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 // Temporary flag to cleanup congruent phis after LSR phi expansion.
81 // It's currently disabled until we can determine whether it's truly useful or
82 // not. The flag should be removed after the v3.0 release.
83 // This is now needed for ivchains.
84 static cl::opt<bool> EnablePhiElim(
85 "enable-lsr-phielim", cl::Hidden, cl::init(true),
86 cl::desc("Enable LSR phi elimination"));
89 // Stress test IV chain generation.
90 static cl::opt<bool> StressIVChain(
91 "stress-ivchain", cl::Hidden, cl::init(false),
92 cl::desc("Stress test LSR IV chains"));
94 static bool StressIVChain = false;
99 /// RegSortData - This class holds data which is used to order reuse candidates.
102 /// UsedByIndices - This represents the set of LSRUse indices which reference
103 /// a particular register.
104 SmallBitVector UsedByIndices;
108 void print(raw_ostream &OS) const;
114 void RegSortData::print(raw_ostream &OS) const {
115 OS << "[NumUses=" << UsedByIndices.count() << ']';
118 void RegSortData::dump() const {
119 print(errs()); errs() << '\n';
124 /// RegUseTracker - Map register candidates to information about how they are
126 class RegUseTracker {
127 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
129 RegUsesTy RegUsesMap;
130 SmallVector<const SCEV *, 16> RegSequence;
133 void CountRegister(const SCEV *Reg, size_t LUIdx);
134 void DropRegister(const SCEV *Reg, size_t LUIdx);
135 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
137 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
139 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
143 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
144 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
145 iterator begin() { return RegSequence.begin(); }
146 iterator end() { return RegSequence.end(); }
147 const_iterator begin() const { return RegSequence.begin(); }
148 const_iterator end() const { return RegSequence.end(); }
154 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
155 std::pair<RegUsesTy::iterator, bool> Pair =
156 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
157 RegSortData &RSD = Pair.first->second;
159 RegSequence.push_back(Reg);
160 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
161 RSD.UsedByIndices.set(LUIdx);
165 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
166 RegUsesTy::iterator It = RegUsesMap.find(Reg);
167 assert(It != RegUsesMap.end());
168 RegSortData &RSD = It->second;
169 assert(RSD.UsedByIndices.size() > LUIdx);
170 RSD.UsedByIndices.reset(LUIdx);
174 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
175 assert(LUIdx <= LastLUIdx);
177 // Update RegUses. The data structure is not optimized for this purpose;
178 // we must iterate through it and update each of the bit vectors.
179 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
181 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
182 if (LUIdx < UsedByIndices.size())
183 UsedByIndices[LUIdx] =
184 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
185 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
190 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
191 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
192 if (I == RegUsesMap.end())
194 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
195 int i = UsedByIndices.find_first();
196 if (i == -1) return false;
197 if ((size_t)i != LUIdx) return true;
198 return UsedByIndices.find_next(i) != -1;
201 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
202 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
203 assert(I != RegUsesMap.end() && "Unknown register!");
204 return I->second.UsedByIndices;
207 void RegUseTracker::clear() {
214 /// Formula - This class holds information that describes a formula for
215 /// computing satisfying a use. It may include broken-out immediates and scaled
218 /// AM - This is used to represent complex addressing, as well as other kinds
219 /// of interesting uses.
220 TargetLowering::AddrMode AM;
222 /// BaseRegs - The list of "base" registers for this use. When this is
223 /// non-empty, AM.HasBaseReg should be set to true.
224 SmallVector<const SCEV *, 2> BaseRegs;
226 /// ScaledReg - The 'scaled' register for this use. This should be non-null
227 /// when AM.Scale is not zero.
228 const SCEV *ScaledReg;
230 /// UnfoldedOffset - An additional constant offset which added near the
231 /// use. This requires a temporary register, but the offset itself can
232 /// live in an add immediate field rather than a register.
233 int64_t UnfoldedOffset;
235 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
237 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
239 unsigned getNumRegs() const;
240 Type *getType() const;
242 void DeleteBaseReg(const SCEV *&S);
244 bool referencesReg(const SCEV *S) const;
245 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
246 const RegUseTracker &RegUses) const;
248 void print(raw_ostream &OS) const;
254 /// DoInitialMatch - Recursion helper for InitialMatch.
255 static void DoInitialMatch(const SCEV *S, Loop *L,
256 SmallVectorImpl<const SCEV *> &Good,
257 SmallVectorImpl<const SCEV *> &Bad,
258 ScalarEvolution &SE) {
259 // Collect expressions which properly dominate the loop header.
260 if (SE.properlyDominates(S, L->getHeader())) {
265 // Look at add operands.
266 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
267 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
269 DoInitialMatch(*I, L, Good, Bad, SE);
273 // Look at addrec operands.
274 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
275 if (!AR->getStart()->isZero()) {
276 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
277 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
278 AR->getStepRecurrence(SE),
279 // FIXME: AR->getNoWrapFlags()
280 AR->getLoop(), SCEV::FlagAnyWrap),
285 // Handle a multiplication by -1 (negation) if it didn't fold.
286 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
287 if (Mul->getOperand(0)->isAllOnesValue()) {
288 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
289 const SCEV *NewMul = SE.getMulExpr(Ops);
291 SmallVector<const SCEV *, 4> MyGood;
292 SmallVector<const SCEV *, 4> MyBad;
293 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
294 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
295 SE.getEffectiveSCEVType(NewMul->getType())));
296 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
297 E = MyGood.end(); I != E; ++I)
298 Good.push_back(SE.getMulExpr(NegOne, *I));
299 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
300 E = MyBad.end(); I != E; ++I)
301 Bad.push_back(SE.getMulExpr(NegOne, *I));
305 // Ok, we can't do anything interesting. Just stuff the whole thing into a
306 // register and hope for the best.
310 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
311 /// attempting to keep all loop-invariant and loop-computable values in a
312 /// single base register.
313 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
314 SmallVector<const SCEV *, 4> Good;
315 SmallVector<const SCEV *, 4> Bad;
316 DoInitialMatch(S, L, Good, Bad, SE);
318 const SCEV *Sum = SE.getAddExpr(Good);
320 BaseRegs.push_back(Sum);
321 AM.HasBaseReg = true;
324 const SCEV *Sum = SE.getAddExpr(Bad);
326 BaseRegs.push_back(Sum);
327 AM.HasBaseReg = true;
331 /// getNumRegs - Return the total number of register operands used by this
332 /// formula. This does not include register uses implied by non-constant
334 unsigned Formula::getNumRegs() const {
335 return !!ScaledReg + BaseRegs.size();
338 /// getType - Return the type of this formula, if it has one, or null
339 /// otherwise. This type is meaningless except for the bit size.
340 Type *Formula::getType() const {
341 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
342 ScaledReg ? ScaledReg->getType() :
343 AM.BaseGV ? AM.BaseGV->getType() :
347 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
348 void Formula::DeleteBaseReg(const SCEV *&S) {
349 if (&S != &BaseRegs.back())
350 std::swap(S, BaseRegs.back());
354 /// referencesReg - Test if this formula references the given register.
355 bool Formula::referencesReg(const SCEV *S) const {
356 return S == ScaledReg ||
357 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
360 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
361 /// which are used by uses other than the use with the given index.
362 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
363 const RegUseTracker &RegUses) const {
365 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
367 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
368 E = BaseRegs.end(); I != E; ++I)
369 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
374 void Formula::print(raw_ostream &OS) const {
377 if (!First) OS << " + "; else First = false;
378 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
380 if (AM.BaseOffs != 0) {
381 if (!First) OS << " + "; else First = false;
384 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
385 E = BaseRegs.end(); I != E; ++I) {
386 if (!First) OS << " + "; else First = false;
387 OS << "reg(" << **I << ')';
389 if (AM.HasBaseReg && BaseRegs.empty()) {
390 if (!First) OS << " + "; else First = false;
391 OS << "**error: HasBaseReg**";
392 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
393 if (!First) OS << " + "; else First = false;
394 OS << "**error: !HasBaseReg**";
397 if (!First) OS << " + "; else First = false;
398 OS << AM.Scale << "*reg(";
405 if (UnfoldedOffset != 0) {
406 if (!First) OS << " + "; else First = false;
407 OS << "imm(" << UnfoldedOffset << ')';
411 void Formula::dump() const {
412 print(errs()); errs() << '\n';
415 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
416 /// without changing its value.
417 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
419 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
420 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
423 /// isAddSExtable - Return true if the given add can be sign-extended
424 /// without changing its value.
425 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
427 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
428 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
431 /// isMulSExtable - Return true if the given mul can be sign-extended
432 /// without changing its value.
433 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
435 IntegerType::get(SE.getContext(),
436 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
437 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
440 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
441 /// and if the remainder is known to be zero, or null otherwise. If
442 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
443 /// to Y, ignoring that the multiplication may overflow, which is useful when
444 /// the result will be used in a context where the most significant bits are
446 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
448 bool IgnoreSignificantBits = false) {
449 // Handle the trivial case, which works for any SCEV type.
451 return SE.getConstant(LHS->getType(), 1);
453 // Handle a few RHS special cases.
454 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
456 const APInt &RA = RC->getValue()->getValue();
457 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
459 if (RA.isAllOnesValue())
460 return SE.getMulExpr(LHS, RC);
461 // Handle x /s 1 as x.
466 // Check for a division of a constant by a constant.
467 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
470 const APInt &LA = C->getValue()->getValue();
471 const APInt &RA = RC->getValue()->getValue();
472 if (LA.srem(RA) != 0)
474 return SE.getConstant(LA.sdiv(RA));
477 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
478 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
479 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
480 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
481 IgnoreSignificantBits);
483 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
484 IgnoreSignificantBits);
485 if (!Start) return 0;
486 // FlagNW is independent of the start value, step direction, and is
487 // preserved with smaller magnitude steps.
488 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
489 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
494 // Distribute the sdiv over add operands, if the add doesn't overflow.
495 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
496 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
497 SmallVector<const SCEV *, 8> Ops;
498 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
500 const SCEV *Op = getExactSDiv(*I, RHS, SE,
501 IgnoreSignificantBits);
505 return SE.getAddExpr(Ops);
510 // Check for a multiply operand that we can pull RHS out of.
511 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
512 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
513 SmallVector<const SCEV *, 4> Ops;
515 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
519 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
520 IgnoreSignificantBits)) {
526 return Found ? SE.getMulExpr(Ops) : 0;
531 // Otherwise we don't know.
535 /// ExtractImmediate - If S involves the addition of a constant integer value,
536 /// return that integer value, and mutate S to point to a new SCEV with that
538 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
539 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
540 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
541 S = SE.getConstant(C->getType(), 0);
542 return C->getValue()->getSExtValue();
544 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
545 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
546 int64_t Result = ExtractImmediate(NewOps.front(), SE);
548 S = SE.getAddExpr(NewOps);
550 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
551 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
552 int64_t Result = ExtractImmediate(NewOps.front(), SE);
554 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
555 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
562 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
563 /// return that symbol, and mutate S to point to a new SCEV with that
565 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
566 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
567 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
568 S = SE.getConstant(GV->getType(), 0);
571 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
572 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
573 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
575 S = SE.getAddExpr(NewOps);
577 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
578 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
579 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
581 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
582 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
589 /// isAddressUse - Returns true if the specified instruction is using the
590 /// specified value as an address.
591 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
592 bool isAddress = isa<LoadInst>(Inst);
593 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
594 if (SI->getOperand(1) == OperandVal)
596 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
597 // Addressing modes can also be folded into prefetches and a variety
599 switch (II->getIntrinsicID()) {
601 case Intrinsic::prefetch:
602 case Intrinsic::x86_sse_storeu_ps:
603 case Intrinsic::x86_sse2_storeu_pd:
604 case Intrinsic::x86_sse2_storeu_dq:
605 case Intrinsic::x86_sse2_storel_dq:
606 if (II->getArgOperand(0) == OperandVal)
614 /// getAccessType - Return the type of the memory being accessed.
615 static Type *getAccessType(const Instruction *Inst) {
616 Type *AccessTy = Inst->getType();
617 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
618 AccessTy = SI->getOperand(0)->getType();
619 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
620 // Addressing modes can also be folded into prefetches and a variety
622 switch (II->getIntrinsicID()) {
624 case Intrinsic::x86_sse_storeu_ps:
625 case Intrinsic::x86_sse2_storeu_pd:
626 case Intrinsic::x86_sse2_storeu_dq:
627 case Intrinsic::x86_sse2_storel_dq:
628 AccessTy = II->getArgOperand(0)->getType();
633 // All pointers have the same requirements, so canonicalize them to an
634 // arbitrary pointer type to minimize variation.
635 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
636 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
637 PTy->getAddressSpace());
642 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
643 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
644 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
645 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
646 if (SE.isSCEVable(PN->getType()) &&
647 (SE.getEffectiveSCEVType(PN->getType()) ==
648 SE.getEffectiveSCEVType(AR->getType())) &&
649 SE.getSCEV(PN) == AR)
655 /// Check if expanding this expression is likely to incur significant cost. This
656 /// is tricky because SCEV doesn't track which expressions are actually computed
657 /// by the current IR.
659 /// We currently allow expansion of IV increments that involve adds,
660 /// multiplication by constants, and AddRecs from existing phis.
662 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
663 /// obvious multiple of the UDivExpr.
664 static bool isHighCostExpansion(const SCEV *S,
665 SmallPtrSet<const SCEV*, 8> &Processed,
666 ScalarEvolution &SE) {
667 // Zero/One operand expressions
668 switch (S->getSCEVType()) {
673 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
676 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
679 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
683 if (!Processed.insert(S))
686 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
687 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
689 if (isHighCostExpansion(*I, Processed, SE))
695 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
696 if (Mul->getNumOperands() == 2) {
697 // Multiplication by a constant is ok
698 if (isa<SCEVConstant>(Mul->getOperand(0)))
699 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
701 // If we have the value of one operand, check if an existing
702 // multiplication already generates this expression.
703 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
704 Value *UVal = U->getValue();
705 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
707 Instruction *User = cast<Instruction>(*UI);
708 if (User->getOpcode() == Instruction::Mul
709 && SE.isSCEVable(User->getType())) {
710 return SE.getSCEV(User) == Mul;
717 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
718 if (isExistingPhi(AR, SE))
722 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
726 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
727 /// specified set are trivially dead, delete them and see if this makes any of
728 /// their operands subsequently dead.
730 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
731 bool Changed = false;
733 while (!DeadInsts.empty()) {
734 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
736 if (I == 0 || !isInstructionTriviallyDead(I))
739 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
740 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
743 DeadInsts.push_back(U);
746 I->eraseFromParent();
755 /// Cost - This class is used to measure and compare candidate formulae.
757 /// TODO: Some of these could be merged. Also, a lexical ordering
758 /// isn't always optimal.
762 unsigned NumBaseAdds;
768 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
771 bool operator<(const Cost &Other) const;
776 // Once any of the metrics loses, they must all remain losers.
778 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
779 | ImmCost | SetupCost) != ~0u)
780 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
781 & ImmCost & SetupCost) == ~0u);
786 assert(isValid() && "invalid cost");
787 return NumRegs == ~0u;
790 void RateFormula(const Formula &F,
791 SmallPtrSet<const SCEV *, 16> &Regs,
792 const DenseSet<const SCEV *> &VisitedRegs,
794 const SmallVectorImpl<int64_t> &Offsets,
795 ScalarEvolution &SE, DominatorTree &DT,
796 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
798 void print(raw_ostream &OS) const;
802 void RateRegister(const SCEV *Reg,
803 SmallPtrSet<const SCEV *, 16> &Regs,
805 ScalarEvolution &SE, DominatorTree &DT);
806 void RatePrimaryRegister(const SCEV *Reg,
807 SmallPtrSet<const SCEV *, 16> &Regs,
809 ScalarEvolution &SE, DominatorTree &DT,
810 SmallPtrSet<const SCEV *, 16> *LoserRegs);
815 /// RateRegister - Tally up interesting quantities from the given register.
816 void Cost::RateRegister(const SCEV *Reg,
817 SmallPtrSet<const SCEV *, 16> &Regs,
819 ScalarEvolution &SE, DominatorTree &DT) {
820 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
821 // If this is an addrec for another loop, don't second-guess its addrec phi
822 // nodes. LSR isn't currently smart enough to reason about more than one
823 // loop at a time. LSR has already run on inner loops, will not run on outer
824 // loops, and cannot be expected to change sibling loops.
825 if (AR->getLoop() != L) {
826 // If the AddRec exists, consider it's register free and leave it alone.
827 if (isExistingPhi(AR, SE))
830 // Otherwise, do not consider this formula at all.
834 AddRecCost += 1; /// TODO: This should be a function of the stride.
836 // Add the step value register, if it needs one.
837 // TODO: The non-affine case isn't precisely modeled here.
838 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
839 if (!Regs.count(AR->getOperand(1))) {
840 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
848 // Rough heuristic; favor registers which don't require extra setup
849 // instructions in the preheader.
850 if (!isa<SCEVUnknown>(Reg) &&
851 !isa<SCEVConstant>(Reg) &&
852 !(isa<SCEVAddRecExpr>(Reg) &&
853 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
854 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
857 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
858 SE.hasComputableLoopEvolution(Reg, L);
861 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
862 /// before, rate it. Optional LoserRegs provides a way to declare any formula
863 /// that refers to one of those regs an instant loser.
864 void Cost::RatePrimaryRegister(const SCEV *Reg,
865 SmallPtrSet<const SCEV *, 16> &Regs,
867 ScalarEvolution &SE, DominatorTree &DT,
868 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
869 if (LoserRegs && LoserRegs->count(Reg)) {
873 if (Regs.insert(Reg)) {
874 RateRegister(Reg, Regs, L, SE, DT);
876 LoserRegs->insert(Reg);
880 void Cost::RateFormula(const Formula &F,
881 SmallPtrSet<const SCEV *, 16> &Regs,
882 const DenseSet<const SCEV *> &VisitedRegs,
884 const SmallVectorImpl<int64_t> &Offsets,
885 ScalarEvolution &SE, DominatorTree &DT,
886 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
887 // Tally up the registers.
888 if (const SCEV *ScaledReg = F.ScaledReg) {
889 if (VisitedRegs.count(ScaledReg)) {
893 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
897 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
898 E = F.BaseRegs.end(); I != E; ++I) {
899 const SCEV *BaseReg = *I;
900 if (VisitedRegs.count(BaseReg)) {
904 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
909 // Determine how many (unfolded) adds we'll need inside the loop.
910 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
911 if (NumBaseParts > 1)
912 NumBaseAdds += NumBaseParts - 1;
914 // Tally up the non-zero immediates.
915 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
916 E = Offsets.end(); I != E; ++I) {
917 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
919 ImmCost += 64; // Handle symbolic values conservatively.
920 // TODO: This should probably be the pointer size.
921 else if (Offset != 0)
922 ImmCost += APInt(64, Offset, true).getMinSignedBits();
924 assert(isValid() && "invalid cost");
927 /// Loose - Set this cost to a losing value.
937 /// operator< - Choose the lower cost.
938 bool Cost::operator<(const Cost &Other) const {
939 if (NumRegs != Other.NumRegs)
940 return NumRegs < Other.NumRegs;
941 if (AddRecCost != Other.AddRecCost)
942 return AddRecCost < Other.AddRecCost;
943 if (NumIVMuls != Other.NumIVMuls)
944 return NumIVMuls < Other.NumIVMuls;
945 if (NumBaseAdds != Other.NumBaseAdds)
946 return NumBaseAdds < Other.NumBaseAdds;
947 if (ImmCost != Other.ImmCost)
948 return ImmCost < Other.ImmCost;
949 if (SetupCost != Other.SetupCost)
950 return SetupCost < Other.SetupCost;
954 void Cost::print(raw_ostream &OS) const {
955 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
957 OS << ", with addrec cost " << AddRecCost;
959 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
960 if (NumBaseAdds != 0)
961 OS << ", plus " << NumBaseAdds << " base add"
962 << (NumBaseAdds == 1 ? "" : "s");
964 OS << ", plus " << ImmCost << " imm cost";
966 OS << ", plus " << SetupCost << " setup cost";
969 void Cost::dump() const {
970 print(errs()); errs() << '\n';
975 /// LSRFixup - An operand value in an instruction which is to be replaced
976 /// with some equivalent, possibly strength-reduced, replacement.
978 /// UserInst - The instruction which will be updated.
979 Instruction *UserInst;
981 /// OperandValToReplace - The operand of the instruction which will
982 /// be replaced. The operand may be used more than once; every instance
983 /// will be replaced.
984 Value *OperandValToReplace;
986 /// PostIncLoops - If this user is to use the post-incremented value of an
987 /// induction variable, this variable is non-null and holds the loop
988 /// associated with the induction variable.
989 PostIncLoopSet PostIncLoops;
991 /// LUIdx - The index of the LSRUse describing the expression which
992 /// this fixup needs, minus an offset (below).
995 /// Offset - A constant offset to be added to the LSRUse expression.
996 /// This allows multiple fixups to share the same LSRUse with different
997 /// offsets, for example in an unrolled loop.
1000 bool isUseFullyOutsideLoop(const Loop *L) const;
1004 void print(raw_ostream &OS) const;
1010 LSRFixup::LSRFixup()
1011 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1013 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1014 /// value outside of the given loop.
1015 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1016 // PHI nodes use their value in their incoming blocks.
1017 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1018 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1019 if (PN->getIncomingValue(i) == OperandValToReplace &&
1020 L->contains(PN->getIncomingBlock(i)))
1025 return !L->contains(UserInst);
1028 void LSRFixup::print(raw_ostream &OS) const {
1030 // Store is common and interesting enough to be worth special-casing.
1031 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1033 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1034 } else if (UserInst->getType()->isVoidTy())
1035 OS << UserInst->getOpcodeName();
1037 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1039 OS << ", OperandValToReplace=";
1040 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1042 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1043 E = PostIncLoops.end(); I != E; ++I) {
1044 OS << ", PostIncLoop=";
1045 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1048 if (LUIdx != ~size_t(0))
1049 OS << ", LUIdx=" << LUIdx;
1052 OS << ", Offset=" << Offset;
1055 void LSRFixup::dump() const {
1056 print(errs()); errs() << '\n';
1061 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1062 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1063 struct UniquifierDenseMapInfo {
1064 static SmallVector<const SCEV *, 2> getEmptyKey() {
1065 SmallVector<const SCEV *, 2> V;
1066 V.push_back(reinterpret_cast<const SCEV *>(-1));
1070 static SmallVector<const SCEV *, 2> getTombstoneKey() {
1071 SmallVector<const SCEV *, 2> V;
1072 V.push_back(reinterpret_cast<const SCEV *>(-2));
1076 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
1077 unsigned Result = 0;
1078 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1079 E = V.end(); I != E; ++I)
1080 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1084 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1085 const SmallVector<const SCEV *, 2> &RHS) {
1090 /// LSRUse - This class holds the state that LSR keeps for each use in
1091 /// IVUsers, as well as uses invented by LSR itself. It includes information
1092 /// about what kinds of things can be folded into the user, information about
1093 /// the user itself, and information about how the use may be satisfied.
1094 /// TODO: Represent multiple users of the same expression in common?
1096 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1099 /// KindType - An enum for a kind of use, indicating what types of
1100 /// scaled and immediate operands it might support.
1102 Basic, ///< A normal use, with no folding.
1103 Special, ///< A special case of basic, allowing -1 scales.
1104 Address, ///< An address use; folding according to TargetLowering
1105 ICmpZero ///< An equality icmp with both operands folded into one.
1106 // TODO: Add a generic icmp too?
1112 SmallVector<int64_t, 8> Offsets;
1116 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1117 /// LSRUse are outside of the loop, in which case some special-case heuristics
1119 bool AllFixupsOutsideLoop;
1121 /// WidestFixupType - This records the widest use type for any fixup using
1122 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1123 /// max fixup widths to be equivalent, because the narrower one may be relying
1124 /// on the implicit truncation to truncate away bogus bits.
1125 Type *WidestFixupType;
1127 /// Formulae - A list of ways to build a value that can satisfy this user.
1128 /// After the list is populated, one of these is selected heuristically and
1129 /// used to formulate a replacement for OperandValToReplace in UserInst.
1130 SmallVector<Formula, 12> Formulae;
1132 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1133 SmallPtrSet<const SCEV *, 4> Regs;
1135 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1136 MinOffset(INT64_MAX),
1137 MaxOffset(INT64_MIN),
1138 AllFixupsOutsideLoop(true),
1139 WidestFixupType(0) {}
1141 bool HasFormulaWithSameRegs(const Formula &F) const;
1142 bool InsertFormula(const Formula &F);
1143 void DeleteFormula(Formula &F);
1144 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1146 void print(raw_ostream &OS) const;
1152 /// HasFormula - Test whether this use as a formula which has the same
1153 /// registers as the given formula.
1154 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1155 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1156 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1157 // Unstable sort by host order ok, because this is only used for uniquifying.
1158 std::sort(Key.begin(), Key.end());
1159 return Uniquifier.count(Key);
1162 /// InsertFormula - If the given formula has not yet been inserted, add it to
1163 /// the list, and return true. Return false otherwise.
1164 bool LSRUse::InsertFormula(const Formula &F) {
1165 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1166 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1167 // Unstable sort by host order ok, because this is only used for uniquifying.
1168 std::sort(Key.begin(), Key.end());
1170 if (!Uniquifier.insert(Key).second)
1173 // Using a register to hold the value of 0 is not profitable.
1174 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1175 "Zero allocated in a scaled register!");
1177 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1178 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1179 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1182 // Add the formula to the list.
1183 Formulae.push_back(F);
1185 // Record registers now being used by this use.
1186 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1191 /// DeleteFormula - Remove the given formula from this use's list.
1192 void LSRUse::DeleteFormula(Formula &F) {
1193 if (&F != &Formulae.back())
1194 std::swap(F, Formulae.back());
1195 Formulae.pop_back();
1198 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1199 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1200 // Now that we've filtered out some formulae, recompute the Regs set.
1201 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1203 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1204 E = Formulae.end(); I != E; ++I) {
1205 const Formula &F = *I;
1206 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1207 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1210 // Update the RegTracker.
1211 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1212 E = OldRegs.end(); I != E; ++I)
1213 if (!Regs.count(*I))
1214 RegUses.DropRegister(*I, LUIdx);
1217 void LSRUse::print(raw_ostream &OS) const {
1218 OS << "LSR Use: Kind=";
1220 case Basic: OS << "Basic"; break;
1221 case Special: OS << "Special"; break;
1222 case ICmpZero: OS << "ICmpZero"; break;
1224 OS << "Address of ";
1225 if (AccessTy->isPointerTy())
1226 OS << "pointer"; // the full pointer type could be really verbose
1231 OS << ", Offsets={";
1232 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1233 E = Offsets.end(); I != E; ++I) {
1235 if (llvm::next(I) != E)
1240 if (AllFixupsOutsideLoop)
1241 OS << ", all-fixups-outside-loop";
1243 if (WidestFixupType)
1244 OS << ", widest fixup type: " << *WidestFixupType;
1247 void LSRUse::dump() const {
1248 print(errs()); errs() << '\n';
1251 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1252 /// be completely folded into the user instruction at isel time. This includes
1253 /// address-mode folding and special icmp tricks.
1254 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1255 LSRUse::KindType Kind, Type *AccessTy,
1256 const TargetLowering *TLI) {
1258 case LSRUse::Address:
1259 // If we have low-level target information, ask the target if it can
1260 // completely fold this address.
1261 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1263 // Otherwise, just guess that reg+reg addressing is legal.
1264 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1266 case LSRUse::ICmpZero:
1267 // There's not even a target hook for querying whether it would be legal to
1268 // fold a GV into an ICmp.
1272 // ICmp only has two operands; don't allow more than two non-trivial parts.
1273 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1276 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1277 // putting the scaled register in the other operand of the icmp.
1278 if (AM.Scale != 0 && AM.Scale != -1)
1281 // If we have low-level target information, ask the target if it can fold an
1282 // integer immediate on an icmp.
1283 if (AM.BaseOffs != 0) {
1284 if (TLI) return TLI->isLegalICmpImmediate(-(uint64_t)AM.BaseOffs);
1291 // Only handle single-register values.
1292 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1294 case LSRUse::Special:
1295 // Only handle -1 scales, or no scale.
1296 return AM.Scale == 0 || AM.Scale == -1;
1299 llvm_unreachable("Invalid LSRUse Kind!");
1302 static bool isLegalUse(TargetLowering::AddrMode AM,
1303 int64_t MinOffset, int64_t MaxOffset,
1304 LSRUse::KindType Kind, Type *AccessTy,
1305 const TargetLowering *TLI) {
1306 // Check for overflow.
1307 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1310 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1311 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1312 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1313 // Check for overflow.
1314 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1317 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1318 return isLegalUse(AM, Kind, AccessTy, TLI);
1323 static bool isAlwaysFoldable(int64_t BaseOffs,
1324 GlobalValue *BaseGV,
1326 LSRUse::KindType Kind, Type *AccessTy,
1327 const TargetLowering *TLI) {
1328 // Fast-path: zero is always foldable.
1329 if (BaseOffs == 0 && !BaseGV) return true;
1331 // Conservatively, create an address with an immediate and a
1332 // base and a scale.
1333 TargetLowering::AddrMode AM;
1334 AM.BaseOffs = BaseOffs;
1336 AM.HasBaseReg = HasBaseReg;
1337 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1339 // Canonicalize a scale of 1 to a base register if the formula doesn't
1340 // already have a base register.
1341 if (!AM.HasBaseReg && AM.Scale == 1) {
1343 AM.HasBaseReg = true;
1346 return isLegalUse(AM, Kind, AccessTy, TLI);
1349 static bool isAlwaysFoldable(const SCEV *S,
1350 int64_t MinOffset, int64_t MaxOffset,
1352 LSRUse::KindType Kind, Type *AccessTy,
1353 const TargetLowering *TLI,
1354 ScalarEvolution &SE) {
1355 // Fast-path: zero is always foldable.
1356 if (S->isZero()) return true;
1358 // Conservatively, create an address with an immediate and a
1359 // base and a scale.
1360 int64_t BaseOffs = ExtractImmediate(S, SE);
1361 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1363 // If there's anything else involved, it's not foldable.
1364 if (!S->isZero()) return false;
1366 // Fast-path: zero is always foldable.
1367 if (BaseOffs == 0 && !BaseGV) return true;
1369 // Conservatively, create an address with an immediate and a
1370 // base and a scale.
1371 TargetLowering::AddrMode AM;
1372 AM.BaseOffs = BaseOffs;
1374 AM.HasBaseReg = HasBaseReg;
1375 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1377 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1382 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1383 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1384 struct UseMapDenseMapInfo {
1385 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1386 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1389 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1390 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1394 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1395 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1396 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1400 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1401 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1406 /// IVInc - An individual increment in a Chain of IV increments.
1407 /// Relate an IV user to an expression that computes the IV it uses from the IV
1408 /// used by the previous link in the Chain.
1410 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1411 /// original IVOperand. The head of the chain's IVOperand is only valid during
1412 /// chain collection, before LSR replaces IV users. During chain generation,
1413 /// IncExpr can be used to find the new IVOperand that computes the same
1416 Instruction *UserInst;
1418 const SCEV *IncExpr;
1420 IVInc(Instruction *U, Value *O, const SCEV *E):
1421 UserInst(U), IVOperand(O), IncExpr(E) {}
1424 // IVChain - The list of IV increments in program order.
1425 // We typically add the head of a chain without finding subsequent links.
1426 typedef SmallVector<IVInc,1> IVChain;
1428 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1429 /// Distinguish between FarUsers that definitely cross IV increments and
1430 /// NearUsers that may be used between IV increments.
1432 SmallPtrSet<Instruction*, 4> FarUsers;
1433 SmallPtrSet<Instruction*, 4> NearUsers;
1436 /// LSRInstance - This class holds state for the main loop strength reduction
1440 ScalarEvolution &SE;
1443 const TargetLowering *const TLI;
1447 /// IVIncInsertPos - This is the insert position that the current loop's
1448 /// induction variable increment should be placed. In simple loops, this is
1449 /// the latch block's terminator. But in more complicated cases, this is a
1450 /// position which will dominate all the in-loop post-increment users.
1451 Instruction *IVIncInsertPos;
1453 /// Factors - Interesting factors between use strides.
1454 SmallSetVector<int64_t, 8> Factors;
1456 /// Types - Interesting use types, to facilitate truncation reuse.
1457 SmallSetVector<Type *, 4> Types;
1459 /// Fixups - The list of operands which are to be replaced.
1460 SmallVector<LSRFixup, 16> Fixups;
1462 /// Uses - The list of interesting uses.
1463 SmallVector<LSRUse, 16> Uses;
1465 /// RegUses - Track which uses use which register candidates.
1466 RegUseTracker RegUses;
1468 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1469 // have more than a few IV increment chains in a loop. Missing a Chain falls
1470 // back to normal LSR behavior for those uses.
1471 static const unsigned MaxChains = 8;
1473 /// IVChainVec - IV users can form a chain of IV increments.
1474 SmallVector<IVChain, MaxChains> IVChainVec;
1476 /// IVIncSet - IV users that belong to profitable IVChains.
1477 SmallPtrSet<Use*, MaxChains> IVIncSet;
1479 void OptimizeShadowIV();
1480 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1481 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1482 void OptimizeLoopTermCond();
1484 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1485 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1486 void FinalizeChain(IVChain &Chain);
1487 void CollectChains();
1488 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1489 SmallVectorImpl<WeakVH> &DeadInsts);
1491 void CollectInterestingTypesAndFactors();
1492 void CollectFixupsAndInitialFormulae();
1494 LSRFixup &getNewFixup() {
1495 Fixups.push_back(LSRFixup());
1496 return Fixups.back();
1499 // Support for sharing of LSRUses between LSRFixups.
1500 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1502 UseMapDenseMapInfo> UseMapTy;
1505 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1506 LSRUse::KindType Kind, Type *AccessTy);
1508 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1509 LSRUse::KindType Kind,
1512 void DeleteUse(LSRUse &LU, size_t LUIdx);
1514 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1516 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1517 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1518 void CountRegisters(const Formula &F, size_t LUIdx);
1519 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1521 void CollectLoopInvariantFixupsAndFormulae();
1523 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1524 unsigned Depth = 0);
1525 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1526 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1527 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1528 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1529 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1530 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1531 void GenerateCrossUseConstantOffsets();
1532 void GenerateAllReuseFormulae();
1534 void FilterOutUndesirableDedicatedRegisters();
1536 size_t EstimateSearchSpaceComplexity() const;
1537 void NarrowSearchSpaceByDetectingSupersets();
1538 void NarrowSearchSpaceByCollapsingUnrolledCode();
1539 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1540 void NarrowSearchSpaceByPickingWinnerRegs();
1541 void NarrowSearchSpaceUsingHeuristics();
1543 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1545 SmallVectorImpl<const Formula *> &Workspace,
1546 const Cost &CurCost,
1547 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1548 DenseSet<const SCEV *> &VisitedRegs) const;
1549 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1551 BasicBlock::iterator
1552 HoistInsertPosition(BasicBlock::iterator IP,
1553 const SmallVectorImpl<Instruction *> &Inputs) const;
1554 BasicBlock::iterator
1555 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1558 SCEVExpander &Rewriter) const;
1560 Value *Expand(const LSRFixup &LF,
1562 BasicBlock::iterator IP,
1563 SCEVExpander &Rewriter,
1564 SmallVectorImpl<WeakVH> &DeadInsts) const;
1565 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1567 SCEVExpander &Rewriter,
1568 SmallVectorImpl<WeakVH> &DeadInsts,
1570 void Rewrite(const LSRFixup &LF,
1572 SCEVExpander &Rewriter,
1573 SmallVectorImpl<WeakVH> &DeadInsts,
1575 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1579 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1581 bool getChanged() const { return Changed; }
1583 void print_factors_and_types(raw_ostream &OS) const;
1584 void print_fixups(raw_ostream &OS) const;
1585 void print_uses(raw_ostream &OS) const;
1586 void print(raw_ostream &OS) const;
1592 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1593 /// inside the loop then try to eliminate the cast operation.
1594 void LSRInstance::OptimizeShadowIV() {
1595 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1596 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1599 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1600 UI != E; /* empty */) {
1601 IVUsers::const_iterator CandidateUI = UI;
1603 Instruction *ShadowUse = CandidateUI->getUser();
1604 Type *DestTy = NULL;
1605 bool IsSigned = false;
1607 /* If shadow use is a int->float cast then insert a second IV
1608 to eliminate this cast.
1610 for (unsigned i = 0; i < n; ++i)
1616 for (unsigned i = 0; i < n; ++i, ++d)
1619 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1621 DestTy = UCast->getDestTy();
1623 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1625 DestTy = SCast->getDestTy();
1627 if (!DestTy) continue;
1630 // If target does not support DestTy natively then do not apply
1631 // this transformation.
1632 EVT DVT = TLI->getValueType(DestTy);
1633 if (!TLI->isTypeLegal(DVT)) continue;
1636 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1638 if (PH->getNumIncomingValues() != 2) continue;
1640 Type *SrcTy = PH->getType();
1641 int Mantissa = DestTy->getFPMantissaWidth();
1642 if (Mantissa == -1) continue;
1643 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1646 unsigned Entry, Latch;
1647 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1655 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1656 if (!Init) continue;
1657 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1658 (double)Init->getSExtValue() :
1659 (double)Init->getZExtValue());
1661 BinaryOperator *Incr =
1662 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1663 if (!Incr) continue;
1664 if (Incr->getOpcode() != Instruction::Add
1665 && Incr->getOpcode() != Instruction::Sub)
1668 /* Initialize new IV, double d = 0.0 in above example. */
1669 ConstantInt *C = NULL;
1670 if (Incr->getOperand(0) == PH)
1671 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1672 else if (Incr->getOperand(1) == PH)
1673 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1679 // Ignore negative constants, as the code below doesn't handle them
1680 // correctly. TODO: Remove this restriction.
1681 if (!C->getValue().isStrictlyPositive()) continue;
1683 /* Add new PHINode. */
1684 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1686 /* create new increment. '++d' in above example. */
1687 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1688 BinaryOperator *NewIncr =
1689 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1690 Instruction::FAdd : Instruction::FSub,
1691 NewPH, CFP, "IV.S.next.", Incr);
1693 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1694 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1696 /* Remove cast operation */
1697 ShadowUse->replaceAllUsesWith(NewPH);
1698 ShadowUse->eraseFromParent();
1704 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1705 /// set the IV user and stride information and return true, otherwise return
1707 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1708 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1709 if (UI->getUser() == Cond) {
1710 // NOTE: we could handle setcc instructions with multiple uses here, but
1711 // InstCombine does it as well for simple uses, it's not clear that it
1712 // occurs enough in real life to handle.
1719 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1720 /// a max computation.
1722 /// This is a narrow solution to a specific, but acute, problem. For loops
1728 /// } while (++i < n);
1730 /// the trip count isn't just 'n', because 'n' might not be positive. And
1731 /// unfortunately this can come up even for loops where the user didn't use
1732 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1733 /// will commonly be lowered like this:
1739 /// } while (++i < n);
1742 /// and then it's possible for subsequent optimization to obscure the if
1743 /// test in such a way that indvars can't find it.
1745 /// When indvars can't find the if test in loops like this, it creates a
1746 /// max expression, which allows it to give the loop a canonical
1747 /// induction variable:
1750 /// max = n < 1 ? 1 : n;
1753 /// } while (++i != max);
1755 /// Canonical induction variables are necessary because the loop passes
1756 /// are designed around them. The most obvious example of this is the
1757 /// LoopInfo analysis, which doesn't remember trip count values. It
1758 /// expects to be able to rediscover the trip count each time it is
1759 /// needed, and it does this using a simple analysis that only succeeds if
1760 /// the loop has a canonical induction variable.
1762 /// However, when it comes time to generate code, the maximum operation
1763 /// can be quite costly, especially if it's inside of an outer loop.
1765 /// This function solves this problem by detecting this type of loop and
1766 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1767 /// the instructions for the maximum computation.
1769 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1770 // Check that the loop matches the pattern we're looking for.
1771 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1772 Cond->getPredicate() != CmpInst::ICMP_NE)
1775 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1776 if (!Sel || !Sel->hasOneUse()) return Cond;
1778 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1779 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1781 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1783 // Add one to the backedge-taken count to get the trip count.
1784 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1785 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1787 // Check for a max calculation that matches the pattern. There's no check
1788 // for ICMP_ULE here because the comparison would be with zero, which
1789 // isn't interesting.
1790 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1791 const SCEVNAryExpr *Max = 0;
1792 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1793 Pred = ICmpInst::ICMP_SLE;
1795 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1796 Pred = ICmpInst::ICMP_SLT;
1798 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1799 Pred = ICmpInst::ICMP_ULT;
1806 // To handle a max with more than two operands, this optimization would
1807 // require additional checking and setup.
1808 if (Max->getNumOperands() != 2)
1811 const SCEV *MaxLHS = Max->getOperand(0);
1812 const SCEV *MaxRHS = Max->getOperand(1);
1814 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1815 // for a comparison with 1. For <= and >=, a comparison with zero.
1817 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1820 // Check the relevant induction variable for conformance to
1822 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1823 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1824 if (!AR || !AR->isAffine() ||
1825 AR->getStart() != One ||
1826 AR->getStepRecurrence(SE) != One)
1829 assert(AR->getLoop() == L &&
1830 "Loop condition operand is an addrec in a different loop!");
1832 // Check the right operand of the select, and remember it, as it will
1833 // be used in the new comparison instruction.
1835 if (ICmpInst::isTrueWhenEqual(Pred)) {
1836 // Look for n+1, and grab n.
1837 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1838 if (isa<ConstantInt>(BO->getOperand(1)) &&
1839 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1840 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1841 NewRHS = BO->getOperand(0);
1842 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1843 if (isa<ConstantInt>(BO->getOperand(1)) &&
1844 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1845 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1846 NewRHS = BO->getOperand(0);
1849 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1850 NewRHS = Sel->getOperand(1);
1851 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1852 NewRHS = Sel->getOperand(2);
1853 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1854 NewRHS = SU->getValue();
1856 // Max doesn't match expected pattern.
1859 // Determine the new comparison opcode. It may be signed or unsigned,
1860 // and the original comparison may be either equality or inequality.
1861 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1862 Pred = CmpInst::getInversePredicate(Pred);
1864 // Ok, everything looks ok to change the condition into an SLT or SGE and
1865 // delete the max calculation.
1867 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1869 // Delete the max calculation instructions.
1870 Cond->replaceAllUsesWith(NewCond);
1871 CondUse->setUser(NewCond);
1872 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1873 Cond->eraseFromParent();
1874 Sel->eraseFromParent();
1875 if (Cmp->use_empty())
1876 Cmp->eraseFromParent();
1880 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1881 /// postinc iv when possible.
1883 LSRInstance::OptimizeLoopTermCond() {
1884 SmallPtrSet<Instruction *, 4> PostIncs;
1886 BasicBlock *LatchBlock = L->getLoopLatch();
1887 SmallVector<BasicBlock*, 8> ExitingBlocks;
1888 L->getExitingBlocks(ExitingBlocks);
1890 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1891 BasicBlock *ExitingBlock = ExitingBlocks[i];
1893 // Get the terminating condition for the loop if possible. If we
1894 // can, we want to change it to use a post-incremented version of its
1895 // induction variable, to allow coalescing the live ranges for the IV into
1896 // one register value.
1898 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1901 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1902 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1905 // Search IVUsesByStride to find Cond's IVUse if there is one.
1906 IVStrideUse *CondUse = 0;
1907 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1908 if (!FindIVUserForCond(Cond, CondUse))
1911 // If the trip count is computed in terms of a max (due to ScalarEvolution
1912 // being unable to find a sufficient guard, for example), change the loop
1913 // comparison to use SLT or ULT instead of NE.
1914 // One consequence of doing this now is that it disrupts the count-down
1915 // optimization. That's not always a bad thing though, because in such
1916 // cases it may still be worthwhile to avoid a max.
1917 Cond = OptimizeMax(Cond, CondUse);
1919 // If this exiting block dominates the latch block, it may also use
1920 // the post-inc value if it won't be shared with other uses.
1921 // Check for dominance.
1922 if (!DT.dominates(ExitingBlock, LatchBlock))
1925 // Conservatively avoid trying to use the post-inc value in non-latch
1926 // exits if there may be pre-inc users in intervening blocks.
1927 if (LatchBlock != ExitingBlock)
1928 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1929 // Test if the use is reachable from the exiting block. This dominator
1930 // query is a conservative approximation of reachability.
1931 if (&*UI != CondUse &&
1932 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1933 // Conservatively assume there may be reuse if the quotient of their
1934 // strides could be a legal scale.
1935 const SCEV *A = IU.getStride(*CondUse, L);
1936 const SCEV *B = IU.getStride(*UI, L);
1937 if (!A || !B) continue;
1938 if (SE.getTypeSizeInBits(A->getType()) !=
1939 SE.getTypeSizeInBits(B->getType())) {
1940 if (SE.getTypeSizeInBits(A->getType()) >
1941 SE.getTypeSizeInBits(B->getType()))
1942 B = SE.getSignExtendExpr(B, A->getType());
1944 A = SE.getSignExtendExpr(A, B->getType());
1946 if (const SCEVConstant *D =
1947 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1948 const ConstantInt *C = D->getValue();
1949 // Stride of one or negative one can have reuse with non-addresses.
1950 if (C->isOne() || C->isAllOnesValue())
1951 goto decline_post_inc;
1952 // Avoid weird situations.
1953 if (C->getValue().getMinSignedBits() >= 64 ||
1954 C->getValue().isMinSignedValue())
1955 goto decline_post_inc;
1956 // Without TLI, assume that any stride might be valid, and so any
1957 // use might be shared.
1959 goto decline_post_inc;
1960 // Check for possible scaled-address reuse.
1961 Type *AccessTy = getAccessType(UI->getUser());
1962 TargetLowering::AddrMode AM;
1963 AM.Scale = C->getSExtValue();
1964 if (TLI->isLegalAddressingMode(AM, AccessTy))
1965 goto decline_post_inc;
1966 AM.Scale = -AM.Scale;
1967 if (TLI->isLegalAddressingMode(AM, AccessTy))
1968 goto decline_post_inc;
1972 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1975 // It's possible for the setcc instruction to be anywhere in the loop, and
1976 // possible for it to have multiple users. If it is not immediately before
1977 // the exiting block branch, move it.
1978 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1979 if (Cond->hasOneUse()) {
1980 Cond->moveBefore(TermBr);
1982 // Clone the terminating condition and insert into the loopend.
1983 ICmpInst *OldCond = Cond;
1984 Cond = cast<ICmpInst>(Cond->clone());
1985 Cond->setName(L->getHeader()->getName() + ".termcond");
1986 ExitingBlock->getInstList().insert(TermBr, Cond);
1988 // Clone the IVUse, as the old use still exists!
1989 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1990 TermBr->replaceUsesOfWith(OldCond, Cond);
1994 // If we get to here, we know that we can transform the setcc instruction to
1995 // use the post-incremented version of the IV, allowing us to coalesce the
1996 // live ranges for the IV correctly.
1997 CondUse->transformToPostInc(L);
2000 PostIncs.insert(Cond);
2004 // Determine an insertion point for the loop induction variable increment. It
2005 // must dominate all the post-inc comparisons we just set up, and it must
2006 // dominate the loop latch edge.
2007 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2008 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2009 E = PostIncs.end(); I != E; ++I) {
2011 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2013 if (BB == (*I)->getParent())
2014 IVIncInsertPos = *I;
2015 else if (BB != IVIncInsertPos->getParent())
2016 IVIncInsertPos = BB->getTerminator();
2020 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2021 /// at the given offset and other details. If so, update the use and
2024 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2025 LSRUse::KindType Kind, Type *AccessTy) {
2026 int64_t NewMinOffset = LU.MinOffset;
2027 int64_t NewMaxOffset = LU.MaxOffset;
2028 Type *NewAccessTy = AccessTy;
2030 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2031 // something conservative, however this can pessimize in the case that one of
2032 // the uses will have all its uses outside the loop, for example.
2033 if (LU.Kind != Kind)
2035 // Conservatively assume HasBaseReg is true for now.
2036 if (NewOffset < LU.MinOffset) {
2037 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
2038 Kind, AccessTy, TLI))
2040 NewMinOffset = NewOffset;
2041 } else if (NewOffset > LU.MaxOffset) {
2042 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
2043 Kind, AccessTy, TLI))
2045 NewMaxOffset = NewOffset;
2047 // Check for a mismatched access type, and fall back conservatively as needed.
2048 // TODO: Be less conservative when the type is similar and can use the same
2049 // addressing modes.
2050 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2051 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2054 LU.MinOffset = NewMinOffset;
2055 LU.MaxOffset = NewMaxOffset;
2056 LU.AccessTy = NewAccessTy;
2057 if (NewOffset != LU.Offsets.back())
2058 LU.Offsets.push_back(NewOffset);
2062 /// getUse - Return an LSRUse index and an offset value for a fixup which
2063 /// needs the given expression, with the given kind and optional access type.
2064 /// Either reuse an existing use or create a new one, as needed.
2065 std::pair<size_t, int64_t>
2066 LSRInstance::getUse(const SCEV *&Expr,
2067 LSRUse::KindType Kind, Type *AccessTy) {
2068 const SCEV *Copy = Expr;
2069 int64_t Offset = ExtractImmediate(Expr, SE);
2071 // Basic uses can't accept any offset, for example.
2072 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
2077 std::pair<UseMapTy::iterator, bool> P =
2078 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2080 // A use already existed with this base.
2081 size_t LUIdx = P.first->second;
2082 LSRUse &LU = Uses[LUIdx];
2083 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2085 return std::make_pair(LUIdx, Offset);
2088 // Create a new use.
2089 size_t LUIdx = Uses.size();
2090 P.first->second = LUIdx;
2091 Uses.push_back(LSRUse(Kind, AccessTy));
2092 LSRUse &LU = Uses[LUIdx];
2094 // We don't need to track redundant offsets, but we don't need to go out
2095 // of our way here to avoid them.
2096 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2097 LU.Offsets.push_back(Offset);
2099 LU.MinOffset = Offset;
2100 LU.MaxOffset = Offset;
2101 return std::make_pair(LUIdx, Offset);
2104 /// DeleteUse - Delete the given use from the Uses list.
2105 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2106 if (&LU != &Uses.back())
2107 std::swap(LU, Uses.back());
2111 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2114 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2115 /// a formula that has the same registers as the given formula.
2117 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2118 const LSRUse &OrigLU) {
2119 // Search all uses for the formula. This could be more clever.
2120 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2121 LSRUse &LU = Uses[LUIdx];
2122 // Check whether this use is close enough to OrigLU, to see whether it's
2123 // worthwhile looking through its formulae.
2124 // Ignore ICmpZero uses because they may contain formulae generated by
2125 // GenerateICmpZeroScales, in which case adding fixup offsets may
2127 if (&LU != &OrigLU &&
2128 LU.Kind != LSRUse::ICmpZero &&
2129 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2130 LU.WidestFixupType == OrigLU.WidestFixupType &&
2131 LU.HasFormulaWithSameRegs(OrigF)) {
2132 // Scan through this use's formulae.
2133 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2134 E = LU.Formulae.end(); I != E; ++I) {
2135 const Formula &F = *I;
2136 // Check to see if this formula has the same registers and symbols
2138 if (F.BaseRegs == OrigF.BaseRegs &&
2139 F.ScaledReg == OrigF.ScaledReg &&
2140 F.AM.BaseGV == OrigF.AM.BaseGV &&
2141 F.AM.Scale == OrigF.AM.Scale &&
2142 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2143 if (F.AM.BaseOffs == 0)
2145 // This is the formula where all the registers and symbols matched;
2146 // there aren't going to be any others. Since we declined it, we
2147 // can skip the rest of the formulae and procede to the next LSRUse.
2154 // Nothing looked good.
2158 void LSRInstance::CollectInterestingTypesAndFactors() {
2159 SmallSetVector<const SCEV *, 4> Strides;
2161 // Collect interesting types and strides.
2162 SmallVector<const SCEV *, 4> Worklist;
2163 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2164 const SCEV *Expr = IU.getExpr(*UI);
2166 // Collect interesting types.
2167 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2169 // Add strides for mentioned loops.
2170 Worklist.push_back(Expr);
2172 const SCEV *S = Worklist.pop_back_val();
2173 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2174 if (AR->getLoop() == L)
2175 Strides.insert(AR->getStepRecurrence(SE));
2176 Worklist.push_back(AR->getStart());
2177 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2178 Worklist.append(Add->op_begin(), Add->op_end());
2180 } while (!Worklist.empty());
2183 // Compute interesting factors from the set of interesting strides.
2184 for (SmallSetVector<const SCEV *, 4>::const_iterator
2185 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2186 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2187 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2188 const SCEV *OldStride = *I;
2189 const SCEV *NewStride = *NewStrideIter;
2191 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2192 SE.getTypeSizeInBits(NewStride->getType())) {
2193 if (SE.getTypeSizeInBits(OldStride->getType()) >
2194 SE.getTypeSizeInBits(NewStride->getType()))
2195 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2197 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2199 if (const SCEVConstant *Factor =
2200 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2202 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2203 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2204 } else if (const SCEVConstant *Factor =
2205 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2208 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2209 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2213 // If all uses use the same type, don't bother looking for truncation-based
2215 if (Types.size() == 1)
2218 DEBUG(print_factors_and_types(dbgs()));
2221 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2222 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2223 /// Instructions to IVStrideUses, we could partially skip this.
2224 static User::op_iterator
2225 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2226 Loop *L, ScalarEvolution &SE) {
2227 for(; OI != OE; ++OI) {
2228 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2229 if (!SE.isSCEVable(Oper->getType()))
2232 if (const SCEVAddRecExpr *AR =
2233 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2234 if (AR->getLoop() == L)
2242 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2243 /// operands, so wrap it in a convenient helper.
2244 static Value *getWideOperand(Value *Oper) {
2245 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2246 return Trunc->getOperand(0);
2250 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2252 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2253 Type *LType = LVal->getType();
2254 Type *RType = RVal->getType();
2255 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2258 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2259 /// NULL for any constant. Returning the expression itself is
2260 /// conservative. Returning a deeper subexpression is more precise and valid as
2261 /// long as it isn't less complex than another subexpression. For expressions
2262 /// involving multiple unscaled values, we need to return the pointer-type
2263 /// SCEVUnknown. This avoids forming chains across objects, such as:
2264 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2266 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2267 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2268 static const SCEV *getExprBase(const SCEV *S) {
2269 switch (S->getSCEVType()) {
2270 default: // uncluding scUnknown.
2275 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2277 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2279 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2281 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2282 // there's nothing more complex.
2283 // FIXME: not sure if we want to recognize negation.
2284 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2285 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2286 E(Add->op_begin()); I != E; ++I) {
2287 const SCEV *SubExpr = *I;
2288 if (SubExpr->getSCEVType() == scAddExpr)
2289 return getExprBase(SubExpr);
2291 if (SubExpr->getSCEVType() != scMulExpr)
2294 return S; // all operands are scaled, be conservative.
2297 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2301 /// Return true if the chain increment is profitable to expand into a loop
2302 /// invariant value, which may require its own register. A profitable chain
2303 /// increment will be an offset relative to the same base. We allow such offsets
2304 /// to potentially be used as chain increment as long as it's not obviously
2305 /// expensive to expand using real instructions.
2307 getProfitableChainIncrement(Value *NextIV, Value *PrevIV,
2308 const IVChain &Chain, Loop *L,
2309 ScalarEvolution &SE, const TargetLowering *TLI) {
2310 // Prune the solution space aggressively by checking that both IV operands
2311 // are expressions that operate on the same unscaled SCEVUnknown. This
2312 // "base" will be canceled by the subsequent getMinusSCEV call. Checking first
2313 // avoids creating extra SCEV expressions.
2314 const SCEV *OperExpr = SE.getSCEV(NextIV);
2315 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2316 if (getExprBase(OperExpr) != getExprBase(PrevExpr) && !StressIVChain)
2319 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2320 if (!SE.isLoopInvariant(IncExpr, L))
2323 // We are not able to expand an increment unless it is loop invariant,
2324 // however, the following checks are purely for profitability.
2328 // Do not replace a constant offset from IV head with a nonconstant IV
2330 if (!isa<SCEVConstant>(IncExpr)) {
2331 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Chain[0].IVOperand));
2332 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2336 SmallPtrSet<const SCEV*, 8> Processed;
2337 if (isHighCostExpansion(IncExpr, Processed, SE))
2343 /// Return true if the number of registers needed for the chain is estimated to
2344 /// be less than the number required for the individual IV users. First prohibit
2345 /// any IV users that keep the IV live across increments (the Users set should
2346 /// be empty). Next count the number and type of increments in the chain.
2348 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2349 /// effectively use postinc addressing modes. Only consider it profitable it the
2350 /// increments can be computed in fewer registers when chained.
2352 /// TODO: Consider IVInc free if it's already used in another chains.
2354 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2355 ScalarEvolution &SE, const TargetLowering *TLI) {
2359 if (Chain.size() <= 2)
2362 if (!Users.empty()) {
2363 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " users:\n";
2364 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2365 E = Users.end(); I != E; ++I) {
2366 dbgs() << " " << **I << "\n";
2370 assert(!Chain.empty() && "empty IV chains are not allowed");
2372 // The chain itself may require a register, so intialize cost to 1.
2375 // A complete chain likely eliminates the need for keeping the original IV in
2376 // a register. LSR does not currently know how to form a complete chain unless
2377 // the header phi already exists.
2378 if (isa<PHINode>(Chain.back().UserInst)
2379 && SE.getSCEV(Chain.back().UserInst) == Chain[0].IncExpr) {
2382 const SCEV *LastIncExpr = 0;
2383 unsigned NumConstIncrements = 0;
2384 unsigned NumVarIncrements = 0;
2385 unsigned NumReusedIncrements = 0;
2386 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2389 if (I->IncExpr->isZero())
2392 // Incrementing by zero or some constant is neutral. We assume constants can
2393 // be folded into an addressing mode or an add's immediate operand.
2394 if (isa<SCEVConstant>(I->IncExpr)) {
2395 ++NumConstIncrements;
2399 if (I->IncExpr == LastIncExpr)
2400 ++NumReusedIncrements;
2404 LastIncExpr = I->IncExpr;
2406 // An IV chain with a single increment is handled by LSR's postinc
2407 // uses. However, a chain with multiple increments requires keeping the IV's
2408 // value live longer than it needs to be if chained.
2409 if (NumConstIncrements > 1)
2412 // Materializing increment expressions in the preheader that didn't exist in
2413 // the original code may cost a register. For example, sign-extended array
2414 // indices can produce ridiculous increments like this:
2415 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2416 cost += NumVarIncrements;
2418 // Reusing variable increments likely saves a register to hold the multiple of
2420 cost -= NumReusedIncrements;
2422 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " Cost: " << cost << "\n");
2427 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2429 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2430 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2431 // When IVs are used as types of varying widths, they are generally converted
2432 // to a wider type with some uses remaining narrow under a (free) trunc.
2433 Value *NextIV = getWideOperand(IVOper);
2435 // Visit all existing chains. Check if its IVOper can be computed as a
2436 // profitable loop invariant increment from the last link in the Chain.
2437 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2438 const SCEV *LastIncExpr = 0;
2439 for (; ChainIdx < NChains; ++ChainIdx) {
2440 Value *PrevIV = getWideOperand(IVChainVec[ChainIdx].back().IVOperand);
2441 if (!isCompatibleIVType(PrevIV, NextIV))
2444 // A phi node terminates a chain.
2445 if (isa<PHINode>(UserInst)
2446 && isa<PHINode>(IVChainVec[ChainIdx].back().UserInst))
2449 if (const SCEV *IncExpr =
2450 getProfitableChainIncrement(NextIV, PrevIV, IVChainVec[ChainIdx],
2452 LastIncExpr = IncExpr;
2456 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2457 // bother for phi nodes, because they must be last in the chain.
2458 if (ChainIdx == NChains) {
2459 if (isa<PHINode>(UserInst))
2461 if (NChains >= MaxChains && !StressIVChain) {
2462 DEBUG(dbgs() << "IV Chain Limit\n");
2465 LastIncExpr = SE.getSCEV(NextIV);
2466 // IVUsers may have skipped over sign/zero extensions. We don't currently
2467 // attempt to form chains involving extensions unless they can be hoisted
2468 // into this loop's AddRec.
2469 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2472 IVChainVec.resize(NChains);
2473 ChainUsersVec.resize(NChains);
2474 DEBUG(dbgs() << "IV Head: (" << *UserInst << ") IV=" << *LastIncExpr
2478 DEBUG(dbgs() << "IV Inc: (" << *UserInst << ") IV+" << *LastIncExpr
2481 // Add this IV user to the end of the chain.
2482 IVChainVec[ChainIdx].push_back(IVInc(UserInst, IVOper, LastIncExpr));
2484 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2485 // This chain's NearUsers become FarUsers.
2486 if (!LastIncExpr->isZero()) {
2487 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2492 // All other uses of IVOperand become near uses of the chain.
2493 // We currently ignore intermediate values within SCEV expressions, assuming
2494 // they will eventually be used be the current chain, or can be computed
2495 // from one of the chain increments. To be more precise we could
2496 // transitively follow its user and only add leaf IV users to the set.
2497 for (Value::use_iterator UseIter = IVOper->use_begin(),
2498 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2499 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2500 if (!OtherUse || OtherUse == UserInst)
2502 if (SE.isSCEVable(OtherUse->getType())
2503 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2504 && IU.isIVUserOrOperand(OtherUse)) {
2507 NearUsers.insert(OtherUse);
2510 // Since this user is part of the chain, it's no longer considered a use
2512 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2515 /// CollectChains - Populate the vector of Chains.
2517 /// This decreases ILP at the architecture level. Targets with ample registers,
2518 /// multiple memory ports, and no register renaming probably don't want
2519 /// this. However, such targets should probably disable LSR altogether.
2521 /// The job of LSR is to make a reasonable choice of induction variables across
2522 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2523 /// ILP *within the loop* if the target wants it.
2525 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2526 /// will not reorder memory operations, it will recognize this as a chain, but
2527 /// will generate redundant IV increments. Ideally this would be corrected later
2528 /// by a smart scheduler:
2534 /// TODO: Walk the entire domtree within this loop, not just the path to the
2535 /// loop latch. This will discover chains on side paths, but requires
2536 /// maintaining multiple copies of the Chains state.
2537 void LSRInstance::CollectChains() {
2538 SmallVector<ChainUsers, 8> ChainUsersVec;
2540 SmallVector<BasicBlock *,8> LatchPath;
2541 BasicBlock *LoopHeader = L->getHeader();
2542 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2543 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2544 LatchPath.push_back(Rung->getBlock());
2546 LatchPath.push_back(LoopHeader);
2548 // Walk the instruction stream from the loop header to the loop latch.
2549 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2550 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2551 BBIter != BBEnd; ++BBIter) {
2552 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2554 // Skip instructions that weren't seen by IVUsers analysis.
2555 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2558 // Ignore users that are part of a SCEV expression. This way we only
2559 // consider leaf IV Users. This effectively rediscovers a portion of
2560 // IVUsers analysis but in program order this time.
2561 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2564 // Remove this instruction from any NearUsers set it may be in.
2565 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2566 ChainIdx < NChains; ++ChainIdx) {
2567 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2569 // Search for operands that can be chained.
2570 SmallPtrSet<Instruction*, 4> UniqueOperands;
2571 User::op_iterator IVOpEnd = I->op_end();
2572 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2573 while (IVOpIter != IVOpEnd) {
2574 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2575 if (UniqueOperands.insert(IVOpInst))
2576 ChainInstruction(I, IVOpInst, ChainUsersVec);
2577 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2579 } // Continue walking down the instructions.
2580 } // Continue walking down the domtree.
2581 // Visit phi backedges to determine if the chain can generate the IV postinc.
2582 for (BasicBlock::iterator I = L->getHeader()->begin();
2583 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2584 if (!SE.isSCEVable(PN->getType()))
2588 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2590 ChainInstruction(PN, IncV, ChainUsersVec);
2592 // Remove any unprofitable chains.
2593 unsigned ChainIdx = 0;
2594 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2595 UsersIdx < NChains; ++UsersIdx) {
2596 if (!isProfitableChain(IVChainVec[UsersIdx],
2597 ChainUsersVec[UsersIdx].FarUsers, SE, TLI))
2599 // Preserve the chain at UsesIdx.
2600 if (ChainIdx != UsersIdx)
2601 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2602 FinalizeChain(IVChainVec[ChainIdx]);
2605 IVChainVec.resize(ChainIdx);
2608 void LSRInstance::FinalizeChain(IVChain &Chain) {
2609 assert(!Chain.empty() && "empty IV chains are not allowed");
2610 DEBUG(dbgs() << "Final Chain: " << *Chain[0].UserInst << "\n");
2612 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2614 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2615 User::op_iterator UseI =
2616 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2617 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2618 IVIncSet.insert(UseI);
2622 /// Return true if the IVInc can be folded into an addressing mode.
2623 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2624 Value *Operand, const TargetLowering *TLI) {
2625 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2626 if (!IncConst || !isAddressUse(UserInst, Operand))
2629 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2632 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2633 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false,
2634 LSRUse::Address, getAccessType(UserInst), TLI))
2640 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2641 /// materialize the IV user's operand from the previous IV user's operand.
2642 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2643 SmallVectorImpl<WeakVH> &DeadInsts) {
2644 // Find the new IVOperand for the head of the chain. It may have been replaced
2646 const IVInc &Head = Chain[0];
2647 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2648 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2651 while (IVOpIter != IVOpEnd) {
2652 IVSrc = getWideOperand(*IVOpIter);
2654 // If this operand computes the expression that the chain needs, we may use
2655 // it. (Check this after setting IVSrc which is used below.)
2657 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2658 // narrow for the chain, so we can no longer use it. We do allow using a
2659 // wider phi, assuming the LSR checked for free truncation. In that case we
2660 // should already have a truncate on this operand such that
2661 // getSCEV(IVSrc) == IncExpr.
2662 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2663 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2666 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2668 if (IVOpIter == IVOpEnd) {
2669 // Gracefully give up on this chain.
2670 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2674 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2675 Type *IVTy = IVSrc->getType();
2676 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2677 const SCEV *LeftOverExpr = 0;
2678 for (IVChain::const_iterator IncI = llvm::next(Chain.begin()),
2679 IncE = Chain.end(); IncI != IncE; ++IncI) {
2681 Instruction *InsertPt = IncI->UserInst;
2682 if (isa<PHINode>(InsertPt))
2683 InsertPt = L->getLoopLatch()->getTerminator();
2685 // IVOper will replace the current IV User's operand. IVSrc is the IV
2686 // value currently held in a register.
2687 Value *IVOper = IVSrc;
2688 if (!IncI->IncExpr->isZero()) {
2689 // IncExpr was the result of subtraction of two narrow values, so must
2691 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2692 LeftOverExpr = LeftOverExpr ?
2693 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2695 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2696 // Expand the IV increment.
2697 Rewriter.clearPostInc();
2698 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2699 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2700 SE.getUnknown(IncV));
2701 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2703 // If an IV increment can't be folded, use it as the next IV value.
2704 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2706 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2711 Type *OperTy = IncI->IVOperand->getType();
2712 if (IVTy != OperTy) {
2713 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2714 "cannot extend a chained IV");
2715 IRBuilder<> Builder(InsertPt);
2716 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2718 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2719 DeadInsts.push_back(IncI->IVOperand);
2721 // If LSR created a new, wider phi, we may also replace its postinc. We only
2722 // do this if we also found a wide value for the head of the chain.
2723 if (isa<PHINode>(Chain.back().UserInst)) {
2724 for (BasicBlock::iterator I = L->getHeader()->begin();
2725 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2726 if (!isCompatibleIVType(Phi, IVSrc))
2728 Instruction *PostIncV = dyn_cast<Instruction>(
2729 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2730 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2732 Value *IVOper = IVSrc;
2733 Type *PostIncTy = PostIncV->getType();
2734 if (IVTy != PostIncTy) {
2735 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2736 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2737 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2738 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2740 Phi->replaceUsesOfWith(PostIncV, IVOper);
2741 DeadInsts.push_back(PostIncV);
2746 void LSRInstance::CollectFixupsAndInitialFormulae() {
2747 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2748 Instruction *UserInst = UI->getUser();
2749 // Skip IV users that are part of profitable IV Chains.
2750 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2751 UI->getOperandValToReplace());
2752 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2753 if (IVIncSet.count(UseI))
2757 LSRFixup &LF = getNewFixup();
2758 LF.UserInst = UserInst;
2759 LF.OperandValToReplace = UI->getOperandValToReplace();
2760 LF.PostIncLoops = UI->getPostIncLoops();
2762 LSRUse::KindType Kind = LSRUse::Basic;
2764 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2765 Kind = LSRUse::Address;
2766 AccessTy = getAccessType(LF.UserInst);
2769 const SCEV *S = IU.getExpr(*UI);
2771 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2772 // (N - i == 0), and this allows (N - i) to be the expression that we work
2773 // with rather than just N or i, so we can consider the register
2774 // requirements for both N and i at the same time. Limiting this code to
2775 // equality icmps is not a problem because all interesting loops use
2776 // equality icmps, thanks to IndVarSimplify.
2777 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2778 if (CI->isEquality()) {
2779 // Swap the operands if needed to put the OperandValToReplace on the
2780 // left, for consistency.
2781 Value *NV = CI->getOperand(1);
2782 if (NV == LF.OperandValToReplace) {
2783 CI->setOperand(1, CI->getOperand(0));
2784 CI->setOperand(0, NV);
2785 NV = CI->getOperand(1);
2789 // x == y --> x - y == 0
2790 const SCEV *N = SE.getSCEV(NV);
2791 if (SE.isLoopInvariant(N, L)) {
2792 // S is normalized, so normalize N before folding it into S
2793 // to keep the result normalized.
2794 N = TransformForPostIncUse(Normalize, N, CI, 0,
2795 LF.PostIncLoops, SE, DT);
2796 Kind = LSRUse::ICmpZero;
2797 S = SE.getMinusSCEV(N, S);
2800 // -1 and the negations of all interesting strides (except the negation
2801 // of -1) are now also interesting.
2802 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2803 if (Factors[i] != -1)
2804 Factors.insert(-(uint64_t)Factors[i]);
2808 // Set up the initial formula for this use.
2809 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2811 LF.Offset = P.second;
2812 LSRUse &LU = Uses[LF.LUIdx];
2813 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2814 if (!LU.WidestFixupType ||
2815 SE.getTypeSizeInBits(LU.WidestFixupType) <
2816 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2817 LU.WidestFixupType = LF.OperandValToReplace->getType();
2819 // If this is the first use of this LSRUse, give it a formula.
2820 if (LU.Formulae.empty()) {
2821 InsertInitialFormula(S, LU, LF.LUIdx);
2822 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2826 DEBUG(print_fixups(dbgs()));
2829 /// InsertInitialFormula - Insert a formula for the given expression into
2830 /// the given use, separating out loop-variant portions from loop-invariant
2831 /// and loop-computable portions.
2833 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2835 F.InitialMatch(S, L, SE);
2836 bool Inserted = InsertFormula(LU, LUIdx, F);
2837 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2840 /// InsertSupplementalFormula - Insert a simple single-register formula for
2841 /// the given expression into the given use.
2843 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2844 LSRUse &LU, size_t LUIdx) {
2846 F.BaseRegs.push_back(S);
2847 F.AM.HasBaseReg = true;
2848 bool Inserted = InsertFormula(LU, LUIdx, F);
2849 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2852 /// CountRegisters - Note which registers are used by the given formula,
2853 /// updating RegUses.
2854 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2856 RegUses.CountRegister(F.ScaledReg, LUIdx);
2857 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2858 E = F.BaseRegs.end(); I != E; ++I)
2859 RegUses.CountRegister(*I, LUIdx);
2862 /// InsertFormula - If the given formula has not yet been inserted, add it to
2863 /// the list, and return true. Return false otherwise.
2864 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2865 if (!LU.InsertFormula(F))
2868 CountRegisters(F, LUIdx);
2872 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2873 /// loop-invariant values which we're tracking. These other uses will pin these
2874 /// values in registers, making them less profitable for elimination.
2875 /// TODO: This currently misses non-constant addrec step registers.
2876 /// TODO: Should this give more weight to users inside the loop?
2878 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2879 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2880 SmallPtrSet<const SCEV *, 8> Inserted;
2882 while (!Worklist.empty()) {
2883 const SCEV *S = Worklist.pop_back_val();
2885 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2886 Worklist.append(N->op_begin(), N->op_end());
2887 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2888 Worklist.push_back(C->getOperand());
2889 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2890 Worklist.push_back(D->getLHS());
2891 Worklist.push_back(D->getRHS());
2892 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2893 if (!Inserted.insert(U)) continue;
2894 const Value *V = U->getValue();
2895 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2896 // Look for instructions defined outside the loop.
2897 if (L->contains(Inst)) continue;
2898 } else if (isa<UndefValue>(V))
2899 // Undef doesn't have a live range, so it doesn't matter.
2901 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2903 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2904 // Ignore non-instructions.
2907 // Ignore instructions in other functions (as can happen with
2909 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2911 // Ignore instructions not dominated by the loop.
2912 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2913 UserInst->getParent() :
2914 cast<PHINode>(UserInst)->getIncomingBlock(
2915 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2916 if (!DT.dominates(L->getHeader(), UseBB))
2918 // Ignore uses which are part of other SCEV expressions, to avoid
2919 // analyzing them multiple times.
2920 if (SE.isSCEVable(UserInst->getType())) {
2921 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2922 // If the user is a no-op, look through to its uses.
2923 if (!isa<SCEVUnknown>(UserS))
2927 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2931 // Ignore icmp instructions which are already being analyzed.
2932 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2933 unsigned OtherIdx = !UI.getOperandNo();
2934 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2935 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2939 LSRFixup &LF = getNewFixup();
2940 LF.UserInst = const_cast<Instruction *>(UserInst);
2941 LF.OperandValToReplace = UI.getUse();
2942 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2944 LF.Offset = P.second;
2945 LSRUse &LU = Uses[LF.LUIdx];
2946 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2947 if (!LU.WidestFixupType ||
2948 SE.getTypeSizeInBits(LU.WidestFixupType) <
2949 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2950 LU.WidestFixupType = LF.OperandValToReplace->getType();
2951 InsertSupplementalFormula(U, LU, LF.LUIdx);
2952 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2959 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2960 /// separate registers. If C is non-null, multiply each subexpression by C.
2961 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2962 SmallVectorImpl<const SCEV *> &Ops,
2964 ScalarEvolution &SE) {
2965 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2966 // Break out add operands.
2967 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2969 CollectSubexprs(*I, C, Ops, L, SE);
2971 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2972 // Split a non-zero base out of an addrec.
2973 if (!AR->getStart()->isZero()) {
2974 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2975 AR->getStepRecurrence(SE),
2977 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2980 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2983 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2984 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2985 if (Mul->getNumOperands() == 2)
2986 if (const SCEVConstant *Op0 =
2987 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2988 CollectSubexprs(Mul->getOperand(1),
2989 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2995 // Otherwise use the value itself, optionally with a scale applied.
2996 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2999 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3001 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3004 // Arbitrarily cap recursion to protect compile time.
3005 if (Depth >= 3) return;
3007 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3008 const SCEV *BaseReg = Base.BaseRegs[i];
3010 SmallVector<const SCEV *, 8> AddOps;
3011 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3013 if (AddOps.size() == 1) continue;
3015 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3016 JE = AddOps.end(); J != JE; ++J) {
3018 // Loop-variant "unknown" values are uninteresting; we won't be able to
3019 // do anything meaningful with them.
3020 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3023 // Don't pull a constant into a register if the constant could be folded
3024 // into an immediate field.
3025 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
3026 Base.getNumRegs() > 1,
3027 LU.Kind, LU.AccessTy, TLI, SE))
3030 // Collect all operands except *J.
3031 SmallVector<const SCEV *, 8> InnerAddOps
3032 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3034 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3036 // Don't leave just a constant behind in a register if the constant could
3037 // be folded into an immediate field.
3038 if (InnerAddOps.size() == 1 &&
3039 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
3040 Base.getNumRegs() > 1,
3041 LU.Kind, LU.AccessTy, TLI, SE))
3044 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3045 if (InnerSum->isZero())
3049 // Add the remaining pieces of the add back into the new formula.
3050 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3051 if (TLI && InnerSumSC &&
3052 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3053 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3054 InnerSumSC->getValue()->getZExtValue())) {
3055 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3056 InnerSumSC->getValue()->getZExtValue();
3057 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3059 F.BaseRegs[i] = InnerSum;
3061 // Add J as its own register, or an unfolded immediate.
3062 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3063 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3064 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3065 SC->getValue()->getZExtValue()))
3066 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3067 SC->getValue()->getZExtValue();
3069 F.BaseRegs.push_back(*J);
3071 if (InsertFormula(LU, LUIdx, F))
3072 // If that formula hadn't been seen before, recurse to find more like
3074 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3079 /// GenerateCombinations - Generate a formula consisting of all of the
3080 /// loop-dominating registers added into a single register.
3081 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3083 // This method is only interesting on a plurality of registers.
3084 if (Base.BaseRegs.size() <= 1) return;
3088 SmallVector<const SCEV *, 4> Ops;
3089 for (SmallVectorImpl<const SCEV *>::const_iterator
3090 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3091 const SCEV *BaseReg = *I;
3092 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3093 !SE.hasComputableLoopEvolution(BaseReg, L))
3094 Ops.push_back(BaseReg);
3096 F.BaseRegs.push_back(BaseReg);
3098 if (Ops.size() > 1) {
3099 const SCEV *Sum = SE.getAddExpr(Ops);
3100 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3101 // opportunity to fold something. For now, just ignore such cases
3102 // rather than proceed with zero in a register.
3103 if (!Sum->isZero()) {
3104 F.BaseRegs.push_back(Sum);
3105 (void)InsertFormula(LU, LUIdx, F);
3110 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3111 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3113 // We can't add a symbolic offset if the address already contains one.
3114 if (Base.AM.BaseGV) return;
3116 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3117 const SCEV *G = Base.BaseRegs[i];
3118 GlobalValue *GV = ExtractSymbol(G, SE);
3119 if (G->isZero() || !GV)
3123 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3124 LU.Kind, LU.AccessTy, TLI))
3127 (void)InsertFormula(LU, LUIdx, F);
3131 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3132 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3134 // TODO: For now, just add the min and max offset, because it usually isn't
3135 // worthwhile looking at everything inbetween.
3136 SmallVector<int64_t, 2> Worklist;
3137 Worklist.push_back(LU.MinOffset);
3138 if (LU.MaxOffset != LU.MinOffset)
3139 Worklist.push_back(LU.MaxOffset);
3141 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3142 const SCEV *G = Base.BaseRegs[i];
3144 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3145 E = Worklist.end(); I != E; ++I) {
3147 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
3148 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
3149 LU.Kind, LU.AccessTy, TLI)) {
3150 // Add the offset to the base register.
3151 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3152 // If it cancelled out, drop the base register, otherwise update it.
3153 if (NewG->isZero()) {
3154 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3155 F.BaseRegs.pop_back();
3157 F.BaseRegs[i] = NewG;
3159 (void)InsertFormula(LU, LUIdx, F);
3163 int64_t Imm = ExtractImmediate(G, SE);
3164 if (G->isZero() || Imm == 0)
3167 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
3168 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3169 LU.Kind, LU.AccessTy, TLI))
3172 (void)InsertFormula(LU, LUIdx, F);
3176 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3177 /// the comparison. For example, x == y -> x*c == y*c.
3178 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3180 if (LU.Kind != LSRUse::ICmpZero) return;
3182 // Determine the integer type for the base formula.
3183 Type *IntTy = Base.getType();
3185 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3187 // Don't do this if there is more than one offset.
3188 if (LU.MinOffset != LU.MaxOffset) return;
3190 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
3192 // Check each interesting stride.
3193 for (SmallSetVector<int64_t, 8>::const_iterator
3194 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3195 int64_t Factor = *I;
3197 // Check that the multiplication doesn't overflow.
3198 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
3200 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
3201 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
3204 // Check that multiplying with the use offset doesn't overflow.
3205 int64_t Offset = LU.MinOffset;
3206 if (Offset == INT64_MIN && Factor == -1)
3208 Offset = (uint64_t)Offset * Factor;
3209 if (Offset / Factor != LU.MinOffset)
3213 F.AM.BaseOffs = NewBaseOffs;
3215 // Check that this scale is legal.
3216 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
3219 // Compensate for the use having MinOffset built into it.
3220 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
3222 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3224 // Check that multiplying with each base register doesn't overflow.
3225 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3226 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3227 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3231 // Check that multiplying with the scaled register doesn't overflow.
3233 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3234 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3238 // Check that multiplying with the unfolded offset doesn't overflow.
3239 if (F.UnfoldedOffset != 0) {
3240 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3242 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3243 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3247 // If we make it here and it's legal, add it.
3248 (void)InsertFormula(LU, LUIdx, F);
3253 /// GenerateScales - Generate stride factor reuse formulae by making use of
3254 /// scaled-offset address modes, for example.
3255 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3256 // Determine the integer type for the base formula.
3257 Type *IntTy = Base.getType();
3260 // If this Formula already has a scaled register, we can't add another one.
3261 if (Base.AM.Scale != 0) return;
3263 // Check each interesting stride.
3264 for (SmallSetVector<int64_t, 8>::const_iterator
3265 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3266 int64_t Factor = *I;
3268 Base.AM.Scale = Factor;
3269 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
3270 // Check whether this scale is going to be legal.
3271 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3272 LU.Kind, LU.AccessTy, TLI)) {
3273 // As a special-case, handle special out-of-loop Basic users specially.
3274 // TODO: Reconsider this special case.
3275 if (LU.Kind == LSRUse::Basic &&
3276 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3277 LSRUse::Special, LU.AccessTy, TLI) &&
3278 LU.AllFixupsOutsideLoop)
3279 LU.Kind = LSRUse::Special;
3283 // For an ICmpZero, negating a solitary base register won't lead to
3285 if (LU.Kind == LSRUse::ICmpZero &&
3286 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
3288 // For each addrec base reg, apply the scale, if possible.
3289 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3290 if (const SCEVAddRecExpr *AR =
3291 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3292 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3293 if (FactorS->isZero())
3295 // Divide out the factor, ignoring high bits, since we'll be
3296 // scaling the value back up in the end.
3297 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3298 // TODO: This could be optimized to avoid all the copying.
3300 F.ScaledReg = Quotient;
3301 F.DeleteBaseReg(F.BaseRegs[i]);
3302 (void)InsertFormula(LU, LUIdx, F);
3308 /// GenerateTruncates - Generate reuse formulae from different IV types.
3309 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3310 // This requires TargetLowering to tell us which truncates are free.
3313 // Don't bother truncating symbolic values.
3314 if (Base.AM.BaseGV) return;
3316 // Determine the integer type for the base formula.
3317 Type *DstTy = Base.getType();
3319 DstTy = SE.getEffectiveSCEVType(DstTy);
3321 for (SmallSetVector<Type *, 4>::const_iterator
3322 I = Types.begin(), E = Types.end(); I != E; ++I) {
3324 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
3327 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3328 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3329 JE = F.BaseRegs.end(); J != JE; ++J)
3330 *J = SE.getAnyExtendExpr(*J, SrcTy);
3332 // TODO: This assumes we've done basic processing on all uses and
3333 // have an idea what the register usage is.
3334 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3337 (void)InsertFormula(LU, LUIdx, F);
3344 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3345 /// defer modifications so that the search phase doesn't have to worry about
3346 /// the data structures moving underneath it.
3350 const SCEV *OrigReg;
3352 WorkItem(size_t LI, int64_t I, const SCEV *R)
3353 : LUIdx(LI), Imm(I), OrigReg(R) {}
3355 void print(raw_ostream &OS) const;
3361 void WorkItem::print(raw_ostream &OS) const {
3362 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3363 << " , add offset " << Imm;
3366 void WorkItem::dump() const {
3367 print(errs()); errs() << '\n';
3370 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3371 /// distance apart and try to form reuse opportunities between them.
3372 void LSRInstance::GenerateCrossUseConstantOffsets() {
3373 // Group the registers by their value without any added constant offset.
3374 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3375 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3377 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3378 SmallVector<const SCEV *, 8> Sequence;
3379 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3381 const SCEV *Reg = *I;
3382 int64_t Imm = ExtractImmediate(Reg, SE);
3383 std::pair<RegMapTy::iterator, bool> Pair =
3384 Map.insert(std::make_pair(Reg, ImmMapTy()));
3386 Sequence.push_back(Reg);
3387 Pair.first->second.insert(std::make_pair(Imm, *I));
3388 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3391 // Now examine each set of registers with the same base value. Build up
3392 // a list of work to do and do the work in a separate step so that we're
3393 // not adding formulae and register counts while we're searching.
3394 SmallVector<WorkItem, 32> WorkItems;
3395 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3396 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3397 E = Sequence.end(); I != E; ++I) {
3398 const SCEV *Reg = *I;
3399 const ImmMapTy &Imms = Map.find(Reg)->second;
3401 // It's not worthwhile looking for reuse if there's only one offset.
3402 if (Imms.size() == 1)
3405 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3406 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3408 dbgs() << ' ' << J->first;
3411 // Examine each offset.
3412 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3414 const SCEV *OrigReg = J->second;
3416 int64_t JImm = J->first;
3417 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3419 if (!isa<SCEVConstant>(OrigReg) &&
3420 UsedByIndicesMap[Reg].count() == 1) {
3421 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3425 // Conservatively examine offsets between this orig reg a few selected
3427 ImmMapTy::const_iterator OtherImms[] = {
3428 Imms.begin(), prior(Imms.end()),
3429 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3431 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3432 ImmMapTy::const_iterator M = OtherImms[i];
3433 if (M == J || M == JE) continue;
3435 // Compute the difference between the two.
3436 int64_t Imm = (uint64_t)JImm - M->first;
3437 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3438 LUIdx = UsedByIndices.find_next(LUIdx))
3439 // Make a memo of this use, offset, and register tuple.
3440 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3441 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3448 UsedByIndicesMap.clear();
3449 UniqueItems.clear();
3451 // Now iterate through the worklist and add new formulae.
3452 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3453 E = WorkItems.end(); I != E; ++I) {
3454 const WorkItem &WI = *I;
3455 size_t LUIdx = WI.LUIdx;
3456 LSRUse &LU = Uses[LUIdx];
3457 int64_t Imm = WI.Imm;
3458 const SCEV *OrigReg = WI.OrigReg;
3460 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3461 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3462 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3464 // TODO: Use a more targeted data structure.
3465 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3466 const Formula &F = LU.Formulae[L];
3467 // Use the immediate in the scaled register.
3468 if (F.ScaledReg == OrigReg) {
3469 int64_t Offs = (uint64_t)F.AM.BaseOffs +
3470 Imm * (uint64_t)F.AM.Scale;
3471 // Don't create 50 + reg(-50).
3472 if (F.referencesReg(SE.getSCEV(
3473 ConstantInt::get(IntTy, -(uint64_t)Offs))))
3476 NewF.AM.BaseOffs = Offs;
3477 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3478 LU.Kind, LU.AccessTy, TLI))
3480 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3482 // If the new scale is a constant in a register, and adding the constant
3483 // value to the immediate would produce a value closer to zero than the
3484 // immediate itself, then the formula isn't worthwhile.
3485 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3486 if (C->getValue()->isNegative() !=
3487 (NewF.AM.BaseOffs < 0) &&
3488 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
3489 .ule(abs64(NewF.AM.BaseOffs)))
3493 (void)InsertFormula(LU, LUIdx, NewF);
3495 // Use the immediate in a base register.
3496 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3497 const SCEV *BaseReg = F.BaseRegs[N];
3498 if (BaseReg != OrigReg)
3501 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
3502 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3503 LU.Kind, LU.AccessTy, TLI)) {
3505 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3508 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3510 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3512 // If the new formula has a constant in a register, and adding the
3513 // constant value to the immediate would produce a value closer to
3514 // zero than the immediate itself, then the formula isn't worthwhile.
3515 for (SmallVectorImpl<const SCEV *>::const_iterator
3516 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3518 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3519 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
3520 abs64(NewF.AM.BaseOffs)) &&
3521 (C->getValue()->getValue() +
3522 NewF.AM.BaseOffs).countTrailingZeros() >=
3523 CountTrailingZeros_64(NewF.AM.BaseOffs))
3527 (void)InsertFormula(LU, LUIdx, NewF);
3536 /// GenerateAllReuseFormulae - Generate formulae for each use.
3538 LSRInstance::GenerateAllReuseFormulae() {
3539 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3540 // queries are more precise.
3541 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3542 LSRUse &LU = Uses[LUIdx];
3543 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3544 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3545 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3546 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3548 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3549 LSRUse &LU = Uses[LUIdx];
3550 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3551 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3552 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3553 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3554 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3555 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3556 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3557 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3559 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3560 LSRUse &LU = Uses[LUIdx];
3561 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3562 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3565 GenerateCrossUseConstantOffsets();
3567 DEBUG(dbgs() << "\n"
3568 "After generating reuse formulae:\n";
3569 print_uses(dbgs()));
3572 /// If there are multiple formulae with the same set of registers used
3573 /// by other uses, pick the best one and delete the others.
3574 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3575 DenseSet<const SCEV *> VisitedRegs;
3576 SmallPtrSet<const SCEV *, 16> Regs;
3577 SmallPtrSet<const SCEV *, 16> LoserRegs;
3579 bool ChangedFormulae = false;
3582 // Collect the best formula for each unique set of shared registers. This
3583 // is reset for each use.
3584 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
3586 BestFormulaeTy BestFormulae;
3588 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3589 LSRUse &LU = Uses[LUIdx];
3590 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3593 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3594 FIdx != NumForms; ++FIdx) {
3595 Formula &F = LU.Formulae[FIdx];
3597 // Some formulas are instant losers. For example, they may depend on
3598 // nonexistent AddRecs from other loops. These need to be filtered
3599 // immediately, otherwise heuristics could choose them over others leading
3600 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3601 // avoids the need to recompute this information across formulae using the
3602 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3603 // the corresponding bad register from the Regs set.
3606 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3608 if (CostF.isLoser()) {
3609 // During initial formula generation, undesirable formulae are generated
3610 // by uses within other loops that have some non-trivial address mode or
3611 // use the postinc form of the IV. LSR needs to provide these formulae
3612 // as the basis of rediscovering the desired formula that uses an AddRec
3613 // corresponding to the existing phi. Once all formulae have been
3614 // generated, these initial losers may be pruned.
3615 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3619 SmallVector<const SCEV *, 2> Key;
3620 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3621 JE = F.BaseRegs.end(); J != JE; ++J) {
3622 const SCEV *Reg = *J;
3623 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3627 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3628 Key.push_back(F.ScaledReg);
3629 // Unstable sort by host order ok, because this is only used for
3631 std::sort(Key.begin(), Key.end());
3633 std::pair<BestFormulaeTy::const_iterator, bool> P =
3634 BestFormulae.insert(std::make_pair(Key, FIdx));
3638 Formula &Best = LU.Formulae[P.first->second];
3642 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3643 if (CostF < CostBest)
3645 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3647 " in favor of formula "; Best.print(dbgs());
3651 ChangedFormulae = true;
3653 LU.DeleteFormula(F);
3659 // Now that we've filtered out some formulae, recompute the Regs set.
3661 LU.RecomputeRegs(LUIdx, RegUses);
3663 // Reset this to prepare for the next use.
3664 BestFormulae.clear();
3667 DEBUG(if (ChangedFormulae) {
3669 "After filtering out undesirable candidates:\n";
3674 // This is a rough guess that seems to work fairly well.
3675 static const size_t ComplexityLimit = UINT16_MAX;
3677 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3678 /// solutions the solver might have to consider. It almost never considers
3679 /// this many solutions because it prune the search space, but the pruning
3680 /// isn't always sufficient.
3681 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3683 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3684 E = Uses.end(); I != E; ++I) {
3685 size_t FSize = I->Formulae.size();
3686 if (FSize >= ComplexityLimit) {
3687 Power = ComplexityLimit;
3691 if (Power >= ComplexityLimit)
3697 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3698 /// of the registers of another formula, it won't help reduce register
3699 /// pressure (though it may not necessarily hurt register pressure); remove
3700 /// it to simplify the system.
3701 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3702 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3703 DEBUG(dbgs() << "The search space is too complex.\n");
3705 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3706 "which use a superset of registers used by other "
3709 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3710 LSRUse &LU = Uses[LUIdx];
3712 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3713 Formula &F = LU.Formulae[i];
3714 // Look for a formula with a constant or GV in a register. If the use
3715 // also has a formula with that same value in an immediate field,
3716 // delete the one that uses a register.
3717 for (SmallVectorImpl<const SCEV *>::const_iterator
3718 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3719 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3721 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3722 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3723 (I - F.BaseRegs.begin()));
3724 if (LU.HasFormulaWithSameRegs(NewF)) {
3725 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3726 LU.DeleteFormula(F);
3732 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3733 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3736 NewF.AM.BaseGV = GV;
3737 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3738 (I - F.BaseRegs.begin()));
3739 if (LU.HasFormulaWithSameRegs(NewF)) {
3740 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3742 LU.DeleteFormula(F);
3753 LU.RecomputeRegs(LUIdx, RegUses);
3756 DEBUG(dbgs() << "After pre-selection:\n";
3757 print_uses(dbgs()));
3761 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3762 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3764 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3765 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3766 DEBUG(dbgs() << "The search space is too complex.\n");
3768 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3769 "separated by a constant offset will use the same "
3772 // This is especially useful for unrolled loops.
3774 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3775 LSRUse &LU = Uses[LUIdx];
3776 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3777 E = LU.Formulae.end(); I != E; ++I) {
3778 const Formula &F = *I;
3779 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3780 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3781 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3782 /*HasBaseReg=*/false,
3783 LU.Kind, LU.AccessTy)) {
3784 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3787 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3789 // Update the relocs to reference the new use.
3790 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3791 E = Fixups.end(); I != E; ++I) {
3792 LSRFixup &Fixup = *I;
3793 if (Fixup.LUIdx == LUIdx) {
3794 Fixup.LUIdx = LUThatHas - &Uses.front();
3795 Fixup.Offset += F.AM.BaseOffs;
3796 // Add the new offset to LUThatHas' offset list.
3797 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3798 LUThatHas->Offsets.push_back(Fixup.Offset);
3799 if (Fixup.Offset > LUThatHas->MaxOffset)
3800 LUThatHas->MaxOffset = Fixup.Offset;
3801 if (Fixup.Offset < LUThatHas->MinOffset)
3802 LUThatHas->MinOffset = Fixup.Offset;
3804 DEBUG(dbgs() << "New fixup has offset "
3805 << Fixup.Offset << '\n');
3807 if (Fixup.LUIdx == NumUses-1)
3808 Fixup.LUIdx = LUIdx;
3811 // Delete formulae from the new use which are no longer legal.
3813 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3814 Formula &F = LUThatHas->Formulae[i];
3815 if (!isLegalUse(F.AM,
3816 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3817 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3818 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3820 LUThatHas->DeleteFormula(F);
3827 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3829 // Delete the old use.
3830 DeleteUse(LU, LUIdx);
3840 DEBUG(dbgs() << "After pre-selection:\n";
3841 print_uses(dbgs()));
3845 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3846 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3847 /// we've done more filtering, as it may be able to find more formulae to
3849 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3850 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3851 DEBUG(dbgs() << "The search space is too complex.\n");
3853 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3854 "undesirable dedicated registers.\n");
3856 FilterOutUndesirableDedicatedRegisters();
3858 DEBUG(dbgs() << "After pre-selection:\n";
3859 print_uses(dbgs()));
3863 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3864 /// to be profitable, and then in any use which has any reference to that
3865 /// register, delete all formulae which do not reference that register.
3866 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3867 // With all other options exhausted, loop until the system is simple
3868 // enough to handle.
3869 SmallPtrSet<const SCEV *, 4> Taken;
3870 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3871 // Ok, we have too many of formulae on our hands to conveniently handle.
3872 // Use a rough heuristic to thin out the list.
3873 DEBUG(dbgs() << "The search space is too complex.\n");
3875 // Pick the register which is used by the most LSRUses, which is likely
3876 // to be a good reuse register candidate.
3877 const SCEV *Best = 0;
3878 unsigned BestNum = 0;
3879 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3881 const SCEV *Reg = *I;
3882 if (Taken.count(Reg))
3887 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3888 if (Count > BestNum) {
3895 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3896 << " will yield profitable reuse.\n");
3899 // In any use with formulae which references this register, delete formulae
3900 // which don't reference it.
3901 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3902 LSRUse &LU = Uses[LUIdx];
3903 if (!LU.Regs.count(Best)) continue;
3906 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3907 Formula &F = LU.Formulae[i];
3908 if (!F.referencesReg(Best)) {
3909 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3910 LU.DeleteFormula(F);
3914 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3920 LU.RecomputeRegs(LUIdx, RegUses);
3923 DEBUG(dbgs() << "After pre-selection:\n";
3924 print_uses(dbgs()));
3928 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3929 /// formulae to choose from, use some rough heuristics to prune down the number
3930 /// of formulae. This keeps the main solver from taking an extraordinary amount
3931 /// of time in some worst-case scenarios.
3932 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3933 NarrowSearchSpaceByDetectingSupersets();
3934 NarrowSearchSpaceByCollapsingUnrolledCode();
3935 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3936 NarrowSearchSpaceByPickingWinnerRegs();
3939 /// SolveRecurse - This is the recursive solver.
3940 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3942 SmallVectorImpl<const Formula *> &Workspace,
3943 const Cost &CurCost,
3944 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3945 DenseSet<const SCEV *> &VisitedRegs) const {
3948 // - use more aggressive filtering
3949 // - sort the formula so that the most profitable solutions are found first
3950 // - sort the uses too
3952 // - don't compute a cost, and then compare. compare while computing a cost
3954 // - track register sets with SmallBitVector
3956 const LSRUse &LU = Uses[Workspace.size()];
3958 // If this use references any register that's already a part of the
3959 // in-progress solution, consider it a requirement that a formula must
3960 // reference that register in order to be considered. This prunes out
3961 // unprofitable searching.
3962 SmallSetVector<const SCEV *, 4> ReqRegs;
3963 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3964 E = CurRegs.end(); I != E; ++I)
3965 if (LU.Regs.count(*I))
3968 SmallPtrSet<const SCEV *, 16> NewRegs;
3970 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3971 E = LU.Formulae.end(); I != E; ++I) {
3972 const Formula &F = *I;
3974 // Ignore formulae which do not use any of the required registers.
3975 bool SatisfiedReqReg = true;
3976 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3977 JE = ReqRegs.end(); J != JE; ++J) {
3978 const SCEV *Reg = *J;
3979 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3980 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3982 SatisfiedReqReg = false;
3986 if (!SatisfiedReqReg) {
3987 // If none of the formulae satisfied the required registers, then we could
3988 // clear ReqRegs and try again. Currently, we simply give up in this case.
3992 // Evaluate the cost of the current formula. If it's already worse than
3993 // the current best, prune the search at that point.
3996 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3997 if (NewCost < SolutionCost) {
3998 Workspace.push_back(&F);
3999 if (Workspace.size() != Uses.size()) {
4000 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4001 NewRegs, VisitedRegs);
4002 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4003 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4005 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4006 dbgs() << ".\n Regs:";
4007 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4008 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4009 dbgs() << ' ' << **I;
4012 SolutionCost = NewCost;
4013 Solution = Workspace;
4015 Workspace.pop_back();
4020 /// Solve - Choose one formula from each use. Return the results in the given
4021 /// Solution vector.
4022 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4023 SmallVector<const Formula *, 8> Workspace;
4025 SolutionCost.Loose();
4027 SmallPtrSet<const SCEV *, 16> CurRegs;
4028 DenseSet<const SCEV *> VisitedRegs;
4029 Workspace.reserve(Uses.size());
4031 // SolveRecurse does all the work.
4032 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4033 CurRegs, VisitedRegs);
4034 if (Solution.empty()) {
4035 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4039 // Ok, we've now made all our decisions.
4040 DEBUG(dbgs() << "\n"
4041 "The chosen solution requires "; SolutionCost.print(dbgs());
4043 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4045 Uses[i].print(dbgs());
4048 Solution[i]->print(dbgs());
4052 assert(Solution.size() == Uses.size() && "Malformed solution!");
4055 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4056 /// the dominator tree far as we can go while still being dominated by the
4057 /// input positions. This helps canonicalize the insert position, which
4058 /// encourages sharing.
4059 BasicBlock::iterator
4060 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4061 const SmallVectorImpl<Instruction *> &Inputs)
4064 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4065 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4068 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4069 if (!Rung) return IP;
4070 Rung = Rung->getIDom();
4071 if (!Rung) return IP;
4072 IDom = Rung->getBlock();
4074 // Don't climb into a loop though.
4075 const Loop *IDomLoop = LI.getLoopFor(IDom);
4076 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4077 if (IDomDepth <= IPLoopDepth &&
4078 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4082 bool AllDominate = true;
4083 Instruction *BetterPos = 0;
4084 Instruction *Tentative = IDom->getTerminator();
4085 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4086 E = Inputs.end(); I != E; ++I) {
4087 Instruction *Inst = *I;
4088 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4089 AllDominate = false;
4092 // Attempt to find an insert position in the middle of the block,
4093 // instead of at the end, so that it can be used for other expansions.
4094 if (IDom == Inst->getParent() &&
4095 (!BetterPos || DT.dominates(BetterPos, Inst)))
4096 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4109 /// AdjustInsertPositionForExpand - Determine an input position which will be
4110 /// dominated by the operands and which will dominate the result.
4111 BasicBlock::iterator
4112 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4115 SCEVExpander &Rewriter) const {
4116 // Collect some instructions which must be dominated by the
4117 // expanding replacement. These must be dominated by any operands that
4118 // will be required in the expansion.
4119 SmallVector<Instruction *, 4> Inputs;
4120 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4121 Inputs.push_back(I);
4122 if (LU.Kind == LSRUse::ICmpZero)
4123 if (Instruction *I =
4124 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4125 Inputs.push_back(I);
4126 if (LF.PostIncLoops.count(L)) {
4127 if (LF.isUseFullyOutsideLoop(L))
4128 Inputs.push_back(L->getLoopLatch()->getTerminator());
4130 Inputs.push_back(IVIncInsertPos);
4132 // The expansion must also be dominated by the increment positions of any
4133 // loops it for which it is using post-inc mode.
4134 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4135 E = LF.PostIncLoops.end(); I != E; ++I) {
4136 const Loop *PIL = *I;
4137 if (PIL == L) continue;
4139 // Be dominated by the loop exit.
4140 SmallVector<BasicBlock *, 4> ExitingBlocks;
4141 PIL->getExitingBlocks(ExitingBlocks);
4142 if (!ExitingBlocks.empty()) {
4143 BasicBlock *BB = ExitingBlocks[0];
4144 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4145 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4146 Inputs.push_back(BB->getTerminator());
4150 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4151 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4152 "Insertion point must be a normal instruction");
4154 // Then, climb up the immediate dominator tree as far as we can go while
4155 // still being dominated by the input positions.
4156 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4158 // Don't insert instructions before PHI nodes.
4159 while (isa<PHINode>(IP)) ++IP;
4161 // Ignore landingpad instructions.
4162 while (isa<LandingPadInst>(IP)) ++IP;
4164 // Ignore debug intrinsics.
4165 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4167 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4168 // IP consistent across expansions and allows the previously inserted
4169 // instructions to be reused by subsequent expansion.
4170 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4175 /// Expand - Emit instructions for the leading candidate expression for this
4176 /// LSRUse (this is called "expanding").
4177 Value *LSRInstance::Expand(const LSRFixup &LF,
4179 BasicBlock::iterator IP,
4180 SCEVExpander &Rewriter,
4181 SmallVectorImpl<WeakVH> &DeadInsts) const {
4182 const LSRUse &LU = Uses[LF.LUIdx];
4184 // Determine an input position which will be dominated by the operands and
4185 // which will dominate the result.
4186 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4188 // Inform the Rewriter if we have a post-increment use, so that it can
4189 // perform an advantageous expansion.
4190 Rewriter.setPostInc(LF.PostIncLoops);
4192 // This is the type that the user actually needs.
4193 Type *OpTy = LF.OperandValToReplace->getType();
4194 // This will be the type that we'll initially expand to.
4195 Type *Ty = F.getType();
4197 // No type known; just expand directly to the ultimate type.
4199 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4200 // Expand directly to the ultimate type if it's the right size.
4202 // This is the type to do integer arithmetic in.
4203 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4205 // Build up a list of operands to add together to form the full base.
4206 SmallVector<const SCEV *, 8> Ops;
4208 // Expand the BaseRegs portion.
4209 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4210 E = F.BaseRegs.end(); I != E; ++I) {
4211 const SCEV *Reg = *I;
4212 assert(!Reg->isZero() && "Zero allocated in a base register!");
4214 // If we're expanding for a post-inc user, make the post-inc adjustment.
4215 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4216 Reg = TransformForPostIncUse(Denormalize, Reg,
4217 LF.UserInst, LF.OperandValToReplace,
4220 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4223 // Flush the operand list to suppress SCEVExpander hoisting.
4225 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4227 Ops.push_back(SE.getUnknown(FullV));
4230 // Expand the ScaledReg portion.
4231 Value *ICmpScaledV = 0;
4232 if (F.AM.Scale != 0) {
4233 const SCEV *ScaledS = F.ScaledReg;
4235 // If we're expanding for a post-inc user, make the post-inc adjustment.
4236 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4237 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4238 LF.UserInst, LF.OperandValToReplace,
4241 if (LU.Kind == LSRUse::ICmpZero) {
4242 // An interesting way of "folding" with an icmp is to use a negated
4243 // scale, which we'll implement by inserting it into the other operand
4245 assert(F.AM.Scale == -1 &&
4246 "The only scale supported by ICmpZero uses is -1!");
4247 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4249 // Otherwise just expand the scaled register and an explicit scale,
4250 // which is expected to be matched as part of the address.
4251 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4252 ScaledS = SE.getMulExpr(ScaledS,
4253 SE.getConstant(ScaledS->getType(), F.AM.Scale));
4254 Ops.push_back(ScaledS);
4256 // Flush the operand list to suppress SCEVExpander hoisting.
4257 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4259 Ops.push_back(SE.getUnknown(FullV));
4263 // Expand the GV portion.
4265 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
4267 // Flush the operand list to suppress SCEVExpander hoisting.
4268 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4270 Ops.push_back(SE.getUnknown(FullV));
4273 // Expand the immediate portion.
4274 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
4276 if (LU.Kind == LSRUse::ICmpZero) {
4277 // The other interesting way of "folding" with an ICmpZero is to use a
4278 // negated immediate.
4280 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4282 Ops.push_back(SE.getUnknown(ICmpScaledV));
4283 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4286 // Just add the immediate values. These again are expected to be matched
4287 // as part of the address.
4288 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4292 // Expand the unfolded offset portion.
4293 int64_t UnfoldedOffset = F.UnfoldedOffset;
4294 if (UnfoldedOffset != 0) {
4295 // Just add the immediate values.
4296 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4300 // Emit instructions summing all the operands.
4301 const SCEV *FullS = Ops.empty() ?
4302 SE.getConstant(IntTy, 0) :
4304 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4306 // We're done expanding now, so reset the rewriter.
4307 Rewriter.clearPostInc();
4309 // An ICmpZero Formula represents an ICmp which we're handling as a
4310 // comparison against zero. Now that we've expanded an expression for that
4311 // form, update the ICmp's other operand.
4312 if (LU.Kind == LSRUse::ICmpZero) {
4313 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4314 DeadInsts.push_back(CI->getOperand(1));
4315 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
4316 "a scale at the same time!");
4317 if (F.AM.Scale == -1) {
4318 if (ICmpScaledV->getType() != OpTy) {
4320 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4322 ICmpScaledV, OpTy, "tmp", CI);
4325 CI->setOperand(1, ICmpScaledV);
4327 assert(F.AM.Scale == 0 &&
4328 "ICmp does not support folding a global value and "
4329 "a scale at the same time!");
4330 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4332 if (C->getType() != OpTy)
4333 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4337 CI->setOperand(1, C);
4344 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4345 /// of their operands effectively happens in their predecessor blocks, so the
4346 /// expression may need to be expanded in multiple places.
4347 void LSRInstance::RewriteForPHI(PHINode *PN,
4350 SCEVExpander &Rewriter,
4351 SmallVectorImpl<WeakVH> &DeadInsts,
4353 DenseMap<BasicBlock *, Value *> Inserted;
4354 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4355 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4356 BasicBlock *BB = PN->getIncomingBlock(i);
4358 // If this is a critical edge, split the edge so that we do not insert
4359 // the code on all predecessor/successor paths. We do this unless this
4360 // is the canonical backedge for this loop, which complicates post-inc
4362 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4363 !isa<IndirectBrInst>(BB->getTerminator())) {
4364 BasicBlock *Parent = PN->getParent();
4365 Loop *PNLoop = LI.getLoopFor(Parent);
4366 if (!PNLoop || Parent != PNLoop->getHeader()) {
4367 // Split the critical edge.
4368 BasicBlock *NewBB = 0;
4369 if (!Parent->isLandingPad()) {
4370 NewBB = SplitCriticalEdge(BB, Parent, P,
4371 /*MergeIdenticalEdges=*/true,
4372 /*DontDeleteUselessPhis=*/true);
4374 SmallVector<BasicBlock*, 2> NewBBs;
4375 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4379 // If PN is outside of the loop and BB is in the loop, we want to
4380 // move the block to be immediately before the PHI block, not
4381 // immediately after BB.
4382 if (L->contains(BB) && !L->contains(PN))
4383 NewBB->moveBefore(PN->getParent());
4385 // Splitting the edge can reduce the number of PHI entries we have.
4386 e = PN->getNumIncomingValues();
4388 i = PN->getBasicBlockIndex(BB);
4392 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4393 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4395 PN->setIncomingValue(i, Pair.first->second);
4397 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4399 // If this is reuse-by-noop-cast, insert the noop cast.
4400 Type *OpTy = LF.OperandValToReplace->getType();
4401 if (FullV->getType() != OpTy)
4403 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4405 FullV, LF.OperandValToReplace->getType(),
4406 "tmp", BB->getTerminator());
4408 PN->setIncomingValue(i, FullV);
4409 Pair.first->second = FullV;
4414 /// Rewrite - Emit instructions for the leading candidate expression for this
4415 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4416 /// the newly expanded value.
4417 void LSRInstance::Rewrite(const LSRFixup &LF,
4419 SCEVExpander &Rewriter,
4420 SmallVectorImpl<WeakVH> &DeadInsts,
4422 // First, find an insertion point that dominates UserInst. For PHI nodes,
4423 // find the nearest block which dominates all the relevant uses.
4424 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4425 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4427 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4429 // If this is reuse-by-noop-cast, insert the noop cast.
4430 Type *OpTy = LF.OperandValToReplace->getType();
4431 if (FullV->getType() != OpTy) {
4433 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4434 FullV, OpTy, "tmp", LF.UserInst);
4438 // Update the user. ICmpZero is handled specially here (for now) because
4439 // Expand may have updated one of the operands of the icmp already, and
4440 // its new value may happen to be equal to LF.OperandValToReplace, in
4441 // which case doing replaceUsesOfWith leads to replacing both operands
4442 // with the same value. TODO: Reorganize this.
4443 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4444 LF.UserInst->setOperand(0, FullV);
4446 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4449 DeadInsts.push_back(LF.OperandValToReplace);
4452 /// ImplementSolution - Rewrite all the fixup locations with new values,
4453 /// following the chosen solution.
4455 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4457 // Keep track of instructions we may have made dead, so that
4458 // we can remove them after we are done working.
4459 SmallVector<WeakVH, 16> DeadInsts;
4461 SCEVExpander Rewriter(SE, "lsr");
4463 Rewriter.setDebugType(DEBUG_TYPE);
4465 Rewriter.disableCanonicalMode();
4466 Rewriter.enableLSRMode();
4467 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4469 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4470 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4471 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4472 if (PHINode *PN = dyn_cast<PHINode>(ChainI->back().UserInst))
4473 Rewriter.setChainedPhi(PN);
4476 // Expand the new value definitions and update the users.
4477 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4478 E = Fixups.end(); I != E; ++I) {
4479 const LSRFixup &Fixup = *I;
4481 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4486 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4487 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4488 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4491 // Clean up after ourselves. This must be done before deleting any
4495 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4498 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
4499 : IU(P->getAnalysis<IVUsers>()),
4500 SE(P->getAnalysis<ScalarEvolution>()),
4501 DT(P->getAnalysis<DominatorTree>()),
4502 LI(P->getAnalysis<LoopInfo>()),
4503 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
4505 // If LoopSimplify form is not available, stay out of trouble.
4506 if (!L->isLoopSimplifyForm())
4509 // If there's no interesting work to be done, bail early.
4510 if (IU.empty()) return;
4513 // All dominating loops must have preheaders, or SCEVExpander may not be able
4514 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4516 // IVUsers analysis should only create users that are dominated by simple loop
4517 // headers. Since this loop should dominate all of its users, its user list
4518 // should be empty if this loop itself is not within a simple loop nest.
4519 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4520 Rung; Rung = Rung->getIDom()) {
4521 BasicBlock *BB = Rung->getBlock();
4522 const Loop *DomLoop = LI.getLoopFor(BB);
4523 if (DomLoop && DomLoop->getHeader() == BB) {
4524 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4529 DEBUG(dbgs() << "\nLSR on loop ";
4530 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4533 // First, perform some low-level loop optimizations.
4535 OptimizeLoopTermCond();
4537 // If loop preparation eliminates all interesting IV users, bail.
4538 if (IU.empty()) return;
4540 // Skip nested loops until we can model them better with formulae.
4542 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4546 // Start collecting data and preparing for the solver.
4548 CollectInterestingTypesAndFactors();
4549 CollectFixupsAndInitialFormulae();
4550 CollectLoopInvariantFixupsAndFormulae();
4552 assert(!Uses.empty() && "IVUsers reported at least one use");
4553 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4554 print_uses(dbgs()));
4556 // Now use the reuse data to generate a bunch of interesting ways
4557 // to formulate the values needed for the uses.
4558 GenerateAllReuseFormulae();
4560 FilterOutUndesirableDedicatedRegisters();
4561 NarrowSearchSpaceUsingHeuristics();
4563 SmallVector<const Formula *, 8> Solution;
4566 // Release memory that is no longer needed.
4571 if (Solution.empty())
4575 // Formulae should be legal.
4576 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4577 E = Uses.end(); I != E; ++I) {
4578 const LSRUse &LU = *I;
4579 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4580 JE = LU.Formulae.end(); J != JE; ++J)
4581 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
4582 LU.Kind, LU.AccessTy, TLI) &&
4583 "Illegal formula generated!");
4587 // Now that we've decided what we want, make it so.
4588 ImplementSolution(Solution, P);
4591 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4592 if (Factors.empty() && Types.empty()) return;
4594 OS << "LSR has identified the following interesting factors and types: ";
4597 for (SmallSetVector<int64_t, 8>::const_iterator
4598 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4599 if (!First) OS << ", ";
4604 for (SmallSetVector<Type *, 4>::const_iterator
4605 I = Types.begin(), E = Types.end(); I != E; ++I) {
4606 if (!First) OS << ", ";
4608 OS << '(' << **I << ')';
4613 void LSRInstance::print_fixups(raw_ostream &OS) const {
4614 OS << "LSR is examining the following fixup sites:\n";
4615 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4616 E = Fixups.end(); I != E; ++I) {
4623 void LSRInstance::print_uses(raw_ostream &OS) const {
4624 OS << "LSR is examining the following uses:\n";
4625 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4626 E = Uses.end(); I != E; ++I) {
4627 const LSRUse &LU = *I;
4631 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4632 JE = LU.Formulae.end(); J != JE; ++J) {
4640 void LSRInstance::print(raw_ostream &OS) const {
4641 print_factors_and_types(OS);
4646 void LSRInstance::dump() const {
4647 print(errs()); errs() << '\n';
4652 class LoopStrengthReduce : public LoopPass {
4653 /// TLI - Keep a pointer of a TargetLowering to consult for determining
4654 /// transformation profitability.
4655 const TargetLowering *const TLI;
4658 static char ID; // Pass ID, replacement for typeid
4659 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
4662 bool runOnLoop(Loop *L, LPPassManager &LPM);
4663 void getAnalysisUsage(AnalysisUsage &AU) const;
4668 char LoopStrengthReduce::ID = 0;
4669 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4670 "Loop Strength Reduction", false, false)
4671 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4672 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4673 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4674 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4675 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4676 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4677 "Loop Strength Reduction", false, false)
4680 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
4681 return new LoopStrengthReduce(TLI);
4684 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
4685 : LoopPass(ID), TLI(tli) {
4686 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4689 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4690 // We split critical edges, so we change the CFG. However, we do update
4691 // many analyses if they are around.
4692 AU.addPreservedID(LoopSimplifyID);
4694 AU.addRequired<LoopInfo>();
4695 AU.addPreserved<LoopInfo>();
4696 AU.addRequiredID(LoopSimplifyID);
4697 AU.addRequired<DominatorTree>();
4698 AU.addPreserved<DominatorTree>();
4699 AU.addRequired<ScalarEvolution>();
4700 AU.addPreserved<ScalarEvolution>();
4701 // Requiring LoopSimplify a second time here prevents IVUsers from running
4702 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4703 AU.addRequiredID(LoopSimplifyID);
4704 AU.addRequired<IVUsers>();
4705 AU.addPreserved<IVUsers>();
4708 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4709 bool Changed = false;
4711 // Run the main LSR transformation.
4712 Changed |= LSRInstance(TLI, L, this).getChanged();
4714 // Remove any extra phis created by processing inner loops.
4715 Changed |= DeleteDeadPHIs(L->getHeader());
4716 if (EnablePhiElim) {
4717 SmallVector<WeakVH, 16> DeadInsts;
4718 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4720 Rewriter.setDebugType(DEBUG_TYPE);
4722 unsigned numFolded = Rewriter.
4723 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI);
4726 DeleteTriviallyDeadInstructions(DeadInsts);
4727 DeleteDeadPHIs(L->getHeader());