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 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
81 /// bail out. This threshold is far beyond the number of users that LSR can
82 /// conceivably solve, so it should not affect generated code, but catches the
83 /// worst cases before LSR burns too much compile time and stack space.
84 static const unsigned MaxIVUsers = 200;
86 // Temporary flag to cleanup congruent phis after LSR phi expansion.
87 // It's currently disabled until we can determine whether it's truly useful or
88 // not. The flag should be removed after the v3.0 release.
89 // This is now needed for ivchains.
90 static cl::opt<bool> EnablePhiElim(
91 "enable-lsr-phielim", cl::Hidden, cl::init(true),
92 cl::desc("Enable LSR phi elimination"));
95 // Stress test IV chain generation.
96 static cl::opt<bool> StressIVChain(
97 "stress-ivchain", cl::Hidden, cl::init(false),
98 cl::desc("Stress test LSR IV chains"));
100 static bool StressIVChain = false;
105 /// RegSortData - This class holds data which is used to order reuse candidates.
108 /// UsedByIndices - This represents the set of LSRUse indices which reference
109 /// a particular register.
110 SmallBitVector UsedByIndices;
114 void print(raw_ostream &OS) const;
120 void RegSortData::print(raw_ostream &OS) const {
121 OS << "[NumUses=" << UsedByIndices.count() << ']';
124 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 // If U is a constant, it may be used by a ConstantExpr.
714 Instruction *User = dyn_cast<Instruction>(*UI);
715 if (User && User->getOpcode() == Instruction::Mul
716 && SE.isSCEVable(User->getType())) {
717 return SE.getSCEV(User) == Mul;
724 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
725 if (isExistingPhi(AR, SE))
729 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
733 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
734 /// specified set are trivially dead, delete them and see if this makes any of
735 /// their operands subsequently dead.
737 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
738 bool Changed = false;
740 while (!DeadInsts.empty()) {
741 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
743 if (I == 0 || !isInstructionTriviallyDead(I))
746 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
747 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
750 DeadInsts.push_back(U);
753 I->eraseFromParent();
762 /// Cost - This class is used to measure and compare candidate formulae.
764 /// TODO: Some of these could be merged. Also, a lexical ordering
765 /// isn't always optimal.
769 unsigned NumBaseAdds;
775 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
778 bool operator<(const Cost &Other) const;
783 // Once any of the metrics loses, they must all remain losers.
785 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
786 | ImmCost | SetupCost) != ~0u)
787 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
788 & ImmCost & SetupCost) == ~0u);
793 assert(isValid() && "invalid cost");
794 return NumRegs == ~0u;
797 void RateFormula(const Formula &F,
798 SmallPtrSet<const SCEV *, 16> &Regs,
799 const DenseSet<const SCEV *> &VisitedRegs,
801 const SmallVectorImpl<int64_t> &Offsets,
802 ScalarEvolution &SE, DominatorTree &DT,
803 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
805 void print(raw_ostream &OS) const;
809 void RateRegister(const SCEV *Reg,
810 SmallPtrSet<const SCEV *, 16> &Regs,
812 ScalarEvolution &SE, DominatorTree &DT);
813 void RatePrimaryRegister(const SCEV *Reg,
814 SmallPtrSet<const SCEV *, 16> &Regs,
816 ScalarEvolution &SE, DominatorTree &DT,
817 SmallPtrSet<const SCEV *, 16> *LoserRegs);
822 /// RateRegister - Tally up interesting quantities from the given register.
823 void Cost::RateRegister(const SCEV *Reg,
824 SmallPtrSet<const SCEV *, 16> &Regs,
826 ScalarEvolution &SE, DominatorTree &DT) {
827 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
828 // If this is an addrec for another loop, don't second-guess its addrec phi
829 // nodes. LSR isn't currently smart enough to reason about more than one
830 // loop at a time. LSR has already run on inner loops, will not run on outer
831 // loops, and cannot be expected to change sibling loops.
832 if (AR->getLoop() != L) {
833 // If the AddRec exists, consider it's register free and leave it alone.
834 if (isExistingPhi(AR, SE))
837 // Otherwise, do not consider this formula at all.
841 AddRecCost += 1; /// TODO: This should be a function of the stride.
843 // Add the step value register, if it needs one.
844 // TODO: The non-affine case isn't precisely modeled here.
845 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
846 if (!Regs.count(AR->getOperand(1))) {
847 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
855 // Rough heuristic; favor registers which don't require extra setup
856 // instructions in the preheader.
857 if (!isa<SCEVUnknown>(Reg) &&
858 !isa<SCEVConstant>(Reg) &&
859 !(isa<SCEVAddRecExpr>(Reg) &&
860 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
861 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
864 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
865 SE.hasComputableLoopEvolution(Reg, L);
868 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
869 /// before, rate it. Optional LoserRegs provides a way to declare any formula
870 /// that refers to one of those regs an instant loser.
871 void Cost::RatePrimaryRegister(const SCEV *Reg,
872 SmallPtrSet<const SCEV *, 16> &Regs,
874 ScalarEvolution &SE, DominatorTree &DT,
875 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
876 if (LoserRegs && LoserRegs->count(Reg)) {
880 if (Regs.insert(Reg)) {
881 RateRegister(Reg, Regs, L, SE, DT);
883 LoserRegs->insert(Reg);
887 void Cost::RateFormula(const Formula &F,
888 SmallPtrSet<const SCEV *, 16> &Regs,
889 const DenseSet<const SCEV *> &VisitedRegs,
891 const SmallVectorImpl<int64_t> &Offsets,
892 ScalarEvolution &SE, DominatorTree &DT,
893 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
894 // Tally up the registers.
895 if (const SCEV *ScaledReg = F.ScaledReg) {
896 if (VisitedRegs.count(ScaledReg)) {
900 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
904 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
905 E = F.BaseRegs.end(); I != E; ++I) {
906 const SCEV *BaseReg = *I;
907 if (VisitedRegs.count(BaseReg)) {
911 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
916 // Determine how many (unfolded) adds we'll need inside the loop.
917 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
918 if (NumBaseParts > 1)
919 NumBaseAdds += NumBaseParts - 1;
921 // Tally up the non-zero immediates.
922 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
923 E = Offsets.end(); I != E; ++I) {
924 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
926 ImmCost += 64; // Handle symbolic values conservatively.
927 // TODO: This should probably be the pointer size.
928 else if (Offset != 0)
929 ImmCost += APInt(64, Offset, true).getMinSignedBits();
931 assert(isValid() && "invalid cost");
934 /// Loose - Set this cost to a losing value.
944 /// operator< - Choose the lower cost.
945 bool Cost::operator<(const Cost &Other) const {
946 if (NumRegs != Other.NumRegs)
947 return NumRegs < Other.NumRegs;
948 if (AddRecCost != Other.AddRecCost)
949 return AddRecCost < Other.AddRecCost;
950 if (NumIVMuls != Other.NumIVMuls)
951 return NumIVMuls < Other.NumIVMuls;
952 if (NumBaseAdds != Other.NumBaseAdds)
953 return NumBaseAdds < Other.NumBaseAdds;
954 if (ImmCost != Other.ImmCost)
955 return ImmCost < Other.ImmCost;
956 if (SetupCost != Other.SetupCost)
957 return SetupCost < Other.SetupCost;
961 void Cost::print(raw_ostream &OS) const {
962 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
964 OS << ", with addrec cost " << AddRecCost;
966 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
967 if (NumBaseAdds != 0)
968 OS << ", plus " << NumBaseAdds << " base add"
969 << (NumBaseAdds == 1 ? "" : "s");
971 OS << ", plus " << ImmCost << " imm cost";
973 OS << ", plus " << SetupCost << " setup cost";
976 void Cost::dump() const {
977 print(errs()); errs() << '\n';
982 /// LSRFixup - An operand value in an instruction which is to be replaced
983 /// with some equivalent, possibly strength-reduced, replacement.
985 /// UserInst - The instruction which will be updated.
986 Instruction *UserInst;
988 /// OperandValToReplace - The operand of the instruction which will
989 /// be replaced. The operand may be used more than once; every instance
990 /// will be replaced.
991 Value *OperandValToReplace;
993 /// PostIncLoops - If this user is to use the post-incremented value of an
994 /// induction variable, this variable is non-null and holds the loop
995 /// associated with the induction variable.
996 PostIncLoopSet PostIncLoops;
998 /// LUIdx - The index of the LSRUse describing the expression which
999 /// this fixup needs, minus an offset (below).
1002 /// Offset - A constant offset to be added to the LSRUse expression.
1003 /// This allows multiple fixups to share the same LSRUse with different
1004 /// offsets, for example in an unrolled loop.
1007 bool isUseFullyOutsideLoop(const Loop *L) const;
1011 void print(raw_ostream &OS) const;
1017 LSRFixup::LSRFixup()
1018 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1020 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1021 /// value outside of the given loop.
1022 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1023 // PHI nodes use their value in their incoming blocks.
1024 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1025 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1026 if (PN->getIncomingValue(i) == OperandValToReplace &&
1027 L->contains(PN->getIncomingBlock(i)))
1032 return !L->contains(UserInst);
1035 void LSRFixup::print(raw_ostream &OS) const {
1037 // Store is common and interesting enough to be worth special-casing.
1038 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1040 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1041 } else if (UserInst->getType()->isVoidTy())
1042 OS << UserInst->getOpcodeName();
1044 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1046 OS << ", OperandValToReplace=";
1047 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1049 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1050 E = PostIncLoops.end(); I != E; ++I) {
1051 OS << ", PostIncLoop=";
1052 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1055 if (LUIdx != ~size_t(0))
1056 OS << ", LUIdx=" << LUIdx;
1059 OS << ", Offset=" << Offset;
1062 void LSRFixup::dump() const {
1063 print(errs()); errs() << '\n';
1068 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1069 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1070 struct UniquifierDenseMapInfo {
1071 static SmallVector<const SCEV *, 2> getEmptyKey() {
1072 SmallVector<const SCEV *, 2> V;
1073 V.push_back(reinterpret_cast<const SCEV *>(-1));
1077 static SmallVector<const SCEV *, 2> getTombstoneKey() {
1078 SmallVector<const SCEV *, 2> V;
1079 V.push_back(reinterpret_cast<const SCEV *>(-2));
1083 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
1084 unsigned Result = 0;
1085 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1086 E = V.end(); I != E; ++I)
1087 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1091 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1092 const SmallVector<const SCEV *, 2> &RHS) {
1097 /// LSRUse - This class holds the state that LSR keeps for each use in
1098 /// IVUsers, as well as uses invented by LSR itself. It includes information
1099 /// about what kinds of things can be folded into the user, information about
1100 /// the user itself, and information about how the use may be satisfied.
1101 /// TODO: Represent multiple users of the same expression in common?
1103 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1106 /// KindType - An enum for a kind of use, indicating what types of
1107 /// scaled and immediate operands it might support.
1109 Basic, ///< A normal use, with no folding.
1110 Special, ///< A special case of basic, allowing -1 scales.
1111 Address, ///< An address use; folding according to TargetLowering
1112 ICmpZero ///< An equality icmp with both operands folded into one.
1113 // TODO: Add a generic icmp too?
1119 SmallVector<int64_t, 8> Offsets;
1123 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1124 /// LSRUse are outside of the loop, in which case some special-case heuristics
1126 bool AllFixupsOutsideLoop;
1128 /// WidestFixupType - This records the widest use type for any fixup using
1129 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1130 /// max fixup widths to be equivalent, because the narrower one may be relying
1131 /// on the implicit truncation to truncate away bogus bits.
1132 Type *WidestFixupType;
1134 /// Formulae - A list of ways to build a value that can satisfy this user.
1135 /// After the list is populated, one of these is selected heuristically and
1136 /// used to formulate a replacement for OperandValToReplace in UserInst.
1137 SmallVector<Formula, 12> Formulae;
1139 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1140 SmallPtrSet<const SCEV *, 4> Regs;
1142 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1143 MinOffset(INT64_MAX),
1144 MaxOffset(INT64_MIN),
1145 AllFixupsOutsideLoop(true),
1146 WidestFixupType(0) {}
1148 bool HasFormulaWithSameRegs(const Formula &F) const;
1149 bool InsertFormula(const Formula &F);
1150 void DeleteFormula(Formula &F);
1151 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1153 void print(raw_ostream &OS) const;
1159 /// HasFormula - Test whether this use as a formula which has the same
1160 /// registers as the given formula.
1161 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1162 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1163 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1164 // Unstable sort by host order ok, because this is only used for uniquifying.
1165 std::sort(Key.begin(), Key.end());
1166 return Uniquifier.count(Key);
1169 /// InsertFormula - If the given formula has not yet been inserted, add it to
1170 /// the list, and return true. Return false otherwise.
1171 bool LSRUse::InsertFormula(const Formula &F) {
1172 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1173 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1174 // Unstable sort by host order ok, because this is only used for uniquifying.
1175 std::sort(Key.begin(), Key.end());
1177 if (!Uniquifier.insert(Key).second)
1180 // Using a register to hold the value of 0 is not profitable.
1181 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1182 "Zero allocated in a scaled register!");
1184 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1185 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1186 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1189 // Add the formula to the list.
1190 Formulae.push_back(F);
1192 // Record registers now being used by this use.
1193 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1198 /// DeleteFormula - Remove the given formula from this use's list.
1199 void LSRUse::DeleteFormula(Formula &F) {
1200 if (&F != &Formulae.back())
1201 std::swap(F, Formulae.back());
1202 Formulae.pop_back();
1205 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1206 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1207 // Now that we've filtered out some formulae, recompute the Regs set.
1208 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1210 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1211 E = Formulae.end(); I != E; ++I) {
1212 const Formula &F = *I;
1213 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1214 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1217 // Update the RegTracker.
1218 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1219 E = OldRegs.end(); I != E; ++I)
1220 if (!Regs.count(*I))
1221 RegUses.DropRegister(*I, LUIdx);
1224 void LSRUse::print(raw_ostream &OS) const {
1225 OS << "LSR Use: Kind=";
1227 case Basic: OS << "Basic"; break;
1228 case Special: OS << "Special"; break;
1229 case ICmpZero: OS << "ICmpZero"; break;
1231 OS << "Address of ";
1232 if (AccessTy->isPointerTy())
1233 OS << "pointer"; // the full pointer type could be really verbose
1238 OS << ", Offsets={";
1239 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1240 E = Offsets.end(); I != E; ++I) {
1242 if (llvm::next(I) != E)
1247 if (AllFixupsOutsideLoop)
1248 OS << ", all-fixups-outside-loop";
1250 if (WidestFixupType)
1251 OS << ", widest fixup type: " << *WidestFixupType;
1254 void LSRUse::dump() const {
1255 print(errs()); errs() << '\n';
1258 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1259 /// be completely folded into the user instruction at isel time. This includes
1260 /// address-mode folding and special icmp tricks.
1261 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1262 LSRUse::KindType Kind, Type *AccessTy,
1263 const TargetLowering *TLI) {
1265 case LSRUse::Address:
1266 // If we have low-level target information, ask the target if it can
1267 // completely fold this address.
1268 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1270 // Otherwise, just guess that reg+reg addressing is legal.
1271 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1273 case LSRUse::ICmpZero:
1274 // There's not even a target hook for querying whether it would be legal to
1275 // fold a GV into an ICmp.
1279 // ICmp only has two operands; don't allow more than two non-trivial parts.
1280 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1283 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1284 // putting the scaled register in the other operand of the icmp.
1285 if (AM.Scale != 0 && AM.Scale != -1)
1288 // If we have low-level target information, ask the target if it can fold an
1289 // integer immediate on an icmp.
1290 if (AM.BaseOffs != 0) {
1294 // ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset
1295 // ICmpZero -1*ScaleReg + Offset => ICmp ScaleReg, Offset
1296 // Offs is the ICmp immediate.
1297 int64_t Offs = AM.BaseOffs;
1299 Offs = -(uint64_t)Offs; // The cast does the right thing with INT64_MIN.
1300 return TLI->isLegalICmpImmediate(Offs);
1303 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1307 // Only handle single-register values.
1308 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1310 case LSRUse::Special:
1311 // Only handle -1 scales, or no scale.
1312 return AM.Scale == 0 || AM.Scale == -1;
1315 llvm_unreachable("Invalid LSRUse Kind!");
1318 static bool isLegalUse(TargetLowering::AddrMode AM,
1319 int64_t MinOffset, int64_t MaxOffset,
1320 LSRUse::KindType Kind, Type *AccessTy,
1321 const TargetLowering *TLI) {
1322 // Check for overflow.
1323 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1326 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1327 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1328 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1329 // Check for overflow.
1330 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1333 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1334 return isLegalUse(AM, Kind, AccessTy, TLI);
1339 static bool isAlwaysFoldable(int64_t BaseOffs,
1340 GlobalValue *BaseGV,
1342 LSRUse::KindType Kind, Type *AccessTy,
1343 const TargetLowering *TLI) {
1344 // Fast-path: zero is always foldable.
1345 if (BaseOffs == 0 && !BaseGV) return true;
1347 // Conservatively, create an address with an immediate and a
1348 // base and a scale.
1349 TargetLowering::AddrMode AM;
1350 AM.BaseOffs = BaseOffs;
1352 AM.HasBaseReg = HasBaseReg;
1353 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1355 // Canonicalize a scale of 1 to a base register if the formula doesn't
1356 // already have a base register.
1357 if (!AM.HasBaseReg && AM.Scale == 1) {
1359 AM.HasBaseReg = true;
1362 return isLegalUse(AM, Kind, AccessTy, TLI);
1365 static bool isAlwaysFoldable(const SCEV *S,
1366 int64_t MinOffset, int64_t MaxOffset,
1368 LSRUse::KindType Kind, Type *AccessTy,
1369 const TargetLowering *TLI,
1370 ScalarEvolution &SE) {
1371 // Fast-path: zero is always foldable.
1372 if (S->isZero()) return true;
1374 // Conservatively, create an address with an immediate and a
1375 // base and a scale.
1376 int64_t BaseOffs = ExtractImmediate(S, SE);
1377 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1379 // If there's anything else involved, it's not foldable.
1380 if (!S->isZero()) return false;
1382 // Fast-path: zero is always foldable.
1383 if (BaseOffs == 0 && !BaseGV) return true;
1385 // Conservatively, create an address with an immediate and a
1386 // base and a scale.
1387 TargetLowering::AddrMode AM;
1388 AM.BaseOffs = BaseOffs;
1390 AM.HasBaseReg = HasBaseReg;
1391 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1393 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1398 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1399 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1400 struct UseMapDenseMapInfo {
1401 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1402 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1405 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1406 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1410 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1411 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1412 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1416 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1417 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1422 /// IVInc - An individual increment in a Chain of IV increments.
1423 /// Relate an IV user to an expression that computes the IV it uses from the IV
1424 /// used by the previous link in the Chain.
1426 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1427 /// original IVOperand. The head of the chain's IVOperand is only valid during
1428 /// chain collection, before LSR replaces IV users. During chain generation,
1429 /// IncExpr can be used to find the new IVOperand that computes the same
1432 Instruction *UserInst;
1434 const SCEV *IncExpr;
1436 IVInc(Instruction *U, Value *O, const SCEV *E):
1437 UserInst(U), IVOperand(O), IncExpr(E) {}
1440 // IVChain - The list of IV increments in program order.
1441 // We typically add the head of a chain without finding subsequent links.
1442 typedef SmallVector<IVInc,1> IVChain;
1444 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1445 /// Distinguish between FarUsers that definitely cross IV increments and
1446 /// NearUsers that may be used between IV increments.
1448 SmallPtrSet<Instruction*, 4> FarUsers;
1449 SmallPtrSet<Instruction*, 4> NearUsers;
1452 /// LSRInstance - This class holds state for the main loop strength reduction
1456 ScalarEvolution &SE;
1459 const TargetLowering *const TLI;
1463 /// IVIncInsertPos - This is the insert position that the current loop's
1464 /// induction variable increment should be placed. In simple loops, this is
1465 /// the latch block's terminator. But in more complicated cases, this is a
1466 /// position which will dominate all the in-loop post-increment users.
1467 Instruction *IVIncInsertPos;
1469 /// Factors - Interesting factors between use strides.
1470 SmallSetVector<int64_t, 8> Factors;
1472 /// Types - Interesting use types, to facilitate truncation reuse.
1473 SmallSetVector<Type *, 4> Types;
1475 /// Fixups - The list of operands which are to be replaced.
1476 SmallVector<LSRFixup, 16> Fixups;
1478 /// Uses - The list of interesting uses.
1479 SmallVector<LSRUse, 16> Uses;
1481 /// RegUses - Track which uses use which register candidates.
1482 RegUseTracker RegUses;
1484 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1485 // have more than a few IV increment chains in a loop. Missing a Chain falls
1486 // back to normal LSR behavior for those uses.
1487 static const unsigned MaxChains = 8;
1489 /// IVChainVec - IV users can form a chain of IV increments.
1490 SmallVector<IVChain, MaxChains> IVChainVec;
1492 /// IVIncSet - IV users that belong to profitable IVChains.
1493 SmallPtrSet<Use*, MaxChains> IVIncSet;
1495 void OptimizeShadowIV();
1496 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1497 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1498 void OptimizeLoopTermCond();
1500 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1501 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1502 void FinalizeChain(IVChain &Chain);
1503 void CollectChains();
1504 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1505 SmallVectorImpl<WeakVH> &DeadInsts);
1507 void CollectInterestingTypesAndFactors();
1508 void CollectFixupsAndInitialFormulae();
1510 LSRFixup &getNewFixup() {
1511 Fixups.push_back(LSRFixup());
1512 return Fixups.back();
1515 // Support for sharing of LSRUses between LSRFixups.
1516 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1518 UseMapDenseMapInfo> UseMapTy;
1521 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1522 LSRUse::KindType Kind, Type *AccessTy);
1524 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1525 LSRUse::KindType Kind,
1528 void DeleteUse(LSRUse &LU, size_t LUIdx);
1530 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1532 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1533 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1534 void CountRegisters(const Formula &F, size_t LUIdx);
1535 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1537 void CollectLoopInvariantFixupsAndFormulae();
1539 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1540 unsigned Depth = 0);
1541 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1542 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1543 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1544 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1545 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1546 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1547 void GenerateCrossUseConstantOffsets();
1548 void GenerateAllReuseFormulae();
1550 void FilterOutUndesirableDedicatedRegisters();
1552 size_t EstimateSearchSpaceComplexity() const;
1553 void NarrowSearchSpaceByDetectingSupersets();
1554 void NarrowSearchSpaceByCollapsingUnrolledCode();
1555 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1556 void NarrowSearchSpaceByPickingWinnerRegs();
1557 void NarrowSearchSpaceUsingHeuristics();
1559 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1561 SmallVectorImpl<const Formula *> &Workspace,
1562 const Cost &CurCost,
1563 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1564 DenseSet<const SCEV *> &VisitedRegs) const;
1565 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1567 BasicBlock::iterator
1568 HoistInsertPosition(BasicBlock::iterator IP,
1569 const SmallVectorImpl<Instruction *> &Inputs) const;
1570 BasicBlock::iterator
1571 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1574 SCEVExpander &Rewriter) const;
1576 Value *Expand(const LSRFixup &LF,
1578 BasicBlock::iterator IP,
1579 SCEVExpander &Rewriter,
1580 SmallVectorImpl<WeakVH> &DeadInsts) const;
1581 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1583 SCEVExpander &Rewriter,
1584 SmallVectorImpl<WeakVH> &DeadInsts,
1586 void Rewrite(const LSRFixup &LF,
1588 SCEVExpander &Rewriter,
1589 SmallVectorImpl<WeakVH> &DeadInsts,
1591 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1595 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1597 bool getChanged() const { return Changed; }
1599 void print_factors_and_types(raw_ostream &OS) const;
1600 void print_fixups(raw_ostream &OS) const;
1601 void print_uses(raw_ostream &OS) const;
1602 void print(raw_ostream &OS) const;
1608 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1609 /// inside the loop then try to eliminate the cast operation.
1610 void LSRInstance::OptimizeShadowIV() {
1611 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1612 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1615 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1616 UI != E; /* empty */) {
1617 IVUsers::const_iterator CandidateUI = UI;
1619 Instruction *ShadowUse = CandidateUI->getUser();
1620 Type *DestTy = NULL;
1621 bool IsSigned = false;
1623 /* If shadow use is a int->float cast then insert a second IV
1624 to eliminate this cast.
1626 for (unsigned i = 0; i < n; ++i)
1632 for (unsigned i = 0; i < n; ++i, ++d)
1635 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1637 DestTy = UCast->getDestTy();
1639 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1641 DestTy = SCast->getDestTy();
1643 if (!DestTy) continue;
1646 // If target does not support DestTy natively then do not apply
1647 // this transformation.
1648 EVT DVT = TLI->getValueType(DestTy);
1649 if (!TLI->isTypeLegal(DVT)) continue;
1652 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1654 if (PH->getNumIncomingValues() != 2) continue;
1656 Type *SrcTy = PH->getType();
1657 int Mantissa = DestTy->getFPMantissaWidth();
1658 if (Mantissa == -1) continue;
1659 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1662 unsigned Entry, Latch;
1663 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1671 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1672 if (!Init) continue;
1673 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1674 (double)Init->getSExtValue() :
1675 (double)Init->getZExtValue());
1677 BinaryOperator *Incr =
1678 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1679 if (!Incr) continue;
1680 if (Incr->getOpcode() != Instruction::Add
1681 && Incr->getOpcode() != Instruction::Sub)
1684 /* Initialize new IV, double d = 0.0 in above example. */
1685 ConstantInt *C = NULL;
1686 if (Incr->getOperand(0) == PH)
1687 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1688 else if (Incr->getOperand(1) == PH)
1689 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1695 // Ignore negative constants, as the code below doesn't handle them
1696 // correctly. TODO: Remove this restriction.
1697 if (!C->getValue().isStrictlyPositive()) continue;
1699 /* Add new PHINode. */
1700 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1702 /* create new increment. '++d' in above example. */
1703 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1704 BinaryOperator *NewIncr =
1705 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1706 Instruction::FAdd : Instruction::FSub,
1707 NewPH, CFP, "IV.S.next.", Incr);
1709 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1710 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1712 /* Remove cast operation */
1713 ShadowUse->replaceAllUsesWith(NewPH);
1714 ShadowUse->eraseFromParent();
1720 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1721 /// set the IV user and stride information and return true, otherwise return
1723 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1724 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1725 if (UI->getUser() == Cond) {
1726 // NOTE: we could handle setcc instructions with multiple uses here, but
1727 // InstCombine does it as well for simple uses, it's not clear that it
1728 // occurs enough in real life to handle.
1735 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1736 /// a max computation.
1738 /// This is a narrow solution to a specific, but acute, problem. For loops
1744 /// } while (++i < n);
1746 /// the trip count isn't just 'n', because 'n' might not be positive. And
1747 /// unfortunately this can come up even for loops where the user didn't use
1748 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1749 /// will commonly be lowered like this:
1755 /// } while (++i < n);
1758 /// and then it's possible for subsequent optimization to obscure the if
1759 /// test in such a way that indvars can't find it.
1761 /// When indvars can't find the if test in loops like this, it creates a
1762 /// max expression, which allows it to give the loop a canonical
1763 /// induction variable:
1766 /// max = n < 1 ? 1 : n;
1769 /// } while (++i != max);
1771 /// Canonical induction variables are necessary because the loop passes
1772 /// are designed around them. The most obvious example of this is the
1773 /// LoopInfo analysis, which doesn't remember trip count values. It
1774 /// expects to be able to rediscover the trip count each time it is
1775 /// needed, and it does this using a simple analysis that only succeeds if
1776 /// the loop has a canonical induction variable.
1778 /// However, when it comes time to generate code, the maximum operation
1779 /// can be quite costly, especially if it's inside of an outer loop.
1781 /// This function solves this problem by detecting this type of loop and
1782 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1783 /// the instructions for the maximum computation.
1785 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1786 // Check that the loop matches the pattern we're looking for.
1787 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1788 Cond->getPredicate() != CmpInst::ICMP_NE)
1791 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1792 if (!Sel || !Sel->hasOneUse()) return Cond;
1794 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1795 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1797 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1799 // Add one to the backedge-taken count to get the trip count.
1800 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1801 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1803 // Check for a max calculation that matches the pattern. There's no check
1804 // for ICMP_ULE here because the comparison would be with zero, which
1805 // isn't interesting.
1806 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1807 const SCEVNAryExpr *Max = 0;
1808 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1809 Pred = ICmpInst::ICMP_SLE;
1811 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1812 Pred = ICmpInst::ICMP_SLT;
1814 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1815 Pred = ICmpInst::ICMP_ULT;
1822 // To handle a max with more than two operands, this optimization would
1823 // require additional checking and setup.
1824 if (Max->getNumOperands() != 2)
1827 const SCEV *MaxLHS = Max->getOperand(0);
1828 const SCEV *MaxRHS = Max->getOperand(1);
1830 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1831 // for a comparison with 1. For <= and >=, a comparison with zero.
1833 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1836 // Check the relevant induction variable for conformance to
1838 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1839 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1840 if (!AR || !AR->isAffine() ||
1841 AR->getStart() != One ||
1842 AR->getStepRecurrence(SE) != One)
1845 assert(AR->getLoop() == L &&
1846 "Loop condition operand is an addrec in a different loop!");
1848 // Check the right operand of the select, and remember it, as it will
1849 // be used in the new comparison instruction.
1851 if (ICmpInst::isTrueWhenEqual(Pred)) {
1852 // Look for n+1, and grab n.
1853 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1854 if (isa<ConstantInt>(BO->getOperand(1)) &&
1855 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1856 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1857 NewRHS = BO->getOperand(0);
1858 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1859 if (isa<ConstantInt>(BO->getOperand(1)) &&
1860 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1861 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1862 NewRHS = BO->getOperand(0);
1865 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1866 NewRHS = Sel->getOperand(1);
1867 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1868 NewRHS = Sel->getOperand(2);
1869 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1870 NewRHS = SU->getValue();
1872 // Max doesn't match expected pattern.
1875 // Determine the new comparison opcode. It may be signed or unsigned,
1876 // and the original comparison may be either equality or inequality.
1877 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1878 Pred = CmpInst::getInversePredicate(Pred);
1880 // Ok, everything looks ok to change the condition into an SLT or SGE and
1881 // delete the max calculation.
1883 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1885 // Delete the max calculation instructions.
1886 Cond->replaceAllUsesWith(NewCond);
1887 CondUse->setUser(NewCond);
1888 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1889 Cond->eraseFromParent();
1890 Sel->eraseFromParent();
1891 if (Cmp->use_empty())
1892 Cmp->eraseFromParent();
1896 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1897 /// postinc iv when possible.
1899 LSRInstance::OptimizeLoopTermCond() {
1900 SmallPtrSet<Instruction *, 4> PostIncs;
1902 BasicBlock *LatchBlock = L->getLoopLatch();
1903 SmallVector<BasicBlock*, 8> ExitingBlocks;
1904 L->getExitingBlocks(ExitingBlocks);
1906 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1907 BasicBlock *ExitingBlock = ExitingBlocks[i];
1909 // Get the terminating condition for the loop if possible. If we
1910 // can, we want to change it to use a post-incremented version of its
1911 // induction variable, to allow coalescing the live ranges for the IV into
1912 // one register value.
1914 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1917 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1918 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1921 // Search IVUsesByStride to find Cond's IVUse if there is one.
1922 IVStrideUse *CondUse = 0;
1923 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1924 if (!FindIVUserForCond(Cond, CondUse))
1927 // If the trip count is computed in terms of a max (due to ScalarEvolution
1928 // being unable to find a sufficient guard, for example), change the loop
1929 // comparison to use SLT or ULT instead of NE.
1930 // One consequence of doing this now is that it disrupts the count-down
1931 // optimization. That's not always a bad thing though, because in such
1932 // cases it may still be worthwhile to avoid a max.
1933 Cond = OptimizeMax(Cond, CondUse);
1935 // If this exiting block dominates the latch block, it may also use
1936 // the post-inc value if it won't be shared with other uses.
1937 // Check for dominance.
1938 if (!DT.dominates(ExitingBlock, LatchBlock))
1941 // Conservatively avoid trying to use the post-inc value in non-latch
1942 // exits if there may be pre-inc users in intervening blocks.
1943 if (LatchBlock != ExitingBlock)
1944 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1945 // Test if the use is reachable from the exiting block. This dominator
1946 // query is a conservative approximation of reachability.
1947 if (&*UI != CondUse &&
1948 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1949 // Conservatively assume there may be reuse if the quotient of their
1950 // strides could be a legal scale.
1951 const SCEV *A = IU.getStride(*CondUse, L);
1952 const SCEV *B = IU.getStride(*UI, L);
1953 if (!A || !B) continue;
1954 if (SE.getTypeSizeInBits(A->getType()) !=
1955 SE.getTypeSizeInBits(B->getType())) {
1956 if (SE.getTypeSizeInBits(A->getType()) >
1957 SE.getTypeSizeInBits(B->getType()))
1958 B = SE.getSignExtendExpr(B, A->getType());
1960 A = SE.getSignExtendExpr(A, B->getType());
1962 if (const SCEVConstant *D =
1963 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1964 const ConstantInt *C = D->getValue();
1965 // Stride of one or negative one can have reuse with non-addresses.
1966 if (C->isOne() || C->isAllOnesValue())
1967 goto decline_post_inc;
1968 // Avoid weird situations.
1969 if (C->getValue().getMinSignedBits() >= 64 ||
1970 C->getValue().isMinSignedValue())
1971 goto decline_post_inc;
1972 // Without TLI, assume that any stride might be valid, and so any
1973 // use might be shared.
1975 goto decline_post_inc;
1976 // Check for possible scaled-address reuse.
1977 Type *AccessTy = getAccessType(UI->getUser());
1978 TargetLowering::AddrMode AM;
1979 AM.Scale = C->getSExtValue();
1980 if (TLI->isLegalAddressingMode(AM, AccessTy))
1981 goto decline_post_inc;
1982 AM.Scale = -AM.Scale;
1983 if (TLI->isLegalAddressingMode(AM, AccessTy))
1984 goto decline_post_inc;
1988 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1991 // It's possible for the setcc instruction to be anywhere in the loop, and
1992 // possible for it to have multiple users. If it is not immediately before
1993 // the exiting block branch, move it.
1994 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1995 if (Cond->hasOneUse()) {
1996 Cond->moveBefore(TermBr);
1998 // Clone the terminating condition and insert into the loopend.
1999 ICmpInst *OldCond = Cond;
2000 Cond = cast<ICmpInst>(Cond->clone());
2001 Cond->setName(L->getHeader()->getName() + ".termcond");
2002 ExitingBlock->getInstList().insert(TermBr, Cond);
2004 // Clone the IVUse, as the old use still exists!
2005 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2006 TermBr->replaceUsesOfWith(OldCond, Cond);
2010 // If we get to here, we know that we can transform the setcc instruction to
2011 // use the post-incremented version of the IV, allowing us to coalesce the
2012 // live ranges for the IV correctly.
2013 CondUse->transformToPostInc(L);
2016 PostIncs.insert(Cond);
2020 // Determine an insertion point for the loop induction variable increment. It
2021 // must dominate all the post-inc comparisons we just set up, and it must
2022 // dominate the loop latch edge.
2023 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2024 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2025 E = PostIncs.end(); I != E; ++I) {
2027 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2029 if (BB == (*I)->getParent())
2030 IVIncInsertPos = *I;
2031 else if (BB != IVIncInsertPos->getParent())
2032 IVIncInsertPos = BB->getTerminator();
2036 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2037 /// at the given offset and other details. If so, update the use and
2040 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2041 LSRUse::KindType Kind, Type *AccessTy) {
2042 int64_t NewMinOffset = LU.MinOffset;
2043 int64_t NewMaxOffset = LU.MaxOffset;
2044 Type *NewAccessTy = AccessTy;
2046 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2047 // something conservative, however this can pessimize in the case that one of
2048 // the uses will have all its uses outside the loop, for example.
2049 if (LU.Kind != Kind)
2051 // Conservatively assume HasBaseReg is true for now.
2052 if (NewOffset < LU.MinOffset) {
2053 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
2054 Kind, AccessTy, TLI))
2056 NewMinOffset = NewOffset;
2057 } else if (NewOffset > LU.MaxOffset) {
2058 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
2059 Kind, AccessTy, TLI))
2061 NewMaxOffset = NewOffset;
2063 // Check for a mismatched access type, and fall back conservatively as needed.
2064 // TODO: Be less conservative when the type is similar and can use the same
2065 // addressing modes.
2066 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2067 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2070 LU.MinOffset = NewMinOffset;
2071 LU.MaxOffset = NewMaxOffset;
2072 LU.AccessTy = NewAccessTy;
2073 if (NewOffset != LU.Offsets.back())
2074 LU.Offsets.push_back(NewOffset);
2078 /// getUse - Return an LSRUse index and an offset value for a fixup which
2079 /// needs the given expression, with the given kind and optional access type.
2080 /// Either reuse an existing use or create a new one, as needed.
2081 std::pair<size_t, int64_t>
2082 LSRInstance::getUse(const SCEV *&Expr,
2083 LSRUse::KindType Kind, Type *AccessTy) {
2084 const SCEV *Copy = Expr;
2085 int64_t Offset = ExtractImmediate(Expr, SE);
2087 // Basic uses can't accept any offset, for example.
2088 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
2093 std::pair<UseMapTy::iterator, bool> P =
2094 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2096 // A use already existed with this base.
2097 size_t LUIdx = P.first->second;
2098 LSRUse &LU = Uses[LUIdx];
2099 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2101 return std::make_pair(LUIdx, Offset);
2104 // Create a new use.
2105 size_t LUIdx = Uses.size();
2106 P.first->second = LUIdx;
2107 Uses.push_back(LSRUse(Kind, AccessTy));
2108 LSRUse &LU = Uses[LUIdx];
2110 // We don't need to track redundant offsets, but we don't need to go out
2111 // of our way here to avoid them.
2112 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2113 LU.Offsets.push_back(Offset);
2115 LU.MinOffset = Offset;
2116 LU.MaxOffset = Offset;
2117 return std::make_pair(LUIdx, Offset);
2120 /// DeleteUse - Delete the given use from the Uses list.
2121 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2122 if (&LU != &Uses.back())
2123 std::swap(LU, Uses.back());
2127 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2130 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2131 /// a formula that has the same registers as the given formula.
2133 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2134 const LSRUse &OrigLU) {
2135 // Search all uses for the formula. This could be more clever.
2136 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2137 LSRUse &LU = Uses[LUIdx];
2138 // Check whether this use is close enough to OrigLU, to see whether it's
2139 // worthwhile looking through its formulae.
2140 // Ignore ICmpZero uses because they may contain formulae generated by
2141 // GenerateICmpZeroScales, in which case adding fixup offsets may
2143 if (&LU != &OrigLU &&
2144 LU.Kind != LSRUse::ICmpZero &&
2145 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2146 LU.WidestFixupType == OrigLU.WidestFixupType &&
2147 LU.HasFormulaWithSameRegs(OrigF)) {
2148 // Scan through this use's formulae.
2149 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2150 E = LU.Formulae.end(); I != E; ++I) {
2151 const Formula &F = *I;
2152 // Check to see if this formula has the same registers and symbols
2154 if (F.BaseRegs == OrigF.BaseRegs &&
2155 F.ScaledReg == OrigF.ScaledReg &&
2156 F.AM.BaseGV == OrigF.AM.BaseGV &&
2157 F.AM.Scale == OrigF.AM.Scale &&
2158 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2159 if (F.AM.BaseOffs == 0)
2161 // This is the formula where all the registers and symbols matched;
2162 // there aren't going to be any others. Since we declined it, we
2163 // can skip the rest of the formulae and procede to the next LSRUse.
2170 // Nothing looked good.
2174 void LSRInstance::CollectInterestingTypesAndFactors() {
2175 SmallSetVector<const SCEV *, 4> Strides;
2177 // Collect interesting types and strides.
2178 SmallVector<const SCEV *, 4> Worklist;
2179 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2180 const SCEV *Expr = IU.getExpr(*UI);
2182 // Collect interesting types.
2183 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2185 // Add strides for mentioned loops.
2186 Worklist.push_back(Expr);
2188 const SCEV *S = Worklist.pop_back_val();
2189 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2190 if (AR->getLoop() == L)
2191 Strides.insert(AR->getStepRecurrence(SE));
2192 Worklist.push_back(AR->getStart());
2193 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2194 Worklist.append(Add->op_begin(), Add->op_end());
2196 } while (!Worklist.empty());
2199 // Compute interesting factors from the set of interesting strides.
2200 for (SmallSetVector<const SCEV *, 4>::const_iterator
2201 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2202 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2203 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2204 const SCEV *OldStride = *I;
2205 const SCEV *NewStride = *NewStrideIter;
2207 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2208 SE.getTypeSizeInBits(NewStride->getType())) {
2209 if (SE.getTypeSizeInBits(OldStride->getType()) >
2210 SE.getTypeSizeInBits(NewStride->getType()))
2211 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2213 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2215 if (const SCEVConstant *Factor =
2216 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2218 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2219 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2220 } else if (const SCEVConstant *Factor =
2221 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2224 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2225 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2229 // If all uses use the same type, don't bother looking for truncation-based
2231 if (Types.size() == 1)
2234 DEBUG(print_factors_and_types(dbgs()));
2237 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2238 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2239 /// Instructions to IVStrideUses, we could partially skip this.
2240 static User::op_iterator
2241 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2242 Loop *L, ScalarEvolution &SE) {
2243 for(; OI != OE; ++OI) {
2244 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2245 if (!SE.isSCEVable(Oper->getType()))
2248 if (const SCEVAddRecExpr *AR =
2249 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2250 if (AR->getLoop() == L)
2258 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2259 /// operands, so wrap it in a convenient helper.
2260 static Value *getWideOperand(Value *Oper) {
2261 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2262 return Trunc->getOperand(0);
2266 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2268 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2269 Type *LType = LVal->getType();
2270 Type *RType = RVal->getType();
2271 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2274 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2275 /// NULL for any constant. Returning the expression itself is
2276 /// conservative. Returning a deeper subexpression is more precise and valid as
2277 /// long as it isn't less complex than another subexpression. For expressions
2278 /// involving multiple unscaled values, we need to return the pointer-type
2279 /// SCEVUnknown. This avoids forming chains across objects, such as:
2280 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2282 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2283 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2284 static const SCEV *getExprBase(const SCEV *S) {
2285 switch (S->getSCEVType()) {
2286 default: // uncluding scUnknown.
2291 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2293 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2295 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2297 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2298 // there's nothing more complex.
2299 // FIXME: not sure if we want to recognize negation.
2300 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2301 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2302 E(Add->op_begin()); I != E; ++I) {
2303 const SCEV *SubExpr = *I;
2304 if (SubExpr->getSCEVType() == scAddExpr)
2305 return getExprBase(SubExpr);
2307 if (SubExpr->getSCEVType() != scMulExpr)
2310 return S; // all operands are scaled, be conservative.
2313 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2317 /// Return true if the chain increment is profitable to expand into a loop
2318 /// invariant value, which may require its own register. A profitable chain
2319 /// increment will be an offset relative to the same base. We allow such offsets
2320 /// to potentially be used as chain increment as long as it's not obviously
2321 /// expensive to expand using real instructions.
2323 getProfitableChainIncrement(Value *NextIV, Value *PrevIV,
2324 const IVChain &Chain, Loop *L,
2325 ScalarEvolution &SE, const TargetLowering *TLI) {
2326 // Prune the solution space aggressively by checking that both IV operands
2327 // are expressions that operate on the same unscaled SCEVUnknown. This
2328 // "base" will be canceled by the subsequent getMinusSCEV call. Checking first
2329 // avoids creating extra SCEV expressions.
2330 const SCEV *OperExpr = SE.getSCEV(NextIV);
2331 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2332 if (getExprBase(OperExpr) != getExprBase(PrevExpr) && !StressIVChain)
2335 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2336 if (!SE.isLoopInvariant(IncExpr, L))
2339 // We are not able to expand an increment unless it is loop invariant,
2340 // however, the following checks are purely for profitability.
2344 // Do not replace a constant offset from IV head with a nonconstant IV
2346 if (!isa<SCEVConstant>(IncExpr)) {
2347 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Chain[0].IVOperand));
2348 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2352 SmallPtrSet<const SCEV*, 8> Processed;
2353 if (isHighCostExpansion(IncExpr, Processed, SE))
2359 /// Return true if the number of registers needed for the chain is estimated to
2360 /// be less than the number required for the individual IV users. First prohibit
2361 /// any IV users that keep the IV live across increments (the Users set should
2362 /// be empty). Next count the number and type of increments in the chain.
2364 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2365 /// effectively use postinc addressing modes. Only consider it profitable it the
2366 /// increments can be computed in fewer registers when chained.
2368 /// TODO: Consider IVInc free if it's already used in another chains.
2370 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2371 ScalarEvolution &SE, const TargetLowering *TLI) {
2375 if (Chain.size() <= 2)
2378 if (!Users.empty()) {
2379 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " users:\n";
2380 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2381 E = Users.end(); I != E; ++I) {
2382 dbgs() << " " << **I << "\n";
2386 assert(!Chain.empty() && "empty IV chains are not allowed");
2388 // The chain itself may require a register, so intialize cost to 1.
2391 // A complete chain likely eliminates the need for keeping the original IV in
2392 // a register. LSR does not currently know how to form a complete chain unless
2393 // the header phi already exists.
2394 if (isa<PHINode>(Chain.back().UserInst)
2395 && SE.getSCEV(Chain.back().UserInst) == Chain[0].IncExpr) {
2398 const SCEV *LastIncExpr = 0;
2399 unsigned NumConstIncrements = 0;
2400 unsigned NumVarIncrements = 0;
2401 unsigned NumReusedIncrements = 0;
2402 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2405 if (I->IncExpr->isZero())
2408 // Incrementing by zero or some constant is neutral. We assume constants can
2409 // be folded into an addressing mode or an add's immediate operand.
2410 if (isa<SCEVConstant>(I->IncExpr)) {
2411 ++NumConstIncrements;
2415 if (I->IncExpr == LastIncExpr)
2416 ++NumReusedIncrements;
2420 LastIncExpr = I->IncExpr;
2422 // An IV chain with a single increment is handled by LSR's postinc
2423 // uses. However, a chain with multiple increments requires keeping the IV's
2424 // value live longer than it needs to be if chained.
2425 if (NumConstIncrements > 1)
2428 // Materializing increment expressions in the preheader that didn't exist in
2429 // the original code may cost a register. For example, sign-extended array
2430 // indices can produce ridiculous increments like this:
2431 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2432 cost += NumVarIncrements;
2434 // Reusing variable increments likely saves a register to hold the multiple of
2436 cost -= NumReusedIncrements;
2438 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " Cost: " << cost << "\n");
2443 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2445 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2446 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2447 // When IVs are used as types of varying widths, they are generally converted
2448 // to a wider type with some uses remaining narrow under a (free) trunc.
2449 Value *NextIV = getWideOperand(IVOper);
2451 // Visit all existing chains. Check if its IVOper can be computed as a
2452 // profitable loop invariant increment from the last link in the Chain.
2453 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2454 const SCEV *LastIncExpr = 0;
2455 for (; ChainIdx < NChains; ++ChainIdx) {
2456 Value *PrevIV = getWideOperand(IVChainVec[ChainIdx].back().IVOperand);
2457 if (!isCompatibleIVType(PrevIV, NextIV))
2460 // A phi node terminates a chain.
2461 if (isa<PHINode>(UserInst)
2462 && isa<PHINode>(IVChainVec[ChainIdx].back().UserInst))
2465 if (const SCEV *IncExpr =
2466 getProfitableChainIncrement(NextIV, PrevIV, IVChainVec[ChainIdx],
2468 LastIncExpr = IncExpr;
2472 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2473 // bother for phi nodes, because they must be last in the chain.
2474 if (ChainIdx == NChains) {
2475 if (isa<PHINode>(UserInst))
2477 if (NChains >= MaxChains && !StressIVChain) {
2478 DEBUG(dbgs() << "IV Chain Limit\n");
2481 LastIncExpr = SE.getSCEV(NextIV);
2482 // IVUsers may have skipped over sign/zero extensions. We don't currently
2483 // attempt to form chains involving extensions unless they can be hoisted
2484 // into this loop's AddRec.
2485 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2488 IVChainVec.resize(NChains);
2489 ChainUsersVec.resize(NChains);
2490 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2491 << ") IV=" << *LastIncExpr << "\n");
2494 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2495 << ") IV+" << *LastIncExpr << "\n");
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 (!OtherUse || OtherUse == UserInst)
2518 if (SE.isSCEVable(OtherUse->getType())
2519 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2520 && IU.isIVUserOrOperand(OtherUse)) {
2523 NearUsers.insert(OtherUse);
2526 // Since this user is part of the chain, it's no longer considered a use
2528 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2531 /// CollectChains - Populate the vector of Chains.
2533 /// This decreases ILP at the architecture level. Targets with ample registers,
2534 /// multiple memory ports, and no register renaming probably don't want
2535 /// this. However, such targets should probably disable LSR altogether.
2537 /// The job of LSR is to make a reasonable choice of induction variables across
2538 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2539 /// ILP *within the loop* if the target wants it.
2541 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2542 /// will not reorder memory operations, it will recognize this as a chain, but
2543 /// will generate redundant IV increments. Ideally this would be corrected later
2544 /// by a smart scheduler:
2550 /// TODO: Walk the entire domtree within this loop, not just the path to the
2551 /// loop latch. This will discover chains on side paths, but requires
2552 /// maintaining multiple copies of the Chains state.
2553 void LSRInstance::CollectChains() {
2554 DEBUG(dbgs() << "Collecting IV Chains.\n");
2555 SmallVector<ChainUsers, 8> ChainUsersVec;
2557 SmallVector<BasicBlock *,8> LatchPath;
2558 BasicBlock *LoopHeader = L->getHeader();
2559 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2560 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2561 LatchPath.push_back(Rung->getBlock());
2563 LatchPath.push_back(LoopHeader);
2565 // Walk the instruction stream from the loop header to the loop latch.
2566 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2567 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2568 BBIter != BBEnd; ++BBIter) {
2569 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2571 // Skip instructions that weren't seen by IVUsers analysis.
2572 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2575 // Ignore users that are part of a SCEV expression. This way we only
2576 // consider leaf IV Users. This effectively rediscovers a portion of
2577 // IVUsers analysis but in program order this time.
2578 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2581 // Remove this instruction from any NearUsers set it may be in.
2582 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2583 ChainIdx < NChains; ++ChainIdx) {
2584 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2586 // Search for operands that can be chained.
2587 SmallPtrSet<Instruction*, 4> UniqueOperands;
2588 User::op_iterator IVOpEnd = I->op_end();
2589 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2590 while (IVOpIter != IVOpEnd) {
2591 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2592 if (UniqueOperands.insert(IVOpInst))
2593 ChainInstruction(I, IVOpInst, ChainUsersVec);
2594 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2596 } // Continue walking down the instructions.
2597 } // Continue walking down the domtree.
2598 // Visit phi backedges to determine if the chain can generate the IV postinc.
2599 for (BasicBlock::iterator I = L->getHeader()->begin();
2600 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2601 if (!SE.isSCEVable(PN->getType()))
2605 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2607 ChainInstruction(PN, IncV, ChainUsersVec);
2609 // Remove any unprofitable chains.
2610 unsigned ChainIdx = 0;
2611 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2612 UsersIdx < NChains; ++UsersIdx) {
2613 if (!isProfitableChain(IVChainVec[UsersIdx],
2614 ChainUsersVec[UsersIdx].FarUsers, SE, TLI))
2616 // Preserve the chain at UsesIdx.
2617 if (ChainIdx != UsersIdx)
2618 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2619 FinalizeChain(IVChainVec[ChainIdx]);
2622 IVChainVec.resize(ChainIdx);
2625 void LSRInstance::FinalizeChain(IVChain &Chain) {
2626 assert(!Chain.empty() && "empty IV chains are not allowed");
2627 DEBUG(dbgs() << "Final Chain: " << *Chain[0].UserInst << "\n");
2629 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2631 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2632 User::op_iterator UseI =
2633 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2634 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2635 IVIncSet.insert(UseI);
2639 /// Return true if the IVInc can be folded into an addressing mode.
2640 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2641 Value *Operand, const TargetLowering *TLI) {
2642 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2643 if (!IncConst || !isAddressUse(UserInst, Operand))
2646 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2649 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2650 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false,
2651 LSRUse::Address, getAccessType(UserInst), TLI))
2657 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2658 /// materialize the IV user's operand from the previous IV user's operand.
2659 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2660 SmallVectorImpl<WeakVH> &DeadInsts) {
2661 // Find the new IVOperand for the head of the chain. It may have been replaced
2663 const IVInc &Head = Chain[0];
2664 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2665 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2668 while (IVOpIter != IVOpEnd) {
2669 IVSrc = getWideOperand(*IVOpIter);
2671 // If this operand computes the expression that the chain needs, we may use
2672 // it. (Check this after setting IVSrc which is used below.)
2674 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2675 // narrow for the chain, so we can no longer use it. We do allow using a
2676 // wider phi, assuming the LSR checked for free truncation. In that case we
2677 // should already have a truncate on this operand such that
2678 // getSCEV(IVSrc) == IncExpr.
2679 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2680 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2683 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2685 if (IVOpIter == IVOpEnd) {
2686 // Gracefully give up on this chain.
2687 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2691 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2692 Type *IVTy = IVSrc->getType();
2693 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2694 const SCEV *LeftOverExpr = 0;
2695 for (IVChain::const_iterator IncI = llvm::next(Chain.begin()),
2696 IncE = Chain.end(); IncI != IncE; ++IncI) {
2698 Instruction *InsertPt = IncI->UserInst;
2699 if (isa<PHINode>(InsertPt))
2700 InsertPt = L->getLoopLatch()->getTerminator();
2702 // IVOper will replace the current IV User's operand. IVSrc is the IV
2703 // value currently held in a register.
2704 Value *IVOper = IVSrc;
2705 if (!IncI->IncExpr->isZero()) {
2706 // IncExpr was the result of subtraction of two narrow values, so must
2708 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2709 LeftOverExpr = LeftOverExpr ?
2710 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2712 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2713 // Expand the IV increment.
2714 Rewriter.clearPostInc();
2715 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2716 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2717 SE.getUnknown(IncV));
2718 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2720 // If an IV increment can't be folded, use it as the next IV value.
2721 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2723 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2728 Type *OperTy = IncI->IVOperand->getType();
2729 if (IVTy != OperTy) {
2730 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2731 "cannot extend a chained IV");
2732 IRBuilder<> Builder(InsertPt);
2733 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2735 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2736 DeadInsts.push_back(IncI->IVOperand);
2738 // If LSR created a new, wider phi, we may also replace its postinc. We only
2739 // do this if we also found a wide value for the head of the chain.
2740 if (isa<PHINode>(Chain.back().UserInst)) {
2741 for (BasicBlock::iterator I = L->getHeader()->begin();
2742 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2743 if (!isCompatibleIVType(Phi, IVSrc))
2745 Instruction *PostIncV = dyn_cast<Instruction>(
2746 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2747 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2749 Value *IVOper = IVSrc;
2750 Type *PostIncTy = PostIncV->getType();
2751 if (IVTy != PostIncTy) {
2752 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2753 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2754 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2755 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2757 Phi->replaceUsesOfWith(PostIncV, IVOper);
2758 DeadInsts.push_back(PostIncV);
2763 void LSRInstance::CollectFixupsAndInitialFormulae() {
2764 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2765 Instruction *UserInst = UI->getUser();
2766 // Skip IV users that are part of profitable IV Chains.
2767 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2768 UI->getOperandValToReplace());
2769 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2770 if (IVIncSet.count(UseI))
2774 LSRFixup &LF = getNewFixup();
2775 LF.UserInst = UserInst;
2776 LF.OperandValToReplace = UI->getOperandValToReplace();
2777 LF.PostIncLoops = UI->getPostIncLoops();
2779 LSRUse::KindType Kind = LSRUse::Basic;
2781 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2782 Kind = LSRUse::Address;
2783 AccessTy = getAccessType(LF.UserInst);
2786 const SCEV *S = IU.getExpr(*UI);
2788 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2789 // (N - i == 0), and this allows (N - i) to be the expression that we work
2790 // with rather than just N or i, so we can consider the register
2791 // requirements for both N and i at the same time. Limiting this code to
2792 // equality icmps is not a problem because all interesting loops use
2793 // equality icmps, thanks to IndVarSimplify.
2794 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2795 if (CI->isEquality()) {
2796 // Swap the operands if needed to put the OperandValToReplace on the
2797 // left, for consistency.
2798 Value *NV = CI->getOperand(1);
2799 if (NV == LF.OperandValToReplace) {
2800 CI->setOperand(1, CI->getOperand(0));
2801 CI->setOperand(0, NV);
2802 NV = CI->getOperand(1);
2806 // x == y --> x - y == 0
2807 const SCEV *N = SE.getSCEV(NV);
2808 if (SE.isLoopInvariant(N, L)) {
2809 // S is normalized, so normalize N before folding it into S
2810 // to keep the result normalized.
2811 N = TransformForPostIncUse(Normalize, N, CI, 0,
2812 LF.PostIncLoops, SE, DT);
2813 Kind = LSRUse::ICmpZero;
2814 S = SE.getMinusSCEV(N, S);
2817 // -1 and the negations of all interesting strides (except the negation
2818 // of -1) are now also interesting.
2819 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2820 if (Factors[i] != -1)
2821 Factors.insert(-(uint64_t)Factors[i]);
2825 // Set up the initial formula for this use.
2826 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2828 LF.Offset = P.second;
2829 LSRUse &LU = Uses[LF.LUIdx];
2830 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2831 if (!LU.WidestFixupType ||
2832 SE.getTypeSizeInBits(LU.WidestFixupType) <
2833 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2834 LU.WidestFixupType = LF.OperandValToReplace->getType();
2836 // If this is the first use of this LSRUse, give it a formula.
2837 if (LU.Formulae.empty()) {
2838 InsertInitialFormula(S, LU, LF.LUIdx);
2839 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2843 DEBUG(print_fixups(dbgs()));
2846 /// InsertInitialFormula - Insert a formula for the given expression into
2847 /// the given use, separating out loop-variant portions from loop-invariant
2848 /// and loop-computable portions.
2850 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2852 F.InitialMatch(S, L, SE);
2853 bool Inserted = InsertFormula(LU, LUIdx, F);
2854 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2857 /// InsertSupplementalFormula - Insert a simple single-register formula for
2858 /// the given expression into the given use.
2860 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2861 LSRUse &LU, size_t LUIdx) {
2863 F.BaseRegs.push_back(S);
2864 F.AM.HasBaseReg = true;
2865 bool Inserted = InsertFormula(LU, LUIdx, F);
2866 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2869 /// CountRegisters - Note which registers are used by the given formula,
2870 /// updating RegUses.
2871 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2873 RegUses.CountRegister(F.ScaledReg, LUIdx);
2874 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2875 E = F.BaseRegs.end(); I != E; ++I)
2876 RegUses.CountRegister(*I, LUIdx);
2879 /// InsertFormula - If the given formula has not yet been inserted, add it to
2880 /// the list, and return true. Return false otherwise.
2881 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2882 if (!LU.InsertFormula(F))
2885 CountRegisters(F, LUIdx);
2889 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2890 /// loop-invariant values which we're tracking. These other uses will pin these
2891 /// values in registers, making them less profitable for elimination.
2892 /// TODO: This currently misses non-constant addrec step registers.
2893 /// TODO: Should this give more weight to users inside the loop?
2895 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2896 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2897 SmallPtrSet<const SCEV *, 8> Inserted;
2899 while (!Worklist.empty()) {
2900 const SCEV *S = Worklist.pop_back_val();
2902 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2903 Worklist.append(N->op_begin(), N->op_end());
2904 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2905 Worklist.push_back(C->getOperand());
2906 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2907 Worklist.push_back(D->getLHS());
2908 Worklist.push_back(D->getRHS());
2909 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2910 if (!Inserted.insert(U)) continue;
2911 const Value *V = U->getValue();
2912 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2913 // Look for instructions defined outside the loop.
2914 if (L->contains(Inst)) continue;
2915 } else if (isa<UndefValue>(V))
2916 // Undef doesn't have a live range, so it doesn't matter.
2918 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2920 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2921 // Ignore non-instructions.
2924 // Ignore instructions in other functions (as can happen with
2926 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2928 // Ignore instructions not dominated by the loop.
2929 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2930 UserInst->getParent() :
2931 cast<PHINode>(UserInst)->getIncomingBlock(
2932 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2933 if (!DT.dominates(L->getHeader(), UseBB))
2935 // Ignore uses which are part of other SCEV expressions, to avoid
2936 // analyzing them multiple times.
2937 if (SE.isSCEVable(UserInst->getType())) {
2938 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2939 // If the user is a no-op, look through to its uses.
2940 if (!isa<SCEVUnknown>(UserS))
2944 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2948 // Ignore icmp instructions which are already being analyzed.
2949 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2950 unsigned OtherIdx = !UI.getOperandNo();
2951 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2952 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2956 LSRFixup &LF = getNewFixup();
2957 LF.UserInst = const_cast<Instruction *>(UserInst);
2958 LF.OperandValToReplace = UI.getUse();
2959 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2961 LF.Offset = P.second;
2962 LSRUse &LU = Uses[LF.LUIdx];
2963 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2964 if (!LU.WidestFixupType ||
2965 SE.getTypeSizeInBits(LU.WidestFixupType) <
2966 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2967 LU.WidestFixupType = LF.OperandValToReplace->getType();
2968 InsertSupplementalFormula(U, LU, LF.LUIdx);
2969 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2976 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2977 /// separate registers. If C is non-null, multiply each subexpression by C.
2978 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2979 SmallVectorImpl<const SCEV *> &Ops,
2981 ScalarEvolution &SE) {
2982 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2983 // Break out add operands.
2984 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2986 CollectSubexprs(*I, C, Ops, L, SE);
2988 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2989 // Split a non-zero base out of an addrec.
2990 if (!AR->getStart()->isZero()) {
2991 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2992 AR->getStepRecurrence(SE),
2994 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2997 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
3000 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3001 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3002 if (Mul->getNumOperands() == 2)
3003 if (const SCEVConstant *Op0 =
3004 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3005 CollectSubexprs(Mul->getOperand(1),
3006 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
3012 // Otherwise use the value itself, optionally with a scale applied.
3013 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
3016 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3018 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3021 // Arbitrarily cap recursion to protect compile time.
3022 if (Depth >= 3) return;
3024 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3025 const SCEV *BaseReg = Base.BaseRegs[i];
3027 SmallVector<const SCEV *, 8> AddOps;
3028 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3030 if (AddOps.size() == 1) continue;
3032 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3033 JE = AddOps.end(); J != JE; ++J) {
3035 // Loop-variant "unknown" values are uninteresting; we won't be able to
3036 // do anything meaningful with them.
3037 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3040 // Don't pull a constant into a register if the constant could be folded
3041 // into an immediate field.
3042 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
3043 Base.getNumRegs() > 1,
3044 LU.Kind, LU.AccessTy, TLI, SE))
3047 // Collect all operands except *J.
3048 SmallVector<const SCEV *, 8> InnerAddOps
3049 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3051 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3053 // Don't leave just a constant behind in a register if the constant could
3054 // be folded into an immediate field.
3055 if (InnerAddOps.size() == 1 &&
3056 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
3057 Base.getNumRegs() > 1,
3058 LU.Kind, LU.AccessTy, TLI, SE))
3061 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3062 if (InnerSum->isZero())
3066 // Add the remaining pieces of the add back into the new formula.
3067 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3068 if (TLI && InnerSumSC &&
3069 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3070 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3071 InnerSumSC->getValue()->getZExtValue())) {
3072 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3073 InnerSumSC->getValue()->getZExtValue();
3074 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3076 F.BaseRegs[i] = InnerSum;
3078 // Add J as its own register, or an unfolded immediate.
3079 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3080 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3081 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3082 SC->getValue()->getZExtValue()))
3083 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3084 SC->getValue()->getZExtValue();
3086 F.BaseRegs.push_back(*J);
3088 if (InsertFormula(LU, LUIdx, F))
3089 // If that formula hadn't been seen before, recurse to find more like
3091 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3096 /// GenerateCombinations - Generate a formula consisting of all of the
3097 /// loop-dominating registers added into a single register.
3098 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3100 // This method is only interesting on a plurality of registers.
3101 if (Base.BaseRegs.size() <= 1) return;
3105 SmallVector<const SCEV *, 4> Ops;
3106 for (SmallVectorImpl<const SCEV *>::const_iterator
3107 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3108 const SCEV *BaseReg = *I;
3109 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3110 !SE.hasComputableLoopEvolution(BaseReg, L))
3111 Ops.push_back(BaseReg);
3113 F.BaseRegs.push_back(BaseReg);
3115 if (Ops.size() > 1) {
3116 const SCEV *Sum = SE.getAddExpr(Ops);
3117 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3118 // opportunity to fold something. For now, just ignore such cases
3119 // rather than proceed with zero in a register.
3120 if (!Sum->isZero()) {
3121 F.BaseRegs.push_back(Sum);
3122 (void)InsertFormula(LU, LUIdx, F);
3127 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3128 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3130 // We can't add a symbolic offset if the address already contains one.
3131 if (Base.AM.BaseGV) return;
3133 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3134 const SCEV *G = Base.BaseRegs[i];
3135 GlobalValue *GV = ExtractSymbol(G, SE);
3136 if (G->isZero() || !GV)
3140 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3141 LU.Kind, LU.AccessTy, TLI))
3144 (void)InsertFormula(LU, LUIdx, F);
3148 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3149 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3151 // TODO: For now, just add the min and max offset, because it usually isn't
3152 // worthwhile looking at everything inbetween.
3153 SmallVector<int64_t, 2> Worklist;
3154 Worklist.push_back(LU.MinOffset);
3155 if (LU.MaxOffset != LU.MinOffset)
3156 Worklist.push_back(LU.MaxOffset);
3158 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3159 const SCEV *G = Base.BaseRegs[i];
3161 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3162 E = Worklist.end(); I != E; ++I) {
3164 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
3165 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
3166 LU.Kind, LU.AccessTy, TLI)) {
3167 // Add the offset to the base register.
3168 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3169 // If it cancelled out, drop the base register, otherwise update it.
3170 if (NewG->isZero()) {
3171 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3172 F.BaseRegs.pop_back();
3174 F.BaseRegs[i] = NewG;
3176 (void)InsertFormula(LU, LUIdx, F);
3180 int64_t Imm = ExtractImmediate(G, SE);
3181 if (G->isZero() || Imm == 0)
3184 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
3185 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3186 LU.Kind, LU.AccessTy, TLI))
3189 (void)InsertFormula(LU, LUIdx, F);
3193 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3194 /// the comparison. For example, x == y -> x*c == y*c.
3195 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3197 if (LU.Kind != LSRUse::ICmpZero) return;
3199 // Determine the integer type for the base formula.
3200 Type *IntTy = Base.getType();
3202 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3204 // Don't do this if there is more than one offset.
3205 if (LU.MinOffset != LU.MaxOffset) return;
3207 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
3209 // Check each interesting stride.
3210 for (SmallSetVector<int64_t, 8>::const_iterator
3211 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3212 int64_t Factor = *I;
3214 // Check that the multiplication doesn't overflow.
3215 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
3217 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
3218 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
3221 // Check that multiplying with the use offset doesn't overflow.
3222 int64_t Offset = LU.MinOffset;
3223 if (Offset == INT64_MIN && Factor == -1)
3225 Offset = (uint64_t)Offset * Factor;
3226 if (Offset / Factor != LU.MinOffset)
3230 F.AM.BaseOffs = NewBaseOffs;
3232 // Check that this scale is legal.
3233 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
3236 // Compensate for the use having MinOffset built into it.
3237 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
3239 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3241 // Check that multiplying with each base register doesn't overflow.
3242 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3243 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3244 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3248 // Check that multiplying with the scaled register doesn't overflow.
3250 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3251 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3255 // Check that multiplying with the unfolded offset doesn't overflow.
3256 if (F.UnfoldedOffset != 0) {
3257 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3259 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3260 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3264 // If we make it here and it's legal, add it.
3265 (void)InsertFormula(LU, LUIdx, F);
3270 /// GenerateScales - Generate stride factor reuse formulae by making use of
3271 /// scaled-offset address modes, for example.
3272 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3273 // Determine the integer type for the base formula.
3274 Type *IntTy = Base.getType();
3277 // If this Formula already has a scaled register, we can't add another one.
3278 if (Base.AM.Scale != 0) return;
3280 // Check each interesting stride.
3281 for (SmallSetVector<int64_t, 8>::const_iterator
3282 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3283 int64_t Factor = *I;
3285 Base.AM.Scale = Factor;
3286 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
3287 // Check whether this scale is going to be legal.
3288 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3289 LU.Kind, LU.AccessTy, TLI)) {
3290 // As a special-case, handle special out-of-loop Basic users specially.
3291 // TODO: Reconsider this special case.
3292 if (LU.Kind == LSRUse::Basic &&
3293 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3294 LSRUse::Special, LU.AccessTy, TLI) &&
3295 LU.AllFixupsOutsideLoop)
3296 LU.Kind = LSRUse::Special;
3300 // For an ICmpZero, negating a solitary base register won't lead to
3302 if (LU.Kind == LSRUse::ICmpZero &&
3303 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
3305 // For each addrec base reg, apply the scale, if possible.
3306 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3307 if (const SCEVAddRecExpr *AR =
3308 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3309 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3310 if (FactorS->isZero())
3312 // Divide out the factor, ignoring high bits, since we'll be
3313 // scaling the value back up in the end.
3314 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3315 // TODO: This could be optimized to avoid all the copying.
3317 F.ScaledReg = Quotient;
3318 F.DeleteBaseReg(F.BaseRegs[i]);
3319 (void)InsertFormula(LU, LUIdx, F);
3325 /// GenerateTruncates - Generate reuse formulae from different IV types.
3326 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3327 // This requires TargetLowering to tell us which truncates are free.
3330 // Don't bother truncating symbolic values.
3331 if (Base.AM.BaseGV) return;
3333 // Determine the integer type for the base formula.
3334 Type *DstTy = Base.getType();
3336 DstTy = SE.getEffectiveSCEVType(DstTy);
3338 for (SmallSetVector<Type *, 4>::const_iterator
3339 I = Types.begin(), E = Types.end(); I != E; ++I) {
3341 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
3344 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3345 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3346 JE = F.BaseRegs.end(); J != JE; ++J)
3347 *J = SE.getAnyExtendExpr(*J, SrcTy);
3349 // TODO: This assumes we've done basic processing on all uses and
3350 // have an idea what the register usage is.
3351 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3354 (void)InsertFormula(LU, LUIdx, F);
3361 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3362 /// defer modifications so that the search phase doesn't have to worry about
3363 /// the data structures moving underneath it.
3367 const SCEV *OrigReg;
3369 WorkItem(size_t LI, int64_t I, const SCEV *R)
3370 : LUIdx(LI), Imm(I), OrigReg(R) {}
3372 void print(raw_ostream &OS) const;
3378 void WorkItem::print(raw_ostream &OS) const {
3379 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3380 << " , add offset " << Imm;
3383 void WorkItem::dump() const {
3384 print(errs()); errs() << '\n';
3387 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3388 /// distance apart and try to form reuse opportunities between them.
3389 void LSRInstance::GenerateCrossUseConstantOffsets() {
3390 // Group the registers by their value without any added constant offset.
3391 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3392 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3394 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3395 SmallVector<const SCEV *, 8> Sequence;
3396 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3398 const SCEV *Reg = *I;
3399 int64_t Imm = ExtractImmediate(Reg, SE);
3400 std::pair<RegMapTy::iterator, bool> Pair =
3401 Map.insert(std::make_pair(Reg, ImmMapTy()));
3403 Sequence.push_back(Reg);
3404 Pair.first->second.insert(std::make_pair(Imm, *I));
3405 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3408 // Now examine each set of registers with the same base value. Build up
3409 // a list of work to do and do the work in a separate step so that we're
3410 // not adding formulae and register counts while we're searching.
3411 SmallVector<WorkItem, 32> WorkItems;
3412 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3413 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3414 E = Sequence.end(); I != E; ++I) {
3415 const SCEV *Reg = *I;
3416 const ImmMapTy &Imms = Map.find(Reg)->second;
3418 // It's not worthwhile looking for reuse if there's only one offset.
3419 if (Imms.size() == 1)
3422 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3423 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3425 dbgs() << ' ' << J->first;
3428 // Examine each offset.
3429 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3431 const SCEV *OrigReg = J->second;
3433 int64_t JImm = J->first;
3434 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3436 if (!isa<SCEVConstant>(OrigReg) &&
3437 UsedByIndicesMap[Reg].count() == 1) {
3438 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3442 // Conservatively examine offsets between this orig reg a few selected
3444 ImmMapTy::const_iterator OtherImms[] = {
3445 Imms.begin(), prior(Imms.end()),
3446 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3448 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3449 ImmMapTy::const_iterator M = OtherImms[i];
3450 if (M == J || M == JE) continue;
3452 // Compute the difference between the two.
3453 int64_t Imm = (uint64_t)JImm - M->first;
3454 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3455 LUIdx = UsedByIndices.find_next(LUIdx))
3456 // Make a memo of this use, offset, and register tuple.
3457 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3458 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3465 UsedByIndicesMap.clear();
3466 UniqueItems.clear();
3468 // Now iterate through the worklist and add new formulae.
3469 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3470 E = WorkItems.end(); I != E; ++I) {
3471 const WorkItem &WI = *I;
3472 size_t LUIdx = WI.LUIdx;
3473 LSRUse &LU = Uses[LUIdx];
3474 int64_t Imm = WI.Imm;
3475 const SCEV *OrigReg = WI.OrigReg;
3477 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3478 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3479 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3481 // TODO: Use a more targeted data structure.
3482 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3483 const Formula &F = LU.Formulae[L];
3484 // Use the immediate in the scaled register.
3485 if (F.ScaledReg == OrigReg) {
3486 int64_t Offs = (uint64_t)F.AM.BaseOffs +
3487 Imm * (uint64_t)F.AM.Scale;
3488 // Don't create 50 + reg(-50).
3489 if (F.referencesReg(SE.getSCEV(
3490 ConstantInt::get(IntTy, -(uint64_t)Offs))))
3493 NewF.AM.BaseOffs = Offs;
3494 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3495 LU.Kind, LU.AccessTy, TLI))
3497 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3499 // If the new scale is a constant in a register, and adding the constant
3500 // value to the immediate would produce a value closer to zero than the
3501 // immediate itself, then the formula isn't worthwhile.
3502 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3503 if (C->getValue()->isNegative() !=
3504 (NewF.AM.BaseOffs < 0) &&
3505 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
3506 .ule(abs64(NewF.AM.BaseOffs)))
3510 (void)InsertFormula(LU, LUIdx, NewF);
3512 // Use the immediate in a base register.
3513 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3514 const SCEV *BaseReg = F.BaseRegs[N];
3515 if (BaseReg != OrigReg)
3518 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
3519 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3520 LU.Kind, LU.AccessTy, TLI)) {
3522 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3525 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3527 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3529 // If the new formula has a constant in a register, and adding the
3530 // constant value to the immediate would produce a value closer to
3531 // zero than the immediate itself, then the formula isn't worthwhile.
3532 for (SmallVectorImpl<const SCEV *>::const_iterator
3533 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3535 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3536 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
3537 abs64(NewF.AM.BaseOffs)) &&
3538 (C->getValue()->getValue() +
3539 NewF.AM.BaseOffs).countTrailingZeros() >=
3540 CountTrailingZeros_64(NewF.AM.BaseOffs))
3544 (void)InsertFormula(LU, LUIdx, NewF);
3553 /// GenerateAllReuseFormulae - Generate formulae for each use.
3555 LSRInstance::GenerateAllReuseFormulae() {
3556 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3557 // queries are more precise.
3558 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3559 LSRUse &LU = Uses[LUIdx];
3560 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3561 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3562 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3563 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3565 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3566 LSRUse &LU = Uses[LUIdx];
3567 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3568 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3569 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3570 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3571 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3572 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3573 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3574 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3576 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3577 LSRUse &LU = Uses[LUIdx];
3578 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3579 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3582 GenerateCrossUseConstantOffsets();
3584 DEBUG(dbgs() << "\n"
3585 "After generating reuse formulae:\n";
3586 print_uses(dbgs()));
3589 /// If there are multiple formulae with the same set of registers used
3590 /// by other uses, pick the best one and delete the others.
3591 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3592 DenseSet<const SCEV *> VisitedRegs;
3593 SmallPtrSet<const SCEV *, 16> Regs;
3594 SmallPtrSet<const SCEV *, 16> LoserRegs;
3596 bool ChangedFormulae = false;
3599 // Collect the best formula for each unique set of shared registers. This
3600 // is reset for each use.
3601 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
3603 BestFormulaeTy BestFormulae;
3605 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3606 LSRUse &LU = Uses[LUIdx];
3607 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3610 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3611 FIdx != NumForms; ++FIdx) {
3612 Formula &F = LU.Formulae[FIdx];
3614 // Some formulas are instant losers. For example, they may depend on
3615 // nonexistent AddRecs from other loops. These need to be filtered
3616 // immediately, otherwise heuristics could choose them over others leading
3617 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3618 // avoids the need to recompute this information across formulae using the
3619 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3620 // the corresponding bad register from the Regs set.
3623 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3625 if (CostF.isLoser()) {
3626 // During initial formula generation, undesirable formulae are generated
3627 // by uses within other loops that have some non-trivial address mode or
3628 // use the postinc form of the IV. LSR needs to provide these formulae
3629 // as the basis of rediscovering the desired formula that uses an AddRec
3630 // corresponding to the existing phi. Once all formulae have been
3631 // generated, these initial losers may be pruned.
3632 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3636 SmallVector<const SCEV *, 2> Key;
3637 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3638 JE = F.BaseRegs.end(); J != JE; ++J) {
3639 const SCEV *Reg = *J;
3640 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3644 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3645 Key.push_back(F.ScaledReg);
3646 // Unstable sort by host order ok, because this is only used for
3648 std::sort(Key.begin(), Key.end());
3650 std::pair<BestFormulaeTy::const_iterator, bool> P =
3651 BestFormulae.insert(std::make_pair(Key, FIdx));
3655 Formula &Best = LU.Formulae[P.first->second];
3659 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3660 if (CostF < CostBest)
3662 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3664 " in favor of formula "; Best.print(dbgs());
3668 ChangedFormulae = true;
3670 LU.DeleteFormula(F);
3676 // Now that we've filtered out some formulae, recompute the Regs set.
3678 LU.RecomputeRegs(LUIdx, RegUses);
3680 // Reset this to prepare for the next use.
3681 BestFormulae.clear();
3684 DEBUG(if (ChangedFormulae) {
3686 "After filtering out undesirable candidates:\n";
3691 // This is a rough guess that seems to work fairly well.
3692 static const size_t ComplexityLimit = UINT16_MAX;
3694 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3695 /// solutions the solver might have to consider. It almost never considers
3696 /// this many solutions because it prune the search space, but the pruning
3697 /// isn't always sufficient.
3698 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3700 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3701 E = Uses.end(); I != E; ++I) {
3702 size_t FSize = I->Formulae.size();
3703 if (FSize >= ComplexityLimit) {
3704 Power = ComplexityLimit;
3708 if (Power >= ComplexityLimit)
3714 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3715 /// of the registers of another formula, it won't help reduce register
3716 /// pressure (though it may not necessarily hurt register pressure); remove
3717 /// it to simplify the system.
3718 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3719 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3720 DEBUG(dbgs() << "The search space is too complex.\n");
3722 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3723 "which use a superset of registers used by other "
3726 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3727 LSRUse &LU = Uses[LUIdx];
3729 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3730 Formula &F = LU.Formulae[i];
3731 // Look for a formula with a constant or GV in a register. If the use
3732 // also has a formula with that same value in an immediate field,
3733 // delete the one that uses a register.
3734 for (SmallVectorImpl<const SCEV *>::const_iterator
3735 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3736 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3738 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3739 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3740 (I - F.BaseRegs.begin()));
3741 if (LU.HasFormulaWithSameRegs(NewF)) {
3742 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3743 LU.DeleteFormula(F);
3749 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3750 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3753 NewF.AM.BaseGV = GV;
3754 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3755 (I - F.BaseRegs.begin()));
3756 if (LU.HasFormulaWithSameRegs(NewF)) {
3757 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3759 LU.DeleteFormula(F);
3770 LU.RecomputeRegs(LUIdx, RegUses);
3773 DEBUG(dbgs() << "After pre-selection:\n";
3774 print_uses(dbgs()));
3778 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3779 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3781 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3782 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3783 DEBUG(dbgs() << "The search space is too complex.\n");
3785 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3786 "separated by a constant offset will use the same "
3789 // This is especially useful for unrolled loops.
3791 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3792 LSRUse &LU = Uses[LUIdx];
3793 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3794 E = LU.Formulae.end(); I != E; ++I) {
3795 const Formula &F = *I;
3796 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3797 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3798 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3799 /*HasBaseReg=*/false,
3800 LU.Kind, LU.AccessTy)) {
3801 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3804 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3806 // Update the relocs to reference the new use.
3807 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3808 E = Fixups.end(); I != E; ++I) {
3809 LSRFixup &Fixup = *I;
3810 if (Fixup.LUIdx == LUIdx) {
3811 Fixup.LUIdx = LUThatHas - &Uses.front();
3812 Fixup.Offset += F.AM.BaseOffs;
3813 // Add the new offset to LUThatHas' offset list.
3814 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3815 LUThatHas->Offsets.push_back(Fixup.Offset);
3816 if (Fixup.Offset > LUThatHas->MaxOffset)
3817 LUThatHas->MaxOffset = Fixup.Offset;
3818 if (Fixup.Offset < LUThatHas->MinOffset)
3819 LUThatHas->MinOffset = Fixup.Offset;
3821 DEBUG(dbgs() << "New fixup has offset "
3822 << Fixup.Offset << '\n');
3824 if (Fixup.LUIdx == NumUses-1)
3825 Fixup.LUIdx = LUIdx;
3828 // Delete formulae from the new use which are no longer legal.
3830 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3831 Formula &F = LUThatHas->Formulae[i];
3832 if (!isLegalUse(F.AM,
3833 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3834 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3835 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3837 LUThatHas->DeleteFormula(F);
3844 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3846 // Delete the old use.
3847 DeleteUse(LU, LUIdx);
3857 DEBUG(dbgs() << "After pre-selection:\n";
3858 print_uses(dbgs()));
3862 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3863 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3864 /// we've done more filtering, as it may be able to find more formulae to
3866 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3867 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3868 DEBUG(dbgs() << "The search space is too complex.\n");
3870 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3871 "undesirable dedicated registers.\n");
3873 FilterOutUndesirableDedicatedRegisters();
3875 DEBUG(dbgs() << "After pre-selection:\n";
3876 print_uses(dbgs()));
3880 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3881 /// to be profitable, and then in any use which has any reference to that
3882 /// register, delete all formulae which do not reference that register.
3883 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3884 // With all other options exhausted, loop until the system is simple
3885 // enough to handle.
3886 SmallPtrSet<const SCEV *, 4> Taken;
3887 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3888 // Ok, we have too many of formulae on our hands to conveniently handle.
3889 // Use a rough heuristic to thin out the list.
3890 DEBUG(dbgs() << "The search space is too complex.\n");
3892 // Pick the register which is used by the most LSRUses, which is likely
3893 // to be a good reuse register candidate.
3894 const SCEV *Best = 0;
3895 unsigned BestNum = 0;
3896 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3898 const SCEV *Reg = *I;
3899 if (Taken.count(Reg))
3904 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3905 if (Count > BestNum) {
3912 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3913 << " will yield profitable reuse.\n");
3916 // In any use with formulae which references this register, delete formulae
3917 // which don't reference it.
3918 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3919 LSRUse &LU = Uses[LUIdx];
3920 if (!LU.Regs.count(Best)) continue;
3923 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3924 Formula &F = LU.Formulae[i];
3925 if (!F.referencesReg(Best)) {
3926 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3927 LU.DeleteFormula(F);
3931 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3937 LU.RecomputeRegs(LUIdx, RegUses);
3940 DEBUG(dbgs() << "After pre-selection:\n";
3941 print_uses(dbgs()));
3945 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3946 /// formulae to choose from, use some rough heuristics to prune down the number
3947 /// of formulae. This keeps the main solver from taking an extraordinary amount
3948 /// of time in some worst-case scenarios.
3949 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3950 NarrowSearchSpaceByDetectingSupersets();
3951 NarrowSearchSpaceByCollapsingUnrolledCode();
3952 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3953 NarrowSearchSpaceByPickingWinnerRegs();
3956 /// SolveRecurse - This is the recursive solver.
3957 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3959 SmallVectorImpl<const Formula *> &Workspace,
3960 const Cost &CurCost,
3961 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3962 DenseSet<const SCEV *> &VisitedRegs) const {
3965 // - use more aggressive filtering
3966 // - sort the formula so that the most profitable solutions are found first
3967 // - sort the uses too
3969 // - don't compute a cost, and then compare. compare while computing a cost
3971 // - track register sets with SmallBitVector
3973 const LSRUse &LU = Uses[Workspace.size()];
3975 // If this use references any register that's already a part of the
3976 // in-progress solution, consider it a requirement that a formula must
3977 // reference that register in order to be considered. This prunes out
3978 // unprofitable searching.
3979 SmallSetVector<const SCEV *, 4> ReqRegs;
3980 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3981 E = CurRegs.end(); I != E; ++I)
3982 if (LU.Regs.count(*I))
3985 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 bool SatisfiedReqReg = true;
3993 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3994 JE = ReqRegs.end(); J != JE; ++J) {
3995 const SCEV *Reg = *J;
3996 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3997 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3999 SatisfiedReqReg = false;
4003 if (!SatisfiedReqReg) {
4004 // If none of the formulae satisfied the required registers, then we could
4005 // clear ReqRegs and try again. Currently, we simply give up in this case.
4009 // Evaluate the cost of the current formula. If it's already worse than
4010 // the current best, prune the search at that point.
4013 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
4014 if (NewCost < SolutionCost) {
4015 Workspace.push_back(&F);
4016 if (Workspace.size() != Uses.size()) {
4017 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4018 NewRegs, VisitedRegs);
4019 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4020 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4022 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4023 dbgs() << ".\n Regs:";
4024 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4025 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4026 dbgs() << ' ' << **I;
4029 SolutionCost = NewCost;
4030 Solution = Workspace;
4032 Workspace.pop_back();
4037 /// Solve - Choose one formula from each use. Return the results in the given
4038 /// Solution vector.
4039 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4040 SmallVector<const Formula *, 8> Workspace;
4042 SolutionCost.Loose();
4044 SmallPtrSet<const SCEV *, 16> CurRegs;
4045 DenseSet<const SCEV *> VisitedRegs;
4046 Workspace.reserve(Uses.size());
4048 // SolveRecurse does all the work.
4049 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4050 CurRegs, VisitedRegs);
4051 if (Solution.empty()) {
4052 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4056 // Ok, we've now made all our decisions.
4057 DEBUG(dbgs() << "\n"
4058 "The chosen solution requires "; SolutionCost.print(dbgs());
4060 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4062 Uses[i].print(dbgs());
4065 Solution[i]->print(dbgs());
4069 assert(Solution.size() == Uses.size() && "Malformed solution!");
4072 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4073 /// the dominator tree far as we can go while still being dominated by the
4074 /// input positions. This helps canonicalize the insert position, which
4075 /// encourages sharing.
4076 BasicBlock::iterator
4077 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4078 const SmallVectorImpl<Instruction *> &Inputs)
4081 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4082 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4085 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4086 if (!Rung) return IP;
4087 Rung = Rung->getIDom();
4088 if (!Rung) return IP;
4089 IDom = Rung->getBlock();
4091 // Don't climb into a loop though.
4092 const Loop *IDomLoop = LI.getLoopFor(IDom);
4093 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4094 if (IDomDepth <= IPLoopDepth &&
4095 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4099 bool AllDominate = true;
4100 Instruction *BetterPos = 0;
4101 Instruction *Tentative = IDom->getTerminator();
4102 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4103 E = Inputs.end(); I != E; ++I) {
4104 Instruction *Inst = *I;
4105 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4106 AllDominate = false;
4109 // Attempt to find an insert position in the middle of the block,
4110 // instead of at the end, so that it can be used for other expansions.
4111 if (IDom == Inst->getParent() &&
4112 (!BetterPos || DT.dominates(BetterPos, Inst)))
4113 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4126 /// AdjustInsertPositionForExpand - Determine an input position which will be
4127 /// dominated by the operands and which will dominate the result.
4128 BasicBlock::iterator
4129 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4132 SCEVExpander &Rewriter) const {
4133 // Collect some instructions which must be dominated by the
4134 // expanding replacement. These must be dominated by any operands that
4135 // will be required in the expansion.
4136 SmallVector<Instruction *, 4> Inputs;
4137 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4138 Inputs.push_back(I);
4139 if (LU.Kind == LSRUse::ICmpZero)
4140 if (Instruction *I =
4141 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4142 Inputs.push_back(I);
4143 if (LF.PostIncLoops.count(L)) {
4144 if (LF.isUseFullyOutsideLoop(L))
4145 Inputs.push_back(L->getLoopLatch()->getTerminator());
4147 Inputs.push_back(IVIncInsertPos);
4149 // The expansion must also be dominated by the increment positions of any
4150 // loops it for which it is using post-inc mode.
4151 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4152 E = LF.PostIncLoops.end(); I != E; ++I) {
4153 const Loop *PIL = *I;
4154 if (PIL == L) continue;
4156 // Be dominated by the loop exit.
4157 SmallVector<BasicBlock *, 4> ExitingBlocks;
4158 PIL->getExitingBlocks(ExitingBlocks);
4159 if (!ExitingBlocks.empty()) {
4160 BasicBlock *BB = ExitingBlocks[0];
4161 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4162 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4163 Inputs.push_back(BB->getTerminator());
4167 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4168 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4169 "Insertion point must be a normal instruction");
4171 // Then, climb up the immediate dominator tree as far as we can go while
4172 // still being dominated by the input positions.
4173 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, 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;
4184 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4185 // IP consistent across expansions and allows the previously inserted
4186 // instructions to be reused by subsequent expansion.
4187 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4192 /// Expand - Emit instructions for the leading candidate expression for this
4193 /// LSRUse (this is called "expanding").
4194 Value *LSRInstance::Expand(const LSRFixup &LF,
4196 BasicBlock::iterator IP,
4197 SCEVExpander &Rewriter,
4198 SmallVectorImpl<WeakVH> &DeadInsts) const {
4199 const LSRUse &LU = Uses[LF.LUIdx];
4201 // Determine an input position which will be dominated by the operands and
4202 // which will dominate the result.
4203 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4205 // Inform the Rewriter if we have a post-increment use, so that it can
4206 // perform an advantageous expansion.
4207 Rewriter.setPostInc(LF.PostIncLoops);
4209 // This is the type that the user actually needs.
4210 Type *OpTy = LF.OperandValToReplace->getType();
4211 // This will be the type that we'll initially expand to.
4212 Type *Ty = F.getType();
4214 // No type known; just expand directly to the ultimate type.
4216 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4217 // Expand directly to the ultimate type if it's the right size.
4219 // This is the type to do integer arithmetic in.
4220 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4222 // Build up a list of operands to add together to form the full base.
4223 SmallVector<const SCEV *, 8> Ops;
4225 // Expand the BaseRegs portion.
4226 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4227 E = F.BaseRegs.end(); I != E; ++I) {
4228 const SCEV *Reg = *I;
4229 assert(!Reg->isZero() && "Zero allocated in a base register!");
4231 // If we're expanding for a post-inc user, make the post-inc adjustment.
4232 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4233 Reg = TransformForPostIncUse(Denormalize, Reg,
4234 LF.UserInst, LF.OperandValToReplace,
4237 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4240 // Flush the operand list to suppress SCEVExpander hoisting.
4242 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4244 Ops.push_back(SE.getUnknown(FullV));
4247 // Expand the ScaledReg portion.
4248 Value *ICmpScaledV = 0;
4249 if (F.AM.Scale != 0) {
4250 const SCEV *ScaledS = F.ScaledReg;
4252 // If we're expanding for a post-inc user, make the post-inc adjustment.
4253 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4254 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4255 LF.UserInst, LF.OperandValToReplace,
4258 if (LU.Kind == LSRUse::ICmpZero) {
4259 // An interesting way of "folding" with an icmp is to use a negated
4260 // scale, which we'll implement by inserting it into the other operand
4262 assert(F.AM.Scale == -1 &&
4263 "The only scale supported by ICmpZero uses is -1!");
4264 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4266 // Otherwise just expand the scaled register and an explicit scale,
4267 // which is expected to be matched as part of the address.
4268 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4269 ScaledS = SE.getMulExpr(ScaledS,
4270 SE.getConstant(ScaledS->getType(), F.AM.Scale));
4271 Ops.push_back(ScaledS);
4273 // Flush the operand list to suppress SCEVExpander hoisting.
4274 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4276 Ops.push_back(SE.getUnknown(FullV));
4280 // Expand the GV portion.
4282 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
4284 // Flush the operand list to suppress SCEVExpander hoisting.
4285 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4287 Ops.push_back(SE.getUnknown(FullV));
4290 // Expand the immediate portion.
4291 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
4293 if (LU.Kind == LSRUse::ICmpZero) {
4294 // The other interesting way of "folding" with an ICmpZero is to use a
4295 // negated immediate.
4297 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4299 Ops.push_back(SE.getUnknown(ICmpScaledV));
4300 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4303 // Just add the immediate values. These again are expected to be matched
4304 // as part of the address.
4305 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4309 // Expand the unfolded offset portion.
4310 int64_t UnfoldedOffset = F.UnfoldedOffset;
4311 if (UnfoldedOffset != 0) {
4312 // Just add the immediate values.
4313 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4317 // Emit instructions summing all the operands.
4318 const SCEV *FullS = Ops.empty() ?
4319 SE.getConstant(IntTy, 0) :
4321 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4323 // We're done expanding now, so reset the rewriter.
4324 Rewriter.clearPostInc();
4326 // An ICmpZero Formula represents an ICmp which we're handling as a
4327 // comparison against zero. Now that we've expanded an expression for that
4328 // form, update the ICmp's other operand.
4329 if (LU.Kind == LSRUse::ICmpZero) {
4330 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4331 DeadInsts.push_back(CI->getOperand(1));
4332 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
4333 "a scale at the same time!");
4334 if (F.AM.Scale == -1) {
4335 if (ICmpScaledV->getType() != OpTy) {
4337 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4339 ICmpScaledV, OpTy, "tmp", CI);
4342 CI->setOperand(1, ICmpScaledV);
4344 assert(F.AM.Scale == 0 &&
4345 "ICmp does not support folding a global value and "
4346 "a scale at the same time!");
4347 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4349 if (C->getType() != OpTy)
4350 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4354 CI->setOperand(1, C);
4361 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4362 /// of their operands effectively happens in their predecessor blocks, so the
4363 /// expression may need to be expanded in multiple places.
4364 void LSRInstance::RewriteForPHI(PHINode *PN,
4367 SCEVExpander &Rewriter,
4368 SmallVectorImpl<WeakVH> &DeadInsts,
4370 DenseMap<BasicBlock *, Value *> Inserted;
4371 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4372 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4373 BasicBlock *BB = PN->getIncomingBlock(i);
4375 // If this is a critical edge, split the edge so that we do not insert
4376 // the code on all predecessor/successor paths. We do this unless this
4377 // is the canonical backedge for this loop, which complicates post-inc
4379 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4380 !isa<IndirectBrInst>(BB->getTerminator())) {
4381 BasicBlock *Parent = PN->getParent();
4382 Loop *PNLoop = LI.getLoopFor(Parent);
4383 if (!PNLoop || Parent != PNLoop->getHeader()) {
4384 // Split the critical edge.
4385 BasicBlock *NewBB = 0;
4386 if (!Parent->isLandingPad()) {
4387 NewBB = SplitCriticalEdge(BB, Parent, P,
4388 /*MergeIdenticalEdges=*/true,
4389 /*DontDeleteUselessPhis=*/true);
4391 SmallVector<BasicBlock*, 2> NewBBs;
4392 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4396 // If PN is outside of the loop and BB is in the loop, we want to
4397 // move the block to be immediately before the PHI block, not
4398 // immediately after BB.
4399 if (L->contains(BB) && !L->contains(PN))
4400 NewBB->moveBefore(PN->getParent());
4402 // Splitting the edge can reduce the number of PHI entries we have.
4403 e = PN->getNumIncomingValues();
4405 i = PN->getBasicBlockIndex(BB);
4409 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4410 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4412 PN->setIncomingValue(i, Pair.first->second);
4414 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4416 // If this is reuse-by-noop-cast, insert the noop cast.
4417 Type *OpTy = LF.OperandValToReplace->getType();
4418 if (FullV->getType() != OpTy)
4420 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4422 FullV, LF.OperandValToReplace->getType(),
4423 "tmp", BB->getTerminator());
4425 PN->setIncomingValue(i, FullV);
4426 Pair.first->second = FullV;
4431 /// Rewrite - Emit instructions for the leading candidate expression for this
4432 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4433 /// the newly expanded value.
4434 void LSRInstance::Rewrite(const LSRFixup &LF,
4436 SCEVExpander &Rewriter,
4437 SmallVectorImpl<WeakVH> &DeadInsts,
4439 // First, find an insertion point that dominates UserInst. For PHI nodes,
4440 // find the nearest block which dominates all the relevant uses.
4441 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4442 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4444 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4446 // If this is reuse-by-noop-cast, insert the noop cast.
4447 Type *OpTy = LF.OperandValToReplace->getType();
4448 if (FullV->getType() != OpTy) {
4450 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4451 FullV, OpTy, "tmp", LF.UserInst);
4455 // Update the user. ICmpZero is handled specially here (for now) because
4456 // Expand may have updated one of the operands of the icmp already, and
4457 // its new value may happen to be equal to LF.OperandValToReplace, in
4458 // which case doing replaceUsesOfWith leads to replacing both operands
4459 // with the same value. TODO: Reorganize this.
4460 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4461 LF.UserInst->setOperand(0, FullV);
4463 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4466 DeadInsts.push_back(LF.OperandValToReplace);
4469 /// ImplementSolution - Rewrite all the fixup locations with new values,
4470 /// following the chosen solution.
4472 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4474 // Keep track of instructions we may have made dead, so that
4475 // we can remove them after we are done working.
4476 SmallVector<WeakVH, 16> DeadInsts;
4478 SCEVExpander Rewriter(SE, "lsr");
4480 Rewriter.setDebugType(DEBUG_TYPE);
4482 Rewriter.disableCanonicalMode();
4483 Rewriter.enableLSRMode();
4484 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4486 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4487 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4488 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4489 if (PHINode *PN = dyn_cast<PHINode>(ChainI->back().UserInst))
4490 Rewriter.setChainedPhi(PN);
4493 // Expand the new value definitions and update the users.
4494 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4495 E = Fixups.end(); I != E; ++I) {
4496 const LSRFixup &Fixup = *I;
4498 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4503 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4504 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4505 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4508 // Clean up after ourselves. This must be done before deleting any
4512 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4515 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
4516 : IU(P->getAnalysis<IVUsers>()),
4517 SE(P->getAnalysis<ScalarEvolution>()),
4518 DT(P->getAnalysis<DominatorTree>()),
4519 LI(P->getAnalysis<LoopInfo>()),
4520 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
4522 // If LoopSimplify form is not available, stay out of trouble.
4523 if (!L->isLoopSimplifyForm())
4526 // If there's no interesting work to be done, bail early.
4527 if (IU.empty()) return;
4529 // If there's too much analysis to be done, bail early. We won't be able to
4530 // model the problem anyway.
4531 unsigned NumUsers = 0;
4532 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4533 if (++NumUsers > MaxIVUsers) {
4534 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4541 // All dominating loops must have preheaders, or SCEVExpander may not be able
4542 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4544 // IVUsers analysis should only create users that are dominated by simple loop
4545 // headers. Since this loop should dominate all of its users, its user list
4546 // should be empty if this loop itself is not within a simple loop nest.
4547 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4548 Rung; Rung = Rung->getIDom()) {
4549 BasicBlock *BB = Rung->getBlock();
4550 const Loop *DomLoop = LI.getLoopFor(BB);
4551 if (DomLoop && DomLoop->getHeader() == BB) {
4552 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4557 DEBUG(dbgs() << "\nLSR on loop ";
4558 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4561 // First, perform some low-level loop optimizations.
4563 OptimizeLoopTermCond();
4565 // If loop preparation eliminates all interesting IV users, bail.
4566 if (IU.empty()) return;
4568 // Skip nested loops until we can model them better with formulae.
4570 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4574 // Start collecting data and preparing for the solver.
4576 CollectInterestingTypesAndFactors();
4577 CollectFixupsAndInitialFormulae();
4578 CollectLoopInvariantFixupsAndFormulae();
4580 assert(!Uses.empty() && "IVUsers reported at least one use");
4581 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4582 print_uses(dbgs()));
4584 // Now use the reuse data to generate a bunch of interesting ways
4585 // to formulate the values needed for the uses.
4586 GenerateAllReuseFormulae();
4588 FilterOutUndesirableDedicatedRegisters();
4589 NarrowSearchSpaceUsingHeuristics();
4591 SmallVector<const Formula *, 8> Solution;
4594 // Release memory that is no longer needed.
4599 if (Solution.empty())
4603 // Formulae should be legal.
4604 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4605 E = Uses.end(); I != E; ++I) {
4606 const LSRUse &LU = *I;
4607 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4608 JE = LU.Formulae.end(); J != JE; ++J)
4609 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
4610 LU.Kind, LU.AccessTy, TLI) &&
4611 "Illegal formula generated!");
4615 // Now that we've decided what we want, make it so.
4616 ImplementSolution(Solution, P);
4619 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4620 if (Factors.empty() && Types.empty()) return;
4622 OS << "LSR has identified the following interesting factors and types: ";
4625 for (SmallSetVector<int64_t, 8>::const_iterator
4626 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4627 if (!First) OS << ", ";
4632 for (SmallSetVector<Type *, 4>::const_iterator
4633 I = Types.begin(), E = Types.end(); I != E; ++I) {
4634 if (!First) OS << ", ";
4636 OS << '(' << **I << ')';
4641 void LSRInstance::print_fixups(raw_ostream &OS) const {
4642 OS << "LSR is examining the following fixup sites:\n";
4643 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4644 E = Fixups.end(); I != E; ++I) {
4651 void LSRInstance::print_uses(raw_ostream &OS) const {
4652 OS << "LSR is examining the following uses:\n";
4653 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4654 E = Uses.end(); I != E; ++I) {
4655 const LSRUse &LU = *I;
4659 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4660 JE = LU.Formulae.end(); J != JE; ++J) {
4668 void LSRInstance::print(raw_ostream &OS) const {
4669 print_factors_and_types(OS);
4674 void LSRInstance::dump() const {
4675 print(errs()); errs() << '\n';
4680 class LoopStrengthReduce : public LoopPass {
4681 /// TLI - Keep a pointer of a TargetLowering to consult for determining
4682 /// transformation profitability.
4683 const TargetLowering *const TLI;
4686 static char ID; // Pass ID, replacement for typeid
4687 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
4690 bool runOnLoop(Loop *L, LPPassManager &LPM);
4691 void getAnalysisUsage(AnalysisUsage &AU) const;
4696 char LoopStrengthReduce::ID = 0;
4697 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4698 "Loop Strength Reduction", false, false)
4699 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4700 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4701 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4702 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4703 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4704 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4705 "Loop Strength Reduction", false, false)
4708 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
4709 return new LoopStrengthReduce(TLI);
4712 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
4713 : LoopPass(ID), TLI(tli) {
4714 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4717 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4718 // We split critical edges, so we change the CFG. However, we do update
4719 // many analyses if they are around.
4720 AU.addPreservedID(LoopSimplifyID);
4722 AU.addRequired<LoopInfo>();
4723 AU.addPreserved<LoopInfo>();
4724 AU.addRequiredID(LoopSimplifyID);
4725 AU.addRequired<DominatorTree>();
4726 AU.addPreserved<DominatorTree>();
4727 AU.addRequired<ScalarEvolution>();
4728 AU.addPreserved<ScalarEvolution>();
4729 // Requiring LoopSimplify a second time here prevents IVUsers from running
4730 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4731 AU.addRequiredID(LoopSimplifyID);
4732 AU.addRequired<IVUsers>();
4733 AU.addPreserved<IVUsers>();
4736 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4737 bool Changed = false;
4739 // Run the main LSR transformation.
4740 Changed |= LSRInstance(TLI, L, this).getChanged();
4742 // Remove any extra phis created by processing inner loops.
4743 Changed |= DeleteDeadPHIs(L->getHeader());
4744 if (EnablePhiElim) {
4745 SmallVector<WeakVH, 16> DeadInsts;
4746 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4748 Rewriter.setDebugType(DEBUG_TYPE);
4750 unsigned numFolded = Rewriter.
4751 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI);
4754 DeleteTriviallyDeadInstructions(DeadInsts);
4755 DeleteDeadPHIs(L->getHeader());