1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Assembly/Writer.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/ADT/SmallBitVector.h"
70 #include "llvm/ADT/SetVector.h"
71 #include "llvm/ADT/DenseSet.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Target/TargetLowering.h"
80 static cl::opt<bool> EnableNested(
81 "enable-lsr-nested", cl::Hidden, cl::desc("Enable LSR on nested loops"));
83 static cl::opt<bool> EnableRetry(
84 "enable-lsr-retry", cl::Hidden, cl::desc("Enable LSR retry"));
86 // Temporary flag to cleanup congruent phis after LSR phi expansion.
87 // It's currently disabled until we can determine whether it's truly useful or
88 // not. The flag should be removed after the v3.0 release.
89 // This is now needed for ivchains.
90 static cl::opt<bool> EnablePhiElim(
91 "enable-lsr-phielim", cl::Hidden, cl::init(true),
92 cl::desc("Enable LSR phi elimination"));
95 // Stress test IV chain generation.
96 static cl::opt<bool> StressIVChain(
97 "stress-ivchain", cl::Hidden, cl::init(false),
98 cl::desc("Stress test LSR IV chains"));
100 static bool StressIVChain = false;
105 /// RegSortData - This class holds data which is used to order reuse candidates.
108 /// UsedByIndices - This represents the set of LSRUse indices which reference
109 /// a particular register.
110 SmallBitVector UsedByIndices;
114 void print(raw_ostream &OS) const;
120 void RegSortData::print(raw_ostream &OS) const {
121 OS << "[NumUses=" << UsedByIndices.count() << ']';
124 void RegSortData::dump() const {
125 print(errs()); errs() << '\n';
130 /// RegUseTracker - Map register candidates to information about how they are
132 class RegUseTracker {
133 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
135 RegUsesTy RegUsesMap;
136 SmallVector<const SCEV *, 16> RegSequence;
139 void CountRegister(const SCEV *Reg, size_t LUIdx);
140 void DropRegister(const SCEV *Reg, size_t LUIdx);
141 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
143 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
145 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
149 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
150 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
151 iterator begin() { return RegSequence.begin(); }
152 iterator end() { return RegSequence.end(); }
153 const_iterator begin() const { return RegSequence.begin(); }
154 const_iterator end() const { return RegSequence.end(); }
160 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
161 std::pair<RegUsesTy::iterator, bool> Pair =
162 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
163 RegSortData &RSD = Pair.first->second;
165 RegSequence.push_back(Reg);
166 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
167 RSD.UsedByIndices.set(LUIdx);
171 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
172 RegUsesTy::iterator It = RegUsesMap.find(Reg);
173 assert(It != RegUsesMap.end());
174 RegSortData &RSD = It->second;
175 assert(RSD.UsedByIndices.size() > LUIdx);
176 RSD.UsedByIndices.reset(LUIdx);
180 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
181 assert(LUIdx <= LastLUIdx);
183 // Update RegUses. The data structure is not optimized for this purpose;
184 // we must iterate through it and update each of the bit vectors.
185 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
187 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
188 if (LUIdx < UsedByIndices.size())
189 UsedByIndices[LUIdx] =
190 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
191 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
196 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
197 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
198 if (I == RegUsesMap.end())
200 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
201 int i = UsedByIndices.find_first();
202 if (i == -1) return false;
203 if ((size_t)i != LUIdx) return true;
204 return UsedByIndices.find_next(i) != -1;
207 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
208 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
209 assert(I != RegUsesMap.end() && "Unknown register!");
210 return I->second.UsedByIndices;
213 void RegUseTracker::clear() {
220 /// Formula - This class holds information that describes a formula for
221 /// computing satisfying a use. It may include broken-out immediates and scaled
224 /// AM - This is used to represent complex addressing, as well as other kinds
225 /// of interesting uses.
226 TargetLowering::AddrMode AM;
228 /// BaseRegs - The list of "base" registers for this use. When this is
229 /// non-empty, AM.HasBaseReg should be set to true.
230 SmallVector<const SCEV *, 2> BaseRegs;
232 /// ScaledReg - The 'scaled' register for this use. This should be non-null
233 /// when AM.Scale is not zero.
234 const SCEV *ScaledReg;
236 /// UnfoldedOffset - An additional constant offset which added near the
237 /// use. This requires a temporary register, but the offset itself can
238 /// live in an add immediate field rather than a register.
239 int64_t UnfoldedOffset;
241 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
243 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
245 unsigned getNumRegs() const;
246 Type *getType() const;
248 void DeleteBaseReg(const SCEV *&S);
250 bool referencesReg(const SCEV *S) const;
251 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
252 const RegUseTracker &RegUses) const;
254 void print(raw_ostream &OS) const;
260 /// DoInitialMatch - Recursion helper for InitialMatch.
261 static void DoInitialMatch(const SCEV *S, Loop *L,
262 SmallVectorImpl<const SCEV *> &Good,
263 SmallVectorImpl<const SCEV *> &Bad,
264 ScalarEvolution &SE) {
265 // Collect expressions which properly dominate the loop header.
266 if (SE.properlyDominates(S, L->getHeader())) {
271 // Look at add operands.
272 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
273 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
275 DoInitialMatch(*I, L, Good, Bad, SE);
279 // Look at addrec operands.
280 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
281 if (!AR->getStart()->isZero()) {
282 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
283 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
284 AR->getStepRecurrence(SE),
285 // FIXME: AR->getNoWrapFlags()
286 AR->getLoop(), SCEV::FlagAnyWrap),
291 // Handle a multiplication by -1 (negation) if it didn't fold.
292 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
293 if (Mul->getOperand(0)->isAllOnesValue()) {
294 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
295 const SCEV *NewMul = SE.getMulExpr(Ops);
297 SmallVector<const SCEV *, 4> MyGood;
298 SmallVector<const SCEV *, 4> MyBad;
299 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
300 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
301 SE.getEffectiveSCEVType(NewMul->getType())));
302 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
303 E = MyGood.end(); I != E; ++I)
304 Good.push_back(SE.getMulExpr(NegOne, *I));
305 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
306 E = MyBad.end(); I != E; ++I)
307 Bad.push_back(SE.getMulExpr(NegOne, *I));
311 // Ok, we can't do anything interesting. Just stuff the whole thing into a
312 // register and hope for the best.
316 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
317 /// attempting to keep all loop-invariant and loop-computable values in a
318 /// single base register.
319 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
320 SmallVector<const SCEV *, 4> Good;
321 SmallVector<const SCEV *, 4> Bad;
322 DoInitialMatch(S, L, Good, Bad, SE);
324 const SCEV *Sum = SE.getAddExpr(Good);
326 BaseRegs.push_back(Sum);
327 AM.HasBaseReg = true;
330 const SCEV *Sum = SE.getAddExpr(Bad);
332 BaseRegs.push_back(Sum);
333 AM.HasBaseReg = true;
337 /// getNumRegs - Return the total number of register operands used by this
338 /// formula. This does not include register uses implied by non-constant
340 unsigned Formula::getNumRegs() const {
341 return !!ScaledReg + BaseRegs.size();
344 /// getType - Return the type of this formula, if it has one, or null
345 /// otherwise. This type is meaningless except for the bit size.
346 Type *Formula::getType() const {
347 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
348 ScaledReg ? ScaledReg->getType() :
349 AM.BaseGV ? AM.BaseGV->getType() :
353 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
354 void Formula::DeleteBaseReg(const SCEV *&S) {
355 if (&S != &BaseRegs.back())
356 std::swap(S, BaseRegs.back());
360 /// referencesReg - Test if this formula references the given register.
361 bool Formula::referencesReg(const SCEV *S) const {
362 return S == ScaledReg ||
363 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
366 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
367 /// which are used by uses other than the use with the given index.
368 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
369 const RegUseTracker &RegUses) const {
371 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
373 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
374 E = BaseRegs.end(); I != E; ++I)
375 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
380 void Formula::print(raw_ostream &OS) const {
383 if (!First) OS << " + "; else First = false;
384 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
386 if (AM.BaseOffs != 0) {
387 if (!First) OS << " + "; else First = false;
390 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
391 E = BaseRegs.end(); I != E; ++I) {
392 if (!First) OS << " + "; else First = false;
393 OS << "reg(" << **I << ')';
395 if (AM.HasBaseReg && BaseRegs.empty()) {
396 if (!First) OS << " + "; else First = false;
397 OS << "**error: HasBaseReg**";
398 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
399 if (!First) OS << " + "; else First = false;
400 OS << "**error: !HasBaseReg**";
403 if (!First) OS << " + "; else First = false;
404 OS << AM.Scale << "*reg(";
411 if (UnfoldedOffset != 0) {
412 if (!First) OS << " + "; else First = false;
413 OS << "imm(" << UnfoldedOffset << ')';
417 void Formula::dump() const {
418 print(errs()); errs() << '\n';
421 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
422 /// without changing its value.
423 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
425 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
426 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
429 /// isAddSExtable - Return true if the given add can be sign-extended
430 /// without changing its value.
431 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
433 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
434 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
437 /// isMulSExtable - Return true if the given mul can be sign-extended
438 /// without changing its value.
439 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
441 IntegerType::get(SE.getContext(),
442 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
443 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
446 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
447 /// and if the remainder is known to be zero, or null otherwise. If
448 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
449 /// to Y, ignoring that the multiplication may overflow, which is useful when
450 /// the result will be used in a context where the most significant bits are
452 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
454 bool IgnoreSignificantBits = false) {
455 // Handle the trivial case, which works for any SCEV type.
457 return SE.getConstant(LHS->getType(), 1);
459 // Handle a few RHS special cases.
460 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
462 const APInt &RA = RC->getValue()->getValue();
463 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
465 if (RA.isAllOnesValue())
466 return SE.getMulExpr(LHS, RC);
467 // Handle x /s 1 as x.
472 // Check for a division of a constant by a constant.
473 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
476 const APInt &LA = C->getValue()->getValue();
477 const APInt &RA = RC->getValue()->getValue();
478 if (LA.srem(RA) != 0)
480 return SE.getConstant(LA.sdiv(RA));
483 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
484 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
485 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
486 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
487 IgnoreSignificantBits);
489 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
490 IgnoreSignificantBits);
491 if (!Start) return 0;
492 // FlagNW is independent of the start value, step direction, and is
493 // preserved with smaller magnitude steps.
494 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
495 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
500 // Distribute the sdiv over add operands, if the add doesn't overflow.
501 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
502 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
503 SmallVector<const SCEV *, 8> Ops;
504 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
506 const SCEV *Op = getExactSDiv(*I, RHS, SE,
507 IgnoreSignificantBits);
511 return SE.getAddExpr(Ops);
516 // Check for a multiply operand that we can pull RHS out of.
517 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
518 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
519 SmallVector<const SCEV *, 4> Ops;
521 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
525 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
526 IgnoreSignificantBits)) {
532 return Found ? SE.getMulExpr(Ops) : 0;
537 // Otherwise we don't know.
541 /// ExtractImmediate - If S involves the addition of a constant integer value,
542 /// return that integer value, and mutate S to point to a new SCEV with that
544 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
545 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
546 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
547 S = SE.getConstant(C->getType(), 0);
548 return C->getValue()->getSExtValue();
550 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
551 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
552 int64_t Result = ExtractImmediate(NewOps.front(), SE);
554 S = SE.getAddExpr(NewOps);
556 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
557 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
558 int64_t Result = ExtractImmediate(NewOps.front(), SE);
560 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
561 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
568 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
569 /// return that symbol, and mutate S to point to a new SCEV with that
571 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
572 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
573 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
574 S = SE.getConstant(GV->getType(), 0);
577 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
578 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
579 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
581 S = SE.getAddExpr(NewOps);
583 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
584 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
585 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
587 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
588 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
595 /// isAddressUse - Returns true if the specified instruction is using the
596 /// specified value as an address.
597 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
598 bool isAddress = isa<LoadInst>(Inst);
599 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
600 if (SI->getOperand(1) == OperandVal)
602 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
603 // Addressing modes can also be folded into prefetches and a variety
605 switch (II->getIntrinsicID()) {
607 case Intrinsic::prefetch:
608 case Intrinsic::x86_sse_storeu_ps:
609 case Intrinsic::x86_sse2_storeu_pd:
610 case Intrinsic::x86_sse2_storeu_dq:
611 case Intrinsic::x86_sse2_storel_dq:
612 if (II->getArgOperand(0) == OperandVal)
620 /// getAccessType - Return the type of the memory being accessed.
621 static Type *getAccessType(const Instruction *Inst) {
622 Type *AccessTy = Inst->getType();
623 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
624 AccessTy = SI->getOperand(0)->getType();
625 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
626 // Addressing modes can also be folded into prefetches and a variety
628 switch (II->getIntrinsicID()) {
630 case Intrinsic::x86_sse_storeu_ps:
631 case Intrinsic::x86_sse2_storeu_pd:
632 case Intrinsic::x86_sse2_storeu_dq:
633 case Intrinsic::x86_sse2_storel_dq:
634 AccessTy = II->getArgOperand(0)->getType();
639 // All pointers have the same requirements, so canonicalize them to an
640 // arbitrary pointer type to minimize variation.
641 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
642 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
643 PTy->getAddressSpace());
648 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
649 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
650 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
651 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
652 if (SE.isSCEVable(PN->getType()) &&
653 (SE.getEffectiveSCEVType(PN->getType()) ==
654 SE.getEffectiveSCEVType(AR->getType())) &&
655 SE.getSCEV(PN) == AR)
661 /// Check if expanding this expression is likely to incur significant cost. This
662 /// is tricky because SCEV doesn't track which expressions are actually computed
663 /// by the current IR.
665 /// We currently allow expansion of IV increments that involve adds,
666 /// multiplication by constants, and AddRecs from existing phis.
668 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
669 /// obvious multiple of the UDivExpr.
670 static bool isHighCostExpansion(const SCEV *S,
671 SmallPtrSet<const SCEV*, 8> &Processed,
672 ScalarEvolution &SE) {
673 // Zero/One operand expressions
674 switch (S->getSCEVType()) {
679 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
682 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
685 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
689 if (!Processed.insert(S))
692 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
693 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
695 if (isHighCostExpansion(*I, Processed, SE))
701 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
702 if (Mul->getNumOperands() == 2) {
703 // Multiplication by a constant is ok
704 if (isa<SCEVConstant>(Mul->getOperand(0)))
705 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
707 // If we have the value of one operand, check if an existing
708 // multiplication already generates this expression.
709 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
710 Value *UVal = U->getValue();
711 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
713 Instruction *User = cast<Instruction>(*UI);
714 if (User->getOpcode() == Instruction::Mul
715 && SE.isSCEVable(User->getType())) {
716 return SE.getSCEV(User) == Mul;
723 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
724 if (isExistingPhi(AR, SE))
728 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
732 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
733 /// specified set are trivially dead, delete them and see if this makes any of
734 /// their operands subsequently dead.
736 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
737 bool Changed = false;
739 while (!DeadInsts.empty()) {
740 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
742 if (I == 0 || !isInstructionTriviallyDead(I))
745 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
746 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
749 DeadInsts.push_back(U);
752 I->eraseFromParent();
761 /// Cost - This class is used to measure and compare candidate formulae.
763 /// TODO: Some of these could be merged. Also, a lexical ordering
764 /// isn't always optimal.
768 unsigned NumBaseAdds;
774 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
777 bool operator<(const Cost &Other) const;
782 // Once any of the metrics loses, they must all remain losers.
784 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
785 | ImmCost | SetupCost) != ~0u)
786 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
787 & ImmCost & SetupCost) == ~0u);
792 assert(isValid() && "invalid cost");
793 return NumRegs == ~0u;
796 void RateFormula(const Formula &F,
797 SmallPtrSet<const SCEV *, 16> &Regs,
798 const DenseSet<const SCEV *> &VisitedRegs,
800 const SmallVectorImpl<int64_t> &Offsets,
801 ScalarEvolution &SE, DominatorTree &DT,
802 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
804 void print(raw_ostream &OS) const;
808 void RateRegister(const SCEV *Reg,
809 SmallPtrSet<const SCEV *, 16> &Regs,
811 ScalarEvolution &SE, DominatorTree &DT);
812 void RatePrimaryRegister(const SCEV *Reg,
813 SmallPtrSet<const SCEV *, 16> &Regs,
815 ScalarEvolution &SE, DominatorTree &DT,
816 SmallPtrSet<const SCEV *, 16> *LoserRegs);
821 /// RateRegister - Tally up interesting quantities from the given register.
822 void Cost::RateRegister(const SCEV *Reg,
823 SmallPtrSet<const SCEV *, 16> &Regs,
825 ScalarEvolution &SE, DominatorTree &DT) {
826 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
827 if (AR->getLoop() == L)
828 AddRecCost += 1; /// TODO: This should be a function of the stride.
830 // If this is an addrec for another loop, don't second-guess its addrec phi
831 // nodes. LSR isn't currently smart enough to reason about more than one
832 // loop at a time. LSR has either already run on inner loops, will not run
833 // on other loops, and cannot be expected to change sibling loops. If the
834 // AddRec exists, consider it's register free and leave it alone. Otherwise,
835 // do not consider this formula at all.
836 else if (!EnableNested || L->contains(AR->getLoop()) ||
837 (!AR->getLoop()->contains(L) &&
838 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
839 if (isExistingPhi(AR, SE))
842 // For !EnableNested, never rewrite IVs in other loops.
847 // If this isn't one of the addrecs that the loop already has, it
848 // would require a costly new phi and add. TODO: This isn't
849 // precisely modeled right now.
851 if (!Regs.count(AR->getStart())) {
852 RateRegister(AR->getStart(), Regs, L, SE, DT);
858 // Add the step value register, if it needs one.
859 // TODO: The non-affine case isn't precisely modeled here.
860 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
861 if (!Regs.count(AR->getOperand(1))) {
862 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
870 // Rough heuristic; favor registers which don't require extra setup
871 // instructions in the preheader.
872 if (!isa<SCEVUnknown>(Reg) &&
873 !isa<SCEVConstant>(Reg) &&
874 !(isa<SCEVAddRecExpr>(Reg) &&
875 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
876 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
879 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
880 SE.hasComputableLoopEvolution(Reg, L);
883 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
884 /// before, rate it. Optional LoserRegs provides a way to declare any formula
885 /// that refers to one of those regs an instant loser.
886 void Cost::RatePrimaryRegister(const SCEV *Reg,
887 SmallPtrSet<const SCEV *, 16> &Regs,
889 ScalarEvolution &SE, DominatorTree &DT,
890 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
891 if (LoserRegs && LoserRegs->count(Reg)) {
895 if (Regs.insert(Reg)) {
896 RateRegister(Reg, Regs, L, SE, DT);
898 LoserRegs->insert(Reg);
902 void Cost::RateFormula(const Formula &F,
903 SmallPtrSet<const SCEV *, 16> &Regs,
904 const DenseSet<const SCEV *> &VisitedRegs,
906 const SmallVectorImpl<int64_t> &Offsets,
907 ScalarEvolution &SE, DominatorTree &DT,
908 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
909 // Tally up the registers.
910 if (const SCEV *ScaledReg = F.ScaledReg) {
911 if (VisitedRegs.count(ScaledReg)) {
915 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
919 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
920 E = F.BaseRegs.end(); I != E; ++I) {
921 const SCEV *BaseReg = *I;
922 if (VisitedRegs.count(BaseReg)) {
926 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
931 // Determine how many (unfolded) adds we'll need inside the loop.
932 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
933 if (NumBaseParts > 1)
934 NumBaseAdds += NumBaseParts - 1;
936 // Tally up the non-zero immediates.
937 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
938 E = Offsets.end(); I != E; ++I) {
939 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
941 ImmCost += 64; // Handle symbolic values conservatively.
942 // TODO: This should probably be the pointer size.
943 else if (Offset != 0)
944 ImmCost += APInt(64, Offset, true).getMinSignedBits();
946 assert(isValid() && "invalid cost");
949 /// Loose - Set this cost to a losing value.
959 /// operator< - Choose the lower cost.
960 bool Cost::operator<(const Cost &Other) const {
961 if (NumRegs != Other.NumRegs)
962 return NumRegs < Other.NumRegs;
963 if (AddRecCost != Other.AddRecCost)
964 return AddRecCost < Other.AddRecCost;
965 if (NumIVMuls != Other.NumIVMuls)
966 return NumIVMuls < Other.NumIVMuls;
967 if (NumBaseAdds != Other.NumBaseAdds)
968 return NumBaseAdds < Other.NumBaseAdds;
969 if (ImmCost != Other.ImmCost)
970 return ImmCost < Other.ImmCost;
971 if (SetupCost != Other.SetupCost)
972 return SetupCost < Other.SetupCost;
976 void Cost::print(raw_ostream &OS) const {
977 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
979 OS << ", with addrec cost " << AddRecCost;
981 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
982 if (NumBaseAdds != 0)
983 OS << ", plus " << NumBaseAdds << " base add"
984 << (NumBaseAdds == 1 ? "" : "s");
986 OS << ", plus " << ImmCost << " imm cost";
988 OS << ", plus " << SetupCost << " setup cost";
991 void Cost::dump() const {
992 print(errs()); errs() << '\n';
997 /// LSRFixup - An operand value in an instruction which is to be replaced
998 /// with some equivalent, possibly strength-reduced, replacement.
1000 /// UserInst - The instruction which will be updated.
1001 Instruction *UserInst;
1003 /// OperandValToReplace - The operand of the instruction which will
1004 /// be replaced. The operand may be used more than once; every instance
1005 /// will be replaced.
1006 Value *OperandValToReplace;
1008 /// PostIncLoops - If this user is to use the post-incremented value of an
1009 /// induction variable, this variable is non-null and holds the loop
1010 /// associated with the induction variable.
1011 PostIncLoopSet PostIncLoops;
1013 /// LUIdx - The index of the LSRUse describing the expression which
1014 /// this fixup needs, minus an offset (below).
1017 /// Offset - A constant offset to be added to the LSRUse expression.
1018 /// This allows multiple fixups to share the same LSRUse with different
1019 /// offsets, for example in an unrolled loop.
1022 bool isUseFullyOutsideLoop(const Loop *L) const;
1026 void print(raw_ostream &OS) const;
1032 LSRFixup::LSRFixup()
1033 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1035 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1036 /// value outside of the given loop.
1037 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1038 // PHI nodes use their value in their incoming blocks.
1039 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1040 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1041 if (PN->getIncomingValue(i) == OperandValToReplace &&
1042 L->contains(PN->getIncomingBlock(i)))
1047 return !L->contains(UserInst);
1050 void LSRFixup::print(raw_ostream &OS) const {
1052 // Store is common and interesting enough to be worth special-casing.
1053 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1055 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1056 } else if (UserInst->getType()->isVoidTy())
1057 OS << UserInst->getOpcodeName();
1059 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1061 OS << ", OperandValToReplace=";
1062 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1064 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1065 E = PostIncLoops.end(); I != E; ++I) {
1066 OS << ", PostIncLoop=";
1067 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1070 if (LUIdx != ~size_t(0))
1071 OS << ", LUIdx=" << LUIdx;
1074 OS << ", Offset=" << Offset;
1077 void LSRFixup::dump() const {
1078 print(errs()); errs() << '\n';
1083 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1084 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1085 struct UniquifierDenseMapInfo {
1086 static SmallVector<const SCEV *, 2> getEmptyKey() {
1087 SmallVector<const SCEV *, 2> V;
1088 V.push_back(reinterpret_cast<const SCEV *>(-1));
1092 static SmallVector<const SCEV *, 2> getTombstoneKey() {
1093 SmallVector<const SCEV *, 2> V;
1094 V.push_back(reinterpret_cast<const SCEV *>(-2));
1098 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
1099 unsigned Result = 0;
1100 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1101 E = V.end(); I != E; ++I)
1102 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1106 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1107 const SmallVector<const SCEV *, 2> &RHS) {
1112 /// LSRUse - This class holds the state that LSR keeps for each use in
1113 /// IVUsers, as well as uses invented by LSR itself. It includes information
1114 /// about what kinds of things can be folded into the user, information about
1115 /// the user itself, and information about how the use may be satisfied.
1116 /// TODO: Represent multiple users of the same expression in common?
1118 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1121 /// KindType - An enum for a kind of use, indicating what types of
1122 /// scaled and immediate operands it might support.
1124 Basic, ///< A normal use, with no folding.
1125 Special, ///< A special case of basic, allowing -1 scales.
1126 Address, ///< An address use; folding according to TargetLowering
1127 ICmpZero ///< An equality icmp with both operands folded into one.
1128 // TODO: Add a generic icmp too?
1134 SmallVector<int64_t, 8> Offsets;
1138 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1139 /// LSRUse are outside of the loop, in which case some special-case heuristics
1141 bool AllFixupsOutsideLoop;
1143 /// WidestFixupType - This records the widest use type for any fixup using
1144 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1145 /// max fixup widths to be equivalent, because the narrower one may be relying
1146 /// on the implicit truncation to truncate away bogus bits.
1147 Type *WidestFixupType;
1149 /// Formulae - A list of ways to build a value that can satisfy this user.
1150 /// After the list is populated, one of these is selected heuristically and
1151 /// used to formulate a replacement for OperandValToReplace in UserInst.
1152 SmallVector<Formula, 12> Formulae;
1154 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1155 SmallPtrSet<const SCEV *, 4> Regs;
1157 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1158 MinOffset(INT64_MAX),
1159 MaxOffset(INT64_MIN),
1160 AllFixupsOutsideLoop(true),
1161 WidestFixupType(0) {}
1163 bool HasFormulaWithSameRegs(const Formula &F) const;
1164 bool InsertFormula(const Formula &F);
1165 void DeleteFormula(Formula &F);
1166 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1168 void print(raw_ostream &OS) const;
1174 /// HasFormula - Test whether this use as a formula which has the same
1175 /// registers as the given formula.
1176 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1177 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1178 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1179 // Unstable sort by host order ok, because this is only used for uniquifying.
1180 std::sort(Key.begin(), Key.end());
1181 return Uniquifier.count(Key);
1184 /// InsertFormula - If the given formula has not yet been inserted, add it to
1185 /// the list, and return true. Return false otherwise.
1186 bool LSRUse::InsertFormula(const Formula &F) {
1187 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1188 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1189 // Unstable sort by host order ok, because this is only used for uniquifying.
1190 std::sort(Key.begin(), Key.end());
1192 if (!Uniquifier.insert(Key).second)
1195 // Using a register to hold the value of 0 is not profitable.
1196 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1197 "Zero allocated in a scaled register!");
1199 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1200 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1201 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1204 // Add the formula to the list.
1205 Formulae.push_back(F);
1207 // Record registers now being used by this use.
1208 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1213 /// DeleteFormula - Remove the given formula from this use's list.
1214 void LSRUse::DeleteFormula(Formula &F) {
1215 if (&F != &Formulae.back())
1216 std::swap(F, Formulae.back());
1217 Formulae.pop_back();
1220 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1221 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1222 // Now that we've filtered out some formulae, recompute the Regs set.
1223 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1225 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1226 E = Formulae.end(); I != E; ++I) {
1227 const Formula &F = *I;
1228 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1229 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1232 // Update the RegTracker.
1233 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1234 E = OldRegs.end(); I != E; ++I)
1235 if (!Regs.count(*I))
1236 RegUses.DropRegister(*I, LUIdx);
1239 void LSRUse::print(raw_ostream &OS) const {
1240 OS << "LSR Use: Kind=";
1242 case Basic: OS << "Basic"; break;
1243 case Special: OS << "Special"; break;
1244 case ICmpZero: OS << "ICmpZero"; break;
1246 OS << "Address of ";
1247 if (AccessTy->isPointerTy())
1248 OS << "pointer"; // the full pointer type could be really verbose
1253 OS << ", Offsets={";
1254 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1255 E = Offsets.end(); I != E; ++I) {
1257 if (llvm::next(I) != E)
1262 if (AllFixupsOutsideLoop)
1263 OS << ", all-fixups-outside-loop";
1265 if (WidestFixupType)
1266 OS << ", widest fixup type: " << *WidestFixupType;
1269 void LSRUse::dump() const {
1270 print(errs()); errs() << '\n';
1273 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1274 /// be completely folded into the user instruction at isel time. This includes
1275 /// address-mode folding and special icmp tricks.
1276 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1277 LSRUse::KindType Kind, Type *AccessTy,
1278 const TargetLowering *TLI) {
1280 case LSRUse::Address:
1281 // If we have low-level target information, ask the target if it can
1282 // completely fold this address.
1283 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1285 // Otherwise, just guess that reg+reg addressing is legal.
1286 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1288 case LSRUse::ICmpZero:
1289 // There's not even a target hook for querying whether it would be legal to
1290 // fold a GV into an ICmp.
1294 // ICmp only has two operands; don't allow more than two non-trivial parts.
1295 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1298 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1299 // putting the scaled register in the other operand of the icmp.
1300 if (AM.Scale != 0 && AM.Scale != -1)
1303 // If we have low-level target information, ask the target if it can fold an
1304 // integer immediate on an icmp.
1305 if (AM.BaseOffs != 0) {
1306 if (TLI) return TLI->isLegalICmpImmediate(-(uint64_t)AM.BaseOffs);
1313 // Only handle single-register values.
1314 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1316 case LSRUse::Special:
1317 // Only handle -1 scales, or no scale.
1318 return AM.Scale == 0 || AM.Scale == -1;
1324 static bool isLegalUse(TargetLowering::AddrMode AM,
1325 int64_t MinOffset, int64_t MaxOffset,
1326 LSRUse::KindType Kind, Type *AccessTy,
1327 const TargetLowering *TLI) {
1328 // Check for overflow.
1329 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1332 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1333 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1334 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1335 // Check for overflow.
1336 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1339 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1340 return isLegalUse(AM, Kind, AccessTy, TLI);
1345 static bool isAlwaysFoldable(int64_t BaseOffs,
1346 GlobalValue *BaseGV,
1348 LSRUse::KindType Kind, Type *AccessTy,
1349 const TargetLowering *TLI) {
1350 // Fast-path: zero is always foldable.
1351 if (BaseOffs == 0 && !BaseGV) return true;
1353 // Conservatively, create an address with an immediate and a
1354 // base and a scale.
1355 TargetLowering::AddrMode AM;
1356 AM.BaseOffs = BaseOffs;
1358 AM.HasBaseReg = HasBaseReg;
1359 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1361 // Canonicalize a scale of 1 to a base register if the formula doesn't
1362 // already have a base register.
1363 if (!AM.HasBaseReg && AM.Scale == 1) {
1365 AM.HasBaseReg = true;
1368 return isLegalUse(AM, Kind, AccessTy, TLI);
1371 static bool isAlwaysFoldable(const SCEV *S,
1372 int64_t MinOffset, int64_t MaxOffset,
1374 LSRUse::KindType Kind, Type *AccessTy,
1375 const TargetLowering *TLI,
1376 ScalarEvolution &SE) {
1377 // Fast-path: zero is always foldable.
1378 if (S->isZero()) return true;
1380 // Conservatively, create an address with an immediate and a
1381 // base and a scale.
1382 int64_t BaseOffs = ExtractImmediate(S, SE);
1383 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1385 // If there's anything else involved, it's not foldable.
1386 if (!S->isZero()) return false;
1388 // Fast-path: zero is always foldable.
1389 if (BaseOffs == 0 && !BaseGV) return true;
1391 // Conservatively, create an address with an immediate and a
1392 // base and a scale.
1393 TargetLowering::AddrMode AM;
1394 AM.BaseOffs = BaseOffs;
1396 AM.HasBaseReg = HasBaseReg;
1397 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1399 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1404 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1405 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1406 struct UseMapDenseMapInfo {
1407 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1408 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1411 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1412 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1416 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1417 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1418 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1422 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1423 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1428 /// IVInc - An individual increment in a Chain of IV increments.
1429 /// Relate an IV user to an expression that computes the IV it uses from the IV
1430 /// used by the previous link in the Chain.
1432 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1433 /// original IVOperand. The head of the chain's IVOperand is only valid during
1434 /// chain collection, before LSR replaces IV users. During chain generation,
1435 /// IncExpr can be used to find the new IVOperand that computes the same
1438 Instruction *UserInst;
1440 const SCEV *IncExpr;
1442 IVInc(Instruction *U, Value *O, const SCEV *E):
1443 UserInst(U), IVOperand(O), IncExpr(E) {}
1446 // IVChain - The list of IV increments in program order.
1447 // We typically add the head of a chain without finding subsequent links.
1448 typedef SmallVector<IVInc,1> IVChain;
1450 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1451 /// Distinguish between FarUsers that definitely cross IV increments and
1452 /// NearUsers that may be used between IV increments.
1454 SmallPtrSet<Instruction*, 4> FarUsers;
1455 SmallPtrSet<Instruction*, 4> NearUsers;
1458 /// LSRInstance - This class holds state for the main loop strength reduction
1462 ScalarEvolution &SE;
1465 const TargetLowering *const TLI;
1469 /// IVIncInsertPos - This is the insert position that the current loop's
1470 /// induction variable increment should be placed. In simple loops, this is
1471 /// the latch block's terminator. But in more complicated cases, this is a
1472 /// position which will dominate all the in-loop post-increment users.
1473 Instruction *IVIncInsertPos;
1475 /// Factors - Interesting factors between use strides.
1476 SmallSetVector<int64_t, 8> Factors;
1478 /// Types - Interesting use types, to facilitate truncation reuse.
1479 SmallSetVector<Type *, 4> Types;
1481 /// Fixups - The list of operands which are to be replaced.
1482 SmallVector<LSRFixup, 16> Fixups;
1484 /// Uses - The list of interesting uses.
1485 SmallVector<LSRUse, 16> Uses;
1487 /// RegUses - Track which uses use which register candidates.
1488 RegUseTracker RegUses;
1490 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1491 // have more than a few IV increment chains in a loop. Missing a Chain falls
1492 // back to normal LSR behavior for those uses.
1493 static const unsigned MaxChains = 8;
1495 /// IVChainVec - IV users can form a chain of IV increments.
1496 SmallVector<IVChain, MaxChains> IVChainVec;
1498 /// IVIncSet - IV users that belong to profitable IVChains.
1499 SmallPtrSet<Use*, MaxChains> IVIncSet;
1501 void OptimizeShadowIV();
1502 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1503 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1504 void OptimizeLoopTermCond();
1506 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1507 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1508 void FinalizeChain(IVChain &Chain);
1509 void CollectChains();
1510 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1511 SmallVectorImpl<WeakVH> &DeadInsts);
1513 void CollectInterestingTypesAndFactors();
1514 void CollectFixupsAndInitialFormulae();
1516 LSRFixup &getNewFixup() {
1517 Fixups.push_back(LSRFixup());
1518 return Fixups.back();
1521 // Support for sharing of LSRUses between LSRFixups.
1522 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1524 UseMapDenseMapInfo> UseMapTy;
1527 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1528 LSRUse::KindType Kind, Type *AccessTy);
1530 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1531 LSRUse::KindType Kind,
1534 void DeleteUse(LSRUse &LU, size_t LUIdx);
1536 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1538 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1539 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1540 void CountRegisters(const Formula &F, size_t LUIdx);
1541 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1543 void CollectLoopInvariantFixupsAndFormulae();
1545 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1546 unsigned Depth = 0);
1547 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1548 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1549 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1550 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1551 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1552 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1553 void GenerateCrossUseConstantOffsets();
1554 void GenerateAllReuseFormulae();
1556 void FilterOutUndesirableDedicatedRegisters();
1558 size_t EstimateSearchSpaceComplexity() const;
1559 void NarrowSearchSpaceByDetectingSupersets();
1560 void NarrowSearchSpaceByCollapsingUnrolledCode();
1561 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1562 void NarrowSearchSpaceByPickingWinnerRegs();
1563 void NarrowSearchSpaceUsingHeuristics();
1565 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1567 SmallVectorImpl<const Formula *> &Workspace,
1568 const Cost &CurCost,
1569 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1570 DenseSet<const SCEV *> &VisitedRegs) const;
1571 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1573 BasicBlock::iterator
1574 HoistInsertPosition(BasicBlock::iterator IP,
1575 const SmallVectorImpl<Instruction *> &Inputs) const;
1576 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1578 const LSRUse &LU) const;
1580 Value *Expand(const LSRFixup &LF,
1582 BasicBlock::iterator IP,
1583 SCEVExpander &Rewriter,
1584 SmallVectorImpl<WeakVH> &DeadInsts) const;
1585 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1587 SCEVExpander &Rewriter,
1588 SmallVectorImpl<WeakVH> &DeadInsts,
1590 void Rewrite(const LSRFixup &LF,
1592 SCEVExpander &Rewriter,
1593 SmallVectorImpl<WeakVH> &DeadInsts,
1595 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1599 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1601 bool getChanged() const { return Changed; }
1603 void print_factors_and_types(raw_ostream &OS) const;
1604 void print_fixups(raw_ostream &OS) const;
1605 void print_uses(raw_ostream &OS) const;
1606 void print(raw_ostream &OS) const;
1612 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1613 /// inside the loop then try to eliminate the cast operation.
1614 void LSRInstance::OptimizeShadowIV() {
1615 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1616 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1619 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1620 UI != E; /* empty */) {
1621 IVUsers::const_iterator CandidateUI = UI;
1623 Instruction *ShadowUse = CandidateUI->getUser();
1624 Type *DestTy = NULL;
1625 bool IsSigned = false;
1627 /* If shadow use is a int->float cast then insert a second IV
1628 to eliminate this cast.
1630 for (unsigned i = 0; i < n; ++i)
1636 for (unsigned i = 0; i < n; ++i, ++d)
1639 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1641 DestTy = UCast->getDestTy();
1643 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1645 DestTy = SCast->getDestTy();
1647 if (!DestTy) continue;
1650 // If target does not support DestTy natively then do not apply
1651 // this transformation.
1652 EVT DVT = TLI->getValueType(DestTy);
1653 if (!TLI->isTypeLegal(DVT)) continue;
1656 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1658 if (PH->getNumIncomingValues() != 2) continue;
1660 Type *SrcTy = PH->getType();
1661 int Mantissa = DestTy->getFPMantissaWidth();
1662 if (Mantissa == -1) continue;
1663 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1666 unsigned Entry, Latch;
1667 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1675 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1676 if (!Init) continue;
1677 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1678 (double)Init->getSExtValue() :
1679 (double)Init->getZExtValue());
1681 BinaryOperator *Incr =
1682 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1683 if (!Incr) continue;
1684 if (Incr->getOpcode() != Instruction::Add
1685 && Incr->getOpcode() != Instruction::Sub)
1688 /* Initialize new IV, double d = 0.0 in above example. */
1689 ConstantInt *C = NULL;
1690 if (Incr->getOperand(0) == PH)
1691 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1692 else if (Incr->getOperand(1) == PH)
1693 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1699 // Ignore negative constants, as the code below doesn't handle them
1700 // correctly. TODO: Remove this restriction.
1701 if (!C->getValue().isStrictlyPositive()) continue;
1703 /* Add new PHINode. */
1704 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1706 /* create new increment. '++d' in above example. */
1707 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1708 BinaryOperator *NewIncr =
1709 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1710 Instruction::FAdd : Instruction::FSub,
1711 NewPH, CFP, "IV.S.next.", Incr);
1713 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1714 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1716 /* Remove cast operation */
1717 ShadowUse->replaceAllUsesWith(NewPH);
1718 ShadowUse->eraseFromParent();
1724 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1725 /// set the IV user and stride information and return true, otherwise return
1727 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1728 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1729 if (UI->getUser() == Cond) {
1730 // NOTE: we could handle setcc instructions with multiple uses here, but
1731 // InstCombine does it as well for simple uses, it's not clear that it
1732 // occurs enough in real life to handle.
1739 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1740 /// a max computation.
1742 /// This is a narrow solution to a specific, but acute, problem. For loops
1748 /// } while (++i < n);
1750 /// the trip count isn't just 'n', because 'n' might not be positive. And
1751 /// unfortunately this can come up even for loops where the user didn't use
1752 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1753 /// will commonly be lowered like this:
1759 /// } while (++i < n);
1762 /// and then it's possible for subsequent optimization to obscure the if
1763 /// test in such a way that indvars can't find it.
1765 /// When indvars can't find the if test in loops like this, it creates a
1766 /// max expression, which allows it to give the loop a canonical
1767 /// induction variable:
1770 /// max = n < 1 ? 1 : n;
1773 /// } while (++i != max);
1775 /// Canonical induction variables are necessary because the loop passes
1776 /// are designed around them. The most obvious example of this is the
1777 /// LoopInfo analysis, which doesn't remember trip count values. It
1778 /// expects to be able to rediscover the trip count each time it is
1779 /// needed, and it does this using a simple analysis that only succeeds if
1780 /// the loop has a canonical induction variable.
1782 /// However, when it comes time to generate code, the maximum operation
1783 /// can be quite costly, especially if it's inside of an outer loop.
1785 /// This function solves this problem by detecting this type of loop and
1786 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1787 /// the instructions for the maximum computation.
1789 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1790 // Check that the loop matches the pattern we're looking for.
1791 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1792 Cond->getPredicate() != CmpInst::ICMP_NE)
1795 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1796 if (!Sel || !Sel->hasOneUse()) return Cond;
1798 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1799 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1801 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1803 // Add one to the backedge-taken count to get the trip count.
1804 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1805 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1807 // Check for a max calculation that matches the pattern. There's no check
1808 // for ICMP_ULE here because the comparison would be with zero, which
1809 // isn't interesting.
1810 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1811 const SCEVNAryExpr *Max = 0;
1812 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1813 Pred = ICmpInst::ICMP_SLE;
1815 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1816 Pred = ICmpInst::ICMP_SLT;
1818 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1819 Pred = ICmpInst::ICMP_ULT;
1826 // To handle a max with more than two operands, this optimization would
1827 // require additional checking and setup.
1828 if (Max->getNumOperands() != 2)
1831 const SCEV *MaxLHS = Max->getOperand(0);
1832 const SCEV *MaxRHS = Max->getOperand(1);
1834 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1835 // for a comparison with 1. For <= and >=, a comparison with zero.
1837 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1840 // Check the relevant induction variable for conformance to
1842 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1843 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1844 if (!AR || !AR->isAffine() ||
1845 AR->getStart() != One ||
1846 AR->getStepRecurrence(SE) != One)
1849 assert(AR->getLoop() == L &&
1850 "Loop condition operand is an addrec in a different loop!");
1852 // Check the right operand of the select, and remember it, as it will
1853 // be used in the new comparison instruction.
1855 if (ICmpInst::isTrueWhenEqual(Pred)) {
1856 // Look for n+1, and grab n.
1857 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1858 if (isa<ConstantInt>(BO->getOperand(1)) &&
1859 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1860 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1861 NewRHS = BO->getOperand(0);
1862 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1863 if (isa<ConstantInt>(BO->getOperand(1)) &&
1864 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1865 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1866 NewRHS = BO->getOperand(0);
1869 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1870 NewRHS = Sel->getOperand(1);
1871 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1872 NewRHS = Sel->getOperand(2);
1873 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1874 NewRHS = SU->getValue();
1876 // Max doesn't match expected pattern.
1879 // Determine the new comparison opcode. It may be signed or unsigned,
1880 // and the original comparison may be either equality or inequality.
1881 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1882 Pred = CmpInst::getInversePredicate(Pred);
1884 // Ok, everything looks ok to change the condition into an SLT or SGE and
1885 // delete the max calculation.
1887 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1889 // Delete the max calculation instructions.
1890 Cond->replaceAllUsesWith(NewCond);
1891 CondUse->setUser(NewCond);
1892 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1893 Cond->eraseFromParent();
1894 Sel->eraseFromParent();
1895 if (Cmp->use_empty())
1896 Cmp->eraseFromParent();
1900 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1901 /// postinc iv when possible.
1903 LSRInstance::OptimizeLoopTermCond() {
1904 SmallPtrSet<Instruction *, 4> PostIncs;
1906 BasicBlock *LatchBlock = L->getLoopLatch();
1907 SmallVector<BasicBlock*, 8> ExitingBlocks;
1908 L->getExitingBlocks(ExitingBlocks);
1910 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1911 BasicBlock *ExitingBlock = ExitingBlocks[i];
1913 // Get the terminating condition for the loop if possible. If we
1914 // can, we want to change it to use a post-incremented version of its
1915 // induction variable, to allow coalescing the live ranges for the IV into
1916 // one register value.
1918 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1921 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1922 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1925 // Search IVUsesByStride to find Cond's IVUse if there is one.
1926 IVStrideUse *CondUse = 0;
1927 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1928 if (!FindIVUserForCond(Cond, CondUse))
1931 // If the trip count is computed in terms of a max (due to ScalarEvolution
1932 // being unable to find a sufficient guard, for example), change the loop
1933 // comparison to use SLT or ULT instead of NE.
1934 // One consequence of doing this now is that it disrupts the count-down
1935 // optimization. That's not always a bad thing though, because in such
1936 // cases it may still be worthwhile to avoid a max.
1937 Cond = OptimizeMax(Cond, CondUse);
1939 // If this exiting block dominates the latch block, it may also use
1940 // the post-inc value if it won't be shared with other uses.
1941 // Check for dominance.
1942 if (!DT.dominates(ExitingBlock, LatchBlock))
1945 // Conservatively avoid trying to use the post-inc value in non-latch
1946 // exits if there may be pre-inc users in intervening blocks.
1947 if (LatchBlock != ExitingBlock)
1948 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1949 // Test if the use is reachable from the exiting block. This dominator
1950 // query is a conservative approximation of reachability.
1951 if (&*UI != CondUse &&
1952 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1953 // Conservatively assume there may be reuse if the quotient of their
1954 // strides could be a legal scale.
1955 const SCEV *A = IU.getStride(*CondUse, L);
1956 const SCEV *B = IU.getStride(*UI, L);
1957 if (!A || !B) continue;
1958 if (SE.getTypeSizeInBits(A->getType()) !=
1959 SE.getTypeSizeInBits(B->getType())) {
1960 if (SE.getTypeSizeInBits(A->getType()) >
1961 SE.getTypeSizeInBits(B->getType()))
1962 B = SE.getSignExtendExpr(B, A->getType());
1964 A = SE.getSignExtendExpr(A, B->getType());
1966 if (const SCEVConstant *D =
1967 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1968 const ConstantInt *C = D->getValue();
1969 // Stride of one or negative one can have reuse with non-addresses.
1970 if (C->isOne() || C->isAllOnesValue())
1971 goto decline_post_inc;
1972 // Avoid weird situations.
1973 if (C->getValue().getMinSignedBits() >= 64 ||
1974 C->getValue().isMinSignedValue())
1975 goto decline_post_inc;
1976 // Without TLI, assume that any stride might be valid, and so any
1977 // use might be shared.
1979 goto decline_post_inc;
1980 // Check for possible scaled-address reuse.
1981 Type *AccessTy = getAccessType(UI->getUser());
1982 TargetLowering::AddrMode AM;
1983 AM.Scale = C->getSExtValue();
1984 if (TLI->isLegalAddressingMode(AM, AccessTy))
1985 goto decline_post_inc;
1986 AM.Scale = -AM.Scale;
1987 if (TLI->isLegalAddressingMode(AM, AccessTy))
1988 goto decline_post_inc;
1992 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1995 // It's possible for the setcc instruction to be anywhere in the loop, and
1996 // possible for it to have multiple users. If it is not immediately before
1997 // the exiting block branch, move it.
1998 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1999 if (Cond->hasOneUse()) {
2000 Cond->moveBefore(TermBr);
2002 // Clone the terminating condition and insert into the loopend.
2003 ICmpInst *OldCond = Cond;
2004 Cond = cast<ICmpInst>(Cond->clone());
2005 Cond->setName(L->getHeader()->getName() + ".termcond");
2006 ExitingBlock->getInstList().insert(TermBr, Cond);
2008 // Clone the IVUse, as the old use still exists!
2009 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2010 TermBr->replaceUsesOfWith(OldCond, Cond);
2014 // If we get to here, we know that we can transform the setcc instruction to
2015 // use the post-incremented version of the IV, allowing us to coalesce the
2016 // live ranges for the IV correctly.
2017 CondUse->transformToPostInc(L);
2020 PostIncs.insert(Cond);
2024 // Determine an insertion point for the loop induction variable increment. It
2025 // must dominate all the post-inc comparisons we just set up, and it must
2026 // dominate the loop latch edge.
2027 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2028 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2029 E = PostIncs.end(); I != E; ++I) {
2031 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2033 if (BB == (*I)->getParent())
2034 IVIncInsertPos = *I;
2035 else if (BB != IVIncInsertPos->getParent())
2036 IVIncInsertPos = BB->getTerminator();
2040 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2041 /// at the given offset and other details. If so, update the use and
2044 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2045 LSRUse::KindType Kind, Type *AccessTy) {
2046 int64_t NewMinOffset = LU.MinOffset;
2047 int64_t NewMaxOffset = LU.MaxOffset;
2048 Type *NewAccessTy = AccessTy;
2050 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2051 // something conservative, however this can pessimize in the case that one of
2052 // the uses will have all its uses outside the loop, for example.
2053 if (LU.Kind != Kind)
2055 // Conservatively assume HasBaseReg is true for now.
2056 if (NewOffset < LU.MinOffset) {
2057 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
2058 Kind, AccessTy, TLI))
2060 NewMinOffset = NewOffset;
2061 } else if (NewOffset > LU.MaxOffset) {
2062 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
2063 Kind, AccessTy, TLI))
2065 NewMaxOffset = NewOffset;
2067 // Check for a mismatched access type, and fall back conservatively as needed.
2068 // TODO: Be less conservative when the type is similar and can use the same
2069 // addressing modes.
2070 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2071 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2074 LU.MinOffset = NewMinOffset;
2075 LU.MaxOffset = NewMaxOffset;
2076 LU.AccessTy = NewAccessTy;
2077 if (NewOffset != LU.Offsets.back())
2078 LU.Offsets.push_back(NewOffset);
2082 /// getUse - Return an LSRUse index and an offset value for a fixup which
2083 /// needs the given expression, with the given kind and optional access type.
2084 /// Either reuse an existing use or create a new one, as needed.
2085 std::pair<size_t, int64_t>
2086 LSRInstance::getUse(const SCEV *&Expr,
2087 LSRUse::KindType Kind, Type *AccessTy) {
2088 const SCEV *Copy = Expr;
2089 int64_t Offset = ExtractImmediate(Expr, SE);
2091 // Basic uses can't accept any offset, for example.
2092 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
2097 std::pair<UseMapTy::iterator, bool> P =
2098 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2100 // A use already existed with this base.
2101 size_t LUIdx = P.first->second;
2102 LSRUse &LU = Uses[LUIdx];
2103 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2105 return std::make_pair(LUIdx, Offset);
2108 // Create a new use.
2109 size_t LUIdx = Uses.size();
2110 P.first->second = LUIdx;
2111 Uses.push_back(LSRUse(Kind, AccessTy));
2112 LSRUse &LU = Uses[LUIdx];
2114 // We don't need to track redundant offsets, but we don't need to go out
2115 // of our way here to avoid them.
2116 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2117 LU.Offsets.push_back(Offset);
2119 LU.MinOffset = Offset;
2120 LU.MaxOffset = Offset;
2121 return std::make_pair(LUIdx, Offset);
2124 /// DeleteUse - Delete the given use from the Uses list.
2125 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2126 if (&LU != &Uses.back())
2127 std::swap(LU, Uses.back());
2131 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2134 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2135 /// a formula that has the same registers as the given formula.
2137 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2138 const LSRUse &OrigLU) {
2139 // Search all uses for the formula. This could be more clever.
2140 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2141 LSRUse &LU = Uses[LUIdx];
2142 // Check whether this use is close enough to OrigLU, to see whether it's
2143 // worthwhile looking through its formulae.
2144 // Ignore ICmpZero uses because they may contain formulae generated by
2145 // GenerateICmpZeroScales, in which case adding fixup offsets may
2147 if (&LU != &OrigLU &&
2148 LU.Kind != LSRUse::ICmpZero &&
2149 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2150 LU.WidestFixupType == OrigLU.WidestFixupType &&
2151 LU.HasFormulaWithSameRegs(OrigF)) {
2152 // Scan through this use's formulae.
2153 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2154 E = LU.Formulae.end(); I != E; ++I) {
2155 const Formula &F = *I;
2156 // Check to see if this formula has the same registers and symbols
2158 if (F.BaseRegs == OrigF.BaseRegs &&
2159 F.ScaledReg == OrigF.ScaledReg &&
2160 F.AM.BaseGV == OrigF.AM.BaseGV &&
2161 F.AM.Scale == OrigF.AM.Scale &&
2162 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2163 if (F.AM.BaseOffs == 0)
2165 // This is the formula where all the registers and symbols matched;
2166 // there aren't going to be any others. Since we declined it, we
2167 // can skip the rest of the formulae and procede to the next LSRUse.
2174 // Nothing looked good.
2178 void LSRInstance::CollectInterestingTypesAndFactors() {
2179 SmallSetVector<const SCEV *, 4> Strides;
2181 // Collect interesting types and strides.
2182 SmallVector<const SCEV *, 4> Worklist;
2183 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2184 const SCEV *Expr = IU.getExpr(*UI);
2186 // Collect interesting types.
2187 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2189 // Add strides for mentioned loops.
2190 Worklist.push_back(Expr);
2192 const SCEV *S = Worklist.pop_back_val();
2193 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2194 if (EnableNested || AR->getLoop() == L)
2195 Strides.insert(AR->getStepRecurrence(SE));
2196 Worklist.push_back(AR->getStart());
2197 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2198 Worklist.append(Add->op_begin(), Add->op_end());
2200 } while (!Worklist.empty());
2203 // Compute interesting factors from the set of interesting strides.
2204 for (SmallSetVector<const SCEV *, 4>::const_iterator
2205 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2206 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2207 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2208 const SCEV *OldStride = *I;
2209 const SCEV *NewStride = *NewStrideIter;
2211 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2212 SE.getTypeSizeInBits(NewStride->getType())) {
2213 if (SE.getTypeSizeInBits(OldStride->getType()) >
2214 SE.getTypeSizeInBits(NewStride->getType()))
2215 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2217 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2219 if (const SCEVConstant *Factor =
2220 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2222 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2223 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2224 } else if (const SCEVConstant *Factor =
2225 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2228 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2229 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2233 // If all uses use the same type, don't bother looking for truncation-based
2235 if (Types.size() == 1)
2238 DEBUG(print_factors_and_types(dbgs()));
2241 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2242 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2243 /// Instructions to IVStrideUses, we could partially skip this.
2244 static User::op_iterator
2245 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2246 Loop *L, ScalarEvolution &SE) {
2247 for(; OI != OE; ++OI) {
2248 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2249 if (!SE.isSCEVable(Oper->getType()))
2252 if (const SCEVAddRecExpr *AR =
2253 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2254 if (AR->getLoop() == L)
2262 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2263 /// operands, so wrap it in a convenient helper.
2264 static Value *getWideOperand(Value *Oper) {
2265 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2266 return Trunc->getOperand(0);
2270 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2272 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2273 Type *LType = LVal->getType();
2274 Type *RType = RVal->getType();
2275 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2278 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2279 /// NULL for any constant. Returning the expression itself is
2280 /// conservative. Returning a deeper subexpression is more precise and valid as
2281 /// long as it isn't less complex than another subexpression. For expressions
2282 /// involving multiple unscaled values, we need to return the pointer-type
2283 /// SCEVUnknown. This avoids forming chains across objects, such as:
2284 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2286 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2287 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2288 static const SCEV *getExprBase(const SCEV *S) {
2289 switch (S->getSCEVType()) {
2290 default: // uncluding scUnknown.
2295 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2297 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2299 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2301 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2302 // there's nothing more complex.
2303 // FIXME: not sure if we want to recognize negation.
2304 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2305 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2306 E(Add->op_begin()); I != E; ++I) {
2307 const SCEV *SubExpr = *I;
2308 if (SubExpr->getSCEVType() == scAddExpr)
2309 return getExprBase(SubExpr);
2311 if (SubExpr->getSCEVType() != scMulExpr)
2314 return S; // all operands are scaled, be conservative.
2317 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2321 /// Return true if the chain increment is profitable to expand into a loop
2322 /// invariant value, which may require its own register. A profitable chain
2323 /// increment will be an offset relative to the same base. We allow such offsets
2324 /// to potentially be used as chain increment as long as it's not obviously
2325 /// expensive to expand using real instructions.
2327 getProfitableChainIncrement(Value *NextIV, Value *PrevIV,
2328 const IVChain &Chain, Loop *L,
2329 ScalarEvolution &SE, const TargetLowering *TLI) {
2330 // Prune the solution space aggressively by checking that both IV operands
2331 // are expressions that operate on the same unscaled SCEVUnknown. This
2332 // "base" will be canceled by the subsequent getMinusSCEV call. Checking first
2333 // avoids creating extra SCEV expressions.
2334 const SCEV *OperExpr = SE.getSCEV(NextIV);
2335 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2336 if (getExprBase(OperExpr) != getExprBase(PrevExpr) && !StressIVChain)
2339 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2340 if (!SE.isLoopInvariant(IncExpr, L))
2343 // We are not able to expand an increment unless it is loop invariant,
2344 // however, the following checks are purely for profitability.
2348 // Do not replace a constant offset from IV head with a nonconstant IV
2350 if (!isa<SCEVConstant>(IncExpr)) {
2351 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Chain[0].IVOperand));
2352 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2356 SmallPtrSet<const SCEV*, 8> Processed;
2357 if (isHighCostExpansion(IncExpr, Processed, SE))
2363 /// Return true if the number of registers needed for the chain is estimated to
2364 /// be less than the number required for the individual IV users. First prohibit
2365 /// any IV users that keep the IV live across increments (the Users set should
2366 /// be empty). Next count the number and type of increments in the chain.
2368 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2369 /// effectively use postinc addressing modes. Only consider it profitable it the
2370 /// increments can be computed in fewer registers when chained.
2372 /// TODO: Consider IVInc free if it's already used in another chains.
2374 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2375 ScalarEvolution &SE, const TargetLowering *TLI) {
2379 if (Chain.size() <= 2)
2382 if (!Users.empty()) {
2383 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " users:\n";
2384 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2385 E = Users.end(); I != E; ++I) {
2386 dbgs() << " " << **I << "\n";
2390 assert(!Chain.empty() && "empty IV chains are not allowed");
2392 // The chain itself may require a register, so intialize cost to 1.
2395 // A complete chain likely eliminates the need for keeping the original IV in
2396 // a register. LSR does not currently know how to form a complete chain unless
2397 // the header phi already exists.
2398 if (isa<PHINode>(Chain.back().UserInst)
2399 && SE.getSCEV(Chain.back().UserInst) == Chain[0].IncExpr) {
2402 const SCEV *LastIncExpr = 0;
2403 unsigned NumConstIncrements = 0;
2404 unsigned NumVarIncrements = 0;
2405 unsigned NumReusedIncrements = 0;
2406 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2409 if (I->IncExpr->isZero())
2412 // Incrementing by zero or some constant is neutral. We assume constants can
2413 // be folded into an addressing mode or an add's immediate operand.
2414 if (isa<SCEVConstant>(I->IncExpr)) {
2415 ++NumConstIncrements;
2419 if (I->IncExpr == LastIncExpr)
2420 ++NumReusedIncrements;
2424 LastIncExpr = I->IncExpr;
2426 // An IV chain with a single increment is handled by LSR's postinc
2427 // uses. However, a chain with multiple increments requires keeping the IV's
2428 // value live longer than it needs to be if chained.
2429 if (NumConstIncrements > 1)
2432 // Materializing increment expressions in the preheader that didn't exist in
2433 // the original code may cost a register. For example, sign-extended array
2434 // indices can produce ridiculous increments like this:
2435 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2436 cost += NumVarIncrements;
2438 // Reusing variable increments likely saves a register to hold the multiple of
2440 cost -= NumReusedIncrements;
2442 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " Cost: " << cost << "\n");
2447 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2449 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2450 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2451 // When IVs are used as types of varying widths, they are generally converted
2452 // to a wider type with some uses remaining narrow under a (free) trunc.
2453 Value *NextIV = getWideOperand(IVOper);
2455 // Visit all existing chains. Check if its IVOper can be computed as a
2456 // profitable loop invariant increment from the last link in the Chain.
2457 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2458 const SCEV *LastIncExpr = 0;
2459 for (; ChainIdx < NChains; ++ChainIdx) {
2460 Value *PrevIV = getWideOperand(IVChainVec[ChainIdx].back().IVOperand);
2461 if (!isCompatibleIVType(PrevIV, NextIV))
2464 // A phi nodes terminates a chain.
2465 if (isa<PHINode>(UserInst)
2466 && isa<PHINode>(IVChainVec[ChainIdx].back().UserInst))
2469 if (const SCEV *IncExpr =
2470 getProfitableChainIncrement(NextIV, PrevIV, IVChainVec[ChainIdx],
2472 LastIncExpr = IncExpr;
2476 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2477 // bother for phi nodes, because they must be last in the chain.
2478 if (ChainIdx == NChains) {
2479 if (isa<PHINode>(UserInst))
2481 if (NChains >= MaxChains && !StressIVChain) {
2482 DEBUG(dbgs() << "IV Chain Limit\n");
2486 IVChainVec.resize(NChains);
2487 ChainUsersVec.resize(NChains);
2488 LastIncExpr = SE.getSCEV(NextIV);
2489 assert(isa<SCEVAddRecExpr>(LastIncExpr) && "expect recurrence at IV user");
2490 DEBUG(dbgs() << "IV Head: (" << *UserInst << ") IV=" << *LastIncExpr
2494 DEBUG(dbgs() << "IV Inc: (" << *UserInst << ") IV+" << *LastIncExpr
2497 // Add this IV user to the end of the chain.
2498 IVChainVec[ChainIdx].push_back(IVInc(UserInst, IVOper, LastIncExpr));
2500 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2501 // This chain's NearUsers become FarUsers.
2502 if (!LastIncExpr->isZero()) {
2503 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2508 // All other uses of IVOperand become near uses of the chain.
2509 // We currently ignore intermediate values within SCEV expressions, assuming
2510 // they will eventually be used be the current chain, or can be computed
2511 // from one of the chain increments. To be more precise we could
2512 // transitively follow its user and only add leaf IV users to the set.
2513 for (Value::use_iterator UseIter = IVOper->use_begin(),
2514 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2515 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2516 if (SE.isSCEVable(OtherUse->getType())
2517 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2518 && IU.isIVUserOrOperand(OtherUse)) {
2521 if (OtherUse && OtherUse != UserInst)
2522 NearUsers.insert(OtherUse);
2525 // Since this user is part of the chain, it's no longer considered a use
2527 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2530 /// CollectChains - Populate the vector of Chains.
2532 /// This decreases ILP at the architecture level. Targets with ample registers,
2533 /// multiple memory ports, and no register renaming probably don't want
2534 /// this. However, such targets should probably disable LSR altogether.
2536 /// The job of LSR is to make a reasonable choice of induction variables across
2537 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2538 /// ILP *within the loop* if the target wants it.
2540 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2541 /// will not reorder memory operations, it will recognize this as a chain, but
2542 /// will generate redundant IV increments. Ideally this would be corrected later
2543 /// by a smart scheduler:
2549 /// TODO: Walk the entire domtree within this loop, not just the path to the
2550 /// loop latch. This will discover chains on side paths, but requires
2551 /// maintaining multiple copies of the Chains state.
2552 void LSRInstance::CollectChains() {
2553 SmallVector<ChainUsers, 8> ChainUsersVec;
2555 SmallVector<BasicBlock *,8> LatchPath;
2556 BasicBlock *LoopHeader = L->getHeader();
2557 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2558 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2559 LatchPath.push_back(Rung->getBlock());
2561 LatchPath.push_back(LoopHeader);
2563 // Walk the instruction stream from the loop header to the loop latch.
2564 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2565 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2566 BBIter != BBEnd; ++BBIter) {
2567 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2569 // Skip instructions that weren't seen by IVUsers analysis.
2570 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2573 // Ignore users that are part of a SCEV expression. This way we only
2574 // consider leaf IV Users. This effectively rediscovers a portion of
2575 // IVUsers analysis but in program order this time.
2576 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2579 // Remove this instruction from any NearUsers set it may be in.
2580 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2581 ChainIdx < NChains; ++ChainIdx) {
2582 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2584 // Search for operands that can be chained.
2585 SmallPtrSet<Instruction*, 4> UniqueOperands;
2586 User::op_iterator IVOpEnd = I->op_end();
2587 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2588 while (IVOpIter != IVOpEnd) {
2589 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2590 if (UniqueOperands.insert(IVOpInst))
2591 ChainInstruction(I, IVOpInst, ChainUsersVec);
2592 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2594 } // Continue walking down the instructions.
2595 } // Continue walking down the domtree.
2596 // Visit phi backedges to determine if the chain can generate the IV postinc.
2597 for (BasicBlock::iterator I = L->getHeader()->begin();
2598 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2599 if (!SE.isSCEVable(PN->getType()))
2603 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2605 ChainInstruction(PN, IncV, ChainUsersVec);
2607 // Remove any unprofitable chains.
2608 unsigned ChainIdx = 0;
2609 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2610 UsersIdx < NChains; ++UsersIdx) {
2611 if (!isProfitableChain(IVChainVec[UsersIdx],
2612 ChainUsersVec[UsersIdx].FarUsers, SE, TLI))
2614 // Preserve the chain at UsesIdx.
2615 if (ChainIdx != UsersIdx)
2616 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2617 FinalizeChain(IVChainVec[ChainIdx]);
2620 IVChainVec.resize(ChainIdx);
2623 void LSRInstance::FinalizeChain(IVChain &Chain) {
2624 assert(!Chain.empty() && "empty IV chains are not allowed");
2625 DEBUG(dbgs() << "Final Chain: " << *Chain[0].UserInst << "\n");
2627 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2629 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2630 User::op_iterator UseI =
2631 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2632 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2633 IVIncSet.insert(UseI);
2637 /// Return true if the IVInc can be folded into an addressing mode.
2638 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2639 Value *Operand, const TargetLowering *TLI) {
2640 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2641 if (!IncConst || !isAddressUse(UserInst, Operand))
2644 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2647 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2648 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false,
2649 LSRUse::Address, getAccessType(UserInst), TLI))
2655 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2656 /// materialize the IV user's operand from the previous IV user's operand.
2657 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2658 SmallVectorImpl<WeakVH> &DeadInsts) {
2659 // Find the new IVOperand for the head of the chain. It may have been replaced
2661 const IVInc &Head = Chain[0];
2662 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2663 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2666 while (IVOpIter != IVOpEnd) {
2667 IVSrc = getWideOperand(*IVOpIter);
2669 // If this operand computes the expression that the chain needs, we may use
2670 // it. (Check this after setting IVSrc which is used below.)
2672 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2673 // narrow for the chain, so we can no longer use it. We do allow using a
2674 // wider phi, assuming the LSR checked for free truncation. In that case we
2675 // should already have a truncate on this operand such that
2676 // getSCEV(IVSrc) == IncExpr.
2677 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2678 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2681 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2683 if (IVOpIter == IVOpEnd) {
2684 // Gracefully give up on this chain.
2685 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2689 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2690 Type *IVTy = IVSrc->getType();
2691 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2692 const SCEV *LeftOverExpr = 0;
2693 for (IVChain::const_iterator IncI = llvm::next(Chain.begin()),
2694 IncE = Chain.end(); IncI != IncE; ++IncI) {
2696 Instruction *InsertPt = IncI->UserInst;
2697 if (isa<PHINode>(InsertPt))
2698 InsertPt = L->getLoopLatch()->getTerminator();
2700 // IVOper will replace the current IV User's operand. IVSrc is the IV
2701 // value currently held in a register.
2702 Value *IVOper = IVSrc;
2703 if (!IncI->IncExpr->isZero()) {
2704 // IncExpr was the result of subtraction of two narrow values, so must
2706 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2707 LeftOverExpr = LeftOverExpr ?
2708 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2710 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2711 // Expand the IV increment.
2712 Rewriter.clearPostInc();
2713 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2714 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2715 SE.getUnknown(IncV));
2716 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2718 // If an IV increment can't be folded, use it as the next IV value.
2719 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2721 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2726 Type *OperTy = IncI->IVOperand->getType();
2727 if (IVTy != OperTy) {
2728 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2729 "cannot extend a chained IV");
2730 IRBuilder<> Builder(InsertPt);
2731 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2733 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2734 DeadInsts.push_back(IncI->IVOperand);
2736 // If LSR created a new, wider phi, we may also replace its postinc. We only
2737 // do this if we also found a wide value for the head of the chain.
2738 if (isa<PHINode>(Chain.back().UserInst)) {
2739 for (BasicBlock::iterator I = L->getHeader()->begin();
2740 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2741 if (!isCompatibleIVType(Phi, IVSrc))
2743 Instruction *PostIncV = dyn_cast<Instruction>(
2744 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2745 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2747 Value *IVOper = IVSrc;
2748 Type *PostIncTy = PostIncV->getType();
2749 if (IVTy != PostIncTy) {
2750 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2751 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2752 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2753 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2755 Phi->replaceUsesOfWith(PostIncV, IVOper);
2756 DeadInsts.push_back(PostIncV);
2761 void LSRInstance::CollectFixupsAndInitialFormulae() {
2762 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2763 Instruction *UserInst = UI->getUser();
2764 // Skip IV users that are part of profitable IV Chains.
2765 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2766 UI->getOperandValToReplace());
2767 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2768 if (IVIncSet.count(UseI))
2772 LSRFixup &LF = getNewFixup();
2773 LF.UserInst = UserInst;
2774 LF.OperandValToReplace = UI->getOperandValToReplace();
2775 LF.PostIncLoops = UI->getPostIncLoops();
2777 LSRUse::KindType Kind = LSRUse::Basic;
2779 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2780 Kind = LSRUse::Address;
2781 AccessTy = getAccessType(LF.UserInst);
2784 const SCEV *S = IU.getExpr(*UI);
2786 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2787 // (N - i == 0), and this allows (N - i) to be the expression that we work
2788 // with rather than just N or i, so we can consider the register
2789 // requirements for both N and i at the same time. Limiting this code to
2790 // equality icmps is not a problem because all interesting loops use
2791 // equality icmps, thanks to IndVarSimplify.
2792 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2793 if (CI->isEquality()) {
2794 // Swap the operands if needed to put the OperandValToReplace on the
2795 // left, for consistency.
2796 Value *NV = CI->getOperand(1);
2797 if (NV == LF.OperandValToReplace) {
2798 CI->setOperand(1, CI->getOperand(0));
2799 CI->setOperand(0, NV);
2800 NV = CI->getOperand(1);
2804 // x == y --> x - y == 0
2805 const SCEV *N = SE.getSCEV(NV);
2806 if (SE.isLoopInvariant(N, L)) {
2807 // S is normalized, so normalize N before folding it into S
2808 // to keep the result normalized.
2809 N = TransformForPostIncUse(Normalize, N, CI, 0,
2810 LF.PostIncLoops, SE, DT);
2811 Kind = LSRUse::ICmpZero;
2812 S = SE.getMinusSCEV(N, S);
2815 // -1 and the negations of all interesting strides (except the negation
2816 // of -1) are now also interesting.
2817 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2818 if (Factors[i] != -1)
2819 Factors.insert(-(uint64_t)Factors[i]);
2823 // Set up the initial formula for this use.
2824 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2826 LF.Offset = P.second;
2827 LSRUse &LU = Uses[LF.LUIdx];
2828 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2829 if (!LU.WidestFixupType ||
2830 SE.getTypeSizeInBits(LU.WidestFixupType) <
2831 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2832 LU.WidestFixupType = LF.OperandValToReplace->getType();
2834 // If this is the first use of this LSRUse, give it a formula.
2835 if (LU.Formulae.empty()) {
2836 InsertInitialFormula(S, LU, LF.LUIdx);
2837 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2841 DEBUG(print_fixups(dbgs()));
2844 /// InsertInitialFormula - Insert a formula for the given expression into
2845 /// the given use, separating out loop-variant portions from loop-invariant
2846 /// and loop-computable portions.
2848 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2850 F.InitialMatch(S, L, SE);
2851 bool Inserted = InsertFormula(LU, LUIdx, F);
2852 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2855 /// InsertSupplementalFormula - Insert a simple single-register formula for
2856 /// the given expression into the given use.
2858 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2859 LSRUse &LU, size_t LUIdx) {
2861 F.BaseRegs.push_back(S);
2862 F.AM.HasBaseReg = true;
2863 bool Inserted = InsertFormula(LU, LUIdx, F);
2864 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2867 /// CountRegisters - Note which registers are used by the given formula,
2868 /// updating RegUses.
2869 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2871 RegUses.CountRegister(F.ScaledReg, LUIdx);
2872 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2873 E = F.BaseRegs.end(); I != E; ++I)
2874 RegUses.CountRegister(*I, LUIdx);
2877 /// InsertFormula - If the given formula has not yet been inserted, add it to
2878 /// the list, and return true. Return false otherwise.
2879 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2880 if (!LU.InsertFormula(F))
2883 CountRegisters(F, LUIdx);
2887 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2888 /// loop-invariant values which we're tracking. These other uses will pin these
2889 /// values in registers, making them less profitable for elimination.
2890 /// TODO: This currently misses non-constant addrec step registers.
2891 /// TODO: Should this give more weight to users inside the loop?
2893 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2894 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2895 SmallPtrSet<const SCEV *, 8> Inserted;
2897 while (!Worklist.empty()) {
2898 const SCEV *S = Worklist.pop_back_val();
2900 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2901 Worklist.append(N->op_begin(), N->op_end());
2902 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2903 Worklist.push_back(C->getOperand());
2904 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2905 Worklist.push_back(D->getLHS());
2906 Worklist.push_back(D->getRHS());
2907 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2908 if (!Inserted.insert(U)) continue;
2909 const Value *V = U->getValue();
2910 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2911 // Look for instructions defined outside the loop.
2912 if (L->contains(Inst)) continue;
2913 } else if (isa<UndefValue>(V))
2914 // Undef doesn't have a live range, so it doesn't matter.
2916 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2918 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2919 // Ignore non-instructions.
2922 // Ignore instructions in other functions (as can happen with
2924 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2926 // Ignore instructions not dominated by the loop.
2927 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2928 UserInst->getParent() :
2929 cast<PHINode>(UserInst)->getIncomingBlock(
2930 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2931 if (!DT.dominates(L->getHeader(), UseBB))
2933 // Ignore uses which are part of other SCEV expressions, to avoid
2934 // analyzing them multiple times.
2935 if (SE.isSCEVable(UserInst->getType())) {
2936 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2937 // If the user is a no-op, look through to its uses.
2938 if (!isa<SCEVUnknown>(UserS))
2942 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2946 // Ignore icmp instructions which are already being analyzed.
2947 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2948 unsigned OtherIdx = !UI.getOperandNo();
2949 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2950 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2954 LSRFixup &LF = getNewFixup();
2955 LF.UserInst = const_cast<Instruction *>(UserInst);
2956 LF.OperandValToReplace = UI.getUse();
2957 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2959 LF.Offset = P.second;
2960 LSRUse &LU = Uses[LF.LUIdx];
2961 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2962 if (!LU.WidestFixupType ||
2963 SE.getTypeSizeInBits(LU.WidestFixupType) <
2964 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2965 LU.WidestFixupType = LF.OperandValToReplace->getType();
2966 InsertSupplementalFormula(U, LU, LF.LUIdx);
2967 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2974 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2975 /// separate registers. If C is non-null, multiply each subexpression by C.
2976 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2977 SmallVectorImpl<const SCEV *> &Ops,
2979 ScalarEvolution &SE) {
2980 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2981 // Break out add operands.
2982 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2984 CollectSubexprs(*I, C, Ops, L, SE);
2986 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2987 // Split a non-zero base out of an addrec.
2988 if (!AR->getStart()->isZero()) {
2989 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2990 AR->getStepRecurrence(SE),
2992 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2995 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2998 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2999 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3000 if (Mul->getNumOperands() == 2)
3001 if (const SCEVConstant *Op0 =
3002 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3003 CollectSubexprs(Mul->getOperand(1),
3004 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
3010 // Otherwise use the value itself, optionally with a scale applied.
3011 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
3014 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3016 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3019 // Arbitrarily cap recursion to protect compile time.
3020 if (Depth >= 3) return;
3022 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3023 const SCEV *BaseReg = Base.BaseRegs[i];
3025 SmallVector<const SCEV *, 8> AddOps;
3026 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3028 if (AddOps.size() == 1) continue;
3030 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3031 JE = AddOps.end(); J != JE; ++J) {
3033 // Loop-variant "unknown" values are uninteresting; we won't be able to
3034 // do anything meaningful with them.
3035 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3038 // Don't pull a constant into a register if the constant could be folded
3039 // into an immediate field.
3040 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
3041 Base.getNumRegs() > 1,
3042 LU.Kind, LU.AccessTy, TLI, SE))
3045 // Collect all operands except *J.
3046 SmallVector<const SCEV *, 8> InnerAddOps
3047 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3049 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3051 // Don't leave just a constant behind in a register if the constant could
3052 // be folded into an immediate field.
3053 if (InnerAddOps.size() == 1 &&
3054 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
3055 Base.getNumRegs() > 1,
3056 LU.Kind, LU.AccessTy, TLI, SE))
3059 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3060 if (InnerSum->isZero())
3064 // Add the remaining pieces of the add back into the new formula.
3065 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3066 if (TLI && InnerSumSC &&
3067 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3068 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3069 InnerSumSC->getValue()->getZExtValue())) {
3070 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3071 InnerSumSC->getValue()->getZExtValue();
3072 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3074 F.BaseRegs[i] = InnerSum;
3076 // Add J as its own register, or an unfolded immediate.
3077 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3078 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3079 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3080 SC->getValue()->getZExtValue()))
3081 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3082 SC->getValue()->getZExtValue();
3084 F.BaseRegs.push_back(*J);
3086 if (InsertFormula(LU, LUIdx, F))
3087 // If that formula hadn't been seen before, recurse to find more like
3089 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3094 /// GenerateCombinations - Generate a formula consisting of all of the
3095 /// loop-dominating registers added into a single register.
3096 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3098 // This method is only interesting on a plurality of registers.
3099 if (Base.BaseRegs.size() <= 1) return;
3103 SmallVector<const SCEV *, 4> Ops;
3104 for (SmallVectorImpl<const SCEV *>::const_iterator
3105 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3106 const SCEV *BaseReg = *I;
3107 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3108 !SE.hasComputableLoopEvolution(BaseReg, L))
3109 Ops.push_back(BaseReg);
3111 F.BaseRegs.push_back(BaseReg);
3113 if (Ops.size() > 1) {
3114 const SCEV *Sum = SE.getAddExpr(Ops);
3115 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3116 // opportunity to fold something. For now, just ignore such cases
3117 // rather than proceed with zero in a register.
3118 if (!Sum->isZero()) {
3119 F.BaseRegs.push_back(Sum);
3120 (void)InsertFormula(LU, LUIdx, F);
3125 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3126 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3128 // We can't add a symbolic offset if the address already contains one.
3129 if (Base.AM.BaseGV) return;
3131 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3132 const SCEV *G = Base.BaseRegs[i];
3133 GlobalValue *GV = ExtractSymbol(G, SE);
3134 if (G->isZero() || !GV)
3138 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3139 LU.Kind, LU.AccessTy, TLI))
3142 (void)InsertFormula(LU, LUIdx, F);
3146 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3147 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3149 // TODO: For now, just add the min and max offset, because it usually isn't
3150 // worthwhile looking at everything inbetween.
3151 SmallVector<int64_t, 2> Worklist;
3152 Worklist.push_back(LU.MinOffset);
3153 if (LU.MaxOffset != LU.MinOffset)
3154 Worklist.push_back(LU.MaxOffset);
3156 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3157 const SCEV *G = Base.BaseRegs[i];
3159 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3160 E = Worklist.end(); I != E; ++I) {
3162 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
3163 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
3164 LU.Kind, LU.AccessTy, TLI)) {
3165 // Add the offset to the base register.
3166 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3167 // If it cancelled out, drop the base register, otherwise update it.
3168 if (NewG->isZero()) {
3169 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3170 F.BaseRegs.pop_back();
3172 F.BaseRegs[i] = NewG;
3174 (void)InsertFormula(LU, LUIdx, F);
3178 int64_t Imm = ExtractImmediate(G, SE);
3179 if (G->isZero() || Imm == 0)
3182 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
3183 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3184 LU.Kind, LU.AccessTy, TLI))
3187 (void)InsertFormula(LU, LUIdx, F);
3191 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3192 /// the comparison. For example, x == y -> x*c == y*c.
3193 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3195 if (LU.Kind != LSRUse::ICmpZero) return;
3197 // Determine the integer type for the base formula.
3198 Type *IntTy = Base.getType();
3200 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3202 // Don't do this if there is more than one offset.
3203 if (LU.MinOffset != LU.MaxOffset) return;
3205 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
3207 // Check each interesting stride.
3208 for (SmallSetVector<int64_t, 8>::const_iterator
3209 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3210 int64_t Factor = *I;
3212 // Check that the multiplication doesn't overflow.
3213 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
3215 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
3216 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
3219 // Check that multiplying with the use offset doesn't overflow.
3220 int64_t Offset = LU.MinOffset;
3221 if (Offset == INT64_MIN && Factor == -1)
3223 Offset = (uint64_t)Offset * Factor;
3224 if (Offset / Factor != LU.MinOffset)
3228 F.AM.BaseOffs = NewBaseOffs;
3230 // Check that this scale is legal.
3231 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
3234 // Compensate for the use having MinOffset built into it.
3235 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
3237 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3239 // Check that multiplying with each base register doesn't overflow.
3240 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3241 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3242 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3246 // Check that multiplying with the scaled register doesn't overflow.
3248 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3249 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3253 // Check that multiplying with the unfolded offset doesn't overflow.
3254 if (F.UnfoldedOffset != 0) {
3255 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3257 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3258 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3262 // If we make it here and it's legal, add it.
3263 (void)InsertFormula(LU, LUIdx, F);
3268 /// GenerateScales - Generate stride factor reuse formulae by making use of
3269 /// scaled-offset address modes, for example.
3270 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3271 // Determine the integer type for the base formula.
3272 Type *IntTy = Base.getType();
3275 // If this Formula already has a scaled register, we can't add another one.
3276 if (Base.AM.Scale != 0) return;
3278 // Check each interesting stride.
3279 for (SmallSetVector<int64_t, 8>::const_iterator
3280 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3281 int64_t Factor = *I;
3283 Base.AM.Scale = Factor;
3284 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
3285 // Check whether this scale is going to be legal.
3286 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3287 LU.Kind, LU.AccessTy, TLI)) {
3288 // As a special-case, handle special out-of-loop Basic users specially.
3289 // TODO: Reconsider this special case.
3290 if (LU.Kind == LSRUse::Basic &&
3291 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3292 LSRUse::Special, LU.AccessTy, TLI) &&
3293 LU.AllFixupsOutsideLoop)
3294 LU.Kind = LSRUse::Special;
3298 // For an ICmpZero, negating a solitary base register won't lead to
3300 if (LU.Kind == LSRUse::ICmpZero &&
3301 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
3303 // For each addrec base reg, apply the scale, if possible.
3304 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3305 if (const SCEVAddRecExpr *AR =
3306 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3307 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3308 if (FactorS->isZero())
3310 // Divide out the factor, ignoring high bits, since we'll be
3311 // scaling the value back up in the end.
3312 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3313 // TODO: This could be optimized to avoid all the copying.
3315 F.ScaledReg = Quotient;
3316 F.DeleteBaseReg(F.BaseRegs[i]);
3317 (void)InsertFormula(LU, LUIdx, F);
3323 /// GenerateTruncates - Generate reuse formulae from different IV types.
3324 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3325 // This requires TargetLowering to tell us which truncates are free.
3328 // Don't bother truncating symbolic values.
3329 if (Base.AM.BaseGV) return;
3331 // Determine the integer type for the base formula.
3332 Type *DstTy = Base.getType();
3334 DstTy = SE.getEffectiveSCEVType(DstTy);
3336 for (SmallSetVector<Type *, 4>::const_iterator
3337 I = Types.begin(), E = Types.end(); I != E; ++I) {
3339 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
3342 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3343 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3344 JE = F.BaseRegs.end(); J != JE; ++J)
3345 *J = SE.getAnyExtendExpr(*J, SrcTy);
3347 // TODO: This assumes we've done basic processing on all uses and
3348 // have an idea what the register usage is.
3349 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3352 (void)InsertFormula(LU, LUIdx, F);
3359 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3360 /// defer modifications so that the search phase doesn't have to worry about
3361 /// the data structures moving underneath it.
3365 const SCEV *OrigReg;
3367 WorkItem(size_t LI, int64_t I, const SCEV *R)
3368 : LUIdx(LI), Imm(I), OrigReg(R) {}
3370 void print(raw_ostream &OS) const;
3376 void WorkItem::print(raw_ostream &OS) const {
3377 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3378 << " , add offset " << Imm;
3381 void WorkItem::dump() const {
3382 print(errs()); errs() << '\n';
3385 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3386 /// distance apart and try to form reuse opportunities between them.
3387 void LSRInstance::GenerateCrossUseConstantOffsets() {
3388 // Group the registers by their value without any added constant offset.
3389 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3390 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3392 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3393 SmallVector<const SCEV *, 8> Sequence;
3394 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3396 const SCEV *Reg = *I;
3397 int64_t Imm = ExtractImmediate(Reg, SE);
3398 std::pair<RegMapTy::iterator, bool> Pair =
3399 Map.insert(std::make_pair(Reg, ImmMapTy()));
3401 Sequence.push_back(Reg);
3402 Pair.first->second.insert(std::make_pair(Imm, *I));
3403 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3406 // Now examine each set of registers with the same base value. Build up
3407 // a list of work to do and do the work in a separate step so that we're
3408 // not adding formulae and register counts while we're searching.
3409 SmallVector<WorkItem, 32> WorkItems;
3410 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3411 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3412 E = Sequence.end(); I != E; ++I) {
3413 const SCEV *Reg = *I;
3414 const ImmMapTy &Imms = Map.find(Reg)->second;
3416 // It's not worthwhile looking for reuse if there's only one offset.
3417 if (Imms.size() == 1)
3420 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3421 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3423 dbgs() << ' ' << J->first;
3426 // Examine each offset.
3427 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3429 const SCEV *OrigReg = J->second;
3431 int64_t JImm = J->first;
3432 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3434 if (!isa<SCEVConstant>(OrigReg) &&
3435 UsedByIndicesMap[Reg].count() == 1) {
3436 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3440 // Conservatively examine offsets between this orig reg a few selected
3442 ImmMapTy::const_iterator OtherImms[] = {
3443 Imms.begin(), prior(Imms.end()),
3444 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3446 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3447 ImmMapTy::const_iterator M = OtherImms[i];
3448 if (M == J || M == JE) continue;
3450 // Compute the difference between the two.
3451 int64_t Imm = (uint64_t)JImm - M->first;
3452 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3453 LUIdx = UsedByIndices.find_next(LUIdx))
3454 // Make a memo of this use, offset, and register tuple.
3455 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3456 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3463 UsedByIndicesMap.clear();
3464 UniqueItems.clear();
3466 // Now iterate through the worklist and add new formulae.
3467 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3468 E = WorkItems.end(); I != E; ++I) {
3469 const WorkItem &WI = *I;
3470 size_t LUIdx = WI.LUIdx;
3471 LSRUse &LU = Uses[LUIdx];
3472 int64_t Imm = WI.Imm;
3473 const SCEV *OrigReg = WI.OrigReg;
3475 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3476 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3477 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3479 // TODO: Use a more targeted data structure.
3480 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3481 const Formula &F = LU.Formulae[L];
3482 // Use the immediate in the scaled register.
3483 if (F.ScaledReg == OrigReg) {
3484 int64_t Offs = (uint64_t)F.AM.BaseOffs +
3485 Imm * (uint64_t)F.AM.Scale;
3486 // Don't create 50 + reg(-50).
3487 if (F.referencesReg(SE.getSCEV(
3488 ConstantInt::get(IntTy, -(uint64_t)Offs))))
3491 NewF.AM.BaseOffs = Offs;
3492 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3493 LU.Kind, LU.AccessTy, TLI))
3495 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3497 // If the new scale is a constant in a register, and adding the constant
3498 // value to the immediate would produce a value closer to zero than the
3499 // immediate itself, then the formula isn't worthwhile.
3500 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3501 if (C->getValue()->isNegative() !=
3502 (NewF.AM.BaseOffs < 0) &&
3503 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
3504 .ule(abs64(NewF.AM.BaseOffs)))
3508 (void)InsertFormula(LU, LUIdx, NewF);
3510 // Use the immediate in a base register.
3511 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3512 const SCEV *BaseReg = F.BaseRegs[N];
3513 if (BaseReg != OrigReg)
3516 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
3517 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3518 LU.Kind, LU.AccessTy, TLI)) {
3520 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3523 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3525 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3527 // If the new formula has a constant in a register, and adding the
3528 // constant value to the immediate would produce a value closer to
3529 // zero than the immediate itself, then the formula isn't worthwhile.
3530 for (SmallVectorImpl<const SCEV *>::const_iterator
3531 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3533 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3534 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
3535 abs64(NewF.AM.BaseOffs)) &&
3536 (C->getValue()->getValue() +
3537 NewF.AM.BaseOffs).countTrailingZeros() >=
3538 CountTrailingZeros_64(NewF.AM.BaseOffs))
3542 (void)InsertFormula(LU, LUIdx, NewF);
3551 /// GenerateAllReuseFormulae - Generate formulae for each use.
3553 LSRInstance::GenerateAllReuseFormulae() {
3554 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3555 // queries are more precise.
3556 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3557 LSRUse &LU = Uses[LUIdx];
3558 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3559 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3560 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3561 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3563 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3564 LSRUse &LU = Uses[LUIdx];
3565 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3566 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3567 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3568 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3569 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3570 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3571 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3572 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3574 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3575 LSRUse &LU = Uses[LUIdx];
3576 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3577 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3580 GenerateCrossUseConstantOffsets();
3582 DEBUG(dbgs() << "\n"
3583 "After generating reuse formulae:\n";
3584 print_uses(dbgs()));
3587 /// If there are multiple formulae with the same set of registers used
3588 /// by other uses, pick the best one and delete the others.
3589 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3590 DenseSet<const SCEV *> VisitedRegs;
3591 SmallPtrSet<const SCEV *, 16> Regs;
3592 SmallPtrSet<const SCEV *, 16> LoserRegs;
3594 bool ChangedFormulae = false;
3597 // Collect the best formula for each unique set of shared registers. This
3598 // is reset for each use.
3599 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
3601 BestFormulaeTy BestFormulae;
3603 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3604 LSRUse &LU = Uses[LUIdx];
3605 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3608 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3609 FIdx != NumForms; ++FIdx) {
3610 Formula &F = LU.Formulae[FIdx];
3612 // Some formulas are instant losers. For example, they may depend on
3613 // nonexistent AddRecs from other loops. These need to be filtered
3614 // immediately, otherwise heuristics could choose them over others leading
3615 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3616 // avoids the need to recompute this information across formulae using the
3617 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3618 // the corresponding bad register from the Regs set.
3621 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3623 if (CostF.isLoser()) {
3624 // During initial formula generation, undesirable formulae are generated
3625 // by uses within other loops that have some non-trivial address mode or
3626 // use the postinc form of the IV. LSR needs to provide these formulae
3627 // as the basis of rediscovering the desired formula that uses an AddRec
3628 // corresponding to the existing phi. Once all formulae have been
3629 // generated, these initial losers may be pruned.
3630 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3634 SmallVector<const SCEV *, 2> Key;
3635 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3636 JE = F.BaseRegs.end(); J != JE; ++J) {
3637 const SCEV *Reg = *J;
3638 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3642 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3643 Key.push_back(F.ScaledReg);
3644 // Unstable sort by host order ok, because this is only used for
3646 std::sort(Key.begin(), Key.end());
3648 std::pair<BestFormulaeTy::const_iterator, bool> P =
3649 BestFormulae.insert(std::make_pair(Key, FIdx));
3653 Formula &Best = LU.Formulae[P.first->second];
3657 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3658 if (CostF < CostBest)
3660 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3662 " in favor of formula "; Best.print(dbgs());
3666 ChangedFormulae = true;
3668 LU.DeleteFormula(F);
3674 // Now that we've filtered out some formulae, recompute the Regs set.
3676 LU.RecomputeRegs(LUIdx, RegUses);
3678 // Reset this to prepare for the next use.
3679 BestFormulae.clear();
3682 DEBUG(if (ChangedFormulae) {
3684 "After filtering out undesirable candidates:\n";
3689 // This is a rough guess that seems to work fairly well.
3690 static const size_t ComplexityLimit = UINT16_MAX;
3692 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3693 /// solutions the solver might have to consider. It almost never considers
3694 /// this many solutions because it prune the search space, but the pruning
3695 /// isn't always sufficient.
3696 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3698 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3699 E = Uses.end(); I != E; ++I) {
3700 size_t FSize = I->Formulae.size();
3701 if (FSize >= ComplexityLimit) {
3702 Power = ComplexityLimit;
3706 if (Power >= ComplexityLimit)
3712 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3713 /// of the registers of another formula, it won't help reduce register
3714 /// pressure (though it may not necessarily hurt register pressure); remove
3715 /// it to simplify the system.
3716 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3717 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3718 DEBUG(dbgs() << "The search space is too complex.\n");
3720 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3721 "which use a superset of registers used by other "
3724 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3725 LSRUse &LU = Uses[LUIdx];
3727 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3728 Formula &F = LU.Formulae[i];
3729 // Look for a formula with a constant or GV in a register. If the use
3730 // also has a formula with that same value in an immediate field,
3731 // delete the one that uses a register.
3732 for (SmallVectorImpl<const SCEV *>::const_iterator
3733 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3734 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3736 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3737 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3738 (I - F.BaseRegs.begin()));
3739 if (LU.HasFormulaWithSameRegs(NewF)) {
3740 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3741 LU.DeleteFormula(F);
3747 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3748 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3751 NewF.AM.BaseGV = GV;
3752 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3753 (I - F.BaseRegs.begin()));
3754 if (LU.HasFormulaWithSameRegs(NewF)) {
3755 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3757 LU.DeleteFormula(F);
3768 LU.RecomputeRegs(LUIdx, RegUses);
3771 DEBUG(dbgs() << "After pre-selection:\n";
3772 print_uses(dbgs()));
3776 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3777 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3779 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3780 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3781 DEBUG(dbgs() << "The search space is too complex.\n");
3783 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3784 "separated by a constant offset will use the same "
3787 // This is especially useful for unrolled loops.
3789 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3790 LSRUse &LU = Uses[LUIdx];
3791 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3792 E = LU.Formulae.end(); I != E; ++I) {
3793 const Formula &F = *I;
3794 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3795 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3796 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3797 /*HasBaseReg=*/false,
3798 LU.Kind, LU.AccessTy)) {
3799 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3802 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3804 // Update the relocs to reference the new use.
3805 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3806 E = Fixups.end(); I != E; ++I) {
3807 LSRFixup &Fixup = *I;
3808 if (Fixup.LUIdx == LUIdx) {
3809 Fixup.LUIdx = LUThatHas - &Uses.front();
3810 Fixup.Offset += F.AM.BaseOffs;
3811 // Add the new offset to LUThatHas' offset list.
3812 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3813 LUThatHas->Offsets.push_back(Fixup.Offset);
3814 if (Fixup.Offset > LUThatHas->MaxOffset)
3815 LUThatHas->MaxOffset = Fixup.Offset;
3816 if (Fixup.Offset < LUThatHas->MinOffset)
3817 LUThatHas->MinOffset = Fixup.Offset;
3819 DEBUG(dbgs() << "New fixup has offset "
3820 << Fixup.Offset << '\n');
3822 if (Fixup.LUIdx == NumUses-1)
3823 Fixup.LUIdx = LUIdx;
3826 // Delete formulae from the new use which are no longer legal.
3828 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3829 Formula &F = LUThatHas->Formulae[i];
3830 if (!isLegalUse(F.AM,
3831 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3832 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3833 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3835 LUThatHas->DeleteFormula(F);
3842 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3844 // Delete the old use.
3845 DeleteUse(LU, LUIdx);
3855 DEBUG(dbgs() << "After pre-selection:\n";
3856 print_uses(dbgs()));
3860 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3861 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3862 /// we've done more filtering, as it may be able to find more formulae to
3864 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3865 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3866 DEBUG(dbgs() << "The search space is too complex.\n");
3868 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3869 "undesirable dedicated registers.\n");
3871 FilterOutUndesirableDedicatedRegisters();
3873 DEBUG(dbgs() << "After pre-selection:\n";
3874 print_uses(dbgs()));
3878 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3879 /// to be profitable, and then in any use which has any reference to that
3880 /// register, delete all formulae which do not reference that register.
3881 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3882 // With all other options exhausted, loop until the system is simple
3883 // enough to handle.
3884 SmallPtrSet<const SCEV *, 4> Taken;
3885 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3886 // Ok, we have too many of formulae on our hands to conveniently handle.
3887 // Use a rough heuristic to thin out the list.
3888 DEBUG(dbgs() << "The search space is too complex.\n");
3890 // Pick the register which is used by the most LSRUses, which is likely
3891 // to be a good reuse register candidate.
3892 const SCEV *Best = 0;
3893 unsigned BestNum = 0;
3894 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3896 const SCEV *Reg = *I;
3897 if (Taken.count(Reg))
3902 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3903 if (Count > BestNum) {
3910 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3911 << " will yield profitable reuse.\n");
3914 // In any use with formulae which references this register, delete formulae
3915 // which don't reference it.
3916 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3917 LSRUse &LU = Uses[LUIdx];
3918 if (!LU.Regs.count(Best)) continue;
3921 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3922 Formula &F = LU.Formulae[i];
3923 if (!F.referencesReg(Best)) {
3924 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3925 LU.DeleteFormula(F);
3929 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3935 LU.RecomputeRegs(LUIdx, RegUses);
3938 DEBUG(dbgs() << "After pre-selection:\n";
3939 print_uses(dbgs()));
3943 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3944 /// formulae to choose from, use some rough heuristics to prune down the number
3945 /// of formulae. This keeps the main solver from taking an extraordinary amount
3946 /// of time in some worst-case scenarios.
3947 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3948 NarrowSearchSpaceByDetectingSupersets();
3949 NarrowSearchSpaceByCollapsingUnrolledCode();
3950 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3951 NarrowSearchSpaceByPickingWinnerRegs();
3954 /// SolveRecurse - This is the recursive solver.
3955 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3957 SmallVectorImpl<const Formula *> &Workspace,
3958 const Cost &CurCost,
3959 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3960 DenseSet<const SCEV *> &VisitedRegs) const {
3963 // - use more aggressive filtering
3964 // - sort the formula so that the most profitable solutions are found first
3965 // - sort the uses too
3967 // - don't compute a cost, and then compare. compare while computing a cost
3969 // - track register sets with SmallBitVector
3971 const LSRUse &LU = Uses[Workspace.size()];
3973 // If this use references any register that's already a part of the
3974 // in-progress solution, consider it a requirement that a formula must
3975 // reference that register in order to be considered. This prunes out
3976 // unprofitable searching.
3977 SmallSetVector<const SCEV *, 4> ReqRegs;
3978 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3979 E = CurRegs.end(); I != E; ++I)
3980 if (LU.Regs.count(*I))
3983 bool AnySatisfiedReqRegs = false;
3984 SmallPtrSet<const SCEV *, 16> NewRegs;
3987 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3988 E = LU.Formulae.end(); I != E; ++I) {
3989 const Formula &F = *I;
3991 // Ignore formulae which do not use any of the required registers.
3992 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3993 JE = ReqRegs.end(); J != JE; ++J) {
3994 const SCEV *Reg = *J;
3995 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3996 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4000 AnySatisfiedReqRegs = true;
4002 // Evaluate the cost of the current formula. If it's already worse than
4003 // the current best, prune the search at that point.
4006 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
4007 if (NewCost < SolutionCost) {
4008 Workspace.push_back(&F);
4009 if (Workspace.size() != Uses.size()) {
4010 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4011 NewRegs, VisitedRegs);
4012 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4013 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4015 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4016 dbgs() << ".\n Regs:";
4017 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4018 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4019 dbgs() << ' ' << **I;
4022 SolutionCost = NewCost;
4023 Solution = Workspace;
4025 Workspace.pop_back();
4030 if (!EnableRetry && !AnySatisfiedReqRegs)
4033 // If none of the formulae had all of the required registers, relax the
4034 // constraint so that we don't exclude all formulae.
4035 if (!AnySatisfiedReqRegs) {
4036 assert(!ReqRegs.empty() && "Solver failed even without required registers");
4042 /// Solve - Choose one formula from each use. Return the results in the given
4043 /// Solution vector.
4044 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4045 SmallVector<const Formula *, 8> Workspace;
4047 SolutionCost.Loose();
4049 SmallPtrSet<const SCEV *, 16> CurRegs;
4050 DenseSet<const SCEV *> VisitedRegs;
4051 Workspace.reserve(Uses.size());
4053 // SolveRecurse does all the work.
4054 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4055 CurRegs, VisitedRegs);
4056 if (Solution.empty()) {
4057 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4061 // Ok, we've now made all our decisions.
4062 DEBUG(dbgs() << "\n"
4063 "The chosen solution requires "; SolutionCost.print(dbgs());
4065 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4067 Uses[i].print(dbgs());
4070 Solution[i]->print(dbgs());
4074 assert(Solution.size() == Uses.size() && "Malformed solution!");
4077 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4078 /// the dominator tree far as we can go while still being dominated by the
4079 /// input positions. This helps canonicalize the insert position, which
4080 /// encourages sharing.
4081 BasicBlock::iterator
4082 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4083 const SmallVectorImpl<Instruction *> &Inputs)
4086 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4087 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4090 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4091 if (!Rung) return IP;
4092 Rung = Rung->getIDom();
4093 if (!Rung) return IP;
4094 IDom = Rung->getBlock();
4096 // Don't climb into a loop though.
4097 const Loop *IDomLoop = LI.getLoopFor(IDom);
4098 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4099 if (IDomDepth <= IPLoopDepth &&
4100 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4104 bool AllDominate = true;
4105 Instruction *BetterPos = 0;
4106 Instruction *Tentative = IDom->getTerminator();
4107 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4108 E = Inputs.end(); I != E; ++I) {
4109 Instruction *Inst = *I;
4110 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4111 AllDominate = false;
4114 // Attempt to find an insert position in the middle of the block,
4115 // instead of at the end, so that it can be used for other expansions.
4116 if (IDom == Inst->getParent() &&
4117 (!BetterPos || DT.dominates(BetterPos, Inst)))
4118 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4131 /// AdjustInsertPositionForExpand - Determine an input position which will be
4132 /// dominated by the operands and which will dominate the result.
4133 BasicBlock::iterator
4134 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
4136 const LSRUse &LU) const {
4137 // Collect some instructions which must be dominated by the
4138 // expanding replacement. These must be dominated by any operands that
4139 // will be required in the expansion.
4140 SmallVector<Instruction *, 4> Inputs;
4141 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4142 Inputs.push_back(I);
4143 if (LU.Kind == LSRUse::ICmpZero)
4144 if (Instruction *I =
4145 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4146 Inputs.push_back(I);
4147 if (LF.PostIncLoops.count(L)) {
4148 if (LF.isUseFullyOutsideLoop(L))
4149 Inputs.push_back(L->getLoopLatch()->getTerminator());
4151 Inputs.push_back(IVIncInsertPos);
4153 // The expansion must also be dominated by the increment positions of any
4154 // loops it for which it is using post-inc mode.
4155 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4156 E = LF.PostIncLoops.end(); I != E; ++I) {
4157 const Loop *PIL = *I;
4158 if (PIL == L) continue;
4160 // Be dominated by the loop exit.
4161 SmallVector<BasicBlock *, 4> ExitingBlocks;
4162 PIL->getExitingBlocks(ExitingBlocks);
4163 if (!ExitingBlocks.empty()) {
4164 BasicBlock *BB = ExitingBlocks[0];
4165 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4166 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4167 Inputs.push_back(BB->getTerminator());
4171 // Then, climb up the immediate dominator tree as far as we can go while
4172 // still being dominated by the input positions.
4173 IP = HoistInsertPosition(IP, Inputs);
4175 // Don't insert instructions before PHI nodes.
4176 while (isa<PHINode>(IP)) ++IP;
4178 // Ignore landingpad instructions.
4179 while (isa<LandingPadInst>(IP)) ++IP;
4181 // Ignore debug intrinsics.
4182 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4187 /// Expand - Emit instructions for the leading candidate expression for this
4188 /// LSRUse (this is called "expanding").
4189 Value *LSRInstance::Expand(const LSRFixup &LF,
4191 BasicBlock::iterator IP,
4192 SCEVExpander &Rewriter,
4193 SmallVectorImpl<WeakVH> &DeadInsts) const {
4194 const LSRUse &LU = Uses[LF.LUIdx];
4196 // Determine an input position which will be dominated by the operands and
4197 // which will dominate the result.
4198 IP = AdjustInsertPositionForExpand(IP, LF, LU);
4200 // Inform the Rewriter if we have a post-increment use, so that it can
4201 // perform an advantageous expansion.
4202 Rewriter.setPostInc(LF.PostIncLoops);
4204 // This is the type that the user actually needs.
4205 Type *OpTy = LF.OperandValToReplace->getType();
4206 // This will be the type that we'll initially expand to.
4207 Type *Ty = F.getType();
4209 // No type known; just expand directly to the ultimate type.
4211 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4212 // Expand directly to the ultimate type if it's the right size.
4214 // This is the type to do integer arithmetic in.
4215 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4217 // Build up a list of operands to add together to form the full base.
4218 SmallVector<const SCEV *, 8> Ops;
4220 // Expand the BaseRegs portion.
4221 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4222 E = F.BaseRegs.end(); I != E; ++I) {
4223 const SCEV *Reg = *I;
4224 assert(!Reg->isZero() && "Zero allocated in a base register!");
4226 // If we're expanding for a post-inc user, make the post-inc adjustment.
4227 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4228 Reg = TransformForPostIncUse(Denormalize, Reg,
4229 LF.UserInst, LF.OperandValToReplace,
4232 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4235 // Flush the operand list to suppress SCEVExpander hoisting.
4237 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4239 Ops.push_back(SE.getUnknown(FullV));
4242 // Expand the ScaledReg portion.
4243 Value *ICmpScaledV = 0;
4244 if (F.AM.Scale != 0) {
4245 const SCEV *ScaledS = F.ScaledReg;
4247 // If we're expanding for a post-inc user, make the post-inc adjustment.
4248 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4249 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4250 LF.UserInst, LF.OperandValToReplace,
4253 if (LU.Kind == LSRUse::ICmpZero) {
4254 // An interesting way of "folding" with an icmp is to use a negated
4255 // scale, which we'll implement by inserting it into the other operand
4257 assert(F.AM.Scale == -1 &&
4258 "The only scale supported by ICmpZero uses is -1!");
4259 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4261 // Otherwise just expand the scaled register and an explicit scale,
4262 // which is expected to be matched as part of the address.
4263 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4264 ScaledS = SE.getMulExpr(ScaledS,
4265 SE.getConstant(ScaledS->getType(), F.AM.Scale));
4266 Ops.push_back(ScaledS);
4268 // Flush the operand list to suppress SCEVExpander hoisting.
4269 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4271 Ops.push_back(SE.getUnknown(FullV));
4275 // Expand the GV portion.
4277 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
4279 // Flush the operand list to suppress SCEVExpander hoisting.
4280 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4282 Ops.push_back(SE.getUnknown(FullV));
4285 // Expand the immediate portion.
4286 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
4288 if (LU.Kind == LSRUse::ICmpZero) {
4289 // The other interesting way of "folding" with an ICmpZero is to use a
4290 // negated immediate.
4292 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4294 Ops.push_back(SE.getUnknown(ICmpScaledV));
4295 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4298 // Just add the immediate values. These again are expected to be matched
4299 // as part of the address.
4300 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4304 // Expand the unfolded offset portion.
4305 int64_t UnfoldedOffset = F.UnfoldedOffset;
4306 if (UnfoldedOffset != 0) {
4307 // Just add the immediate values.
4308 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4312 // Emit instructions summing all the operands.
4313 const SCEV *FullS = Ops.empty() ?
4314 SE.getConstant(IntTy, 0) :
4316 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4318 // We're done expanding now, so reset the rewriter.
4319 Rewriter.clearPostInc();
4321 // An ICmpZero Formula represents an ICmp which we're handling as a
4322 // comparison against zero. Now that we've expanded an expression for that
4323 // form, update the ICmp's other operand.
4324 if (LU.Kind == LSRUse::ICmpZero) {
4325 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4326 DeadInsts.push_back(CI->getOperand(1));
4327 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
4328 "a scale at the same time!");
4329 if (F.AM.Scale == -1) {
4330 if (ICmpScaledV->getType() != OpTy) {
4332 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4334 ICmpScaledV, OpTy, "tmp", CI);
4337 CI->setOperand(1, ICmpScaledV);
4339 assert(F.AM.Scale == 0 &&
4340 "ICmp does not support folding a global value and "
4341 "a scale at the same time!");
4342 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4344 if (C->getType() != OpTy)
4345 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4349 CI->setOperand(1, C);
4356 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4357 /// of their operands effectively happens in their predecessor blocks, so the
4358 /// expression may need to be expanded in multiple places.
4359 void LSRInstance::RewriteForPHI(PHINode *PN,
4362 SCEVExpander &Rewriter,
4363 SmallVectorImpl<WeakVH> &DeadInsts,
4365 DenseMap<BasicBlock *, Value *> Inserted;
4366 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4367 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4368 BasicBlock *BB = PN->getIncomingBlock(i);
4370 // If this is a critical edge, split the edge so that we do not insert
4371 // the code on all predecessor/successor paths. We do this unless this
4372 // is the canonical backedge for this loop, which complicates post-inc
4374 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4375 !isa<IndirectBrInst>(BB->getTerminator())) {
4376 BasicBlock *Parent = PN->getParent();
4377 Loop *PNLoop = LI.getLoopFor(Parent);
4378 if (!PNLoop || Parent != PNLoop->getHeader()) {
4379 // Split the critical edge.
4380 BasicBlock *NewBB = 0;
4381 if (!Parent->isLandingPad()) {
4382 NewBB = SplitCriticalEdge(BB, Parent, P,
4383 /*MergeIdenticalEdges=*/true,
4384 /*DontDeleteUselessPhis=*/true);
4386 SmallVector<BasicBlock*, 2> NewBBs;
4387 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4391 // If PN is outside of the loop and BB is in the loop, we want to
4392 // move the block to be immediately before the PHI block, not
4393 // immediately after BB.
4394 if (L->contains(BB) && !L->contains(PN))
4395 NewBB->moveBefore(PN->getParent());
4397 // Splitting the edge can reduce the number of PHI entries we have.
4398 e = PN->getNumIncomingValues();
4400 i = PN->getBasicBlockIndex(BB);
4404 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4405 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4407 PN->setIncomingValue(i, Pair.first->second);
4409 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4411 // If this is reuse-by-noop-cast, insert the noop cast.
4412 Type *OpTy = LF.OperandValToReplace->getType();
4413 if (FullV->getType() != OpTy)
4415 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4417 FullV, LF.OperandValToReplace->getType(),
4418 "tmp", BB->getTerminator());
4420 PN->setIncomingValue(i, FullV);
4421 Pair.first->second = FullV;
4426 /// Rewrite - Emit instructions for the leading candidate expression for this
4427 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4428 /// the newly expanded value.
4429 void LSRInstance::Rewrite(const LSRFixup &LF,
4431 SCEVExpander &Rewriter,
4432 SmallVectorImpl<WeakVH> &DeadInsts,
4434 // First, find an insertion point that dominates UserInst. For PHI nodes,
4435 // find the nearest block which dominates all the relevant uses.
4436 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4437 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4439 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4441 // If this is reuse-by-noop-cast, insert the noop cast.
4442 Type *OpTy = LF.OperandValToReplace->getType();
4443 if (FullV->getType() != OpTy) {
4445 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4446 FullV, OpTy, "tmp", LF.UserInst);
4450 // Update the user. ICmpZero is handled specially here (for now) because
4451 // Expand may have updated one of the operands of the icmp already, and
4452 // its new value may happen to be equal to LF.OperandValToReplace, in
4453 // which case doing replaceUsesOfWith leads to replacing both operands
4454 // with the same value. TODO: Reorganize this.
4455 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4456 LF.UserInst->setOperand(0, FullV);
4458 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4461 DeadInsts.push_back(LF.OperandValToReplace);
4464 /// ImplementSolution - Rewrite all the fixup locations with new values,
4465 /// following the chosen solution.
4467 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4469 // Keep track of instructions we may have made dead, so that
4470 // we can remove them after we are done working.
4471 SmallVector<WeakVH, 16> DeadInsts;
4473 SCEVExpander Rewriter(SE, "lsr");
4475 Rewriter.setDebugType(DEBUG_TYPE);
4477 Rewriter.disableCanonicalMode();
4478 Rewriter.enableLSRMode();
4479 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4481 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4482 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4483 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4484 if (PHINode *PN = dyn_cast<PHINode>(ChainI->back().UserInst))
4485 Rewriter.setChainedPhi(PN);
4488 // Expand the new value definitions and update the users.
4489 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4490 E = Fixups.end(); I != E; ++I) {
4491 const LSRFixup &Fixup = *I;
4493 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4498 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4499 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4500 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4503 // Clean up after ourselves. This must be done before deleting any
4507 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4510 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
4511 : IU(P->getAnalysis<IVUsers>()),
4512 SE(P->getAnalysis<ScalarEvolution>()),
4513 DT(P->getAnalysis<DominatorTree>()),
4514 LI(P->getAnalysis<LoopInfo>()),
4515 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
4517 // If LoopSimplify form is not available, stay out of trouble.
4518 if (!L->isLoopSimplifyForm())
4521 // All outer loops must have preheaders, or SCEVExpander may not be able to
4522 // materialize an AddRecExpr whose Start is an outer AddRecExpr.
4523 for (const Loop *OuterLoop = L; (OuterLoop = OuterLoop->getParentLoop());) {
4524 if (!OuterLoop->getLoopPreheader())
4527 // If there's no interesting work to be done, bail early.
4528 if (IU.empty()) return;
4530 DEBUG(dbgs() << "\nLSR on loop ";
4531 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4534 // First, perform some low-level loop optimizations.
4536 OptimizeLoopTermCond();
4538 // If loop preparation eliminates all interesting IV users, bail.
4539 if (IU.empty()) return;
4541 // Skip nested loops until we can model them better with formulae.
4542 if (!EnableNested && !L->empty()) {
4543 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4547 // Start collecting data and preparing for the solver.
4549 CollectInterestingTypesAndFactors();
4550 CollectFixupsAndInitialFormulae();
4551 CollectLoopInvariantFixupsAndFormulae();
4553 assert(!Uses.empty() && "IVUsers reported at least one use");
4554 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4555 print_uses(dbgs()));
4557 // Now use the reuse data to generate a bunch of interesting ways
4558 // to formulate the values needed for the uses.
4559 GenerateAllReuseFormulae();
4561 FilterOutUndesirableDedicatedRegisters();
4562 NarrowSearchSpaceUsingHeuristics();
4564 SmallVector<const Formula *, 8> Solution;
4567 // Release memory that is no longer needed.
4572 if (Solution.empty())
4576 // Formulae should be legal.
4577 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4578 E = Uses.end(); I != E; ++I) {
4579 const LSRUse &LU = *I;
4580 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4581 JE = LU.Formulae.end(); J != JE; ++J)
4582 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
4583 LU.Kind, LU.AccessTy, TLI) &&
4584 "Illegal formula generated!");
4588 // Now that we've decided what we want, make it so.
4589 ImplementSolution(Solution, P);
4592 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4593 if (Factors.empty() && Types.empty()) return;
4595 OS << "LSR has identified the following interesting factors and types: ";
4598 for (SmallSetVector<int64_t, 8>::const_iterator
4599 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4600 if (!First) OS << ", ";
4605 for (SmallSetVector<Type *, 4>::const_iterator
4606 I = Types.begin(), E = Types.end(); I != E; ++I) {
4607 if (!First) OS << ", ";
4609 OS << '(' << **I << ')';
4614 void LSRInstance::print_fixups(raw_ostream &OS) const {
4615 OS << "LSR is examining the following fixup sites:\n";
4616 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4617 E = Fixups.end(); I != E; ++I) {
4624 void LSRInstance::print_uses(raw_ostream &OS) const {
4625 OS << "LSR is examining the following uses:\n";
4626 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4627 E = Uses.end(); I != E; ++I) {
4628 const LSRUse &LU = *I;
4632 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4633 JE = LU.Formulae.end(); J != JE; ++J) {
4641 void LSRInstance::print(raw_ostream &OS) const {
4642 print_factors_and_types(OS);
4647 void LSRInstance::dump() const {
4648 print(errs()); errs() << '\n';
4653 class LoopStrengthReduce : public LoopPass {
4654 /// TLI - Keep a pointer of a TargetLowering to consult for determining
4655 /// transformation profitability.
4656 const TargetLowering *const TLI;
4659 static char ID; // Pass ID, replacement for typeid
4660 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
4663 bool runOnLoop(Loop *L, LPPassManager &LPM);
4664 void getAnalysisUsage(AnalysisUsage &AU) const;
4669 char LoopStrengthReduce::ID = 0;
4670 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4671 "Loop Strength Reduction", false, false)
4672 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4673 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4674 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4675 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4676 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4677 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4678 "Loop Strength Reduction", false, false)
4681 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
4682 return new LoopStrengthReduce(TLI);
4685 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
4686 : LoopPass(ID), TLI(tli) {
4687 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4690 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4691 // We split critical edges, so we change the CFG. However, we do update
4692 // many analyses if they are around.
4693 AU.addPreservedID(LoopSimplifyID);
4695 AU.addRequired<LoopInfo>();
4696 AU.addPreserved<LoopInfo>();
4697 AU.addRequiredID(LoopSimplifyID);
4698 AU.addRequired<DominatorTree>();
4699 AU.addPreserved<DominatorTree>();
4700 AU.addRequired<ScalarEvolution>();
4701 AU.addPreserved<ScalarEvolution>();
4702 // Requiring LoopSimplify a second time here prevents IVUsers from running
4703 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4704 AU.addRequiredID(LoopSimplifyID);
4705 AU.addRequired<IVUsers>();
4706 AU.addPreserved<IVUsers>();
4709 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4710 bool Changed = false;
4712 // Run the main LSR transformation.
4713 Changed |= LSRInstance(TLI, L, this).getChanged();
4715 // Remove any extra phis created by processing inner loops.
4716 Changed |= DeleteDeadPHIs(L->getHeader());
4717 if (EnablePhiElim) {
4718 SmallVector<WeakVH, 16> DeadInsts;
4719 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4721 Rewriter.setDebugType(DEBUG_TYPE);
4723 unsigned numFolded = Rewriter.
4724 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI);
4727 DeleteTriviallyDeadInstructions(DeadInsts);
4728 DeleteDeadPHIs(L->getHeader());