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/Support/CommandLine.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/ValueHandle.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
76 #include "llvm/Transforms/Utils/Local.h"
80 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
81 /// bail out. This threshold is far beyond the number of users that LSR can
82 /// conceivably solve, so it should not affect generated code, but catches the
83 /// worst cases before LSR burns too much compile time and stack space.
84 static const unsigned MaxIVUsers = 200;
86 // Temporary flag to cleanup congruent phis after LSR phi expansion.
87 // It's currently disabled until we can determine whether it's truly useful or
88 // not. The flag should be removed after the v3.0 release.
89 // This is now needed for ivchains.
90 static cl::opt<bool> EnablePhiElim(
91 "enable-lsr-phielim", cl::Hidden, cl::init(true),
92 cl::desc("Enable LSR phi elimination"));
95 // Stress test IV chain generation.
96 static cl::opt<bool> StressIVChain(
97 "stress-ivchain", cl::Hidden, cl::init(false),
98 cl::desc("Stress test LSR IV chains"));
100 static bool StressIVChain = false;
105 /// RegSortData - This class holds data which is used to order reuse candidates.
108 /// UsedByIndices - This represents the set of LSRUse indices which reference
109 /// a particular register.
110 SmallBitVector UsedByIndices;
114 void print(raw_ostream &OS) const;
120 void RegSortData::print(raw_ostream &OS) const {
121 OS << "[NumUses=" << UsedByIndices.count() << ']';
124 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
125 void RegSortData::dump() const {
126 print(errs()); errs() << '\n';
132 /// RegUseTracker - Map register candidates to information about how they are
134 class RegUseTracker {
135 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
137 RegUsesTy RegUsesMap;
138 SmallVector<const SCEV *, 16> RegSequence;
141 void CountRegister(const SCEV *Reg, size_t LUIdx);
142 void DropRegister(const SCEV *Reg, size_t LUIdx);
143 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
145 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
147 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
151 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
152 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
153 iterator begin() { return RegSequence.begin(); }
154 iterator end() { return RegSequence.end(); }
155 const_iterator begin() const { return RegSequence.begin(); }
156 const_iterator end() const { return RegSequence.end(); }
162 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
163 std::pair<RegUsesTy::iterator, bool> Pair =
164 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
165 RegSortData &RSD = Pair.first->second;
167 RegSequence.push_back(Reg);
168 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
169 RSD.UsedByIndices.set(LUIdx);
173 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
174 RegUsesTy::iterator It = RegUsesMap.find(Reg);
175 assert(It != RegUsesMap.end());
176 RegSortData &RSD = It->second;
177 assert(RSD.UsedByIndices.size() > LUIdx);
178 RSD.UsedByIndices.reset(LUIdx);
182 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
183 assert(LUIdx <= LastLUIdx);
185 // Update RegUses. The data structure is not optimized for this purpose;
186 // we must iterate through it and update each of the bit vectors.
187 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
189 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
190 if (LUIdx < UsedByIndices.size())
191 UsedByIndices[LUIdx] =
192 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
193 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
198 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
199 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
200 if (I == RegUsesMap.end())
202 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
203 int i = UsedByIndices.find_first();
204 if (i == -1) return false;
205 if ((size_t)i != LUIdx) return true;
206 return UsedByIndices.find_next(i) != -1;
209 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
210 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
211 assert(I != RegUsesMap.end() && "Unknown register!");
212 return I->second.UsedByIndices;
215 void RegUseTracker::clear() {
222 /// Formula - This class holds information that describes a formula for
223 /// computing satisfying a use. It may include broken-out immediates and scaled
226 /// Global base address used for complex addressing.
229 /// Base offset for complex addressing.
232 /// Whether any complex addressing has a base register.
235 /// The scale of any complex addressing.
238 /// BaseRegs - The list of "base" registers for this use. When this is
240 SmallVector<const SCEV *, 4> BaseRegs;
242 /// ScaledReg - The 'scaled' register for this use. This should be non-null
243 /// when Scale is not zero.
244 const SCEV *ScaledReg;
246 /// UnfoldedOffset - An additional constant offset which added near the
247 /// use. This requires a temporary register, but the offset itself can
248 /// live in an add immediate field rather than a register.
249 int64_t UnfoldedOffset;
252 : BaseGV(0), BaseOffset(0), HasBaseReg(false), Scale(0), ScaledReg(0),
255 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
257 unsigned getNumRegs() const;
258 Type *getType() const;
260 void DeleteBaseReg(const SCEV *&S);
262 bool referencesReg(const SCEV *S) const;
263 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
264 const RegUseTracker &RegUses) const;
266 void print(raw_ostream &OS) const;
272 /// DoInitialMatch - Recursion helper for InitialMatch.
273 static void DoInitialMatch(const SCEV *S, Loop *L,
274 SmallVectorImpl<const SCEV *> &Good,
275 SmallVectorImpl<const SCEV *> &Bad,
276 ScalarEvolution &SE) {
277 // Collect expressions which properly dominate the loop header.
278 if (SE.properlyDominates(S, L->getHeader())) {
283 // Look at add operands.
284 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
285 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
287 DoInitialMatch(*I, L, Good, Bad, SE);
291 // Look at addrec operands.
292 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
293 if (!AR->getStart()->isZero()) {
294 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
295 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
296 AR->getStepRecurrence(SE),
297 // FIXME: AR->getNoWrapFlags()
298 AR->getLoop(), SCEV::FlagAnyWrap),
303 // Handle a multiplication by -1 (negation) if it didn't fold.
304 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
305 if (Mul->getOperand(0)->isAllOnesValue()) {
306 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
307 const SCEV *NewMul = SE.getMulExpr(Ops);
309 SmallVector<const SCEV *, 4> MyGood;
310 SmallVector<const SCEV *, 4> MyBad;
311 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
312 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
313 SE.getEffectiveSCEVType(NewMul->getType())));
314 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
315 E = MyGood.end(); I != E; ++I)
316 Good.push_back(SE.getMulExpr(NegOne, *I));
317 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
318 E = MyBad.end(); I != E; ++I)
319 Bad.push_back(SE.getMulExpr(NegOne, *I));
323 // Ok, we can't do anything interesting. Just stuff the whole thing into a
324 // register and hope for the best.
328 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
329 /// attempting to keep all loop-invariant and loop-computable values in a
330 /// single base register.
331 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
332 SmallVector<const SCEV *, 4> Good;
333 SmallVector<const SCEV *, 4> Bad;
334 DoInitialMatch(S, L, Good, Bad, SE);
336 const SCEV *Sum = SE.getAddExpr(Good);
338 BaseRegs.push_back(Sum);
342 const SCEV *Sum = SE.getAddExpr(Bad);
344 BaseRegs.push_back(Sum);
349 /// getNumRegs - Return the total number of register operands used by this
350 /// formula. This does not include register uses implied by non-constant
352 unsigned Formula::getNumRegs() const {
353 return !!ScaledReg + BaseRegs.size();
356 /// getType - Return the type of this formula, if it has one, or null
357 /// otherwise. This type is meaningless except for the bit size.
358 Type *Formula::getType() const {
359 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
360 ScaledReg ? ScaledReg->getType() :
361 BaseGV ? BaseGV->getType() :
365 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
366 void Formula::DeleteBaseReg(const SCEV *&S) {
367 if (&S != &BaseRegs.back())
368 std::swap(S, BaseRegs.back());
372 /// referencesReg - Test if this formula references the given register.
373 bool Formula::referencesReg(const SCEV *S) const {
374 return S == ScaledReg ||
375 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
378 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
379 /// which are used by uses other than the use with the given index.
380 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
381 const RegUseTracker &RegUses) const {
383 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
385 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
386 E = BaseRegs.end(); I != E; ++I)
387 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
392 void Formula::print(raw_ostream &OS) const {
395 if (!First) OS << " + "; else First = false;
396 BaseGV->printAsOperand(OS, /*PrintType=*/false);
398 if (BaseOffset != 0) {
399 if (!First) OS << " + "; else First = false;
402 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
403 E = BaseRegs.end(); I != E; ++I) {
404 if (!First) OS << " + "; else First = false;
405 OS << "reg(" << **I << ')';
407 if (HasBaseReg && BaseRegs.empty()) {
408 if (!First) OS << " + "; else First = false;
409 OS << "**error: HasBaseReg**";
410 } else if (!HasBaseReg && !BaseRegs.empty()) {
411 if (!First) OS << " + "; else First = false;
412 OS << "**error: !HasBaseReg**";
415 if (!First) OS << " + "; else First = false;
416 OS << Scale << "*reg(";
423 if (UnfoldedOffset != 0) {
424 if (!First) OS << " + "; else First = false;
425 OS << "imm(" << UnfoldedOffset << ')';
429 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
430 void Formula::dump() const {
431 print(errs()); errs() << '\n';
435 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
436 /// without changing its value.
437 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
439 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
440 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
443 /// isAddSExtable - Return true if the given add can be sign-extended
444 /// without changing its value.
445 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
447 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
448 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
451 /// isMulSExtable - Return true if the given mul can be sign-extended
452 /// without changing its value.
453 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
455 IntegerType::get(SE.getContext(),
456 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
457 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
460 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
461 /// and if the remainder is known to be zero, or null otherwise. If
462 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
463 /// to Y, ignoring that the multiplication may overflow, which is useful when
464 /// the result will be used in a context where the most significant bits are
466 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
468 bool IgnoreSignificantBits = false) {
469 // Handle the trivial case, which works for any SCEV type.
471 return SE.getConstant(LHS->getType(), 1);
473 // Handle a few RHS special cases.
474 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
476 const APInt &RA = RC->getValue()->getValue();
477 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
479 if (RA.isAllOnesValue())
480 return SE.getMulExpr(LHS, RC);
481 // Handle x /s 1 as x.
486 // Check for a division of a constant by a constant.
487 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
490 const APInt &LA = C->getValue()->getValue();
491 const APInt &RA = RC->getValue()->getValue();
492 if (LA.srem(RA) != 0)
494 return SE.getConstant(LA.sdiv(RA));
497 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
498 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
499 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
500 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
501 IgnoreSignificantBits);
503 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
504 IgnoreSignificantBits);
505 if (!Start) return 0;
506 // FlagNW is independent of the start value, step direction, and is
507 // preserved with smaller magnitude steps.
508 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
509 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
514 // Distribute the sdiv over add operands, if the add doesn't overflow.
515 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
516 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
517 SmallVector<const SCEV *, 8> Ops;
518 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
520 const SCEV *Op = getExactSDiv(*I, RHS, SE,
521 IgnoreSignificantBits);
525 return SE.getAddExpr(Ops);
530 // Check for a multiply operand that we can pull RHS out of.
531 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
532 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
533 SmallVector<const SCEV *, 4> Ops;
535 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
539 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
540 IgnoreSignificantBits)) {
546 return Found ? SE.getMulExpr(Ops) : 0;
551 // Otherwise we don't know.
555 /// ExtractImmediate - If S involves the addition of a constant integer value,
556 /// return that integer value, and mutate S to point to a new SCEV with that
558 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
559 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
560 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
561 S = SE.getConstant(C->getType(), 0);
562 return C->getValue()->getSExtValue();
564 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
565 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
566 int64_t Result = ExtractImmediate(NewOps.front(), SE);
568 S = SE.getAddExpr(NewOps);
570 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
571 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
572 int64_t Result = ExtractImmediate(NewOps.front(), SE);
574 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
575 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
582 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
583 /// return that symbol, and mutate S to point to a new SCEV with that
585 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
586 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
587 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
588 S = SE.getConstant(GV->getType(), 0);
591 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
592 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
593 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
595 S = SE.getAddExpr(NewOps);
597 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
598 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
599 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
601 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
602 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
609 /// isAddressUse - Returns true if the specified instruction is using the
610 /// specified value as an address.
611 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
612 bool isAddress = isa<LoadInst>(Inst);
613 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
614 if (SI->getOperand(1) == OperandVal)
616 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
617 // Addressing modes can also be folded into prefetches and a variety
619 switch (II->getIntrinsicID()) {
621 case Intrinsic::prefetch:
622 case Intrinsic::x86_sse_storeu_ps:
623 case Intrinsic::x86_sse2_storeu_pd:
624 case Intrinsic::x86_sse2_storeu_dq:
625 case Intrinsic::x86_sse2_storel_dq:
626 if (II->getArgOperand(0) == OperandVal)
634 /// getAccessType - Return the type of the memory being accessed.
635 static Type *getAccessType(const Instruction *Inst) {
636 Type *AccessTy = Inst->getType();
637 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
638 AccessTy = SI->getOperand(0)->getType();
639 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
640 // Addressing modes can also be folded into prefetches and a variety
642 switch (II->getIntrinsicID()) {
644 case Intrinsic::x86_sse_storeu_ps:
645 case Intrinsic::x86_sse2_storeu_pd:
646 case Intrinsic::x86_sse2_storeu_dq:
647 case Intrinsic::x86_sse2_storel_dq:
648 AccessTy = II->getArgOperand(0)->getType();
653 // All pointers have the same requirements, so canonicalize them to an
654 // arbitrary pointer type to minimize variation.
655 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
656 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
657 PTy->getAddressSpace());
662 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
663 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
664 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
665 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
666 if (SE.isSCEVable(PN->getType()) &&
667 (SE.getEffectiveSCEVType(PN->getType()) ==
668 SE.getEffectiveSCEVType(AR->getType())) &&
669 SE.getSCEV(PN) == AR)
675 /// Check if expanding this expression is likely to incur significant cost. This
676 /// is tricky because SCEV doesn't track which expressions are actually computed
677 /// by the current IR.
679 /// We currently allow expansion of IV increments that involve adds,
680 /// multiplication by constants, and AddRecs from existing phis.
682 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
683 /// obvious multiple of the UDivExpr.
684 static bool isHighCostExpansion(const SCEV *S,
685 SmallPtrSet<const SCEV*, 8> &Processed,
686 ScalarEvolution &SE) {
687 // Zero/One operand expressions
688 switch (S->getSCEVType()) {
693 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
696 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
699 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
703 if (!Processed.insert(S))
706 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
707 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
709 if (isHighCostExpansion(*I, Processed, SE))
715 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
716 if (Mul->getNumOperands() == 2) {
717 // Multiplication by a constant is ok
718 if (isa<SCEVConstant>(Mul->getOperand(0)))
719 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
721 // If we have the value of one operand, check if an existing
722 // multiplication already generates this expression.
723 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
724 Value *UVal = U->getValue();
725 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
727 // If U is a constant, it may be used by a ConstantExpr.
728 Instruction *User = dyn_cast<Instruction>(*UI);
729 if (User && User->getOpcode() == Instruction::Mul
730 && SE.isSCEVable(User->getType())) {
731 return SE.getSCEV(User) == Mul;
738 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
739 if (isExistingPhi(AR, SE))
743 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
747 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
748 /// specified set are trivially dead, delete them and see if this makes any of
749 /// their operands subsequently dead.
751 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
752 bool Changed = false;
754 while (!DeadInsts.empty()) {
755 Value *V = DeadInsts.pop_back_val();
756 Instruction *I = dyn_cast_or_null<Instruction>(V);
758 if (I == 0 || !isInstructionTriviallyDead(I))
761 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
762 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
765 DeadInsts.push_back(U);
768 I->eraseFromParent();
778 // Check if it is legal to fold 2 base registers.
779 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
781 // Get the cost of the scaling factor used in F for LU.
782 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
783 const LSRUse &LU, const Formula &F);
787 /// Cost - This class is used to measure and compare candidate formulae.
789 /// TODO: Some of these could be merged. Also, a lexical ordering
790 /// isn't always optimal.
794 unsigned NumBaseAdds;
801 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
802 SetupCost(0), ScaleCost(0) {}
804 bool operator<(const Cost &Other) const;
809 // Once any of the metrics loses, they must all remain losers.
811 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
812 | ImmCost | SetupCost | ScaleCost) != ~0u)
813 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
814 & ImmCost & SetupCost & ScaleCost) == ~0u);
819 assert(isValid() && "invalid cost");
820 return NumRegs == ~0u;
823 void RateFormula(const TargetTransformInfo &TTI,
825 SmallPtrSet<const SCEV *, 16> &Regs,
826 const DenseSet<const SCEV *> &VisitedRegs,
828 const SmallVectorImpl<int64_t> &Offsets,
829 ScalarEvolution &SE, DominatorTree &DT,
831 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
833 void print(raw_ostream &OS) const;
837 void RateRegister(const SCEV *Reg,
838 SmallPtrSet<const SCEV *, 16> &Regs,
840 ScalarEvolution &SE, DominatorTree &DT);
841 void RatePrimaryRegister(const SCEV *Reg,
842 SmallPtrSet<const SCEV *, 16> &Regs,
844 ScalarEvolution &SE, DominatorTree &DT,
845 SmallPtrSet<const SCEV *, 16> *LoserRegs);
850 /// RateRegister - Tally up interesting quantities from the given register.
851 void Cost::RateRegister(const SCEV *Reg,
852 SmallPtrSet<const SCEV *, 16> &Regs,
854 ScalarEvolution &SE, DominatorTree &DT) {
855 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
856 // If this is an addrec for another loop, don't second-guess its addrec phi
857 // nodes. LSR isn't currently smart enough to reason about more than one
858 // loop at a time. LSR has already run on inner loops, will not run on outer
859 // loops, and cannot be expected to change sibling loops.
860 if (AR->getLoop() != L) {
861 // If the AddRec exists, consider it's register free and leave it alone.
862 if (isExistingPhi(AR, SE))
865 // Otherwise, do not consider this formula at all.
869 AddRecCost += 1; /// TODO: This should be a function of the stride.
871 // Add the step value register, if it needs one.
872 // TODO: The non-affine case isn't precisely modeled here.
873 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
874 if (!Regs.count(AR->getOperand(1))) {
875 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
883 // Rough heuristic; favor registers which don't require extra setup
884 // instructions in the preheader.
885 if (!isa<SCEVUnknown>(Reg) &&
886 !isa<SCEVConstant>(Reg) &&
887 !(isa<SCEVAddRecExpr>(Reg) &&
888 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
889 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
892 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
893 SE.hasComputableLoopEvolution(Reg, L);
896 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
897 /// before, rate it. Optional LoserRegs provides a way to declare any formula
898 /// that refers to one of those regs an instant loser.
899 void Cost::RatePrimaryRegister(const SCEV *Reg,
900 SmallPtrSet<const SCEV *, 16> &Regs,
902 ScalarEvolution &SE, DominatorTree &DT,
903 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
904 if (LoserRegs && LoserRegs->count(Reg)) {
908 if (Regs.insert(Reg)) {
909 RateRegister(Reg, Regs, L, SE, DT);
910 if (LoserRegs && isLoser())
911 LoserRegs->insert(Reg);
915 void Cost::RateFormula(const TargetTransformInfo &TTI,
917 SmallPtrSet<const SCEV *, 16> &Regs,
918 const DenseSet<const SCEV *> &VisitedRegs,
920 const SmallVectorImpl<int64_t> &Offsets,
921 ScalarEvolution &SE, DominatorTree &DT,
923 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
924 // Tally up the registers.
925 if (const SCEV *ScaledReg = F.ScaledReg) {
926 if (VisitedRegs.count(ScaledReg)) {
930 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
934 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
935 E = F.BaseRegs.end(); I != E; ++I) {
936 const SCEV *BaseReg = *I;
937 if (VisitedRegs.count(BaseReg)) {
941 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
946 // Determine how many (unfolded) adds we'll need inside the loop.
947 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
948 if (NumBaseParts > 1)
949 // Do not count the base and a possible second register if the target
950 // allows to fold 2 registers.
951 NumBaseAdds += NumBaseParts - (1 + isLegal2RegAMUse(TTI, LU, F));
953 // Accumulate non-free scaling amounts.
954 ScaleCost += getScalingFactorCost(TTI, LU, F);
956 // Tally up the non-zero immediates.
957 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
958 E = Offsets.end(); I != E; ++I) {
959 int64_t Offset = (uint64_t)*I + F.BaseOffset;
961 ImmCost += 64; // Handle symbolic values conservatively.
962 // TODO: This should probably be the pointer size.
963 else if (Offset != 0)
964 ImmCost += APInt(64, Offset, true).getMinSignedBits();
966 assert(isValid() && "invalid cost");
969 /// Lose - Set this cost to a losing value.
980 /// operator< - Choose the lower cost.
981 bool Cost::operator<(const Cost &Other) const {
982 if (NumRegs != Other.NumRegs)
983 return NumRegs < Other.NumRegs;
984 if (AddRecCost != Other.AddRecCost)
985 return AddRecCost < Other.AddRecCost;
986 if (NumIVMuls != Other.NumIVMuls)
987 return NumIVMuls < Other.NumIVMuls;
988 if (NumBaseAdds != Other.NumBaseAdds)
989 return NumBaseAdds < Other.NumBaseAdds;
990 if (ScaleCost != Other.ScaleCost)
991 return ScaleCost < Other.ScaleCost;
992 if (ImmCost != Other.ImmCost)
993 return ImmCost < Other.ImmCost;
994 if (SetupCost != Other.SetupCost)
995 return SetupCost < Other.SetupCost;
999 void Cost::print(raw_ostream &OS) const {
1000 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
1001 if (AddRecCost != 0)
1002 OS << ", with addrec cost " << AddRecCost;
1004 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
1005 if (NumBaseAdds != 0)
1006 OS << ", plus " << NumBaseAdds << " base add"
1007 << (NumBaseAdds == 1 ? "" : "s");
1009 OS << ", plus " << ScaleCost << " scale cost";
1011 OS << ", plus " << ImmCost << " imm cost";
1013 OS << ", plus " << SetupCost << " setup cost";
1016 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1017 void Cost::dump() const {
1018 print(errs()); errs() << '\n';
1024 /// LSRFixup - An operand value in an instruction which is to be replaced
1025 /// with some equivalent, possibly strength-reduced, replacement.
1027 /// UserInst - The instruction which will be updated.
1028 Instruction *UserInst;
1030 /// OperandValToReplace - The operand of the instruction which will
1031 /// be replaced. The operand may be used more than once; every instance
1032 /// will be replaced.
1033 Value *OperandValToReplace;
1035 /// PostIncLoops - If this user is to use the post-incremented value of an
1036 /// induction variable, this variable is non-null and holds the loop
1037 /// associated with the induction variable.
1038 PostIncLoopSet PostIncLoops;
1040 /// LUIdx - The index of the LSRUse describing the expression which
1041 /// this fixup needs, minus an offset (below).
1044 /// Offset - A constant offset to be added to the LSRUse expression.
1045 /// This allows multiple fixups to share the same LSRUse with different
1046 /// offsets, for example in an unrolled loop.
1049 bool isUseFullyOutsideLoop(const Loop *L) const;
1053 void print(raw_ostream &OS) const;
1059 LSRFixup::LSRFixup()
1060 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1062 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1063 /// value outside of the given loop.
1064 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1065 // PHI nodes use their value in their incoming blocks.
1066 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1067 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1068 if (PN->getIncomingValue(i) == OperandValToReplace &&
1069 L->contains(PN->getIncomingBlock(i)))
1074 return !L->contains(UserInst);
1077 void LSRFixup::print(raw_ostream &OS) const {
1079 // Store is common and interesting enough to be worth special-casing.
1080 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1082 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1083 } else if (UserInst->getType()->isVoidTy())
1084 OS << UserInst->getOpcodeName();
1086 UserInst->printAsOperand(OS, /*PrintType=*/false);
1088 OS << ", OperandValToReplace=";
1089 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1091 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1092 E = PostIncLoops.end(); I != E; ++I) {
1093 OS << ", PostIncLoop=";
1094 (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1097 if (LUIdx != ~size_t(0))
1098 OS << ", LUIdx=" << LUIdx;
1101 OS << ", Offset=" << Offset;
1104 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1105 void LSRFixup::dump() const {
1106 print(errs()); errs() << '\n';
1112 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1113 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1114 struct UniquifierDenseMapInfo {
1115 static SmallVector<const SCEV *, 4> getEmptyKey() {
1116 SmallVector<const SCEV *, 4> V;
1117 V.push_back(reinterpret_cast<const SCEV *>(-1));
1121 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1122 SmallVector<const SCEV *, 4> V;
1123 V.push_back(reinterpret_cast<const SCEV *>(-2));
1127 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1128 unsigned Result = 0;
1129 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1130 E = V.end(); I != E; ++I)
1131 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1135 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1136 const SmallVector<const SCEV *, 4> &RHS) {
1141 /// LSRUse - This class holds the state that LSR keeps for each use in
1142 /// IVUsers, as well as uses invented by LSR itself. It includes information
1143 /// about what kinds of things can be folded into the user, information about
1144 /// the user itself, and information about how the use may be satisfied.
1145 /// TODO: Represent multiple users of the same expression in common?
1147 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1150 /// KindType - An enum for a kind of use, indicating what types of
1151 /// scaled and immediate operands it might support.
1153 Basic, ///< A normal use, with no folding.
1154 Special, ///< A special case of basic, allowing -1 scales.
1155 Address, ///< An address use; folding according to TargetLowering
1156 ICmpZero ///< An equality icmp with both operands folded into one.
1157 // TODO: Add a generic icmp too?
1163 SmallVector<int64_t, 8> Offsets;
1167 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1168 /// LSRUse are outside of the loop, in which case some special-case heuristics
1170 bool AllFixupsOutsideLoop;
1172 /// RigidFormula is set to true to guarantee that this use will be associated
1173 /// with a single formula--the one that initially matched. Some SCEV
1174 /// expressions cannot be expanded. This allows LSR to consider the registers
1175 /// used by those expressions without the need to expand them later after
1176 /// changing the formula.
1179 /// WidestFixupType - This records the widest use type for any fixup using
1180 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1181 /// max fixup widths to be equivalent, because the narrower one may be relying
1182 /// on the implicit truncation to truncate away bogus bits.
1183 Type *WidestFixupType;
1185 /// Formulae - A list of ways to build a value that can satisfy this user.
1186 /// After the list is populated, one of these is selected heuristically and
1187 /// used to formulate a replacement for OperandValToReplace in UserInst.
1188 SmallVector<Formula, 12> Formulae;
1190 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1191 SmallPtrSet<const SCEV *, 4> Regs;
1193 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1194 MinOffset(INT64_MAX),
1195 MaxOffset(INT64_MIN),
1196 AllFixupsOutsideLoop(true),
1197 RigidFormula(false),
1198 WidestFixupType(0) {}
1200 bool HasFormulaWithSameRegs(const Formula &F) const;
1201 bool InsertFormula(const Formula &F);
1202 void DeleteFormula(Formula &F);
1203 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1205 void print(raw_ostream &OS) const;
1211 /// HasFormula - Test whether this use as a formula which has the same
1212 /// registers as the given formula.
1213 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1214 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1215 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1216 // Unstable sort by host order ok, because this is only used for uniquifying.
1217 std::sort(Key.begin(), Key.end());
1218 return Uniquifier.count(Key);
1221 /// InsertFormula - If the given formula has not yet been inserted, add it to
1222 /// the list, and return true. Return false otherwise.
1223 bool LSRUse::InsertFormula(const Formula &F) {
1224 if (!Formulae.empty() && RigidFormula)
1227 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1228 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1229 // Unstable sort by host order ok, because this is only used for uniquifying.
1230 std::sort(Key.begin(), Key.end());
1232 if (!Uniquifier.insert(Key).second)
1235 // Using a register to hold the value of 0 is not profitable.
1236 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1237 "Zero allocated in a scaled register!");
1239 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1240 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1241 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1244 // Add the formula to the list.
1245 Formulae.push_back(F);
1247 // Record registers now being used by this use.
1248 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1253 /// DeleteFormula - Remove the given formula from this use's list.
1254 void LSRUse::DeleteFormula(Formula &F) {
1255 if (&F != &Formulae.back())
1256 std::swap(F, Formulae.back());
1257 Formulae.pop_back();
1260 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1261 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1262 // Now that we've filtered out some formulae, recompute the Regs set.
1263 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1265 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1266 E = Formulae.end(); I != E; ++I) {
1267 const Formula &F = *I;
1268 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1269 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1272 // Update the RegTracker.
1273 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1274 E = OldRegs.end(); I != E; ++I)
1275 if (!Regs.count(*I))
1276 RegUses.DropRegister(*I, LUIdx);
1279 void LSRUse::print(raw_ostream &OS) const {
1280 OS << "LSR Use: Kind=";
1282 case Basic: OS << "Basic"; break;
1283 case Special: OS << "Special"; break;
1284 case ICmpZero: OS << "ICmpZero"; break;
1286 OS << "Address of ";
1287 if (AccessTy->isPointerTy())
1288 OS << "pointer"; // the full pointer type could be really verbose
1293 OS << ", Offsets={";
1294 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1295 E = Offsets.end(); I != E; ++I) {
1297 if (llvm::next(I) != E)
1302 if (AllFixupsOutsideLoop)
1303 OS << ", all-fixups-outside-loop";
1305 if (WidestFixupType)
1306 OS << ", widest fixup type: " << *WidestFixupType;
1309 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1310 void LSRUse::dump() const {
1311 print(errs()); errs() << '\n';
1315 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1316 /// be completely folded into the user instruction at isel time. This includes
1317 /// address-mode folding and special icmp tricks.
1318 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1319 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1320 bool HasBaseReg, int64_t Scale) {
1322 case LSRUse::Address:
1323 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1325 // Otherwise, just guess that reg+reg addressing is legal.
1328 case LSRUse::ICmpZero:
1329 // There's not even a target hook for querying whether it would be legal to
1330 // fold a GV into an ICmp.
1334 // ICmp only has two operands; don't allow more than two non-trivial parts.
1335 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1338 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1339 // putting the scaled register in the other operand of the icmp.
1340 if (Scale != 0 && Scale != -1)
1343 // If we have low-level target information, ask the target if it can fold an
1344 // integer immediate on an icmp.
1345 if (BaseOffset != 0) {
1347 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1348 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1349 // Offs is the ICmp immediate.
1351 // The cast does the right thing with INT64_MIN.
1352 BaseOffset = -(uint64_t)BaseOffset;
1353 return TTI.isLegalICmpImmediate(BaseOffset);
1356 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1360 // Only handle single-register values.
1361 return !BaseGV && Scale == 0 && BaseOffset == 0;
1363 case LSRUse::Special:
1364 // Special case Basic to handle -1 scales.
1365 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1368 llvm_unreachable("Invalid LSRUse Kind!");
1371 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1372 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1373 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1375 // Check for overflow.
1376 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1379 MinOffset = (uint64_t)BaseOffset + MinOffset;
1380 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1383 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1385 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1387 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1390 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1391 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1393 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1394 F.BaseOffset, F.HasBaseReg, F.Scale);
1397 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
1399 // If F is used as an Addressing Mode, it may fold one Base plus one
1400 // scaled register. If the scaled register is nil, do as if another
1401 // element of the base regs is a 1-scaled register.
1402 // This is possible if BaseRegs has at least 2 registers.
1404 // If this is not an address calculation, this is not an addressing mode
1406 if (LU.Kind != LSRUse::Address)
1409 // F is already scaled.
1413 // We need to keep one register for the base and one to scale.
1414 if (F.BaseRegs.size() < 2)
1417 return isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
1418 F.BaseGV, F.BaseOffset, F.HasBaseReg, 1);
1421 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1422 const LSRUse &LU, const Formula &F) {
1425 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1426 LU.AccessTy, F) && "Illegal formula in use.");
1429 case LSRUse::Address: {
1430 // Check the scaling factor cost with both the min and max offsets.
1431 int ScaleCostMinOffset =
1432 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1433 F.BaseOffset + LU.MinOffset,
1434 F.HasBaseReg, F.Scale);
1435 int ScaleCostMaxOffset =
1436 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1437 F.BaseOffset + LU.MaxOffset,
1438 F.HasBaseReg, F.Scale);
1440 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1441 "Legal addressing mode has an illegal cost!");
1442 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1444 case LSRUse::ICmpZero:
1445 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
1446 // Therefore, return 0 in case F.Scale == -1.
1447 return F.Scale != -1;
1450 case LSRUse::Special:
1454 llvm_unreachable("Invalid LSRUse Kind!");
1457 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1458 LSRUse::KindType Kind, Type *AccessTy,
1459 GlobalValue *BaseGV, int64_t BaseOffset,
1461 // Fast-path: zero is always foldable.
1462 if (BaseOffset == 0 && !BaseGV) return true;
1464 // Conservatively, create an address with an immediate and a
1465 // base and a scale.
1466 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1468 // Canonicalize a scale of 1 to a base register if the formula doesn't
1469 // already have a base register.
1470 if (!HasBaseReg && Scale == 1) {
1475 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1478 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1479 ScalarEvolution &SE, int64_t MinOffset,
1480 int64_t MaxOffset, LSRUse::KindType Kind,
1481 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1482 // Fast-path: zero is always foldable.
1483 if (S->isZero()) return true;
1485 // Conservatively, create an address with an immediate and a
1486 // base and a scale.
1487 int64_t BaseOffset = ExtractImmediate(S, SE);
1488 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1490 // If there's anything else involved, it's not foldable.
1491 if (!S->isZero()) return false;
1493 // Fast-path: zero is always foldable.
1494 if (BaseOffset == 0 && !BaseGV) return true;
1496 // Conservatively, create an address with an immediate and a
1497 // base and a scale.
1498 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1500 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1501 BaseOffset, HasBaseReg, Scale);
1506 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1507 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1508 struct UseMapDenseMapInfo {
1509 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1510 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1513 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1514 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1518 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1519 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1520 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1524 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1525 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1530 /// IVInc - An individual increment in a Chain of IV increments.
1531 /// Relate an IV user to an expression that computes the IV it uses from the IV
1532 /// used by the previous link in the Chain.
1534 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1535 /// original IVOperand. The head of the chain's IVOperand is only valid during
1536 /// chain collection, before LSR replaces IV users. During chain generation,
1537 /// IncExpr can be used to find the new IVOperand that computes the same
1540 Instruction *UserInst;
1542 const SCEV *IncExpr;
1544 IVInc(Instruction *U, Value *O, const SCEV *E):
1545 UserInst(U), IVOperand(O), IncExpr(E) {}
1548 // IVChain - The list of IV increments in program order.
1549 // We typically add the head of a chain without finding subsequent links.
1551 SmallVector<IVInc,1> Incs;
1552 const SCEV *ExprBase;
1554 IVChain() : ExprBase(0) {}
1556 IVChain(const IVInc &Head, const SCEV *Base)
1557 : Incs(1, Head), ExprBase(Base) {}
1559 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1561 // begin - return the first increment in the chain.
1562 const_iterator begin() const {
1563 assert(!Incs.empty());
1564 return llvm::next(Incs.begin());
1566 const_iterator end() const {
1570 // hasIncs - Returns true if this chain contains any increments.
1571 bool hasIncs() const { return Incs.size() >= 2; }
1573 // add - Add an IVInc to the end of this chain.
1574 void add(const IVInc &X) { Incs.push_back(X); }
1576 // tailUserInst - Returns the last UserInst in the chain.
1577 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1579 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1581 bool isProfitableIncrement(const SCEV *OperExpr,
1582 const SCEV *IncExpr,
1586 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1587 /// Distinguish between FarUsers that definitely cross IV increments and
1588 /// NearUsers that may be used between IV increments.
1590 SmallPtrSet<Instruction*, 4> FarUsers;
1591 SmallPtrSet<Instruction*, 4> NearUsers;
1594 /// LSRInstance - This class holds state for the main loop strength reduction
1598 ScalarEvolution &SE;
1601 const TargetTransformInfo &TTI;
1605 /// IVIncInsertPos - This is the insert position that the current loop's
1606 /// induction variable increment should be placed. In simple loops, this is
1607 /// the latch block's terminator. But in more complicated cases, this is a
1608 /// position which will dominate all the in-loop post-increment users.
1609 Instruction *IVIncInsertPos;
1611 /// Factors - Interesting factors between use strides.
1612 SmallSetVector<int64_t, 8> Factors;
1614 /// Types - Interesting use types, to facilitate truncation reuse.
1615 SmallSetVector<Type *, 4> Types;
1617 /// Fixups - The list of operands which are to be replaced.
1618 SmallVector<LSRFixup, 16> Fixups;
1620 /// Uses - The list of interesting uses.
1621 SmallVector<LSRUse, 16> Uses;
1623 /// RegUses - Track which uses use which register candidates.
1624 RegUseTracker RegUses;
1626 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1627 // have more than a few IV increment chains in a loop. Missing a Chain falls
1628 // back to normal LSR behavior for those uses.
1629 static const unsigned MaxChains = 8;
1631 /// IVChainVec - IV users can form a chain of IV increments.
1632 SmallVector<IVChain, MaxChains> IVChainVec;
1634 /// IVIncSet - IV users that belong to profitable IVChains.
1635 SmallPtrSet<Use*, MaxChains> IVIncSet;
1637 void OptimizeShadowIV();
1638 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1639 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1640 void OptimizeLoopTermCond();
1642 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1643 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1644 void FinalizeChain(IVChain &Chain);
1645 void CollectChains();
1646 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1647 SmallVectorImpl<WeakVH> &DeadInsts);
1649 void CollectInterestingTypesAndFactors();
1650 void CollectFixupsAndInitialFormulae();
1652 LSRFixup &getNewFixup() {
1653 Fixups.push_back(LSRFixup());
1654 return Fixups.back();
1657 // Support for sharing of LSRUses between LSRFixups.
1658 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1660 UseMapDenseMapInfo> UseMapTy;
1663 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1664 LSRUse::KindType Kind, Type *AccessTy);
1666 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1667 LSRUse::KindType Kind,
1670 void DeleteUse(LSRUse &LU, size_t LUIdx);
1672 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1674 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1675 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1676 void CountRegisters(const Formula &F, size_t LUIdx);
1677 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1679 void CollectLoopInvariantFixupsAndFormulae();
1681 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1682 unsigned Depth = 0);
1683 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1684 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1685 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1686 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1687 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1688 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1689 void GenerateCrossUseConstantOffsets();
1690 void GenerateAllReuseFormulae();
1692 void FilterOutUndesirableDedicatedRegisters();
1694 size_t EstimateSearchSpaceComplexity() const;
1695 void NarrowSearchSpaceByDetectingSupersets();
1696 void NarrowSearchSpaceByCollapsingUnrolledCode();
1697 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1698 void NarrowSearchSpaceByPickingWinnerRegs();
1699 void NarrowSearchSpaceUsingHeuristics();
1701 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1703 SmallVectorImpl<const Formula *> &Workspace,
1704 const Cost &CurCost,
1705 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1706 DenseSet<const SCEV *> &VisitedRegs) const;
1707 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1709 BasicBlock::iterator
1710 HoistInsertPosition(BasicBlock::iterator IP,
1711 const SmallVectorImpl<Instruction *> &Inputs) const;
1712 BasicBlock::iterator
1713 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1716 SCEVExpander &Rewriter) const;
1718 Value *Expand(const LSRFixup &LF,
1720 BasicBlock::iterator IP,
1721 SCEVExpander &Rewriter,
1722 SmallVectorImpl<WeakVH> &DeadInsts) const;
1723 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1725 SCEVExpander &Rewriter,
1726 SmallVectorImpl<WeakVH> &DeadInsts,
1728 void Rewrite(const LSRFixup &LF,
1730 SCEVExpander &Rewriter,
1731 SmallVectorImpl<WeakVH> &DeadInsts,
1733 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1737 LSRInstance(Loop *L, Pass *P);
1739 bool getChanged() const { return Changed; }
1741 void print_factors_and_types(raw_ostream &OS) const;
1742 void print_fixups(raw_ostream &OS) const;
1743 void print_uses(raw_ostream &OS) const;
1744 void print(raw_ostream &OS) const;
1750 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1751 /// inside the loop then try to eliminate the cast operation.
1752 void LSRInstance::OptimizeShadowIV() {
1753 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1754 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1757 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1758 UI != E; /* empty */) {
1759 IVUsers::const_iterator CandidateUI = UI;
1761 Instruction *ShadowUse = CandidateUI->getUser();
1763 bool IsSigned = false;
1765 /* If shadow use is a int->float cast then insert a second IV
1766 to eliminate this cast.
1768 for (unsigned i = 0; i < n; ++i)
1774 for (unsigned i = 0; i < n; ++i, ++d)
1777 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1779 DestTy = UCast->getDestTy();
1781 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1783 DestTy = SCast->getDestTy();
1785 if (!DestTy) continue;
1787 // If target does not support DestTy natively then do not apply
1788 // this transformation.
1789 if (!TTI.isTypeLegal(DestTy)) continue;
1791 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1793 if (PH->getNumIncomingValues() != 2) continue;
1795 Type *SrcTy = PH->getType();
1796 int Mantissa = DestTy->getFPMantissaWidth();
1797 if (Mantissa == -1) continue;
1798 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1801 unsigned Entry, Latch;
1802 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1810 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1811 if (!Init) continue;
1812 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1813 (double)Init->getSExtValue() :
1814 (double)Init->getZExtValue());
1816 BinaryOperator *Incr =
1817 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1818 if (!Incr) continue;
1819 if (Incr->getOpcode() != Instruction::Add
1820 && Incr->getOpcode() != Instruction::Sub)
1823 /* Initialize new IV, double d = 0.0 in above example. */
1825 if (Incr->getOperand(0) == PH)
1826 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1827 else if (Incr->getOperand(1) == PH)
1828 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1834 // Ignore negative constants, as the code below doesn't handle them
1835 // correctly. TODO: Remove this restriction.
1836 if (!C->getValue().isStrictlyPositive()) continue;
1838 /* Add new PHINode. */
1839 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1841 /* create new increment. '++d' in above example. */
1842 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1843 BinaryOperator *NewIncr =
1844 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1845 Instruction::FAdd : Instruction::FSub,
1846 NewPH, CFP, "IV.S.next.", Incr);
1848 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1849 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1851 /* Remove cast operation */
1852 ShadowUse->replaceAllUsesWith(NewPH);
1853 ShadowUse->eraseFromParent();
1859 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1860 /// set the IV user and stride information and return true, otherwise return
1862 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1863 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1864 if (UI->getUser() == Cond) {
1865 // NOTE: we could handle setcc instructions with multiple uses here, but
1866 // InstCombine does it as well for simple uses, it's not clear that it
1867 // occurs enough in real life to handle.
1874 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1875 /// a max computation.
1877 /// This is a narrow solution to a specific, but acute, problem. For loops
1883 /// } while (++i < n);
1885 /// the trip count isn't just 'n', because 'n' might not be positive. And
1886 /// unfortunately this can come up even for loops where the user didn't use
1887 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1888 /// will commonly be lowered like this:
1894 /// } while (++i < n);
1897 /// and then it's possible for subsequent optimization to obscure the if
1898 /// test in such a way that indvars can't find it.
1900 /// When indvars can't find the if test in loops like this, it creates a
1901 /// max expression, which allows it to give the loop a canonical
1902 /// induction variable:
1905 /// max = n < 1 ? 1 : n;
1908 /// } while (++i != max);
1910 /// Canonical induction variables are necessary because the loop passes
1911 /// are designed around them. The most obvious example of this is the
1912 /// LoopInfo analysis, which doesn't remember trip count values. It
1913 /// expects to be able to rediscover the trip count each time it is
1914 /// needed, and it does this using a simple analysis that only succeeds if
1915 /// the loop has a canonical induction variable.
1917 /// However, when it comes time to generate code, the maximum operation
1918 /// can be quite costly, especially if it's inside of an outer loop.
1920 /// This function solves this problem by detecting this type of loop and
1921 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1922 /// the instructions for the maximum computation.
1924 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1925 // Check that the loop matches the pattern we're looking for.
1926 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1927 Cond->getPredicate() != CmpInst::ICMP_NE)
1930 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1931 if (!Sel || !Sel->hasOneUse()) return Cond;
1933 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1934 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1936 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1938 // Add one to the backedge-taken count to get the trip count.
1939 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1940 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1942 // Check for a max calculation that matches the pattern. There's no check
1943 // for ICMP_ULE here because the comparison would be with zero, which
1944 // isn't interesting.
1945 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1946 const SCEVNAryExpr *Max = 0;
1947 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1948 Pred = ICmpInst::ICMP_SLE;
1950 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1951 Pred = ICmpInst::ICMP_SLT;
1953 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1954 Pred = ICmpInst::ICMP_ULT;
1961 // To handle a max with more than two operands, this optimization would
1962 // require additional checking and setup.
1963 if (Max->getNumOperands() != 2)
1966 const SCEV *MaxLHS = Max->getOperand(0);
1967 const SCEV *MaxRHS = Max->getOperand(1);
1969 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1970 // for a comparison with 1. For <= and >=, a comparison with zero.
1972 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1975 // Check the relevant induction variable for conformance to
1977 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1978 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1979 if (!AR || !AR->isAffine() ||
1980 AR->getStart() != One ||
1981 AR->getStepRecurrence(SE) != One)
1984 assert(AR->getLoop() == L &&
1985 "Loop condition operand is an addrec in a different loop!");
1987 // Check the right operand of the select, and remember it, as it will
1988 // be used in the new comparison instruction.
1990 if (ICmpInst::isTrueWhenEqual(Pred)) {
1991 // Look for n+1, and grab n.
1992 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1993 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1994 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1995 NewRHS = BO->getOperand(0);
1996 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1997 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1998 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1999 NewRHS = BO->getOperand(0);
2002 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2003 NewRHS = Sel->getOperand(1);
2004 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2005 NewRHS = Sel->getOperand(2);
2006 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2007 NewRHS = SU->getValue();
2009 // Max doesn't match expected pattern.
2012 // Determine the new comparison opcode. It may be signed or unsigned,
2013 // and the original comparison may be either equality or inequality.
2014 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2015 Pred = CmpInst::getInversePredicate(Pred);
2017 // Ok, everything looks ok to change the condition into an SLT or SGE and
2018 // delete the max calculation.
2020 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2022 // Delete the max calculation instructions.
2023 Cond->replaceAllUsesWith(NewCond);
2024 CondUse->setUser(NewCond);
2025 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2026 Cond->eraseFromParent();
2027 Sel->eraseFromParent();
2028 if (Cmp->use_empty())
2029 Cmp->eraseFromParent();
2033 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2034 /// postinc iv when possible.
2036 LSRInstance::OptimizeLoopTermCond() {
2037 SmallPtrSet<Instruction *, 4> PostIncs;
2039 BasicBlock *LatchBlock = L->getLoopLatch();
2040 SmallVector<BasicBlock*, 8> ExitingBlocks;
2041 L->getExitingBlocks(ExitingBlocks);
2043 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2044 BasicBlock *ExitingBlock = ExitingBlocks[i];
2046 // Get the terminating condition for the loop if possible. If we
2047 // can, we want to change it to use a post-incremented version of its
2048 // induction variable, to allow coalescing the live ranges for the IV into
2049 // one register value.
2051 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2054 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2055 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2058 // Search IVUsesByStride to find Cond's IVUse if there is one.
2059 IVStrideUse *CondUse = 0;
2060 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2061 if (!FindIVUserForCond(Cond, CondUse))
2064 // If the trip count is computed in terms of a max (due to ScalarEvolution
2065 // being unable to find a sufficient guard, for example), change the loop
2066 // comparison to use SLT or ULT instead of NE.
2067 // One consequence of doing this now is that it disrupts the count-down
2068 // optimization. That's not always a bad thing though, because in such
2069 // cases it may still be worthwhile to avoid a max.
2070 Cond = OptimizeMax(Cond, CondUse);
2072 // If this exiting block dominates the latch block, it may also use
2073 // the post-inc value if it won't be shared with other uses.
2074 // Check for dominance.
2075 if (!DT.dominates(ExitingBlock, LatchBlock))
2078 // Conservatively avoid trying to use the post-inc value in non-latch
2079 // exits if there may be pre-inc users in intervening blocks.
2080 if (LatchBlock != ExitingBlock)
2081 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2082 // Test if the use is reachable from the exiting block. This dominator
2083 // query is a conservative approximation of reachability.
2084 if (&*UI != CondUse &&
2085 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2086 // Conservatively assume there may be reuse if the quotient of their
2087 // strides could be a legal scale.
2088 const SCEV *A = IU.getStride(*CondUse, L);
2089 const SCEV *B = IU.getStride(*UI, L);
2090 if (!A || !B) continue;
2091 if (SE.getTypeSizeInBits(A->getType()) !=
2092 SE.getTypeSizeInBits(B->getType())) {
2093 if (SE.getTypeSizeInBits(A->getType()) >
2094 SE.getTypeSizeInBits(B->getType()))
2095 B = SE.getSignExtendExpr(B, A->getType());
2097 A = SE.getSignExtendExpr(A, B->getType());
2099 if (const SCEVConstant *D =
2100 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2101 const ConstantInt *C = D->getValue();
2102 // Stride of one or negative one can have reuse with non-addresses.
2103 if (C->isOne() || C->isAllOnesValue())
2104 goto decline_post_inc;
2105 // Avoid weird situations.
2106 if (C->getValue().getMinSignedBits() >= 64 ||
2107 C->getValue().isMinSignedValue())
2108 goto decline_post_inc;
2109 // Check for possible scaled-address reuse.
2110 Type *AccessTy = getAccessType(UI->getUser());
2111 int64_t Scale = C->getSExtValue();
2112 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2114 /*HasBaseReg=*/ false, Scale))
2115 goto decline_post_inc;
2117 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2119 /*HasBaseReg=*/ false, Scale))
2120 goto decline_post_inc;
2124 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2127 // It's possible for the setcc instruction to be anywhere in the loop, and
2128 // possible for it to have multiple users. If it is not immediately before
2129 // the exiting block branch, move it.
2130 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2131 if (Cond->hasOneUse()) {
2132 Cond->moveBefore(TermBr);
2134 // Clone the terminating condition and insert into the loopend.
2135 ICmpInst *OldCond = Cond;
2136 Cond = cast<ICmpInst>(Cond->clone());
2137 Cond->setName(L->getHeader()->getName() + ".termcond");
2138 ExitingBlock->getInstList().insert(TermBr, Cond);
2140 // Clone the IVUse, as the old use still exists!
2141 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2142 TermBr->replaceUsesOfWith(OldCond, Cond);
2146 // If we get to here, we know that we can transform the setcc instruction to
2147 // use the post-incremented version of the IV, allowing us to coalesce the
2148 // live ranges for the IV correctly.
2149 CondUse->transformToPostInc(L);
2152 PostIncs.insert(Cond);
2156 // Determine an insertion point for the loop induction variable increment. It
2157 // must dominate all the post-inc comparisons we just set up, and it must
2158 // dominate the loop latch edge.
2159 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2160 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2161 E = PostIncs.end(); I != E; ++I) {
2163 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2165 if (BB == (*I)->getParent())
2166 IVIncInsertPos = *I;
2167 else if (BB != IVIncInsertPos->getParent())
2168 IVIncInsertPos = BB->getTerminator();
2172 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2173 /// at the given offset and other details. If so, update the use and
2176 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2177 LSRUse::KindType Kind, Type *AccessTy) {
2178 int64_t NewMinOffset = LU.MinOffset;
2179 int64_t NewMaxOffset = LU.MaxOffset;
2180 Type *NewAccessTy = AccessTy;
2182 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2183 // something conservative, however this can pessimize in the case that one of
2184 // the uses will have all its uses outside the loop, for example.
2185 if (LU.Kind != Kind)
2187 // Conservatively assume HasBaseReg is true for now.
2188 if (NewOffset < LU.MinOffset) {
2189 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2190 LU.MaxOffset - NewOffset, HasBaseReg))
2192 NewMinOffset = NewOffset;
2193 } else if (NewOffset > LU.MaxOffset) {
2194 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2195 NewOffset - LU.MinOffset, HasBaseReg))
2197 NewMaxOffset = NewOffset;
2199 // Check for a mismatched access type, and fall back conservatively as needed.
2200 // TODO: Be less conservative when the type is similar and can use the same
2201 // addressing modes.
2202 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2203 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2206 LU.MinOffset = NewMinOffset;
2207 LU.MaxOffset = NewMaxOffset;
2208 LU.AccessTy = NewAccessTy;
2209 if (NewOffset != LU.Offsets.back())
2210 LU.Offsets.push_back(NewOffset);
2214 /// getUse - Return an LSRUse index and an offset value for a fixup which
2215 /// needs the given expression, with the given kind and optional access type.
2216 /// Either reuse an existing use or create a new one, as needed.
2217 std::pair<size_t, int64_t>
2218 LSRInstance::getUse(const SCEV *&Expr,
2219 LSRUse::KindType Kind, Type *AccessTy) {
2220 const SCEV *Copy = Expr;
2221 int64_t Offset = ExtractImmediate(Expr, SE);
2223 // Basic uses can't accept any offset, for example.
2224 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2225 Offset, /*HasBaseReg=*/ true)) {
2230 std::pair<UseMapTy::iterator, bool> P =
2231 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2233 // A use already existed with this base.
2234 size_t LUIdx = P.first->second;
2235 LSRUse &LU = Uses[LUIdx];
2236 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2238 return std::make_pair(LUIdx, Offset);
2241 // Create a new use.
2242 size_t LUIdx = Uses.size();
2243 P.first->second = LUIdx;
2244 Uses.push_back(LSRUse(Kind, AccessTy));
2245 LSRUse &LU = Uses[LUIdx];
2247 // We don't need to track redundant offsets, but we don't need to go out
2248 // of our way here to avoid them.
2249 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2250 LU.Offsets.push_back(Offset);
2252 LU.MinOffset = Offset;
2253 LU.MaxOffset = Offset;
2254 return std::make_pair(LUIdx, Offset);
2257 /// DeleteUse - Delete the given use from the Uses list.
2258 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2259 if (&LU != &Uses.back())
2260 std::swap(LU, Uses.back());
2264 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2267 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2268 /// a formula that has the same registers as the given formula.
2270 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2271 const LSRUse &OrigLU) {
2272 // Search all uses for the formula. This could be more clever.
2273 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2274 LSRUse &LU = Uses[LUIdx];
2275 // Check whether this use is close enough to OrigLU, to see whether it's
2276 // worthwhile looking through its formulae.
2277 // Ignore ICmpZero uses because they may contain formulae generated by
2278 // GenerateICmpZeroScales, in which case adding fixup offsets may
2280 if (&LU != &OrigLU &&
2281 LU.Kind != LSRUse::ICmpZero &&
2282 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2283 LU.WidestFixupType == OrigLU.WidestFixupType &&
2284 LU.HasFormulaWithSameRegs(OrigF)) {
2285 // Scan through this use's formulae.
2286 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2287 E = LU.Formulae.end(); I != E; ++I) {
2288 const Formula &F = *I;
2289 // Check to see if this formula has the same registers and symbols
2291 if (F.BaseRegs == OrigF.BaseRegs &&
2292 F.ScaledReg == OrigF.ScaledReg &&
2293 F.BaseGV == OrigF.BaseGV &&
2294 F.Scale == OrigF.Scale &&
2295 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2296 if (F.BaseOffset == 0)
2298 // This is the formula where all the registers and symbols matched;
2299 // there aren't going to be any others. Since we declined it, we
2300 // can skip the rest of the formulae and proceed to the next LSRUse.
2307 // Nothing looked good.
2311 void LSRInstance::CollectInterestingTypesAndFactors() {
2312 SmallSetVector<const SCEV *, 4> Strides;
2314 // Collect interesting types and strides.
2315 SmallVector<const SCEV *, 4> Worklist;
2316 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2317 const SCEV *Expr = IU.getExpr(*UI);
2319 // Collect interesting types.
2320 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2322 // Add strides for mentioned loops.
2323 Worklist.push_back(Expr);
2325 const SCEV *S = Worklist.pop_back_val();
2326 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2327 if (AR->getLoop() == L)
2328 Strides.insert(AR->getStepRecurrence(SE));
2329 Worklist.push_back(AR->getStart());
2330 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2331 Worklist.append(Add->op_begin(), Add->op_end());
2333 } while (!Worklist.empty());
2336 // Compute interesting factors from the set of interesting strides.
2337 for (SmallSetVector<const SCEV *, 4>::const_iterator
2338 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2339 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2340 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2341 const SCEV *OldStride = *I;
2342 const SCEV *NewStride = *NewStrideIter;
2344 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2345 SE.getTypeSizeInBits(NewStride->getType())) {
2346 if (SE.getTypeSizeInBits(OldStride->getType()) >
2347 SE.getTypeSizeInBits(NewStride->getType()))
2348 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2350 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2352 if (const SCEVConstant *Factor =
2353 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2355 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2356 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2357 } else if (const SCEVConstant *Factor =
2358 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2361 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2362 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2366 // If all uses use the same type, don't bother looking for truncation-based
2368 if (Types.size() == 1)
2371 DEBUG(print_factors_and_types(dbgs()));
2374 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2375 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2376 /// Instructions to IVStrideUses, we could partially skip this.
2377 static User::op_iterator
2378 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2379 Loop *L, ScalarEvolution &SE) {
2380 for(; OI != OE; ++OI) {
2381 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2382 if (!SE.isSCEVable(Oper->getType()))
2385 if (const SCEVAddRecExpr *AR =
2386 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2387 if (AR->getLoop() == L)
2395 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2396 /// operands, so wrap it in a convenient helper.
2397 static Value *getWideOperand(Value *Oper) {
2398 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2399 return Trunc->getOperand(0);
2403 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2405 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2406 Type *LType = LVal->getType();
2407 Type *RType = RVal->getType();
2408 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2411 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2412 /// NULL for any constant. Returning the expression itself is
2413 /// conservative. Returning a deeper subexpression is more precise and valid as
2414 /// long as it isn't less complex than another subexpression. For expressions
2415 /// involving multiple unscaled values, we need to return the pointer-type
2416 /// SCEVUnknown. This avoids forming chains across objects, such as:
2417 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2419 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2420 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2421 static const SCEV *getExprBase(const SCEV *S) {
2422 switch (S->getSCEVType()) {
2423 default: // uncluding scUnknown.
2428 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2430 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2432 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2434 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2435 // there's nothing more complex.
2436 // FIXME: not sure if we want to recognize negation.
2437 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2438 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2439 E(Add->op_begin()); I != E; ++I) {
2440 const SCEV *SubExpr = *I;
2441 if (SubExpr->getSCEVType() == scAddExpr)
2442 return getExprBase(SubExpr);
2444 if (SubExpr->getSCEVType() != scMulExpr)
2447 return S; // all operands are scaled, be conservative.
2450 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2454 /// Return true if the chain increment is profitable to expand into a loop
2455 /// invariant value, which may require its own register. A profitable chain
2456 /// increment will be an offset relative to the same base. We allow such offsets
2457 /// to potentially be used as chain increment as long as it's not obviously
2458 /// expensive to expand using real instructions.
2459 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2460 const SCEV *IncExpr,
2461 ScalarEvolution &SE) {
2462 // Aggressively form chains when -stress-ivchain.
2466 // Do not replace a constant offset from IV head with a nonconstant IV
2468 if (!isa<SCEVConstant>(IncExpr)) {
2469 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2470 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2474 SmallPtrSet<const SCEV*, 8> Processed;
2475 return !isHighCostExpansion(IncExpr, Processed, SE);
2478 /// Return true if the number of registers needed for the chain is estimated to
2479 /// be less than the number required for the individual IV users. First prohibit
2480 /// any IV users that keep the IV live across increments (the Users set should
2481 /// be empty). Next count the number and type of increments in the chain.
2483 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2484 /// effectively use postinc addressing modes. Only consider it profitable it the
2485 /// increments can be computed in fewer registers when chained.
2487 /// TODO: Consider IVInc free if it's already used in another chains.
2489 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2490 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2494 if (!Chain.hasIncs())
2497 if (!Users.empty()) {
2498 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2499 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2500 E = Users.end(); I != E; ++I) {
2501 dbgs() << " " << **I << "\n";
2505 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2507 // The chain itself may require a register, so intialize cost to 1.
2510 // A complete chain likely eliminates the need for keeping the original IV in
2511 // a register. LSR does not currently know how to form a complete chain unless
2512 // the header phi already exists.
2513 if (isa<PHINode>(Chain.tailUserInst())
2514 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2517 const SCEV *LastIncExpr = 0;
2518 unsigned NumConstIncrements = 0;
2519 unsigned NumVarIncrements = 0;
2520 unsigned NumReusedIncrements = 0;
2521 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2524 if (I->IncExpr->isZero())
2527 // Incrementing by zero or some constant is neutral. We assume constants can
2528 // be folded into an addressing mode or an add's immediate operand.
2529 if (isa<SCEVConstant>(I->IncExpr)) {
2530 ++NumConstIncrements;
2534 if (I->IncExpr == LastIncExpr)
2535 ++NumReusedIncrements;
2539 LastIncExpr = I->IncExpr;
2541 // An IV chain with a single increment is handled by LSR's postinc
2542 // uses. However, a chain with multiple increments requires keeping the IV's
2543 // value live longer than it needs to be if chained.
2544 if (NumConstIncrements > 1)
2547 // Materializing increment expressions in the preheader that didn't exist in
2548 // the original code may cost a register. For example, sign-extended array
2549 // indices can produce ridiculous increments like this:
2550 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2551 cost += NumVarIncrements;
2553 // Reusing variable increments likely saves a register to hold the multiple of
2555 cost -= NumReusedIncrements;
2557 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2563 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2565 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2566 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2567 // When IVs are used as types of varying widths, they are generally converted
2568 // to a wider type with some uses remaining narrow under a (free) trunc.
2569 Value *const NextIV = getWideOperand(IVOper);
2570 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2571 const SCEV *const OperExprBase = getExprBase(OperExpr);
2573 // Visit all existing chains. Check if its IVOper can be computed as a
2574 // profitable loop invariant increment from the last link in the Chain.
2575 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2576 const SCEV *LastIncExpr = 0;
2577 for (; ChainIdx < NChains; ++ChainIdx) {
2578 IVChain &Chain = IVChainVec[ChainIdx];
2580 // Prune the solution space aggressively by checking that both IV operands
2581 // are expressions that operate on the same unscaled SCEVUnknown. This
2582 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2583 // first avoids creating extra SCEV expressions.
2584 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2587 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2588 if (!isCompatibleIVType(PrevIV, NextIV))
2591 // A phi node terminates a chain.
2592 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2595 // The increment must be loop-invariant so it can be kept in a register.
2596 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2597 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2598 if (!SE.isLoopInvariant(IncExpr, L))
2601 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2602 LastIncExpr = IncExpr;
2606 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2607 // bother for phi nodes, because they must be last in the chain.
2608 if (ChainIdx == NChains) {
2609 if (isa<PHINode>(UserInst))
2611 if (NChains >= MaxChains && !StressIVChain) {
2612 DEBUG(dbgs() << "IV Chain Limit\n");
2615 LastIncExpr = OperExpr;
2616 // IVUsers may have skipped over sign/zero extensions. We don't currently
2617 // attempt to form chains involving extensions unless they can be hoisted
2618 // into this loop's AddRec.
2619 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2622 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2624 ChainUsersVec.resize(NChains);
2625 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2626 << ") IV=" << *LastIncExpr << "\n");
2628 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2629 << ") IV+" << *LastIncExpr << "\n");
2630 // Add this IV user to the end of the chain.
2631 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2633 IVChain &Chain = IVChainVec[ChainIdx];
2635 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2636 // This chain's NearUsers become FarUsers.
2637 if (!LastIncExpr->isZero()) {
2638 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2643 // All other uses of IVOperand become near uses of the chain.
2644 // We currently ignore intermediate values within SCEV expressions, assuming
2645 // they will eventually be used be the current chain, or can be computed
2646 // from one of the chain increments. To be more precise we could
2647 // transitively follow its user and only add leaf IV users to the set.
2648 for (Value::use_iterator UseIter = IVOper->use_begin(),
2649 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2650 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2653 // Uses in the chain will no longer be uses if the chain is formed.
2654 // Include the head of the chain in this iteration (not Chain.begin()).
2655 IVChain::const_iterator IncIter = Chain.Incs.begin();
2656 IVChain::const_iterator IncEnd = Chain.Incs.end();
2657 for( ; IncIter != IncEnd; ++IncIter) {
2658 if (IncIter->UserInst == OtherUse)
2661 if (IncIter != IncEnd)
2664 if (SE.isSCEVable(OtherUse->getType())
2665 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2666 && IU.isIVUserOrOperand(OtherUse)) {
2669 NearUsers.insert(OtherUse);
2672 // Since this user is part of the chain, it's no longer considered a use
2674 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2677 /// CollectChains - Populate the vector of Chains.
2679 /// This decreases ILP at the architecture level. Targets with ample registers,
2680 /// multiple memory ports, and no register renaming probably don't want
2681 /// this. However, such targets should probably disable LSR altogether.
2683 /// The job of LSR is to make a reasonable choice of induction variables across
2684 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2685 /// ILP *within the loop* if the target wants it.
2687 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2688 /// will not reorder memory operations, it will recognize this as a chain, but
2689 /// will generate redundant IV increments. Ideally this would be corrected later
2690 /// by a smart scheduler:
2696 /// TODO: Walk the entire domtree within this loop, not just the path to the
2697 /// loop latch. This will discover chains on side paths, but requires
2698 /// maintaining multiple copies of the Chains state.
2699 void LSRInstance::CollectChains() {
2700 DEBUG(dbgs() << "Collecting IV Chains.\n");
2701 SmallVector<ChainUsers, 8> ChainUsersVec;
2703 SmallVector<BasicBlock *,8> LatchPath;
2704 BasicBlock *LoopHeader = L->getHeader();
2705 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2706 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2707 LatchPath.push_back(Rung->getBlock());
2709 LatchPath.push_back(LoopHeader);
2711 // Walk the instruction stream from the loop header to the loop latch.
2712 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2713 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2714 BBIter != BBEnd; ++BBIter) {
2715 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2717 // Skip instructions that weren't seen by IVUsers analysis.
2718 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2721 // Ignore users that are part of a SCEV expression. This way we only
2722 // consider leaf IV Users. This effectively rediscovers a portion of
2723 // IVUsers analysis but in program order this time.
2724 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2727 // Remove this instruction from any NearUsers set it may be in.
2728 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2729 ChainIdx < NChains; ++ChainIdx) {
2730 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2732 // Search for operands that can be chained.
2733 SmallPtrSet<Instruction*, 4> UniqueOperands;
2734 User::op_iterator IVOpEnd = I->op_end();
2735 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2736 while (IVOpIter != IVOpEnd) {
2737 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2738 if (UniqueOperands.insert(IVOpInst))
2739 ChainInstruction(I, IVOpInst, ChainUsersVec);
2740 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2742 } // Continue walking down the instructions.
2743 } // Continue walking down the domtree.
2744 // Visit phi backedges to determine if the chain can generate the IV postinc.
2745 for (BasicBlock::iterator I = L->getHeader()->begin();
2746 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2747 if (!SE.isSCEVable(PN->getType()))
2751 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2753 ChainInstruction(PN, IncV, ChainUsersVec);
2755 // Remove any unprofitable chains.
2756 unsigned ChainIdx = 0;
2757 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2758 UsersIdx < NChains; ++UsersIdx) {
2759 if (!isProfitableChain(IVChainVec[UsersIdx],
2760 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2762 // Preserve the chain at UsesIdx.
2763 if (ChainIdx != UsersIdx)
2764 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2765 FinalizeChain(IVChainVec[ChainIdx]);
2768 IVChainVec.resize(ChainIdx);
2771 void LSRInstance::FinalizeChain(IVChain &Chain) {
2772 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2773 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2775 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2777 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2778 User::op_iterator UseI =
2779 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2780 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2781 IVIncSet.insert(UseI);
2785 /// Return true if the IVInc can be folded into an addressing mode.
2786 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2787 Value *Operand, const TargetTransformInfo &TTI) {
2788 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2789 if (!IncConst || !isAddressUse(UserInst, Operand))
2792 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2795 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2796 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2797 getAccessType(UserInst), /*BaseGV=*/ 0,
2798 IncOffset, /*HaseBaseReg=*/ false))
2804 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2805 /// materialize the IV user's operand from the previous IV user's operand.
2806 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2807 SmallVectorImpl<WeakVH> &DeadInsts) {
2808 // Find the new IVOperand for the head of the chain. It may have been replaced
2810 const IVInc &Head = Chain.Incs[0];
2811 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2812 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2813 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2816 while (IVOpIter != IVOpEnd) {
2817 IVSrc = getWideOperand(*IVOpIter);
2819 // If this operand computes the expression that the chain needs, we may use
2820 // it. (Check this after setting IVSrc which is used below.)
2822 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2823 // narrow for the chain, so we can no longer use it. We do allow using a
2824 // wider phi, assuming the LSR checked for free truncation. In that case we
2825 // should already have a truncate on this operand such that
2826 // getSCEV(IVSrc) == IncExpr.
2827 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2828 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2831 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2833 if (IVOpIter == IVOpEnd) {
2834 // Gracefully give up on this chain.
2835 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2839 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2840 Type *IVTy = IVSrc->getType();
2841 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2842 const SCEV *LeftOverExpr = 0;
2843 for (IVChain::const_iterator IncI = Chain.begin(),
2844 IncE = Chain.end(); IncI != IncE; ++IncI) {
2846 Instruction *InsertPt = IncI->UserInst;
2847 if (isa<PHINode>(InsertPt))
2848 InsertPt = L->getLoopLatch()->getTerminator();
2850 // IVOper will replace the current IV User's operand. IVSrc is the IV
2851 // value currently held in a register.
2852 Value *IVOper = IVSrc;
2853 if (!IncI->IncExpr->isZero()) {
2854 // IncExpr was the result of subtraction of two narrow values, so must
2856 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2857 LeftOverExpr = LeftOverExpr ?
2858 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2860 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2861 // Expand the IV increment.
2862 Rewriter.clearPostInc();
2863 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2864 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2865 SE.getUnknown(IncV));
2866 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2868 // If an IV increment can't be folded, use it as the next IV value.
2869 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2871 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2876 Type *OperTy = IncI->IVOperand->getType();
2877 if (IVTy != OperTy) {
2878 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2879 "cannot extend a chained IV");
2880 IRBuilder<> Builder(InsertPt);
2881 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2883 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2884 DeadInsts.push_back(IncI->IVOperand);
2886 // If LSR created a new, wider phi, we may also replace its postinc. We only
2887 // do this if we also found a wide value for the head of the chain.
2888 if (isa<PHINode>(Chain.tailUserInst())) {
2889 for (BasicBlock::iterator I = L->getHeader()->begin();
2890 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2891 if (!isCompatibleIVType(Phi, IVSrc))
2893 Instruction *PostIncV = dyn_cast<Instruction>(
2894 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2895 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2897 Value *IVOper = IVSrc;
2898 Type *PostIncTy = PostIncV->getType();
2899 if (IVTy != PostIncTy) {
2900 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2901 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2902 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2903 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2905 Phi->replaceUsesOfWith(PostIncV, IVOper);
2906 DeadInsts.push_back(PostIncV);
2911 void LSRInstance::CollectFixupsAndInitialFormulae() {
2912 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2913 Instruction *UserInst = UI->getUser();
2914 // Skip IV users that are part of profitable IV Chains.
2915 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2916 UI->getOperandValToReplace());
2917 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2918 if (IVIncSet.count(UseI))
2922 LSRFixup &LF = getNewFixup();
2923 LF.UserInst = UserInst;
2924 LF.OperandValToReplace = UI->getOperandValToReplace();
2925 LF.PostIncLoops = UI->getPostIncLoops();
2927 LSRUse::KindType Kind = LSRUse::Basic;
2929 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2930 Kind = LSRUse::Address;
2931 AccessTy = getAccessType(LF.UserInst);
2934 const SCEV *S = IU.getExpr(*UI);
2936 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2937 // (N - i == 0), and this allows (N - i) to be the expression that we work
2938 // with rather than just N or i, so we can consider the register
2939 // requirements for both N and i at the same time. Limiting this code to
2940 // equality icmps is not a problem because all interesting loops use
2941 // equality icmps, thanks to IndVarSimplify.
2942 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2943 if (CI->isEquality()) {
2944 // Swap the operands if needed to put the OperandValToReplace on the
2945 // left, for consistency.
2946 Value *NV = CI->getOperand(1);
2947 if (NV == LF.OperandValToReplace) {
2948 CI->setOperand(1, CI->getOperand(0));
2949 CI->setOperand(0, NV);
2950 NV = CI->getOperand(1);
2954 // x == y --> x - y == 0
2955 const SCEV *N = SE.getSCEV(NV);
2956 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2957 // S is normalized, so normalize N before folding it into S
2958 // to keep the result normalized.
2959 N = TransformForPostIncUse(Normalize, N, CI, 0,
2960 LF.PostIncLoops, SE, DT);
2961 Kind = LSRUse::ICmpZero;
2962 S = SE.getMinusSCEV(N, S);
2965 // -1 and the negations of all interesting strides (except the negation
2966 // of -1) are now also interesting.
2967 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2968 if (Factors[i] != -1)
2969 Factors.insert(-(uint64_t)Factors[i]);
2973 // Set up the initial formula for this use.
2974 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2976 LF.Offset = P.second;
2977 LSRUse &LU = Uses[LF.LUIdx];
2978 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2979 if (!LU.WidestFixupType ||
2980 SE.getTypeSizeInBits(LU.WidestFixupType) <
2981 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2982 LU.WidestFixupType = LF.OperandValToReplace->getType();
2984 // If this is the first use of this LSRUse, give it a formula.
2985 if (LU.Formulae.empty()) {
2986 InsertInitialFormula(S, LU, LF.LUIdx);
2987 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2991 DEBUG(print_fixups(dbgs()));
2994 /// InsertInitialFormula - Insert a formula for the given expression into
2995 /// the given use, separating out loop-variant portions from loop-invariant
2996 /// and loop-computable portions.
2998 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2999 // Mark uses whose expressions cannot be expanded.
3000 if (!isSafeToExpand(S, SE))
3001 LU.RigidFormula = true;
3004 F.InitialMatch(S, L, SE);
3005 bool Inserted = InsertFormula(LU, LUIdx, F);
3006 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3009 /// InsertSupplementalFormula - Insert a simple single-register formula for
3010 /// the given expression into the given use.
3012 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3013 LSRUse &LU, size_t LUIdx) {
3015 F.BaseRegs.push_back(S);
3016 F.HasBaseReg = true;
3017 bool Inserted = InsertFormula(LU, LUIdx, F);
3018 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3021 /// CountRegisters - Note which registers are used by the given formula,
3022 /// updating RegUses.
3023 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3025 RegUses.CountRegister(F.ScaledReg, LUIdx);
3026 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3027 E = F.BaseRegs.end(); I != E; ++I)
3028 RegUses.CountRegister(*I, LUIdx);
3031 /// InsertFormula - If the given formula has not yet been inserted, add it to
3032 /// the list, and return true. Return false otherwise.
3033 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3034 if (!LU.InsertFormula(F))
3037 CountRegisters(F, LUIdx);
3041 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3042 /// loop-invariant values which we're tracking. These other uses will pin these
3043 /// values in registers, making them less profitable for elimination.
3044 /// TODO: This currently misses non-constant addrec step registers.
3045 /// TODO: Should this give more weight to users inside the loop?
3047 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3048 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3049 SmallPtrSet<const SCEV *, 8> Inserted;
3051 while (!Worklist.empty()) {
3052 const SCEV *S = Worklist.pop_back_val();
3054 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3055 Worklist.append(N->op_begin(), N->op_end());
3056 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3057 Worklist.push_back(C->getOperand());
3058 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3059 Worklist.push_back(D->getLHS());
3060 Worklist.push_back(D->getRHS());
3061 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3062 if (!Inserted.insert(U)) continue;
3063 const Value *V = U->getValue();
3064 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3065 // Look for instructions defined outside the loop.
3066 if (L->contains(Inst)) continue;
3067 } else if (isa<UndefValue>(V))
3068 // Undef doesn't have a live range, so it doesn't matter.
3070 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
3072 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
3073 // Ignore non-instructions.
3076 // Ignore instructions in other functions (as can happen with
3078 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3080 // Ignore instructions not dominated by the loop.
3081 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3082 UserInst->getParent() :
3083 cast<PHINode>(UserInst)->getIncomingBlock(
3084 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
3085 if (!DT.dominates(L->getHeader(), UseBB))
3087 // Ignore uses which are part of other SCEV expressions, to avoid
3088 // analyzing them multiple times.
3089 if (SE.isSCEVable(UserInst->getType())) {
3090 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3091 // If the user is a no-op, look through to its uses.
3092 if (!isa<SCEVUnknown>(UserS))
3096 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3100 // Ignore icmp instructions which are already being analyzed.
3101 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3102 unsigned OtherIdx = !UI.getOperandNo();
3103 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3104 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3108 LSRFixup &LF = getNewFixup();
3109 LF.UserInst = const_cast<Instruction *>(UserInst);
3110 LF.OperandValToReplace = UI.getUse();
3111 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
3113 LF.Offset = P.second;
3114 LSRUse &LU = Uses[LF.LUIdx];
3115 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3116 if (!LU.WidestFixupType ||
3117 SE.getTypeSizeInBits(LU.WidestFixupType) <
3118 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3119 LU.WidestFixupType = LF.OperandValToReplace->getType();
3120 InsertSupplementalFormula(U, LU, LF.LUIdx);
3121 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3128 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3129 /// separate registers. If C is non-null, multiply each subexpression by C.
3131 /// Return remainder expression after factoring the subexpressions captured by
3132 /// Ops. If Ops is complete, return NULL.
3133 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3134 SmallVectorImpl<const SCEV *> &Ops,
3136 ScalarEvolution &SE,
3137 unsigned Depth = 0) {
3138 // Arbitrarily cap recursion to protect compile time.
3142 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3143 // Break out add operands.
3144 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3146 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3148 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3151 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3152 // Split a non-zero base out of an addrec.
3153 if (AR->getStart()->isZero())
3156 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3157 C, Ops, L, SE, Depth+1);
3158 // Split the non-zero AddRec unless it is part of a nested recurrence that
3159 // does not pertain to this loop.
3160 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3161 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3164 if (Remainder != AR->getStart()) {
3166 Remainder = SE.getConstant(AR->getType(), 0);
3167 return SE.getAddRecExpr(Remainder,
3168 AR->getStepRecurrence(SE),
3170 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3173 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3174 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3175 if (Mul->getNumOperands() != 2)
3177 if (const SCEVConstant *Op0 =
3178 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3179 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3180 const SCEV *Remainder =
3181 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3183 Ops.push_back(SE.getMulExpr(C, Remainder));
3190 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3192 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3195 // Arbitrarily cap recursion to protect compile time.
3196 if (Depth >= 3) return;
3198 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3199 const SCEV *BaseReg = Base.BaseRegs[i];
3201 SmallVector<const SCEV *, 8> AddOps;
3202 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3204 AddOps.push_back(Remainder);
3206 if (AddOps.size() == 1) continue;
3208 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3209 JE = AddOps.end(); J != JE; ++J) {
3211 // Loop-variant "unknown" values are uninteresting; we won't be able to
3212 // do anything meaningful with them.
3213 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3216 // Don't pull a constant into a register if the constant could be folded
3217 // into an immediate field.
3218 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3219 LU.AccessTy, *J, Base.getNumRegs() > 1))
3222 // Collect all operands except *J.
3223 SmallVector<const SCEV *, 8> InnerAddOps
3224 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3226 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3228 // Don't leave just a constant behind in a register if the constant could
3229 // be folded into an immediate field.
3230 if (InnerAddOps.size() == 1 &&
3231 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3232 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3235 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3236 if (InnerSum->isZero())
3240 // Add the remaining pieces of the add back into the new formula.
3241 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3243 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3244 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3245 InnerSumSC->getValue()->getZExtValue())) {
3246 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3247 InnerSumSC->getValue()->getZExtValue();
3248 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3250 F.BaseRegs[i] = InnerSum;
3252 // Add J as its own register, or an unfolded immediate.
3253 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3254 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3255 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3256 SC->getValue()->getZExtValue()))
3257 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3258 SC->getValue()->getZExtValue();
3260 F.BaseRegs.push_back(*J);
3262 if (InsertFormula(LU, LUIdx, F))
3263 // If that formula hadn't been seen before, recurse to find more like
3265 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3270 /// GenerateCombinations - Generate a formula consisting of all of the
3271 /// loop-dominating registers added into a single register.
3272 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3274 // This method is only interesting on a plurality of registers.
3275 if (Base.BaseRegs.size() <= 1) return;
3279 SmallVector<const SCEV *, 4> Ops;
3280 for (SmallVectorImpl<const SCEV *>::const_iterator
3281 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3282 const SCEV *BaseReg = *I;
3283 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3284 !SE.hasComputableLoopEvolution(BaseReg, L))
3285 Ops.push_back(BaseReg);
3287 F.BaseRegs.push_back(BaseReg);
3289 if (Ops.size() > 1) {
3290 const SCEV *Sum = SE.getAddExpr(Ops);
3291 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3292 // opportunity to fold something. For now, just ignore such cases
3293 // rather than proceed with zero in a register.
3294 if (!Sum->isZero()) {
3295 F.BaseRegs.push_back(Sum);
3296 (void)InsertFormula(LU, LUIdx, F);
3301 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3302 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3304 // We can't add a symbolic offset if the address already contains one.
3305 if (Base.BaseGV) return;
3307 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3308 const SCEV *G = Base.BaseRegs[i];
3309 GlobalValue *GV = ExtractSymbol(G, SE);
3310 if (G->isZero() || !GV)
3314 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3317 (void)InsertFormula(LU, LUIdx, F);
3321 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3322 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3324 // TODO: For now, just add the min and max offset, because it usually isn't
3325 // worthwhile looking at everything inbetween.
3326 SmallVector<int64_t, 2> Worklist;
3327 Worklist.push_back(LU.MinOffset);
3328 if (LU.MaxOffset != LU.MinOffset)
3329 Worklist.push_back(LU.MaxOffset);
3331 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3332 const SCEV *G = Base.BaseRegs[i];
3334 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3335 E = Worklist.end(); I != E; ++I) {
3337 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3338 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3340 // Add the offset to the base register.
3341 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3342 // If it cancelled out, drop the base register, otherwise update it.
3343 if (NewG->isZero()) {
3344 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3345 F.BaseRegs.pop_back();
3347 F.BaseRegs[i] = NewG;
3349 (void)InsertFormula(LU, LUIdx, F);
3353 int64_t Imm = ExtractImmediate(G, SE);
3354 if (G->isZero() || Imm == 0)
3357 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3358 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3361 (void)InsertFormula(LU, LUIdx, F);
3365 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3366 /// the comparison. For example, x == y -> x*c == y*c.
3367 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3369 if (LU.Kind != LSRUse::ICmpZero) return;
3371 // Determine the integer type for the base formula.
3372 Type *IntTy = Base.getType();
3374 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3376 // Don't do this if there is more than one offset.
3377 if (LU.MinOffset != LU.MaxOffset) return;
3379 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3381 // Check each interesting stride.
3382 for (SmallSetVector<int64_t, 8>::const_iterator
3383 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3384 int64_t Factor = *I;
3386 // Check that the multiplication doesn't overflow.
3387 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3389 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3390 if (NewBaseOffset / Factor != Base.BaseOffset)
3392 // If the offset will be truncated at this use, check that it is in bounds.
3393 if (!IntTy->isPointerTy() &&
3394 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3397 // Check that multiplying with the use offset doesn't overflow.
3398 int64_t Offset = LU.MinOffset;
3399 if (Offset == INT64_MIN && Factor == -1)
3401 Offset = (uint64_t)Offset * Factor;
3402 if (Offset / Factor != LU.MinOffset)
3404 // If the offset will be truncated at this use, check that it is in bounds.
3405 if (!IntTy->isPointerTy() &&
3406 !ConstantInt::isValueValidForType(IntTy, Offset))
3410 F.BaseOffset = NewBaseOffset;
3412 // Check that this scale is legal.
3413 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3416 // Compensate for the use having MinOffset built into it.
3417 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3419 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3421 // Check that multiplying with each base register doesn't overflow.
3422 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3423 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3424 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3428 // Check that multiplying with the scaled register doesn't overflow.
3430 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3431 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3435 // Check that multiplying with the unfolded offset doesn't overflow.
3436 if (F.UnfoldedOffset != 0) {
3437 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3439 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3440 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3442 // If the offset will be truncated, check that it is in bounds.
3443 if (!IntTy->isPointerTy() &&
3444 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3448 // If we make it here and it's legal, add it.
3449 (void)InsertFormula(LU, LUIdx, F);
3454 /// GenerateScales - Generate stride factor reuse formulae by making use of
3455 /// scaled-offset address modes, for example.
3456 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3457 // Determine the integer type for the base formula.
3458 Type *IntTy = Base.getType();
3461 // If this Formula already has a scaled register, we can't add another one.
3462 if (Base.Scale != 0) return;
3464 // Check each interesting stride.
3465 for (SmallSetVector<int64_t, 8>::const_iterator
3466 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3467 int64_t Factor = *I;
3469 Base.Scale = Factor;
3470 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3471 // Check whether this scale is going to be legal.
3472 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3474 // As a special-case, handle special out-of-loop Basic users specially.
3475 // TODO: Reconsider this special case.
3476 if (LU.Kind == LSRUse::Basic &&
3477 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3478 LU.AccessTy, Base) &&
3479 LU.AllFixupsOutsideLoop)
3480 LU.Kind = LSRUse::Special;
3484 // For an ICmpZero, negating a solitary base register won't lead to
3486 if (LU.Kind == LSRUse::ICmpZero &&
3487 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3489 // For each addrec base reg, apply the scale, if possible.
3490 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3491 if (const SCEVAddRecExpr *AR =
3492 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3493 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3494 if (FactorS->isZero())
3496 // Divide out the factor, ignoring high bits, since we'll be
3497 // scaling the value back up in the end.
3498 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3499 // TODO: This could be optimized to avoid all the copying.
3501 F.ScaledReg = Quotient;
3502 F.DeleteBaseReg(F.BaseRegs[i]);
3503 (void)InsertFormula(LU, LUIdx, F);
3509 /// GenerateTruncates - Generate reuse formulae from different IV types.
3510 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3511 // Don't bother truncating symbolic values.
3512 if (Base.BaseGV) return;
3514 // Determine the integer type for the base formula.
3515 Type *DstTy = Base.getType();
3517 DstTy = SE.getEffectiveSCEVType(DstTy);
3519 for (SmallSetVector<Type *, 4>::const_iterator
3520 I = Types.begin(), E = Types.end(); I != E; ++I) {
3522 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3525 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3526 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3527 JE = F.BaseRegs.end(); J != JE; ++J)
3528 *J = SE.getAnyExtendExpr(*J, SrcTy);
3530 // TODO: This assumes we've done basic processing on all uses and
3531 // have an idea what the register usage is.
3532 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3535 (void)InsertFormula(LU, LUIdx, F);
3542 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3543 /// defer modifications so that the search phase doesn't have to worry about
3544 /// the data structures moving underneath it.
3548 const SCEV *OrigReg;
3550 WorkItem(size_t LI, int64_t I, const SCEV *R)
3551 : LUIdx(LI), Imm(I), OrigReg(R) {}
3553 void print(raw_ostream &OS) const;
3559 void WorkItem::print(raw_ostream &OS) const {
3560 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3561 << " , add offset " << Imm;
3564 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3565 void WorkItem::dump() const {
3566 print(errs()); errs() << '\n';
3570 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3571 /// distance apart and try to form reuse opportunities between them.
3572 void LSRInstance::GenerateCrossUseConstantOffsets() {
3573 // Group the registers by their value without any added constant offset.
3574 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3575 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3577 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3578 SmallVector<const SCEV *, 8> Sequence;
3579 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3581 const SCEV *Reg = *I;
3582 int64_t Imm = ExtractImmediate(Reg, SE);
3583 std::pair<RegMapTy::iterator, bool> Pair =
3584 Map.insert(std::make_pair(Reg, ImmMapTy()));
3586 Sequence.push_back(Reg);
3587 Pair.first->second.insert(std::make_pair(Imm, *I));
3588 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3591 // Now examine each set of registers with the same base value. Build up
3592 // a list of work to do and do the work in a separate step so that we're
3593 // not adding formulae and register counts while we're searching.
3594 SmallVector<WorkItem, 32> WorkItems;
3595 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3596 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3597 E = Sequence.end(); I != E; ++I) {
3598 const SCEV *Reg = *I;
3599 const ImmMapTy &Imms = Map.find(Reg)->second;
3601 // It's not worthwhile looking for reuse if there's only one offset.
3602 if (Imms.size() == 1)
3605 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3606 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3608 dbgs() << ' ' << J->first;
3611 // Examine each offset.
3612 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3614 const SCEV *OrigReg = J->second;
3616 int64_t JImm = J->first;
3617 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3619 if (!isa<SCEVConstant>(OrigReg) &&
3620 UsedByIndicesMap[Reg].count() == 1) {
3621 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3625 // Conservatively examine offsets between this orig reg a few selected
3627 ImmMapTy::const_iterator OtherImms[] = {
3628 Imms.begin(), prior(Imms.end()),
3629 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3631 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3632 ImmMapTy::const_iterator M = OtherImms[i];
3633 if (M == J || M == JE) continue;
3635 // Compute the difference between the two.
3636 int64_t Imm = (uint64_t)JImm - M->first;
3637 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3638 LUIdx = UsedByIndices.find_next(LUIdx))
3639 // Make a memo of this use, offset, and register tuple.
3640 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3641 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3648 UsedByIndicesMap.clear();
3649 UniqueItems.clear();
3651 // Now iterate through the worklist and add new formulae.
3652 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3653 E = WorkItems.end(); I != E; ++I) {
3654 const WorkItem &WI = *I;
3655 size_t LUIdx = WI.LUIdx;
3656 LSRUse &LU = Uses[LUIdx];
3657 int64_t Imm = WI.Imm;
3658 const SCEV *OrigReg = WI.OrigReg;
3660 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3661 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3662 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3664 // TODO: Use a more targeted data structure.
3665 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3666 const Formula &F = LU.Formulae[L];
3667 // Use the immediate in the scaled register.
3668 if (F.ScaledReg == OrigReg) {
3669 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3670 // Don't create 50 + reg(-50).
3671 if (F.referencesReg(SE.getSCEV(
3672 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3675 NewF.BaseOffset = Offset;
3676 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3679 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3681 // If the new scale is a constant in a register, and adding the constant
3682 // value to the immediate would produce a value closer to zero than the
3683 // immediate itself, then the formula isn't worthwhile.
3684 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3685 if (C->getValue()->isNegative() !=
3686 (NewF.BaseOffset < 0) &&
3687 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3688 .ule(abs64(NewF.BaseOffset)))
3692 (void)InsertFormula(LU, LUIdx, NewF);
3694 // Use the immediate in a base register.
3695 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3696 const SCEV *BaseReg = F.BaseRegs[N];
3697 if (BaseReg != OrigReg)
3700 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3701 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3702 LU.Kind, LU.AccessTy, NewF)) {
3703 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3706 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3708 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3710 // If the new formula has a constant in a register, and adding the
3711 // constant value to the immediate would produce a value closer to
3712 // zero than the immediate itself, then the formula isn't worthwhile.
3713 for (SmallVectorImpl<const SCEV *>::const_iterator
3714 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3716 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3717 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3718 abs64(NewF.BaseOffset)) &&
3719 (C->getValue()->getValue() +
3720 NewF.BaseOffset).countTrailingZeros() >=
3721 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3725 (void)InsertFormula(LU, LUIdx, NewF);
3734 /// GenerateAllReuseFormulae - Generate formulae for each use.
3736 LSRInstance::GenerateAllReuseFormulae() {
3737 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3738 // queries are more precise.
3739 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3740 LSRUse &LU = Uses[LUIdx];
3741 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3742 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3743 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3744 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3746 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3747 LSRUse &LU = Uses[LUIdx];
3748 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3749 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3750 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3751 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3752 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3753 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3754 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3755 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3757 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3758 LSRUse &LU = Uses[LUIdx];
3759 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3760 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3763 GenerateCrossUseConstantOffsets();
3765 DEBUG(dbgs() << "\n"
3766 "After generating reuse formulae:\n";
3767 print_uses(dbgs()));
3770 /// If there are multiple formulae with the same set of registers used
3771 /// by other uses, pick the best one and delete the others.
3772 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3773 DenseSet<const SCEV *> VisitedRegs;
3774 SmallPtrSet<const SCEV *, 16> Regs;
3775 SmallPtrSet<const SCEV *, 16> LoserRegs;
3777 bool ChangedFormulae = false;
3780 // Collect the best formula for each unique set of shared registers. This
3781 // is reset for each use.
3782 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3784 BestFormulaeTy BestFormulae;
3786 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3787 LSRUse &LU = Uses[LUIdx];
3788 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3791 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3792 FIdx != NumForms; ++FIdx) {
3793 Formula &F = LU.Formulae[FIdx];
3795 // Some formulas are instant losers. For example, they may depend on
3796 // nonexistent AddRecs from other loops. These need to be filtered
3797 // immediately, otherwise heuristics could choose them over others leading
3798 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3799 // avoids the need to recompute this information across formulae using the
3800 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3801 // the corresponding bad register from the Regs set.
3804 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3806 if (CostF.isLoser()) {
3807 // During initial formula generation, undesirable formulae are generated
3808 // by uses within other loops that have some non-trivial address mode or
3809 // use the postinc form of the IV. LSR needs to provide these formulae
3810 // as the basis of rediscovering the desired formula that uses an AddRec
3811 // corresponding to the existing phi. Once all formulae have been
3812 // generated, these initial losers may be pruned.
3813 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3817 SmallVector<const SCEV *, 4> Key;
3818 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3819 JE = F.BaseRegs.end(); J != JE; ++J) {
3820 const SCEV *Reg = *J;
3821 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3825 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3826 Key.push_back(F.ScaledReg);
3827 // Unstable sort by host order ok, because this is only used for
3829 std::sort(Key.begin(), Key.end());
3831 std::pair<BestFormulaeTy::const_iterator, bool> P =
3832 BestFormulae.insert(std::make_pair(Key, FIdx));
3836 Formula &Best = LU.Formulae[P.first->second];
3840 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3842 if (CostF < CostBest)
3844 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3846 " in favor of formula "; Best.print(dbgs());
3850 ChangedFormulae = true;
3852 LU.DeleteFormula(F);
3858 // Now that we've filtered out some formulae, recompute the Regs set.
3860 LU.RecomputeRegs(LUIdx, RegUses);
3862 // Reset this to prepare for the next use.
3863 BestFormulae.clear();
3866 DEBUG(if (ChangedFormulae) {
3868 "After filtering out undesirable candidates:\n";
3873 // This is a rough guess that seems to work fairly well.
3874 static const size_t ComplexityLimit = UINT16_MAX;
3876 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3877 /// solutions the solver might have to consider. It almost never considers
3878 /// this many solutions because it prune the search space, but the pruning
3879 /// isn't always sufficient.
3880 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3882 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3883 E = Uses.end(); I != E; ++I) {
3884 size_t FSize = I->Formulae.size();
3885 if (FSize >= ComplexityLimit) {
3886 Power = ComplexityLimit;
3890 if (Power >= ComplexityLimit)
3896 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3897 /// of the registers of another formula, it won't help reduce register
3898 /// pressure (though it may not necessarily hurt register pressure); remove
3899 /// it to simplify the system.
3900 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3901 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3902 DEBUG(dbgs() << "The search space is too complex.\n");
3904 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3905 "which use a superset of registers used by other "
3908 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3909 LSRUse &LU = Uses[LUIdx];
3911 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3912 Formula &F = LU.Formulae[i];
3913 // Look for a formula with a constant or GV in a register. If the use
3914 // also has a formula with that same value in an immediate field,
3915 // delete the one that uses a register.
3916 for (SmallVectorImpl<const SCEV *>::const_iterator
3917 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3918 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3920 NewF.BaseOffset += C->getValue()->getSExtValue();
3921 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3922 (I - F.BaseRegs.begin()));
3923 if (LU.HasFormulaWithSameRegs(NewF)) {
3924 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3925 LU.DeleteFormula(F);
3931 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3932 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3936 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3937 (I - F.BaseRegs.begin()));
3938 if (LU.HasFormulaWithSameRegs(NewF)) {
3939 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3941 LU.DeleteFormula(F);
3952 LU.RecomputeRegs(LUIdx, RegUses);
3955 DEBUG(dbgs() << "After pre-selection:\n";
3956 print_uses(dbgs()));
3960 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3961 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3963 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3964 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
3967 DEBUG(dbgs() << "The search space is too complex.\n"
3968 "Narrowing the search space by assuming that uses separated "
3969 "by a constant offset will use the same registers.\n");
3971 // This is especially useful for unrolled loops.
3973 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3974 LSRUse &LU = Uses[LUIdx];
3975 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3976 E = LU.Formulae.end(); I != E; ++I) {
3977 const Formula &F = *I;
3978 if (F.BaseOffset == 0 || F.Scale != 0)
3981 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
3985 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
3986 LU.Kind, LU.AccessTy))
3989 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
3991 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3993 // Update the relocs to reference the new use.
3994 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3995 E = Fixups.end(); I != E; ++I) {
3996 LSRFixup &Fixup = *I;
3997 if (Fixup.LUIdx == LUIdx) {
3998 Fixup.LUIdx = LUThatHas - &Uses.front();
3999 Fixup.Offset += F.BaseOffset;
4000 // Add the new offset to LUThatHas' offset list.
4001 if (LUThatHas->Offsets.back() != Fixup.Offset) {
4002 LUThatHas->Offsets.push_back(Fixup.Offset);
4003 if (Fixup.Offset > LUThatHas->MaxOffset)
4004 LUThatHas->MaxOffset = Fixup.Offset;
4005 if (Fixup.Offset < LUThatHas->MinOffset)
4006 LUThatHas->MinOffset = Fixup.Offset;
4008 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4010 if (Fixup.LUIdx == NumUses-1)
4011 Fixup.LUIdx = LUIdx;
4014 // Delete formulae from the new use which are no longer legal.
4016 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4017 Formula &F = LUThatHas->Formulae[i];
4018 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4019 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4020 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4022 LUThatHas->DeleteFormula(F);
4030 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4032 // Delete the old use.
4033 DeleteUse(LU, LUIdx);
4040 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4043 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4044 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4045 /// we've done more filtering, as it may be able to find more formulae to
4047 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4048 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4049 DEBUG(dbgs() << "The search space is too complex.\n");
4051 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4052 "undesirable dedicated registers.\n");
4054 FilterOutUndesirableDedicatedRegisters();
4056 DEBUG(dbgs() << "After pre-selection:\n";
4057 print_uses(dbgs()));
4061 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4062 /// to be profitable, and then in any use which has any reference to that
4063 /// register, delete all formulae which do not reference that register.
4064 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4065 // With all other options exhausted, loop until the system is simple
4066 // enough to handle.
4067 SmallPtrSet<const SCEV *, 4> Taken;
4068 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4069 // Ok, we have too many of formulae on our hands to conveniently handle.
4070 // Use a rough heuristic to thin out the list.
4071 DEBUG(dbgs() << "The search space is too complex.\n");
4073 // Pick the register which is used by the most LSRUses, which is likely
4074 // to be a good reuse register candidate.
4075 const SCEV *Best = 0;
4076 unsigned BestNum = 0;
4077 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4079 const SCEV *Reg = *I;
4080 if (Taken.count(Reg))
4085 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4086 if (Count > BestNum) {
4093 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4094 << " will yield profitable reuse.\n");
4097 // In any use with formulae which references this register, delete formulae
4098 // which don't reference it.
4099 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4100 LSRUse &LU = Uses[LUIdx];
4101 if (!LU.Regs.count(Best)) continue;
4104 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4105 Formula &F = LU.Formulae[i];
4106 if (!F.referencesReg(Best)) {
4107 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4108 LU.DeleteFormula(F);
4112 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4118 LU.RecomputeRegs(LUIdx, RegUses);
4121 DEBUG(dbgs() << "After pre-selection:\n";
4122 print_uses(dbgs()));
4126 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4127 /// formulae to choose from, use some rough heuristics to prune down the number
4128 /// of formulae. This keeps the main solver from taking an extraordinary amount
4129 /// of time in some worst-case scenarios.
4130 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4131 NarrowSearchSpaceByDetectingSupersets();
4132 NarrowSearchSpaceByCollapsingUnrolledCode();
4133 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4134 NarrowSearchSpaceByPickingWinnerRegs();
4137 /// SolveRecurse - This is the recursive solver.
4138 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4140 SmallVectorImpl<const Formula *> &Workspace,
4141 const Cost &CurCost,
4142 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4143 DenseSet<const SCEV *> &VisitedRegs) const {
4146 // - use more aggressive filtering
4147 // - sort the formula so that the most profitable solutions are found first
4148 // - sort the uses too
4150 // - don't compute a cost, and then compare. compare while computing a cost
4152 // - track register sets with SmallBitVector
4154 const LSRUse &LU = Uses[Workspace.size()];
4156 // If this use references any register that's already a part of the
4157 // in-progress solution, consider it a requirement that a formula must
4158 // reference that register in order to be considered. This prunes out
4159 // unprofitable searching.
4160 SmallSetVector<const SCEV *, 4> ReqRegs;
4161 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4162 E = CurRegs.end(); I != E; ++I)
4163 if (LU.Regs.count(*I))
4166 SmallPtrSet<const SCEV *, 16> NewRegs;
4168 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4169 E = LU.Formulae.end(); I != E; ++I) {
4170 const Formula &F = *I;
4172 // Ignore formulae which do not use any of the required registers.
4173 bool SatisfiedReqReg = true;
4174 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4175 JE = ReqRegs.end(); J != JE; ++J) {
4176 const SCEV *Reg = *J;
4177 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4178 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4180 SatisfiedReqReg = false;
4184 if (!SatisfiedReqReg) {
4185 // If none of the formulae satisfied the required registers, then we could
4186 // clear ReqRegs and try again. Currently, we simply give up in this case.
4190 // Evaluate the cost of the current formula. If it's already worse than
4191 // the current best, prune the search at that point.
4194 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4196 if (NewCost < SolutionCost) {
4197 Workspace.push_back(&F);
4198 if (Workspace.size() != Uses.size()) {
4199 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4200 NewRegs, VisitedRegs);
4201 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4202 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4204 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4205 dbgs() << ".\n Regs:";
4206 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4207 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4208 dbgs() << ' ' << **I;
4211 SolutionCost = NewCost;
4212 Solution = Workspace;
4214 Workspace.pop_back();
4219 /// Solve - Choose one formula from each use. Return the results in the given
4220 /// Solution vector.
4221 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4222 SmallVector<const Formula *, 8> Workspace;
4224 SolutionCost.Lose();
4226 SmallPtrSet<const SCEV *, 16> CurRegs;
4227 DenseSet<const SCEV *> VisitedRegs;
4228 Workspace.reserve(Uses.size());
4230 // SolveRecurse does all the work.
4231 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4232 CurRegs, VisitedRegs);
4233 if (Solution.empty()) {
4234 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4238 // Ok, we've now made all our decisions.
4239 DEBUG(dbgs() << "\n"
4240 "The chosen solution requires "; SolutionCost.print(dbgs());
4242 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4244 Uses[i].print(dbgs());
4247 Solution[i]->print(dbgs());
4251 assert(Solution.size() == Uses.size() && "Malformed solution!");
4254 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4255 /// the dominator tree far as we can go while still being dominated by the
4256 /// input positions. This helps canonicalize the insert position, which
4257 /// encourages sharing.
4258 BasicBlock::iterator
4259 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4260 const SmallVectorImpl<Instruction *> &Inputs)
4263 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4264 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4267 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4268 if (!Rung) return IP;
4269 Rung = Rung->getIDom();
4270 if (!Rung) return IP;
4271 IDom = Rung->getBlock();
4273 // Don't climb into a loop though.
4274 const Loop *IDomLoop = LI.getLoopFor(IDom);
4275 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4276 if (IDomDepth <= IPLoopDepth &&
4277 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4281 bool AllDominate = true;
4282 Instruction *BetterPos = 0;
4283 Instruction *Tentative = IDom->getTerminator();
4284 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4285 E = Inputs.end(); I != E; ++I) {
4286 Instruction *Inst = *I;
4287 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4288 AllDominate = false;
4291 // Attempt to find an insert position in the middle of the block,
4292 // instead of at the end, so that it can be used for other expansions.
4293 if (IDom == Inst->getParent() &&
4294 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4295 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4308 /// AdjustInsertPositionForExpand - Determine an input position which will be
4309 /// dominated by the operands and which will dominate the result.
4310 BasicBlock::iterator
4311 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4314 SCEVExpander &Rewriter) const {
4315 // Collect some instructions which must be dominated by the
4316 // expanding replacement. These must be dominated by any operands that
4317 // will be required in the expansion.
4318 SmallVector<Instruction *, 4> Inputs;
4319 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4320 Inputs.push_back(I);
4321 if (LU.Kind == LSRUse::ICmpZero)
4322 if (Instruction *I =
4323 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4324 Inputs.push_back(I);
4325 if (LF.PostIncLoops.count(L)) {
4326 if (LF.isUseFullyOutsideLoop(L))
4327 Inputs.push_back(L->getLoopLatch()->getTerminator());
4329 Inputs.push_back(IVIncInsertPos);
4331 // The expansion must also be dominated by the increment positions of any
4332 // loops it for which it is using post-inc mode.
4333 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4334 E = LF.PostIncLoops.end(); I != E; ++I) {
4335 const Loop *PIL = *I;
4336 if (PIL == L) continue;
4338 // Be dominated by the loop exit.
4339 SmallVector<BasicBlock *, 4> ExitingBlocks;
4340 PIL->getExitingBlocks(ExitingBlocks);
4341 if (!ExitingBlocks.empty()) {
4342 BasicBlock *BB = ExitingBlocks[0];
4343 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4344 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4345 Inputs.push_back(BB->getTerminator());
4349 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4350 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4351 "Insertion point must be a normal instruction");
4353 // Then, climb up the immediate dominator tree as far as we can go while
4354 // still being dominated by the input positions.
4355 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4357 // Don't insert instructions before PHI nodes.
4358 while (isa<PHINode>(IP)) ++IP;
4360 // Ignore landingpad instructions.
4361 while (isa<LandingPadInst>(IP)) ++IP;
4363 // Ignore debug intrinsics.
4364 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4366 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4367 // IP consistent across expansions and allows the previously inserted
4368 // instructions to be reused by subsequent expansion.
4369 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4374 /// Expand - Emit instructions for the leading candidate expression for this
4375 /// LSRUse (this is called "expanding").
4376 Value *LSRInstance::Expand(const LSRFixup &LF,
4378 BasicBlock::iterator IP,
4379 SCEVExpander &Rewriter,
4380 SmallVectorImpl<WeakVH> &DeadInsts) const {
4381 const LSRUse &LU = Uses[LF.LUIdx];
4382 if (LU.RigidFormula)
4383 return LF.OperandValToReplace;
4385 // Determine an input position which will be dominated by the operands and
4386 // which will dominate the result.
4387 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4389 // Inform the Rewriter if we have a post-increment use, so that it can
4390 // perform an advantageous expansion.
4391 Rewriter.setPostInc(LF.PostIncLoops);
4393 // This is the type that the user actually needs.
4394 Type *OpTy = LF.OperandValToReplace->getType();
4395 // This will be the type that we'll initially expand to.
4396 Type *Ty = F.getType();
4398 // No type known; just expand directly to the ultimate type.
4400 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4401 // Expand directly to the ultimate type if it's the right size.
4403 // This is the type to do integer arithmetic in.
4404 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4406 // Build up a list of operands to add together to form the full base.
4407 SmallVector<const SCEV *, 8> Ops;
4409 // Expand the BaseRegs portion.
4410 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4411 E = F.BaseRegs.end(); I != E; ++I) {
4412 const SCEV *Reg = *I;
4413 assert(!Reg->isZero() && "Zero allocated in a base register!");
4415 // If we're expanding for a post-inc user, make the post-inc adjustment.
4416 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4417 Reg = TransformForPostIncUse(Denormalize, Reg,
4418 LF.UserInst, LF.OperandValToReplace,
4421 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4424 // Expand the ScaledReg portion.
4425 Value *ICmpScaledV = 0;
4427 const SCEV *ScaledS = F.ScaledReg;
4429 // If we're expanding for a post-inc user, make the post-inc adjustment.
4430 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4431 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4432 LF.UserInst, LF.OperandValToReplace,
4435 if (LU.Kind == LSRUse::ICmpZero) {
4436 // An interesting way of "folding" with an icmp is to use a negated
4437 // scale, which we'll implement by inserting it into the other operand
4439 assert(F.Scale == -1 &&
4440 "The only scale supported by ICmpZero uses is -1!");
4441 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4443 // Otherwise just expand the scaled register and an explicit scale,
4444 // which is expected to be matched as part of the address.
4446 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4447 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4448 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4450 Ops.push_back(SE.getUnknown(FullV));
4452 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4453 ScaledS = SE.getMulExpr(ScaledS,
4454 SE.getConstant(ScaledS->getType(), F.Scale));
4455 Ops.push_back(ScaledS);
4459 // Expand the GV portion.
4461 // Flush the operand list to suppress SCEVExpander hoisting.
4463 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4465 Ops.push_back(SE.getUnknown(FullV));
4467 Ops.push_back(SE.getUnknown(F.BaseGV));
4470 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4471 // unfolded offsets. LSR assumes they both live next to their uses.
4473 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4475 Ops.push_back(SE.getUnknown(FullV));
4478 // Expand the immediate portion.
4479 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4481 if (LU.Kind == LSRUse::ICmpZero) {
4482 // The other interesting way of "folding" with an ICmpZero is to use a
4483 // negated immediate.
4485 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4487 Ops.push_back(SE.getUnknown(ICmpScaledV));
4488 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4491 // Just add the immediate values. These again are expected to be matched
4492 // as part of the address.
4493 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4497 // Expand the unfolded offset portion.
4498 int64_t UnfoldedOffset = F.UnfoldedOffset;
4499 if (UnfoldedOffset != 0) {
4500 // Just add the immediate values.
4501 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4505 // Emit instructions summing all the operands.
4506 const SCEV *FullS = Ops.empty() ?
4507 SE.getConstant(IntTy, 0) :
4509 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4511 // We're done expanding now, so reset the rewriter.
4512 Rewriter.clearPostInc();
4514 // An ICmpZero Formula represents an ICmp which we're handling as a
4515 // comparison against zero. Now that we've expanded an expression for that
4516 // form, update the ICmp's other operand.
4517 if (LU.Kind == LSRUse::ICmpZero) {
4518 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4519 DeadInsts.push_back(CI->getOperand(1));
4520 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4521 "a scale at the same time!");
4522 if (F.Scale == -1) {
4523 if (ICmpScaledV->getType() != OpTy) {
4525 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4527 ICmpScaledV, OpTy, "tmp", CI);
4530 CI->setOperand(1, ICmpScaledV);
4532 assert(F.Scale == 0 &&
4533 "ICmp does not support folding a global value and "
4534 "a scale at the same time!");
4535 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4537 if (C->getType() != OpTy)
4538 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4542 CI->setOperand(1, C);
4549 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4550 /// of their operands effectively happens in their predecessor blocks, so the
4551 /// expression may need to be expanded in multiple places.
4552 void LSRInstance::RewriteForPHI(PHINode *PN,
4555 SCEVExpander &Rewriter,
4556 SmallVectorImpl<WeakVH> &DeadInsts,
4558 DenseMap<BasicBlock *, Value *> Inserted;
4559 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4560 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4561 BasicBlock *BB = PN->getIncomingBlock(i);
4563 // If this is a critical edge, split the edge so that we do not insert
4564 // the code on all predecessor/successor paths. We do this unless this
4565 // is the canonical backedge for this loop, which complicates post-inc
4567 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4568 !isa<IndirectBrInst>(BB->getTerminator())) {
4569 BasicBlock *Parent = PN->getParent();
4570 Loop *PNLoop = LI.getLoopFor(Parent);
4571 if (!PNLoop || Parent != PNLoop->getHeader()) {
4572 // Split the critical edge.
4573 BasicBlock *NewBB = 0;
4574 if (!Parent->isLandingPad()) {
4575 NewBB = SplitCriticalEdge(BB, Parent, P,
4576 /*MergeIdenticalEdges=*/true,
4577 /*DontDeleteUselessPhis=*/true);
4579 SmallVector<BasicBlock*, 2> NewBBs;
4580 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4583 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4584 // phi predecessors are identical. The simple thing to do is skip
4585 // splitting in this case rather than complicate the API.
4587 // If PN is outside of the loop and BB is in the loop, we want to
4588 // move the block to be immediately before the PHI block, not
4589 // immediately after BB.
4590 if (L->contains(BB) && !L->contains(PN))
4591 NewBB->moveBefore(PN->getParent());
4593 // Splitting the edge can reduce the number of PHI entries we have.
4594 e = PN->getNumIncomingValues();
4596 i = PN->getBasicBlockIndex(BB);
4601 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4602 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4604 PN->setIncomingValue(i, Pair.first->second);
4606 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4608 // If this is reuse-by-noop-cast, insert the noop cast.
4609 Type *OpTy = LF.OperandValToReplace->getType();
4610 if (FullV->getType() != OpTy)
4612 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4614 FullV, LF.OperandValToReplace->getType(),
4615 "tmp", BB->getTerminator());
4617 PN->setIncomingValue(i, FullV);
4618 Pair.first->second = FullV;
4623 /// Rewrite - Emit instructions for the leading candidate expression for this
4624 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4625 /// the newly expanded value.
4626 void LSRInstance::Rewrite(const LSRFixup &LF,
4628 SCEVExpander &Rewriter,
4629 SmallVectorImpl<WeakVH> &DeadInsts,
4631 // First, find an insertion point that dominates UserInst. For PHI nodes,
4632 // find the nearest block which dominates all the relevant uses.
4633 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4634 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4636 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4638 // If this is reuse-by-noop-cast, insert the noop cast.
4639 Type *OpTy = LF.OperandValToReplace->getType();
4640 if (FullV->getType() != OpTy) {
4642 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4643 FullV, OpTy, "tmp", LF.UserInst);
4647 // Update the user. ICmpZero is handled specially here (for now) because
4648 // Expand may have updated one of the operands of the icmp already, and
4649 // its new value may happen to be equal to LF.OperandValToReplace, in
4650 // which case doing replaceUsesOfWith leads to replacing both operands
4651 // with the same value. TODO: Reorganize this.
4652 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4653 LF.UserInst->setOperand(0, FullV);
4655 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4658 DeadInsts.push_back(LF.OperandValToReplace);
4661 /// ImplementSolution - Rewrite all the fixup locations with new values,
4662 /// following the chosen solution.
4664 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4666 // Keep track of instructions we may have made dead, so that
4667 // we can remove them after we are done working.
4668 SmallVector<WeakVH, 16> DeadInsts;
4670 SCEVExpander Rewriter(SE, "lsr");
4672 Rewriter.setDebugType(DEBUG_TYPE);
4674 Rewriter.disableCanonicalMode();
4675 Rewriter.enableLSRMode();
4676 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4678 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4679 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4680 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4681 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4682 Rewriter.setChainedPhi(PN);
4685 // Expand the new value definitions and update the users.
4686 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4687 E = Fixups.end(); I != E; ++I) {
4688 const LSRFixup &Fixup = *I;
4690 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4695 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4696 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4697 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4700 // Clean up after ourselves. This must be done before deleting any
4704 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4707 LSRInstance::LSRInstance(Loop *L, Pass *P)
4708 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4709 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4710 LI(P->getAnalysis<LoopInfo>()),
4711 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4713 // If LoopSimplify form is not available, stay out of trouble.
4714 if (!L->isLoopSimplifyForm())
4717 // If there's no interesting work to be done, bail early.
4718 if (IU.empty()) return;
4720 // If there's too much analysis to be done, bail early. We won't be able to
4721 // model the problem anyway.
4722 unsigned NumUsers = 0;
4723 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4724 if (++NumUsers > MaxIVUsers) {
4725 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4732 // All dominating loops must have preheaders, or SCEVExpander may not be able
4733 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4735 // IVUsers analysis should only create users that are dominated by simple loop
4736 // headers. Since this loop should dominate all of its users, its user list
4737 // should be empty if this loop itself is not within a simple loop nest.
4738 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4739 Rung; Rung = Rung->getIDom()) {
4740 BasicBlock *BB = Rung->getBlock();
4741 const Loop *DomLoop = LI.getLoopFor(BB);
4742 if (DomLoop && DomLoop->getHeader() == BB) {
4743 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4748 DEBUG(dbgs() << "\nLSR on loop ";
4749 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4752 // First, perform some low-level loop optimizations.
4754 OptimizeLoopTermCond();
4756 // If loop preparation eliminates all interesting IV users, bail.
4757 if (IU.empty()) return;
4759 // Skip nested loops until we can model them better with formulae.
4761 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4765 // Start collecting data and preparing for the solver.
4767 CollectInterestingTypesAndFactors();
4768 CollectFixupsAndInitialFormulae();
4769 CollectLoopInvariantFixupsAndFormulae();
4771 assert(!Uses.empty() && "IVUsers reported at least one use");
4772 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4773 print_uses(dbgs()));
4775 // Now use the reuse data to generate a bunch of interesting ways
4776 // to formulate the values needed for the uses.
4777 GenerateAllReuseFormulae();
4779 FilterOutUndesirableDedicatedRegisters();
4780 NarrowSearchSpaceUsingHeuristics();
4782 SmallVector<const Formula *, 8> Solution;
4785 // Release memory that is no longer needed.
4790 if (Solution.empty())
4794 // Formulae should be legal.
4795 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4797 const LSRUse &LU = *I;
4798 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4799 JE = LU.Formulae.end();
4801 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4802 *J) && "Illegal formula generated!");
4806 // Now that we've decided what we want, make it so.
4807 ImplementSolution(Solution, P);
4810 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4811 if (Factors.empty() && Types.empty()) return;
4813 OS << "LSR has identified the following interesting factors and types: ";
4816 for (SmallSetVector<int64_t, 8>::const_iterator
4817 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4818 if (!First) OS << ", ";
4823 for (SmallSetVector<Type *, 4>::const_iterator
4824 I = Types.begin(), E = Types.end(); I != E; ++I) {
4825 if (!First) OS << ", ";
4827 OS << '(' << **I << ')';
4832 void LSRInstance::print_fixups(raw_ostream &OS) const {
4833 OS << "LSR is examining the following fixup sites:\n";
4834 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4835 E = Fixups.end(); I != E; ++I) {
4842 void LSRInstance::print_uses(raw_ostream &OS) const {
4843 OS << "LSR is examining the following uses:\n";
4844 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4845 E = Uses.end(); I != E; ++I) {
4846 const LSRUse &LU = *I;
4850 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4851 JE = LU.Formulae.end(); J != JE; ++J) {
4859 void LSRInstance::print(raw_ostream &OS) const {
4860 print_factors_and_types(OS);
4865 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4866 void LSRInstance::dump() const {
4867 print(errs()); errs() << '\n';
4873 class LoopStrengthReduce : public LoopPass {
4875 static char ID; // Pass ID, replacement for typeid
4876 LoopStrengthReduce();
4879 bool runOnLoop(Loop *L, LPPassManager &LPM);
4880 void getAnalysisUsage(AnalysisUsage &AU) const;
4885 char LoopStrengthReduce::ID = 0;
4886 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4887 "Loop Strength Reduction", false, false)
4888 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4889 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4890 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4891 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4892 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4893 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4894 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4895 "Loop Strength Reduction", false, false)
4898 Pass *llvm::createLoopStrengthReducePass() {
4899 return new LoopStrengthReduce();
4902 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4903 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4906 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4907 // We split critical edges, so we change the CFG. However, we do update
4908 // many analyses if they are around.
4909 AU.addPreservedID(LoopSimplifyID);
4911 AU.addRequired<LoopInfo>();
4912 AU.addPreserved<LoopInfo>();
4913 AU.addRequiredID(LoopSimplifyID);
4914 AU.addRequired<DominatorTreeWrapperPass>();
4915 AU.addPreserved<DominatorTreeWrapperPass>();
4916 AU.addRequired<ScalarEvolution>();
4917 AU.addPreserved<ScalarEvolution>();
4918 // Requiring LoopSimplify a second time here prevents IVUsers from running
4919 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4920 AU.addRequiredID(LoopSimplifyID);
4921 AU.addRequired<IVUsers>();
4922 AU.addPreserved<IVUsers>();
4923 AU.addRequired<TargetTransformInfo>();
4926 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4927 if (skipOptnoneFunction(L))
4930 bool Changed = false;
4932 // Run the main LSR transformation.
4933 Changed |= LSRInstance(L, this).getChanged();
4935 // Remove any extra phis created by processing inner loops.
4936 Changed |= DeleteDeadPHIs(L->getHeader());
4937 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4938 SmallVector<WeakVH, 16> DeadInsts;
4939 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4941 Rewriter.setDebugType(DEBUG_TYPE);
4943 unsigned numFolded = Rewriter.replaceCongruentIVs(
4944 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
4945 &getAnalysis<TargetTransformInfo>());
4948 DeleteTriviallyDeadInstructions(DeadInsts);
4949 DeleteDeadPHIs(L->getHeader());