable candidates:\n";
3861 // This is a rough guess that seems to work fairly well.
3862 static const size_t ComplexityLimit = UINT16_MAX;
3864 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3865 /// solutions the solver might have to consider. It almost never considers
3866 /// this many solutions because it prune the search space, but the pruning
3867 /// isn't always sufficient.
3868 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3870 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3871 E = Uses.end(); I != E; ++I) {
3872 size_t FSize = I->Formulae.size();
3873 if (FSize >= ComplexityLimit) {
3874 Power = ComplexityLimit;
3878 if (Power >= ComplexityLimit)
3884 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3885 /// of the registers of another formula, it won't help reduce register
3886 /// pressure (though it may not necessarily hurt register pressure); remove
3887 /// it to simplify the system.
3888 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3889 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3890 DEBUG(dbgs() << "The search space is too complex.\n");
3892 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3893 "which use a superset of registers used by other "
3896 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3897 LSRUse &LU = Uses[LUIdx];
3899 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3900 Formula &F = LU.Formulae[i];
3901 // Look for a formula with a constant or GV in a register. If the use
3902 // also has a formula with that same value in an immediate field,
3903 // delete the one that uses a register.
3904 for (SmallVectorImpl<const SCEV *>::const_iterator
3905 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3906 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3908 NewF.BaseOffset += C->getValue()->getSExtValue();
3909 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3910 (I - F.BaseRegs.begin()));
3911 if (LU.HasFormulaWithSameRegs(NewF)) {
3912 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3913 LU.DeleteFormula(F);
3919 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3920 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3924 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3925 (I - F.BaseRegs.begin()));
3926 if (LU.HasFormulaWithSameRegs(NewF)) {
3927 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3929 LU.DeleteFormula(F);
3940 LU.RecomputeRegs(LUIdx, RegUses);
3943 DEBUG(dbgs() << "After pre-selection:\n";
3944 print_uses(dbgs()));
3948 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3949 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3951 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3952 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
3955 DEBUG(dbgs() << "The search space is too complex.\n"
3956 "Narrowing the search space by assuming that uses separated "
3957 "by a constant offset will use the same registers.\n");
3959 // This is especially useful for unrolled loops.
3961 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3962 LSRUse &LU = Uses[LUIdx];
3963 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3964 E = LU.Formulae.end(); I != E; ++I) {
3965 const Formula &F = *I;
3966 if (F.BaseOffset == 0 || F.Scale != 0)
3969 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
3973 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
3974 LU.Kind, LU.AccessTy))
3977 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
3979 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3981 // Update the relocs to reference the new use.
3982 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3983 E = Fixups.end(); I != E; ++I) {
3984 LSRFixup &Fixup = *I;
3985 if (Fixup.LUIdx == LUIdx) {
3986 Fixup.LUIdx = LUThatHas - &Uses.front();
3987 Fixup.Offset += F.BaseOffset;
3988 // Add the new offset to LUThatHas' offset list.
3989 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3990 LUThatHas->Offsets.push_back(Fixup.Offset);
3991 if (Fixup.Offset > LUThatHas->MaxOffset)
3992 LUThatHas->MaxOffset = Fixup.Offset;
3993 if (Fixup.Offset < LUThatHas->MinOffset)
3994 LUThatHas->MinOffset = Fixup.Offset;
3996 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
3998 if (Fixup.LUIdx == NumUses-1)
3999 Fixup.LUIdx = LUIdx;
4002 // Delete formulae from the new use which are no longer legal.
4004 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4005 Formula &F = LUThatHas->Formulae[i];
4006 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4007 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4008 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4010 LUThatHas->DeleteFormula(F);
4018 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4020 // Delete the old use.
4021 DeleteUse(LU, LUIdx);
4028 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4031 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4032 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4033 /// we've done more filtering, as it may be able to find more formulae to
4035 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4036 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4037 DEBUG(dbgs() << "The search space is too complex.\n");
4039 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4040 "undesirable dedicated registers.\n");
4042 FilterOutUndesirableDedicatedRegisters();
4044 DEBUG(dbgs() << "After pre-selection:\n";
4045 print_uses(dbgs()));
4049 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4050 /// to be profitable, and then in any use which has any reference to that
4051 /// register, delete all formulae which do not reference that register.
4052 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4053 // With all other options exhausted, loop until the system is simple
4054 // enough to handle.
4055 SmallPtrSet<const SCEV *, 4> Taken;
4056 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4057 // Ok, we have too many of formulae on our hands to conveniently handle.
4058 // Use a rough heuristic to thin out the list.
4059 DEBUG(dbgs() << "The search space is too complex.\n");
4061 // Pick the register which is used by the most LSRUses, which is likely
4062 // to be a good reuse register candidate.
4063 const SCEV *Best = 0;
4064 unsigned BestNum = 0;
4065 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4067 const SCEV *Reg = *I;
4068 if (Taken.count(Reg))
4073 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4074 if (Count > BestNum) {
4081 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4082 << " will yield profitable reuse.\n");
4085 // In any use with formulae which references this register, delete formulae
4086 // which don't reference it.
4087 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4088 LSRUse &LU = Uses[LUIdx];
4089 if (!LU.Regs.count(Best)) continue;
4092 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4093 Formula &F = LU.Formulae[i];
4094 if (!F.referencesReg(Best)) {
4095 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4096 LU.DeleteFormula(F);
4100 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4106 LU.RecomputeRegs(LUIdx, RegUses);
4109 DEBUG(dbgs() << "After pre-selection:\n";
4110 print_uses(dbgs()));
4114 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4115 /// formulae to choose from, use some rough heuristics to prune down the number
4116 /// of formulae. This keeps the main solver from taking an extraordinary amount
4117 /// of time in some worst-case scenarios.
4118 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4119 NarrowSearchSpaceByDetectingSupersets();
4120 NarrowSearchSpaceByCollapsingUnrolledCode();
4121 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4122 NarrowSearchSpaceByPickingWinnerRegs();
4125 /// SolveRecurse - This is the recursive solver.
4126 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4128 SmallVectorImpl<const Formula *> &Workspace,
4129 const Cost &CurCost,
4130 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4131 DenseSet<const SCEV *> &VisitedRegs) const {
4134 // - use more aggressive filtering
4135 // - sort the formula so that the most profitable solutions are found first
4136 // - sort the uses too
4138 // - don't compute a cost, and then compare. compare while computing a cost
4140 // - track register sets with SmallBitVector
4142 const LSRUse &LU = Uses[Workspace.size()];
4144 // If this use references any register that's already a part of the
4145 // in-progress solution, consider it a requirement that a formula must
4146 // reference that register in order to be considered. This prunes out
4147 // unprofitable searching.
4148 SmallSetVector<const SCEV *, 4> ReqRegs;
4149 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4150 E = CurRegs.end(); I != E; ++I)
4151 if (LU.Regs.count(*I))
4154 SmallPtrSet<const SCEV *, 16> NewRegs;
4156 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4157 E = LU.Formulae.end(); I != E; ++I) {
4158 const Formula &F = *I;
4160 // Ignore formulae which do not use any of the required registers.
4161 bool SatisfiedReqReg = true;
4162 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4163 JE = ReqRegs.end(); J != JE; ++J) {
4164 const SCEV *Reg = *J;
4165 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4166 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4168 SatisfiedReqReg = false;
4172 if (!SatisfiedReqReg) {
4173 // If none of the formulae satisfied the required registers, then we could
4174 // clear ReqRegs and try again. Currently, we simply give up in this case.
4178 // Evaluate the cost of the current formula. If it's already worse than
4179 // the current best, prune the search at that point.
4182 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4184 if (NewCost < SolutionCost) {
4185 Workspace.push_back(&F);
4186 if (Workspace.size() != Uses.size()) {
4187 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4188 NewRegs, VisitedRegs);
4189 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4190 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4192 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4193 dbgs() << ".\n Regs:";
4194 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4195 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4196 dbgs() << ' ' << **I;
4199 SolutionCost = NewCost;
4200 Solution = Workspace;
4202 Workspace.pop_back();
4207 /// Solve - Choose one formula from each use. Return the results in the given
4208 /// Solution vector.
4209 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4210 SmallVector<const Formula *, 8> Workspace;
4212 SolutionCost.Lose();
4214 SmallPtrSet<const SCEV *, 16> CurRegs;
4215 DenseSet<const SCEV *> VisitedRegs;
4216 Workspace.reserve(Uses.size());
4218 // SolveRecurse does all the work.
4219 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4220 CurRegs, VisitedRegs);
4221 if (Solution.empty()) {
4222 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4226 // Ok, we've now made all our decisions.
4227 DEBUG(dbgs() << "\n"
4228 "The chosen solution requires "; SolutionCost.print(dbgs());
4230 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4232 Uses[i].print(dbgs());
4235 Solution[i]->print(dbgs());
4239 assert(Solution.size() == Uses.size() && "Malformed solution!");
4242 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4243 /// the dominator tree far as we can go while still being dominated by the
4244 /// input positions. This helps canonicalize the insert position, which
4245 /// encourages sharing.
4246 BasicBlock::iterator
4247 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4248 const SmallVectorImpl<Instruction *> &Inputs)
4251 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4252 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4255 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4256 if (!Rung) return IP;
4257 Rung = Rung->getIDom();
4258 if (!Rung) return IP;
4259 IDom = Rung->getBlock();
4261 // Don't climb into a loop though.
4262 const Loop *IDomLoop = LI.getLoopFor(IDom);
4263 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4264 if (IDomDepth <= IPLoopDepth &&
4265 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4269 bool AllDominate = true;
4270 Instruction *BetterPos = 0;
4271 Instruction *Tentative = IDom->getTerminator();
4272 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4273 E = Inputs.end(); I != E; ++I) {
4274 Instruction *Inst = *I;
4275 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4276 AllDominate = false;
4279 // Attempt to find an insert position in the middle of the block,
4280 // instead of at the end, so that it can be used for other expansions.
4281 if (IDom == Inst->getParent() &&
4282 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4283 BetterPos = std::next(BasicBlock::iterator(Inst));
4296 /// AdjustInsertPositionForExpand - Determine an input position which will be
4297 /// dominated by the operands and which will dominate the result.
4298 BasicBlock::iterator
4299 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4302 SCEVExpander &Rewriter) const {
4303 // Collect some instructions which must be dominated by the
4304 // expanding replacement. These must be dominated by any operands that
4305 // will be required in the expansion.
4306 SmallVector<Instruction *, 4> Inputs;
4307 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4308 Inputs.push_back(I);
4309 if (LU.Kind == LSRUse::ICmpZero)
4310 if (Instruction *I =
4311 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4312 Inputs.push_back(I);
4313 if (LF.PostIncLoops.count(L)) {
4314 if (LF.isUseFullyOutsideLoop(L))
4315 Inputs.push_back(L->getLoopLatch()->getTerminator());
4317 Inputs.push_back(IVIncInsertPos);
4319 // The expansion must also be dominated by the increment positions of any
4320 // loops it for which it is using post-inc mode.
4321 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4322 E = LF.PostIncLoops.end(); I != E; ++I) {
4323 const Loop *PIL = *I;
4324 if (PIL == L) continue;
4326 // Be dominated by the loop exit.
4327 SmallVector<BasicBlock *, 4> ExitingBlocks;
4328 PIL->getExitingBlocks(ExitingBlocks);
4329 if (!ExitingBlocks.empty()) {
4330 BasicBlock *BB = ExitingBlocks[0];
4331 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4332 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4333 Inputs.push_back(BB->getTerminator());
4337 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4338 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4339 "Insertion point must be a normal instruction");
4341 // Then, climb up the immediate dominator tree as far as we can go while
4342 // still being dominated by the input positions.
4343 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4345 // Don't insert instructions before PHI nodes.
4346 while (isa<PHINode>(IP)) ++IP;
4348 // Ignore landingpad instructions.
4349 while (isa<LandingPadInst>(IP)) ++IP;
4351 // Ignore debug intrinsics.
4352 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4354 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4355 // IP consistent across expansions and allows the previously inserted
4356 // instructions to be reused by subsequent expansion.
4357 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4362 /// Expand - Emit instructions for the leading candidate expression for this
4363 /// LSRUse (this is called "expanding").
4364 Value *LSRInstance::Expand(const LSRFixup &LF,
4366 BasicBlock::iterator IP,
4367 SCEVExpander &Rewriter,
4368 SmallVectorImpl<WeakVH> &DeadInsts) const {
4369 const LSRUse &LU = Uses[LF.LUIdx];
4370 if (LU.RigidFormula)
4371 return LF.OperandValToReplace;
4373 // Determine an input position which will be dominated by the operands and
4374 // which will dominate the result.
4375 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4377 // Inform the Rewriter if we have a post-increment use, so that it can
4378 // perform an advantageous expansion.
4379 Rewriter.setPostInc(LF.PostIncLoops);
4381 // This is the type that the user actually needs.
4382 Type *OpTy = LF.OperandValToReplace->getType();
4383 // This will be the type that we'll initially expand to.
4384 Type *Ty = F.getType();
4386 // No type known; just expand directly to the ultimate type.
4388 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4389 // Expand directly to the ultimate type if it's the right size.
4391 // This is the type to do integer arithmetic in.
4392 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4394 // Build up a list of operands to add together to form the full base.
4395 SmallVector<const SCEV *, 8> Ops;
4397 // Expand the BaseRegs portion.
4398 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4399 E = F.BaseRegs.end(); I != E; ++I) {
4400 const SCEV *Reg = *I;
4401 assert(!Reg->isZero() && "Zero allocated in a base register!");
4403 // If we're expanding for a post-inc user, make the post-inc adjustment.
4404 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4405 Reg = TransformForPostIncUse(Denormalize, Reg,
4406 LF.UserInst, LF.OperandValToReplace,
4409 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4412 // Expand the ScaledReg portion.
4413 Value *ICmpScaledV = 0;
4415 const SCEV *ScaledS = F.ScaledReg;
4417 // If we're expanding for a post-inc user, make the post-inc adjustment.
4418 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4419 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4420 LF.UserInst, LF.OperandValToReplace,
4423 if (LU.Kind == LSRUse::ICmpZero) {
4424 // An interesting way of "folding" with an icmp is to use a negated
4425 // scale, which we'll implement by inserting it into the other operand
4427 assert(F.Scale == -1 &&
4428 "The only scale supported by ICmpZero uses is -1!");
4429 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4431 // Otherwise just expand the scaled register and an explicit scale,
4432 // which is expected to be matched as part of the address.
4434 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4435 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4436 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4438 Ops.push_back(SE.getUnknown(FullV));
4440 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4441 ScaledS = SE.getMulExpr(ScaledS,
4442 SE.getConstant(ScaledS->getType(), F.Scale));
4443 Ops.push_back(ScaledS);
4447 // Expand the GV portion.
4449 // Flush the operand list to suppress SCEVExpander hoisting.
4451 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4453 Ops.push_back(SE.getUnknown(FullV));
4455 Ops.push_back(SE.getUnknown(F.BaseGV));
4458 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4459 // unfolded offsets. LSR assumes they both live next to their uses.
4461 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4463 Ops.push_back(SE.getUnknown(FullV));
4466 // Expand the immediate portion.
4467 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4469 if (LU.Kind == LSRUse::ICmpZero) {
4470 // The other interesting way of "folding" with an ICmpZero is to use a
4471 // negated immediate.
4473 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4475 Ops.push_back(SE.getUnknown(ICmpScaledV));
4476 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4479 // Just add the immediate values. These again are expected to be matched
4480 // as part of the address.
4481 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4485 // Expand the unfolded offset portion.
4486 int64_t UnfoldedOffset = F.UnfoldedOffset;
4487 if (UnfoldedOffset != 0) {
4488 // Just add the immediate values.
4489 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4493 // Emit instructions summing all the operands.
4494 const SCEV *FullS = Ops.empty() ?
4495 SE.getConstant(IntTy, 0) :
4497 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4499 // We're done expanding now, so reset the rewriter.
4500 Rewriter.clearPostInc();
4502 // An ICmpZero Formula represents an ICmp which we're handling as a
4503 // comparison against zero. Now that we've expanded an expression for that
4504 // form, update the ICmp's other operand.
4505 if (LU.Kind == LSRUse::ICmpZero) {
4506 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4507 DeadInsts.push_back(CI->getOperand(1));
4508 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4509 "a scale at the same time!");
4510 if (F.Scale == -1) {
4511 if (ICmpScaledV->getType() != OpTy) {
4513 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4515 ICmpScaledV, OpTy, "tmp", CI);
4518 CI->setOperand(1, ICmpScaledV);
4520 assert(F.Scale == 0 &&
4521 "ICmp does not support folding a global value and "
4522 "a scale at the same time!");
4523 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4525 if (C->getType() != OpTy)
4526 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4530 CI->setOperand(1, C);
4537 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4538 /// of their operands effectively happens in their predecessor blocks, so the
4539 /// expression may need to be expanded in multiple places.
4540 void LSRInstance::RewriteForPHI(PHINode *PN,
4543 SCEVExpander &Rewriter,
4544 SmallVectorImpl<WeakVH> &DeadInsts,
4546 DenseMap<BasicBlock *, Value *> Inserted;
4547 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4548 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4549 BasicBlock *BB = PN->getIncomingBlock(i);
4551 // If this is a critical edge, split the edge so that we do not insert
4552 // the code on all predecessor/successor paths. We do this unless this
4553 // is the canonical backedge for this loop, which complicates post-inc
4555 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4556 !isa<IndirectBrInst>(BB->getTerminator())) {
4557 BasicBlock *Parent = PN->getParent();
4558 Loop *PNLoop = LI.getLoopFor(Parent);
4559 if (!PNLoop || Parent != PNLoop->getHeader()) {
4560 // Split the critical edge.
4561 BasicBlock *NewBB = 0;
4562 if (!Parent->isLandingPad()) {
4563 NewBB = SplitCriticalEdge(BB, Parent, P,
4564 /*MergeIdenticalEdges=*/true,
4565 /*DontDeleteUselessPhis=*/true);
4567 SmallVector<BasicBlock*, 2> NewBBs;
4568 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4571 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4572 // phi predecessors are identical. The simple thing to do is skip
4573 // splitting in this case rather than complicate the API.
4575 // If PN is outside of the loop and BB is in the loop, we want to
4576 // move the block to be immediately before the PHI block, not
4577 // immediately after BB.
4578 if (L->contains(BB) && !L->contains(PN))
4579 NewBB->moveBefore(PN->getParent());
4581 // Splitting the edge can reduce the number of PHI entries we have.
4582 e = PN->getNumIncomingValues();
4584 i = PN->getBasicBlockIndex(BB);
4589 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4590 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4592 PN->setIncomingValue(i, Pair.first->second);
4594 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4596 // If this is reuse-by-noop-cast, insert the noop cast.
4597 Type *OpTy = LF.OperandValToReplace->getType();
4598 if (FullV->getType() != OpTy)
4600 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4602 FullV, LF.OperandValToReplace->getType(),
4603 "tmp", BB->getTerminator());
4605 PN->setIncomingValue(i, FullV);
4606 Pair.first->second = FullV;
4611 /// Rewrite - Emit instructions for the leading candidate expression for this
4612 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4613 /// the newly expanded value.
4614 void LSRInstance::Rewrite(const LSRFixup &LF,
4616 SCEVExpander &Rewriter,
4617 SmallVectorImpl<WeakVH> &DeadInsts,
4619 // First, find an insertion point that dominates UserInst. For PHI nodes,
4620 // find the nearest block which dominates all the relevant uses.
4621 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4622 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4624 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4626 // If this is reuse-by-noop-cast, insert the noop cast.
4627 Type *OpTy = LF.OperandValToReplace->getType();
4628 if (FullV->getType() != OpTy) {
4630 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4631 FullV, OpTy, "tmp", LF.UserInst);
4635 // Update the user. ICmpZero is handled specially here (for now) because
4636 // Expand may have updated one of the operands of the icmp already, and
4637 // its new value may happen to be equal to LF.OperandValToReplace, in
4638 // which case doing replaceUsesOfWith leads to replacing both operands
4639 // with the same value. TODO: Reorganize this.
4640 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4641 LF.UserInst->setOperand(0, FullV);
4643 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4646 DeadInsts.push_back(LF.OperandValToReplace);
4649 /// ImplementSolution - Rewrite all the fixup locations with new values,
4650 /// following the chosen solution.
4652 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4654 // Keep track of instructions we may have made dead, so that
4655 // we can remove them after we are done working.
4656 SmallVector<WeakVH, 16> DeadInsts;
4658 SCEVExpander Rewriter(SE, "lsr");
4660 Rewriter.setDebugType(DEBUG_TYPE);
4662 Rewriter.disableCanonicalMode();
4663 Rewriter.enableLSRMode();
4664 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4666 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4667 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4668 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4669 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4670 Rewriter.setChainedPhi(PN);
4673 // Expand the new value definitions and update the users.
4674 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4675 E = Fixups.end(); I != E; ++I) {
4676 const LSRFixup &Fixup = *I;
4678 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4683 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4684 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4685 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4688 // Clean up after ourselves. This must be done before deleting any
4692 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4695 LSRInstance::LSRInstance(Loop *L, Pass *P)
4696 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4697 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4698 LI(P->getAnalysis<LoopInfo>()),
4699 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4701 // If LoopSimplify form is not available, stay out of trouble.
4702 if (!L->isLoopSimplifyForm())
4705 // If there's no interesting work to be done, bail early.
4706 if (IU.empty()) return;
4708 // If there's too much analysis to be done, bail early. We won't be able to
4709 // model the problem anyway.
4710 unsigned NumUsers = 0;
4711 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4712 if (++NumUsers > MaxIVUsers) {
4713 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4720 // All dominating loops must have preheaders, or SCEVExpander may not be able
4721 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4723 // IVUsers analysis should only create users that are dominated by simple loop
4724 // headers. Since this loop should dominate all of its users, its user list
4725 // should be empty if this loop itself is not within a simple loop nest.
4726 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4727 Rung; Rung = Rung->getIDom()) {
4728 BasicBlock *BB = Rung->getBlock();
4729 const Loop *DomLoop = LI.getLoopFor(BB);
4730 if (DomLoop && DomLoop->getHeader() == BB) {
4731 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4736 DEBUG(dbgs() << "\nLSR on loop ";
4737 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4740 // First, perform some low-level loop optimizations.
4742 OptimizeLoopTermCond();
4744 // If loop preparation eliminates all interesting IV users, bail.
4745 if (IU.empty()) return;
4747 // Skip nested loops until we can model them better with formulae.
4749 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4753 // Start collecting data and preparing for the solver.
4755 CollectInterestingTypesAndFactors();
4756 CollectFixupsAndInitialFormulae();
4757 CollectLoopInvariantFixupsAndFormulae();
4759 assert(!Uses.empty() && "IVUsers reported at least one use");
4760 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4761 print_uses(dbgs()));
4763 // Now use the reuse data to generate a bunch of interesting ways
4764 // to formulate the values needed for the uses.
4765 GenerateAllReuseFormulae();
4767 FilterOutUndesirableDedicatedRegisters();
4768 NarrowSearchSpaceUsingHeuristics();
4770 SmallVector<const Formula *, 8> Solution;
4773 // Release memory that is no longer needed.
4778 if (Solution.empty())
4782 // Formulae should be legal.
4783 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4785 const LSRUse &LU = *I;
4786 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4787 JE = LU.Formulae.end();
4789 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4790 *J) && "Illegal formula generated!");
4794 // Now that we've decided what we want, make it so.
4795 ImplementSolution(Solution, P);
4798 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4799 if (Factors.empty() && Types.empty()) return;
4801 OS << "LSR has identified the following interesting factors and types: ";
4804 for (SmallSetVector<int64_t, 8>::const_iterator
4805 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4806 if (!First) OS << ", ";
4811 for (SmallSetVector<Type *, 4>::const_iterator
4812 I = Types.begin(), E = Types.end(); I != E; ++I) {
4813 if (!First) OS << ", ";
4815 OS << '(' << **I << ')';
4820 void LSRInstance::print_fixups(raw_ostream &OS) const {
4821 OS << "LSR is examining the following fixup sites:\n";
4822 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4823 E = Fixups.end(); I != E; ++I) {
4830 void LSRInstance::print_uses(raw_ostream &OS) const {
4831 OS << "LSR is examining the following uses:\n";
4832 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4833 E = Uses.end(); I != E; ++I) {
4834 const LSRUse &LU = *I;
4838 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4839 JE = LU.Formulae.end(); J != JE; ++J) {
4847 void LSRInstance::print(raw_ostream &OS) const {
4848 print_factors_and_types(OS);
4853 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4854 void LSRInstance::dump() const {
4855 print(errs()); errs() << '\n';
4861 class LoopStrengthReduce : public LoopPass {
4863 static char ID; // Pass ID, replacement for typeid
4864 LoopStrengthReduce();
4867 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
4868 void getAnalysisUsage(AnalysisUsage &AU) const override;
4873 char LoopStrengthReduce::ID = 0;
4874 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4875 "Loop Strength Reduction", false, false)
4876 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4877 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4878 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4879 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4880 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4881 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4882 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4883 "Loop Strength Reduction", false, false)
4886 Pass *llvm::createLoopStrengthReducePass() {
4887 return new LoopStrengthReduce();
4890 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4891 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4894 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4895 // We split critical edges, so we change the CFG. However, we do update
4896 // many analyses if they are around.
4897 AU.addPreservedID(LoopSimplifyID);
4899 AU.addRequired<LoopInfo>();
4900 AU.addPreserved<LoopInfo>();
4901 AU.addRequiredID(LoopSimplifyID);
4902 AU.addRequired<DominatorTreeWrapperPass>();
4903 AU.addPreserved<DominatorTreeWrapperPass>();
4904 AU.addRequired<ScalarEvolution>();
4905 AU.addPreserved<ScalarEvolution>();
4906 // Requiring LoopSimplify a second time here prevents IVUsers from running
4907 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4908 AU.addRequiredID(LoopSimplifyID);
4909 AU.addRequired<IVUsers>();
4910 AU.addPreserved<IVUsers>();
4911 AU.addRequired<TargetTransformInfo>();
4914 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4915 if (skipOptnoneFunction(L))
4918 bool Changed = false;
4920 // Run the main LSR transformation.
4921 Changed |= LSRInstance(L, this).getChanged();
4923 // Remove any extra phis created by processing inner loops.
4924 Changed |= DeleteDeadPHIs(L->getHeader());
4925 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4926 SmallVector<WeakVH, 16> DeadInsts;
4927 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4929 Rewriter.setDebugType(DEBUG_TYPE);
4931 unsigned numFolded = Rewriter.replaceCongruentIVs(
4932 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
4933 &getAnalysis<TargetTransformInfo>());
4936 DeleteTriviallyDeadInstructions(DeadInsts);
4937 DeleteDeadPHIs(L->getHeader());