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 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/AddressingMode.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/Scalar.h"
68 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
69 #include "llvm/Transforms/Utils/Local.h"
70 #include "llvm/TargetTransformInfo.h"
71 #include "llvm/ADT/SmallBitVector.h"
72 #include "llvm/ADT/SetVector.h"
73 #include "llvm/ADT/DenseSet.h"
74 #include "llvm/Support/Debug.h"
75 #include "llvm/Support/CommandLine.h"
76 #include "llvm/Support/ValueHandle.h"
77 #include "llvm/Support/raw_ostream.h"
81 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
82 /// bail out. This threshold is far beyond the number of users that LSR can
83 /// conceivably solve, so it should not affect generated code, but catches the
84 /// worst cases before LSR burns too much compile time and stack space.
85 static const unsigned MaxIVUsers = 200;
87 // Temporary flag to cleanup congruent phis after LSR phi expansion.
88 // It's currently disabled until we can determine whether it's truly useful or
89 // not. The flag should be removed after the v3.0 release.
90 // This is now needed for ivchains.
91 static cl::opt<bool> EnablePhiElim(
92 "enable-lsr-phielim", cl::Hidden, cl::init(true),
93 cl::desc("Enable LSR phi elimination"));
96 // Stress test IV chain generation.
97 static cl::opt<bool> StressIVChain(
98 "stress-ivchain", cl::Hidden, cl::init(false),
99 cl::desc("Stress test LSR IV chains"));
101 static bool StressIVChain = false;
106 /// RegSortData - This class holds data which is used to order reuse candidates.
109 /// UsedByIndices - This represents the set of LSRUse indices which reference
110 /// a particular register.
111 SmallBitVector UsedByIndices;
115 void print(raw_ostream &OS) const;
121 void RegSortData::print(raw_ostream &OS) const {
122 OS << "[NumUses=" << UsedByIndices.count() << ']';
125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
126 void RegSortData::dump() const {
127 print(errs()); errs() << '\n';
133 /// RegUseTracker - Map register candidates to information about how they are
135 class RegUseTracker {
136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
138 RegUsesTy RegUsesMap;
139 SmallVector<const SCEV *, 16> RegSequence;
142 void CountRegister(const SCEV *Reg, size_t LUIdx);
143 void DropRegister(const SCEV *Reg, size_t LUIdx);
144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
152 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
154 iterator begin() { return RegSequence.begin(); }
155 iterator end() { return RegSequence.end(); }
156 const_iterator begin() const { return RegSequence.begin(); }
157 const_iterator end() const { return RegSequence.end(); }
163 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
164 std::pair<RegUsesTy::iterator, bool> Pair =
165 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
166 RegSortData &RSD = Pair.first->second;
168 RegSequence.push_back(Reg);
169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
170 RSD.UsedByIndices.set(LUIdx);
174 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
175 RegUsesTy::iterator It = RegUsesMap.find(Reg);
176 assert(It != RegUsesMap.end());
177 RegSortData &RSD = It->second;
178 assert(RSD.UsedByIndices.size() > LUIdx);
179 RSD.UsedByIndices.reset(LUIdx);
183 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
184 assert(LUIdx <= LastLUIdx);
186 // Update RegUses. The data structure is not optimized for this purpose;
187 // we must iterate through it and update each of the bit vectors.
188 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
190 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
191 if (LUIdx < UsedByIndices.size())
192 UsedByIndices[LUIdx] =
193 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
194 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
199 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
200 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
201 if (I == RegUsesMap.end())
203 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
204 int i = UsedByIndices.find_first();
205 if (i == -1) return false;
206 if ((size_t)i != LUIdx) return true;
207 return UsedByIndices.find_next(i) != -1;
210 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
211 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
212 assert(I != RegUsesMap.end() && "Unknown register!");
213 return I->second.UsedByIndices;
216 void RegUseTracker::clear() {
223 /// Formula - This class holds information that describes a formula for
224 /// computing satisfying a use. It may include broken-out immediates and scaled
227 /// AM - This is used to represent complex addressing, as well as other kinds
228 /// of interesting uses.
231 /// BaseRegs - The list of "base" registers for this use. When this is
232 /// non-empty, AM.HasBaseReg should be set to true.
233 SmallVector<const SCEV *, 2> BaseRegs;
235 /// ScaledReg - The 'scaled' register for this use. This should be non-null
236 /// when AM.Scale is not zero.
237 const SCEV *ScaledReg;
239 /// UnfoldedOffset - An additional constant offset which added near the
240 /// use. This requires a temporary register, but the offset itself can
241 /// live in an add immediate field rather than a register.
242 int64_t UnfoldedOffset;
244 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
246 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
248 unsigned getNumRegs() const;
249 Type *getType() const;
251 void DeleteBaseReg(const SCEV *&S);
253 bool referencesReg(const SCEV *S) const;
254 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
255 const RegUseTracker &RegUses) const;
257 void print(raw_ostream &OS) const;
263 /// DoInitialMatch - Recursion helper for InitialMatch.
264 static void DoInitialMatch(const SCEV *S, Loop *L,
265 SmallVectorImpl<const SCEV *> &Good,
266 SmallVectorImpl<const SCEV *> &Bad,
267 ScalarEvolution &SE) {
268 // Collect expressions which properly dominate the loop header.
269 if (SE.properlyDominates(S, L->getHeader())) {
274 // Look at add operands.
275 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
276 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
278 DoInitialMatch(*I, L, Good, Bad, SE);
282 // Look at addrec operands.
283 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
284 if (!AR->getStart()->isZero()) {
285 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
286 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
287 AR->getStepRecurrence(SE),
288 // FIXME: AR->getNoWrapFlags()
289 AR->getLoop(), SCEV::FlagAnyWrap),
294 // Handle a multiplication by -1 (negation) if it didn't fold.
295 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
296 if (Mul->getOperand(0)->isAllOnesValue()) {
297 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
298 const SCEV *NewMul = SE.getMulExpr(Ops);
300 SmallVector<const SCEV *, 4> MyGood;
301 SmallVector<const SCEV *, 4> MyBad;
302 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
303 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
304 SE.getEffectiveSCEVType(NewMul->getType())));
305 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
306 E = MyGood.end(); I != E; ++I)
307 Good.push_back(SE.getMulExpr(NegOne, *I));
308 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
309 E = MyBad.end(); I != E; ++I)
310 Bad.push_back(SE.getMulExpr(NegOne, *I));
314 // Ok, we can't do anything interesting. Just stuff the whole thing into a
315 // register and hope for the best.
319 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
320 /// attempting to keep all loop-invariant and loop-computable values in a
321 /// single base register.
322 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
323 SmallVector<const SCEV *, 4> Good;
324 SmallVector<const SCEV *, 4> Bad;
325 DoInitialMatch(S, L, Good, Bad, SE);
327 const SCEV *Sum = SE.getAddExpr(Good);
329 BaseRegs.push_back(Sum);
330 AM.HasBaseReg = true;
333 const SCEV *Sum = SE.getAddExpr(Bad);
335 BaseRegs.push_back(Sum);
336 AM.HasBaseReg = true;
340 /// getNumRegs - Return the total number of register operands used by this
341 /// formula. This does not include register uses implied by non-constant
343 unsigned Formula::getNumRegs() const {
344 return !!ScaledReg + BaseRegs.size();
347 /// getType - Return the type of this formula, if it has one, or null
348 /// otherwise. This type is meaningless except for the bit size.
349 Type *Formula::getType() const {
350 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
351 ScaledReg ? ScaledReg->getType() :
352 AM.BaseGV ? AM.BaseGV->getType() :
356 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
357 void Formula::DeleteBaseReg(const SCEV *&S) {
358 if (&S != &BaseRegs.back())
359 std::swap(S, BaseRegs.back());
363 /// referencesReg - Test if this formula references the given register.
364 bool Formula::referencesReg(const SCEV *S) const {
365 return S == ScaledReg ||
366 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
369 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
370 /// which are used by uses other than the use with the given index.
371 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
372 const RegUseTracker &RegUses) const {
374 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
376 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
377 E = BaseRegs.end(); I != E; ++I)
378 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
383 void Formula::print(raw_ostream &OS) const {
386 if (!First) OS << " + "; else First = false;
387 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
389 if (AM.BaseOffs != 0) {
390 if (!First) OS << " + "; else First = false;
393 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
394 E = BaseRegs.end(); I != E; ++I) {
395 if (!First) OS << " + "; else First = false;
396 OS << "reg(" << **I << ')';
398 if (AM.HasBaseReg && BaseRegs.empty()) {
399 if (!First) OS << " + "; else First = false;
400 OS << "**error: HasBaseReg**";
401 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
402 if (!First) OS << " + "; else First = false;
403 OS << "**error: !HasBaseReg**";
406 if (!First) OS << " + "; else First = false;
407 OS << AM.Scale << "*reg(";
414 if (UnfoldedOffset != 0) {
415 if (!First) OS << " + "; else First = false;
416 OS << "imm(" << UnfoldedOffset << ')';
420 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
421 void Formula::dump() const {
422 print(errs()); errs() << '\n';
426 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
427 /// without changing its value.
428 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
430 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
431 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
434 /// isAddSExtable - Return true if the given add can be sign-extended
435 /// without changing its value.
436 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
438 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
439 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
442 /// isMulSExtable - Return true if the given mul can be sign-extended
443 /// without changing its value.
444 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
446 IntegerType::get(SE.getContext(),
447 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
448 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
451 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
452 /// and if the remainder is known to be zero, or null otherwise. If
453 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
454 /// to Y, ignoring that the multiplication may overflow, which is useful when
455 /// the result will be used in a context where the most significant bits are
457 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
459 bool IgnoreSignificantBits = false) {
460 // Handle the trivial case, which works for any SCEV type.
462 return SE.getConstant(LHS->getType(), 1);
464 // Handle a few RHS special cases.
465 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
467 const APInt &RA = RC->getValue()->getValue();
468 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
470 if (RA.isAllOnesValue())
471 return SE.getMulExpr(LHS, RC);
472 // Handle x /s 1 as x.
477 // Check for a division of a constant by a constant.
478 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
481 const APInt &LA = C->getValue()->getValue();
482 const APInt &RA = RC->getValue()->getValue();
483 if (LA.srem(RA) != 0)
485 return SE.getConstant(LA.sdiv(RA));
488 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
489 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
490 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
491 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
492 IgnoreSignificantBits);
494 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
495 IgnoreSignificantBits);
496 if (!Start) return 0;
497 // FlagNW is independent of the start value, step direction, and is
498 // preserved with smaller magnitude steps.
499 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
500 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
505 // Distribute the sdiv over add operands, if the add doesn't overflow.
506 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
507 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
508 SmallVector<const SCEV *, 8> Ops;
509 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
511 const SCEV *Op = getExactSDiv(*I, RHS, SE,
512 IgnoreSignificantBits);
516 return SE.getAddExpr(Ops);
521 // Check for a multiply operand that we can pull RHS out of.
522 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
523 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
524 SmallVector<const SCEV *, 4> Ops;
526 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
530 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
531 IgnoreSignificantBits)) {
537 return Found ? SE.getMulExpr(Ops) : 0;
542 // Otherwise we don't know.
546 /// ExtractImmediate - If S involves the addition of a constant integer value,
547 /// return that integer value, and mutate S to point to a new SCEV with that
549 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
550 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
551 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
552 S = SE.getConstant(C->getType(), 0);
553 return C->getValue()->getSExtValue();
555 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
556 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
557 int64_t Result = ExtractImmediate(NewOps.front(), SE);
559 S = SE.getAddExpr(NewOps);
561 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
562 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
563 int64_t Result = ExtractImmediate(NewOps.front(), SE);
565 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
566 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
573 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
574 /// return that symbol, and mutate S to point to a new SCEV with that
576 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
577 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
578 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
579 S = SE.getConstant(GV->getType(), 0);
582 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
583 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
584 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
586 S = SE.getAddExpr(NewOps);
588 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
589 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
590 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
592 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
593 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
600 /// isAddressUse - Returns true if the specified instruction is using the
601 /// specified value as an address.
602 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
603 bool isAddress = isa<LoadInst>(Inst);
604 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
605 if (SI->getOperand(1) == OperandVal)
607 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
608 // Addressing modes can also be folded into prefetches and a variety
610 switch (II->getIntrinsicID()) {
612 case Intrinsic::prefetch:
613 case Intrinsic::x86_sse_storeu_ps:
614 case Intrinsic::x86_sse2_storeu_pd:
615 case Intrinsic::x86_sse2_storeu_dq:
616 case Intrinsic::x86_sse2_storel_dq:
617 if (II->getArgOperand(0) == OperandVal)
625 /// getAccessType - Return the type of the memory being accessed.
626 static Type *getAccessType(const Instruction *Inst) {
627 Type *AccessTy = Inst->getType();
628 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
629 AccessTy = SI->getOperand(0)->getType();
630 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
631 // Addressing modes can also be folded into prefetches and a variety
633 switch (II->getIntrinsicID()) {
635 case Intrinsic::x86_sse_storeu_ps:
636 case Intrinsic::x86_sse2_storeu_pd:
637 case Intrinsic::x86_sse2_storeu_dq:
638 case Intrinsic::x86_sse2_storel_dq:
639 AccessTy = II->getArgOperand(0)->getType();
644 // All pointers have the same requirements, so canonicalize them to an
645 // arbitrary pointer type to minimize variation.
646 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
647 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
648 PTy->getAddressSpace());
653 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
654 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
655 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
656 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
657 if (SE.isSCEVable(PN->getType()) &&
658 (SE.getEffectiveSCEVType(PN->getType()) ==
659 SE.getEffectiveSCEVType(AR->getType())) &&
660 SE.getSCEV(PN) == AR)
666 /// Check if expanding this expression is likely to incur significant cost. This
667 /// is tricky because SCEV doesn't track which expressions are actually computed
668 /// by the current IR.
670 /// We currently allow expansion of IV increments that involve adds,
671 /// multiplication by constants, and AddRecs from existing phis.
673 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
674 /// obvious multiple of the UDivExpr.
675 static bool isHighCostExpansion(const SCEV *S,
676 SmallPtrSet<const SCEV*, 8> &Processed,
677 ScalarEvolution &SE) {
678 // Zero/One operand expressions
679 switch (S->getSCEVType()) {
684 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
687 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
690 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
694 if (!Processed.insert(S))
697 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
698 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
700 if (isHighCostExpansion(*I, Processed, SE))
706 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
707 if (Mul->getNumOperands() == 2) {
708 // Multiplication by a constant is ok
709 if (isa<SCEVConstant>(Mul->getOperand(0)))
710 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
712 // If we have the value of one operand, check if an existing
713 // multiplication already generates this expression.
714 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
715 Value *UVal = U->getValue();
716 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
718 // If U is a constant, it may be used by a ConstantExpr.
719 Instruction *User = dyn_cast<Instruction>(*UI);
720 if (User && User->getOpcode() == Instruction::Mul
721 && SE.isSCEVable(User->getType())) {
722 return SE.getSCEV(User) == Mul;
729 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
730 if (isExistingPhi(AR, SE))
734 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
738 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
739 /// specified set are trivially dead, delete them and see if this makes any of
740 /// their operands subsequently dead.
742 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
743 bool Changed = false;
745 while (!DeadInsts.empty()) {
746 Value *V = DeadInsts.pop_back_val();
747 Instruction *I = dyn_cast_or_null<Instruction>(V);
749 if (I == 0 || !isInstructionTriviallyDead(I))
752 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
753 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
756 DeadInsts.push_back(U);
759 I->eraseFromParent();
768 /// Cost - This class is used to measure and compare candidate formulae.
770 /// TODO: Some of these could be merged. Also, a lexical ordering
771 /// isn't always optimal.
775 unsigned NumBaseAdds;
781 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
784 bool operator<(const Cost &Other) const;
789 // Once any of the metrics loses, they must all remain losers.
791 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
792 | ImmCost | SetupCost) != ~0u)
793 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
794 & ImmCost & SetupCost) == ~0u);
799 assert(isValid() && "invalid cost");
800 return NumRegs == ~0u;
803 void RateFormula(const Formula &F,
804 SmallPtrSet<const SCEV *, 16> &Regs,
805 const DenseSet<const SCEV *> &VisitedRegs,
807 const SmallVectorImpl<int64_t> &Offsets,
808 ScalarEvolution &SE, DominatorTree &DT,
809 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
811 void print(raw_ostream &OS) const;
815 void RateRegister(const SCEV *Reg,
816 SmallPtrSet<const SCEV *, 16> &Regs,
818 ScalarEvolution &SE, DominatorTree &DT);
819 void RatePrimaryRegister(const SCEV *Reg,
820 SmallPtrSet<const SCEV *, 16> &Regs,
822 ScalarEvolution &SE, DominatorTree &DT,
823 SmallPtrSet<const SCEV *, 16> *LoserRegs);
828 /// RateRegister - Tally up interesting quantities from the given register.
829 void Cost::RateRegister(const SCEV *Reg,
830 SmallPtrSet<const SCEV *, 16> &Regs,
832 ScalarEvolution &SE, DominatorTree &DT) {
833 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
834 // If this is an addrec for another loop, don't second-guess its addrec phi
835 // nodes. LSR isn't currently smart enough to reason about more than one
836 // loop at a time. LSR has already run on inner loops, will not run on outer
837 // loops, and cannot be expected to change sibling loops.
838 if (AR->getLoop() != L) {
839 // If the AddRec exists, consider it's register free and leave it alone.
840 if (isExistingPhi(AR, SE))
843 // Otherwise, do not consider this formula at all.
847 AddRecCost += 1; /// TODO: This should be a function of the stride.
849 // Add the step value register, if it needs one.
850 // TODO: The non-affine case isn't precisely modeled here.
851 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
852 if (!Regs.count(AR->getOperand(1))) {
853 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
861 // Rough heuristic; favor registers which don't require extra setup
862 // instructions in the preheader.
863 if (!isa<SCEVUnknown>(Reg) &&
864 !isa<SCEVConstant>(Reg) &&
865 !(isa<SCEVAddRecExpr>(Reg) &&
866 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
867 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
870 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
871 SE.hasComputableLoopEvolution(Reg, L);
874 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
875 /// before, rate it. Optional LoserRegs provides a way to declare any formula
876 /// that refers to one of those regs an instant loser.
877 void Cost::RatePrimaryRegister(const SCEV *Reg,
878 SmallPtrSet<const SCEV *, 16> &Regs,
880 ScalarEvolution &SE, DominatorTree &DT,
881 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
882 if (LoserRegs && LoserRegs->count(Reg)) {
886 if (Regs.insert(Reg)) {
887 RateRegister(Reg, Regs, L, SE, DT);
889 LoserRegs->insert(Reg);
893 void Cost::RateFormula(const Formula &F,
894 SmallPtrSet<const SCEV *, 16> &Regs,
895 const DenseSet<const SCEV *> &VisitedRegs,
897 const SmallVectorImpl<int64_t> &Offsets,
898 ScalarEvolution &SE, DominatorTree &DT,
899 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
900 // Tally up the registers.
901 if (const SCEV *ScaledReg = F.ScaledReg) {
902 if (VisitedRegs.count(ScaledReg)) {
906 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
910 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
911 E = F.BaseRegs.end(); I != E; ++I) {
912 const SCEV *BaseReg = *I;
913 if (VisitedRegs.count(BaseReg)) {
917 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
922 // Determine how many (unfolded) adds we'll need inside the loop.
923 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
924 if (NumBaseParts > 1)
925 NumBaseAdds += NumBaseParts - 1;
927 // Tally up the non-zero immediates.
928 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
929 E = Offsets.end(); I != E; ++I) {
930 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
932 ImmCost += 64; // Handle symbolic values conservatively.
933 // TODO: This should probably be the pointer size.
934 else if (Offset != 0)
935 ImmCost += APInt(64, Offset, true).getMinSignedBits();
937 assert(isValid() && "invalid cost");
940 /// Loose - Set this cost to a losing value.
950 /// operator< - Choose the lower cost.
951 bool Cost::operator<(const Cost &Other) const {
952 if (NumRegs != Other.NumRegs)
953 return NumRegs < Other.NumRegs;
954 if (AddRecCost != Other.AddRecCost)
955 return AddRecCost < Other.AddRecCost;
956 if (NumIVMuls != Other.NumIVMuls)
957 return NumIVMuls < Other.NumIVMuls;
958 if (NumBaseAdds != Other.NumBaseAdds)
959 return NumBaseAdds < Other.NumBaseAdds;
960 if (ImmCost != Other.ImmCost)
961 return ImmCost < Other.ImmCost;
962 if (SetupCost != Other.SetupCost)
963 return SetupCost < Other.SetupCost;
967 void Cost::print(raw_ostream &OS) const {
968 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
970 OS << ", with addrec cost " << AddRecCost;
972 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
973 if (NumBaseAdds != 0)
974 OS << ", plus " << NumBaseAdds << " base add"
975 << (NumBaseAdds == 1 ? "" : "s");
977 OS << ", plus " << ImmCost << " imm cost";
979 OS << ", plus " << SetupCost << " setup cost";
982 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
983 void Cost::dump() const {
984 print(errs()); errs() << '\n';
990 /// LSRFixup - An operand value in an instruction which is to be replaced
991 /// with some equivalent, possibly strength-reduced, replacement.
993 /// UserInst - The instruction which will be updated.
994 Instruction *UserInst;
996 /// OperandValToReplace - The operand of the instruction which will
997 /// be replaced. The operand may be used more than once; every instance
998 /// will be replaced.
999 Value *OperandValToReplace;
1001 /// PostIncLoops - If this user is to use the post-incremented value of an
1002 /// induction variable, this variable is non-null and holds the loop
1003 /// associated with the induction variable.
1004 PostIncLoopSet PostIncLoops;
1006 /// LUIdx - The index of the LSRUse describing the expression which
1007 /// this fixup needs, minus an offset (below).
1010 /// Offset - A constant offset to be added to the LSRUse expression.
1011 /// This allows multiple fixups to share the same LSRUse with different
1012 /// offsets, for example in an unrolled loop.
1015 bool isUseFullyOutsideLoop(const Loop *L) const;
1019 void print(raw_ostream &OS) const;
1025 LSRFixup::LSRFixup()
1026 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1028 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1029 /// value outside of the given loop.
1030 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1031 // PHI nodes use their value in their incoming blocks.
1032 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1033 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1034 if (PN->getIncomingValue(i) == OperandValToReplace &&
1035 L->contains(PN->getIncomingBlock(i)))
1040 return !L->contains(UserInst);
1043 void LSRFixup::print(raw_ostream &OS) const {
1045 // Store is common and interesting enough to be worth special-casing.
1046 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1048 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1049 } else if (UserInst->getType()->isVoidTy())
1050 OS << UserInst->getOpcodeName();
1052 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1054 OS << ", OperandValToReplace=";
1055 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1057 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1058 E = PostIncLoops.end(); I != E; ++I) {
1059 OS << ", PostIncLoop=";
1060 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1063 if (LUIdx != ~size_t(0))
1064 OS << ", LUIdx=" << LUIdx;
1067 OS << ", Offset=" << Offset;
1070 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1071 void LSRFixup::dump() const {
1072 print(errs()); errs() << '\n';
1078 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1079 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1080 struct UniquifierDenseMapInfo {
1081 static SmallVector<const SCEV *, 2> getEmptyKey() {
1082 SmallVector<const SCEV *, 2> V;
1083 V.push_back(reinterpret_cast<const SCEV *>(-1));
1087 static SmallVector<const SCEV *, 2> getTombstoneKey() {
1088 SmallVector<const SCEV *, 2> V;
1089 V.push_back(reinterpret_cast<const SCEV *>(-2));
1093 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
1094 unsigned Result = 0;
1095 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1096 E = V.end(); I != E; ++I)
1097 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1101 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1102 const SmallVector<const SCEV *, 2> &RHS) {
1107 /// LSRUse - This class holds the state that LSR keeps for each use in
1108 /// IVUsers, as well as uses invented by LSR itself. It includes information
1109 /// about what kinds of things can be folded into the user, information about
1110 /// the user itself, and information about how the use may be satisfied.
1111 /// TODO: Represent multiple users of the same expression in common?
1113 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1116 /// KindType - An enum for a kind of use, indicating what types of
1117 /// scaled and immediate operands it might support.
1119 Basic, ///< A normal use, with no folding.
1120 Special, ///< A special case of basic, allowing -1 scales.
1121 Address, ///< An address use; folding according to ScalarTargetTransformInfo.
1122 ICmpZero ///< An equality icmp with both operands folded into one.
1123 // TODO: Add a generic icmp too?
1129 SmallVector<int64_t, 8> Offsets;
1133 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1134 /// LSRUse are outside of the loop, in which case some special-case heuristics
1136 bool AllFixupsOutsideLoop;
1138 /// WidestFixupType - This records the widest use type for any fixup using
1139 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1140 /// max fixup widths to be equivalent, because the narrower one may be relying
1141 /// on the implicit truncation to truncate away bogus bits.
1142 Type *WidestFixupType;
1144 /// Formulae - A list of ways to build a value that can satisfy this user.
1145 /// After the list is populated, one of these is selected heuristically and
1146 /// used to formulate a replacement for OperandValToReplace in UserInst.
1147 SmallVector<Formula, 12> Formulae;
1149 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1150 SmallPtrSet<const SCEV *, 4> Regs;
1152 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1153 MinOffset(INT64_MAX),
1154 MaxOffset(INT64_MIN),
1155 AllFixupsOutsideLoop(true),
1156 WidestFixupType(0) {}
1158 bool HasFormulaWithSameRegs(const Formula &F) const;
1159 bool InsertFormula(const Formula &F);
1160 void DeleteFormula(Formula &F);
1161 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1163 void print(raw_ostream &OS) const;
1169 /// HasFormula - Test whether this use as a formula which has the same
1170 /// registers as the given formula.
1171 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1172 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1173 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1174 // Unstable sort by host order ok, because this is only used for uniquifying.
1175 std::sort(Key.begin(), Key.end());
1176 return Uniquifier.count(Key);
1179 /// InsertFormula - If the given formula has not yet been inserted, add it to
1180 /// the list, and return true. Return false otherwise.
1181 bool LSRUse::InsertFormula(const Formula &F) {
1182 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1183 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1184 // Unstable sort by host order ok, because this is only used for uniquifying.
1185 std::sort(Key.begin(), Key.end());
1187 if (!Uniquifier.insert(Key).second)
1190 // Using a register to hold the value of 0 is not profitable.
1191 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1192 "Zero allocated in a scaled register!");
1194 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1195 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1196 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1199 // Add the formula to the list.
1200 Formulae.push_back(F);
1202 // Record registers now being used by this use.
1203 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1208 /// DeleteFormula - Remove the given formula from this use's list.
1209 void LSRUse::DeleteFormula(Formula &F) {
1210 if (&F != &Formulae.back())
1211 std::swap(F, Formulae.back());
1212 Formulae.pop_back();
1215 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1216 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1217 // Now that we've filtered out some formulae, recompute the Regs set.
1218 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1220 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1221 E = Formulae.end(); I != E; ++I) {
1222 const Formula &F = *I;
1223 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1224 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1227 // Update the RegTracker.
1228 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1229 E = OldRegs.end(); I != E; ++I)
1230 if (!Regs.count(*I))
1231 RegUses.DropRegister(*I, LUIdx);
1234 void LSRUse::print(raw_ostream &OS) const {
1235 OS << "LSR Use: Kind=";
1237 case Basic: OS << "Basic"; break;
1238 case Special: OS << "Special"; break;
1239 case ICmpZero: OS << "ICmpZero"; break;
1241 OS << "Address of ";
1242 if (AccessTy->isPointerTy())
1243 OS << "pointer"; // the full pointer type could be really verbose
1248 OS << ", Offsets={";
1249 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1250 E = Offsets.end(); I != E; ++I) {
1252 if (llvm::next(I) != E)
1257 if (AllFixupsOutsideLoop)
1258 OS << ", all-fixups-outside-loop";
1260 if (WidestFixupType)
1261 OS << ", widest fixup type: " << *WidestFixupType;
1264 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1265 void LSRUse::dump() const {
1266 print(errs()); errs() << '\n';
1270 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1271 /// be completely folded into the user instruction at isel time. This includes
1272 /// address-mode folding and special icmp tricks.
1273 static bool isLegalUse(const AddrMode &AM,
1274 LSRUse::KindType Kind, Type *AccessTy,
1275 const ScalarTargetTransformInfo *STTI) {
1277 case LSRUse::Address:
1278 // If we have low-level target information, ask the target if it can
1279 // completely fold this address.
1280 if (STTI) return STTI->isLegalAddressingMode(AM, AccessTy);
1282 // Otherwise, just guess that reg+reg addressing is legal.
1283 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1285 case LSRUse::ICmpZero:
1286 // There's not even a target hook for querying whether it would be legal to
1287 // fold a GV into an ICmp.
1291 // ICmp only has two operands; don't allow more than two non-trivial parts.
1292 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1295 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1296 // putting the scaled register in the other operand of the icmp.
1297 if (AM.Scale != 0 && AM.Scale != -1)
1300 // If we have low-level target information, ask the target if it can fold an
1301 // integer immediate on an icmp.
1302 if (AM.BaseOffs != 0) {
1306 // ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset
1307 // ICmpZero -1*ScaleReg + Offset => ICmp ScaleReg, Offset
1308 // Offs is the ICmp immediate.
1309 int64_t Offs = AM.BaseOffs;
1311 Offs = -(uint64_t)Offs; // The cast does the right thing with INT64_MIN.
1312 return STTI->isLegalICmpImmediate(Offs);
1315 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1319 // Only handle single-register values.
1320 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1322 case LSRUse::Special:
1323 // Special case Basic to handle -1 scales.
1324 return !AM.BaseGV && (AM.Scale == 0 || AM.Scale == -1) && AM.BaseOffs == 0;
1327 llvm_unreachable("Invalid LSRUse Kind!");
1330 static bool isLegalUse(AddrMode AM,
1331 int64_t MinOffset, int64_t MaxOffset,
1332 LSRUse::KindType Kind, Type *AccessTy,
1333 const ScalarTargetTransformInfo *LTTI) {
1334 // Check for overflow.
1335 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1338 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1339 if (isLegalUse(AM, Kind, AccessTy, LTTI)) {
1340 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1341 // Check for overflow.
1342 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1345 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1346 return isLegalUse(AM, Kind, AccessTy, LTTI);
1351 static bool isAlwaysFoldable(int64_t BaseOffs,
1352 GlobalValue *BaseGV,
1354 LSRUse::KindType Kind, Type *AccessTy,
1355 const ScalarTargetTransformInfo *LTTI) {
1356 // Fast-path: zero is always foldable.
1357 if (BaseOffs == 0 && !BaseGV) return true;
1359 // Conservatively, create an address with an immediate and a
1360 // base and a scale.
1362 AM.BaseOffs = BaseOffs;
1364 AM.HasBaseReg = HasBaseReg;
1365 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1367 // Canonicalize a scale of 1 to a base register if the formula doesn't
1368 // already have a base register.
1369 if (!AM.HasBaseReg && AM.Scale == 1) {
1371 AM.HasBaseReg = true;
1374 return isLegalUse(AM, Kind, AccessTy, LTTI);
1377 static bool isAlwaysFoldable(const SCEV *S,
1378 int64_t MinOffset, int64_t MaxOffset,
1380 LSRUse::KindType Kind, Type *AccessTy,
1381 const ScalarTargetTransformInfo *LTTI,
1382 ScalarEvolution &SE) {
1383 // Fast-path: zero is always foldable.
1384 if (S->isZero()) return true;
1386 // Conservatively, create an address with an immediate and a
1387 // base and a scale.
1388 int64_t BaseOffs = ExtractImmediate(S, SE);
1389 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1391 // If there's anything else involved, it's not foldable.
1392 if (!S->isZero()) return false;
1394 // Fast-path: zero is always foldable.
1395 if (BaseOffs == 0 && !BaseGV) return true;
1397 // Conservatively, create an address with an immediate and a
1398 // base and a scale.
1400 AM.BaseOffs = BaseOffs;
1402 AM.HasBaseReg = HasBaseReg;
1403 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1405 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, LTTI);
1410 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1411 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1412 struct UseMapDenseMapInfo {
1413 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1414 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1417 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1418 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1422 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1423 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1424 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1428 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1429 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1434 /// IVInc - An individual increment in a Chain of IV increments.
1435 /// Relate an IV user to an expression that computes the IV it uses from the IV
1436 /// used by the previous link in the Chain.
1438 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1439 /// original IVOperand. The head of the chain's IVOperand is only valid during
1440 /// chain collection, before LSR replaces IV users. During chain generation,
1441 /// IncExpr can be used to find the new IVOperand that computes the same
1444 Instruction *UserInst;
1446 const SCEV *IncExpr;
1448 IVInc(Instruction *U, Value *O, const SCEV *E):
1449 UserInst(U), IVOperand(O), IncExpr(E) {}
1452 // IVChain - The list of IV increments in program order.
1453 // We typically add the head of a chain without finding subsequent links.
1455 SmallVector<IVInc,1> Incs;
1456 const SCEV *ExprBase;
1458 IVChain() : ExprBase(0) {}
1460 IVChain(const IVInc &Head, const SCEV *Base)
1461 : Incs(1, Head), ExprBase(Base) {}
1463 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1465 // begin - return the first increment in the chain.
1466 const_iterator begin() const {
1467 assert(!Incs.empty());
1468 return llvm::next(Incs.begin());
1470 const_iterator end() const {
1474 // hasIncs - Returns true if this chain contains any increments.
1475 bool hasIncs() const { return Incs.size() >= 2; }
1477 // add - Add an IVInc to the end of this chain.
1478 void add(const IVInc &X) { Incs.push_back(X); }
1480 // tailUserInst - Returns the last UserInst in the chain.
1481 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1483 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1485 bool isProfitableIncrement(const SCEV *OperExpr,
1486 const SCEV *IncExpr,
1490 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1491 /// Distinguish between FarUsers that definitely cross IV increments and
1492 /// NearUsers that may be used between IV increments.
1494 SmallPtrSet<Instruction*, 4> FarUsers;
1495 SmallPtrSet<Instruction*, 4> NearUsers;
1498 /// LSRInstance - This class holds state for the main loop strength reduction
1502 ScalarEvolution &SE;
1505 const ScalarTargetTransformInfo *const STTI;
1509 /// IVIncInsertPos - This is the insert position that the current loop's
1510 /// induction variable increment should be placed. In simple loops, this is
1511 /// the latch block's terminator. But in more complicated cases, this is a
1512 /// position which will dominate all the in-loop post-increment users.
1513 Instruction *IVIncInsertPos;
1515 /// Factors - Interesting factors between use strides.
1516 SmallSetVector<int64_t, 8> Factors;
1518 /// Types - Interesting use types, to facilitate truncation reuse.
1519 SmallSetVector<Type *, 4> Types;
1521 /// Fixups - The list of operands which are to be replaced.
1522 SmallVector<LSRFixup, 16> Fixups;
1524 /// Uses - The list of interesting uses.
1525 SmallVector<LSRUse, 16> Uses;
1527 /// RegUses - Track which uses use which register candidates.
1528 RegUseTracker RegUses;
1530 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1531 // have more than a few IV increment chains in a loop. Missing a Chain falls
1532 // back to normal LSR behavior for those uses.
1533 static const unsigned MaxChains = 8;
1535 /// IVChainVec - IV users can form a chain of IV increments.
1536 SmallVector<IVChain, MaxChains> IVChainVec;
1538 /// IVIncSet - IV users that belong to profitable IVChains.
1539 SmallPtrSet<Use*, MaxChains> IVIncSet;
1541 void OptimizeShadowIV();
1542 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1543 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1544 void OptimizeLoopTermCond();
1546 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1547 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1548 void FinalizeChain(IVChain &Chain);
1549 void CollectChains();
1550 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1551 SmallVectorImpl<WeakVH> &DeadInsts);
1553 void CollectInterestingTypesAndFactors();
1554 void CollectFixupsAndInitialFormulae();
1556 LSRFixup &getNewFixup() {
1557 Fixups.push_back(LSRFixup());
1558 return Fixups.back();
1561 // Support for sharing of LSRUses between LSRFixups.
1562 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1564 UseMapDenseMapInfo> UseMapTy;
1567 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1568 LSRUse::KindType Kind, Type *AccessTy);
1570 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1571 LSRUse::KindType Kind,
1574 void DeleteUse(LSRUse &LU, size_t LUIdx);
1576 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1578 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1579 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1580 void CountRegisters(const Formula &F, size_t LUIdx);
1581 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1583 void CollectLoopInvariantFixupsAndFormulae();
1585 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1586 unsigned Depth = 0);
1587 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1588 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1589 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1590 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1591 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1592 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1593 void GenerateCrossUseConstantOffsets();
1594 void GenerateAllReuseFormulae();
1596 void FilterOutUndesirableDedicatedRegisters();
1598 size_t EstimateSearchSpaceComplexity() const;
1599 void NarrowSearchSpaceByDetectingSupersets();
1600 void NarrowSearchSpaceByCollapsingUnrolledCode();
1601 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1602 void NarrowSearchSpaceByPickingWinnerRegs();
1603 void NarrowSearchSpaceUsingHeuristics();
1605 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1607 SmallVectorImpl<const Formula *> &Workspace,
1608 const Cost &CurCost,
1609 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1610 DenseSet<const SCEV *> &VisitedRegs) const;
1611 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1613 BasicBlock::iterator
1614 HoistInsertPosition(BasicBlock::iterator IP,
1615 const SmallVectorImpl<Instruction *> &Inputs) const;
1616 BasicBlock::iterator
1617 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1620 SCEVExpander &Rewriter) const;
1622 Value *Expand(const LSRFixup &LF,
1624 BasicBlock::iterator IP,
1625 SCEVExpander &Rewriter,
1626 SmallVectorImpl<WeakVH> &DeadInsts) const;
1627 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1629 SCEVExpander &Rewriter,
1630 SmallVectorImpl<WeakVH> &DeadInsts,
1632 void Rewrite(const LSRFixup &LF,
1634 SCEVExpander &Rewriter,
1635 SmallVectorImpl<WeakVH> &DeadInsts,
1637 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1641 LSRInstance(const ScalarTargetTransformInfo *ltti, Loop *l, Pass *P);
1643 bool getChanged() const { return Changed; }
1645 void print_factors_and_types(raw_ostream &OS) const;
1646 void print_fixups(raw_ostream &OS) const;
1647 void print_uses(raw_ostream &OS) const;
1648 void print(raw_ostream &OS) const;
1654 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1655 /// inside the loop then try to eliminate the cast operation.
1656 void LSRInstance::OptimizeShadowIV() {
1657 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1658 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1661 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1662 UI != E; /* empty */) {
1663 IVUsers::const_iterator CandidateUI = UI;
1665 Instruction *ShadowUse = CandidateUI->getUser();
1666 Type *DestTy = NULL;
1667 bool IsSigned = false;
1669 /* If shadow use is a int->float cast then insert a second IV
1670 to eliminate this cast.
1672 for (unsigned i = 0; i < n; ++i)
1678 for (unsigned i = 0; i < n; ++i, ++d)
1681 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1683 DestTy = UCast->getDestTy();
1685 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1687 DestTy = SCast->getDestTy();
1689 if (!DestTy) continue;
1692 // If target does not support DestTy natively then do not apply
1693 // this transformation.
1694 if (!STTI->isTypeLegal(DestTy)) continue;
1697 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1699 if (PH->getNumIncomingValues() != 2) continue;
1701 Type *SrcTy = PH->getType();
1702 int Mantissa = DestTy->getFPMantissaWidth();
1703 if (Mantissa == -1) continue;
1704 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1707 unsigned Entry, Latch;
1708 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1716 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1717 if (!Init) continue;
1718 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1719 (double)Init->getSExtValue() :
1720 (double)Init->getZExtValue());
1722 BinaryOperator *Incr =
1723 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1724 if (!Incr) continue;
1725 if (Incr->getOpcode() != Instruction::Add
1726 && Incr->getOpcode() != Instruction::Sub)
1729 /* Initialize new IV, double d = 0.0 in above example. */
1730 ConstantInt *C = NULL;
1731 if (Incr->getOperand(0) == PH)
1732 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1733 else if (Incr->getOperand(1) == PH)
1734 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1740 // Ignore negative constants, as the code below doesn't handle them
1741 // correctly. TODO: Remove this restriction.
1742 if (!C->getValue().isStrictlyPositive()) continue;
1744 /* Add new PHINode. */
1745 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1747 /* create new increment. '++d' in above example. */
1748 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1749 BinaryOperator *NewIncr =
1750 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1751 Instruction::FAdd : Instruction::FSub,
1752 NewPH, CFP, "IV.S.next.", Incr);
1754 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1755 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1757 /* Remove cast operation */
1758 ShadowUse->replaceAllUsesWith(NewPH);
1759 ShadowUse->eraseFromParent();
1765 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1766 /// set the IV user and stride information and return true, otherwise return
1768 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1769 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1770 if (UI->getUser() == Cond) {
1771 // NOTE: we could handle setcc instructions with multiple uses here, but
1772 // InstCombine does it as well for simple uses, it's not clear that it
1773 // occurs enough in real life to handle.
1780 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1781 /// a max computation.
1783 /// This is a narrow solution to a specific, but acute, problem. For loops
1789 /// } while (++i < n);
1791 /// the trip count isn't just 'n', because 'n' might not be positive. And
1792 /// unfortunately this can come up even for loops where the user didn't use
1793 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1794 /// will commonly be lowered like this:
1800 /// } while (++i < n);
1803 /// and then it's possible for subsequent optimization to obscure the if
1804 /// test in such a way that indvars can't find it.
1806 /// When indvars can't find the if test in loops like this, it creates a
1807 /// max expression, which allows it to give the loop a canonical
1808 /// induction variable:
1811 /// max = n < 1 ? 1 : n;
1814 /// } while (++i != max);
1816 /// Canonical induction variables are necessary because the loop passes
1817 /// are designed around them. The most obvious example of this is the
1818 /// LoopInfo analysis, which doesn't remember trip count values. It
1819 /// expects to be able to rediscover the trip count each time it is
1820 /// needed, and it does this using a simple analysis that only succeeds if
1821 /// the loop has a canonical induction variable.
1823 /// However, when it comes time to generate code, the maximum operation
1824 /// can be quite costly, especially if it's inside of an outer loop.
1826 /// This function solves this problem by detecting this type of loop and
1827 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1828 /// the instructions for the maximum computation.
1830 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1831 // Check that the loop matches the pattern we're looking for.
1832 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1833 Cond->getPredicate() != CmpInst::ICMP_NE)
1836 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1837 if (!Sel || !Sel->hasOneUse()) return Cond;
1839 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1840 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1842 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1844 // Add one to the backedge-taken count to get the trip count.
1845 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1846 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1848 // Check for a max calculation that matches the pattern. There's no check
1849 // for ICMP_ULE here because the comparison would be with zero, which
1850 // isn't interesting.
1851 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1852 const SCEVNAryExpr *Max = 0;
1853 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1854 Pred = ICmpInst::ICMP_SLE;
1856 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1857 Pred = ICmpInst::ICMP_SLT;
1859 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1860 Pred = ICmpInst::ICMP_ULT;
1867 // To handle a max with more than two operands, this optimization would
1868 // require additional checking and setup.
1869 if (Max->getNumOperands() != 2)
1872 const SCEV *MaxLHS = Max->getOperand(0);
1873 const SCEV *MaxRHS = Max->getOperand(1);
1875 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1876 // for a comparison with 1. For <= and >=, a comparison with zero.
1878 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1881 // Check the relevant induction variable for conformance to
1883 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1884 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1885 if (!AR || !AR->isAffine() ||
1886 AR->getStart() != One ||
1887 AR->getStepRecurrence(SE) != One)
1890 assert(AR->getLoop() == L &&
1891 "Loop condition operand is an addrec in a different loop!");
1893 // Check the right operand of the select, and remember it, as it will
1894 // be used in the new comparison instruction.
1896 if (ICmpInst::isTrueWhenEqual(Pred)) {
1897 // Look for n+1, and grab n.
1898 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1899 if (isa<ConstantInt>(BO->getOperand(1)) &&
1900 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1901 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1902 NewRHS = BO->getOperand(0);
1903 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1904 if (isa<ConstantInt>(BO->getOperand(1)) &&
1905 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1906 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1907 NewRHS = BO->getOperand(0);
1910 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1911 NewRHS = Sel->getOperand(1);
1912 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1913 NewRHS = Sel->getOperand(2);
1914 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1915 NewRHS = SU->getValue();
1917 // Max doesn't match expected pattern.
1920 // Determine the new comparison opcode. It may be signed or unsigned,
1921 // and the original comparison may be either equality or inequality.
1922 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1923 Pred = CmpInst::getInversePredicate(Pred);
1925 // Ok, everything looks ok to change the condition into an SLT or SGE and
1926 // delete the max calculation.
1928 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1930 // Delete the max calculation instructions.
1931 Cond->replaceAllUsesWith(NewCond);
1932 CondUse->setUser(NewCond);
1933 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1934 Cond->eraseFromParent();
1935 Sel->eraseFromParent();
1936 if (Cmp->use_empty())
1937 Cmp->eraseFromParent();
1941 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1942 /// postinc iv when possible.
1944 LSRInstance::OptimizeLoopTermCond() {
1945 SmallPtrSet<Instruction *, 4> PostIncs;
1947 BasicBlock *LatchBlock = L->getLoopLatch();
1948 SmallVector<BasicBlock*, 8> ExitingBlocks;
1949 L->getExitingBlocks(ExitingBlocks);
1951 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1952 BasicBlock *ExitingBlock = ExitingBlocks[i];
1954 // Get the terminating condition for the loop if possible. If we
1955 // can, we want to change it to use a post-incremented version of its
1956 // induction variable, to allow coalescing the live ranges for the IV into
1957 // one register value.
1959 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1962 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1963 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1966 // Search IVUsesByStride to find Cond's IVUse if there is one.
1967 IVStrideUse *CondUse = 0;
1968 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1969 if (!FindIVUserForCond(Cond, CondUse))
1972 // If the trip count is computed in terms of a max (due to ScalarEvolution
1973 // being unable to find a sufficient guard, for example), change the loop
1974 // comparison to use SLT or ULT instead of NE.
1975 // One consequence of doing this now is that it disrupts the count-down
1976 // optimization. That's not always a bad thing though, because in such
1977 // cases it may still be worthwhile to avoid a max.
1978 Cond = OptimizeMax(Cond, CondUse);
1980 // If this exiting block dominates the latch block, it may also use
1981 // the post-inc value if it won't be shared with other uses.
1982 // Check for dominance.
1983 if (!DT.dominates(ExitingBlock, LatchBlock))
1986 // Conservatively avoid trying to use the post-inc value in non-latch
1987 // exits if there may be pre-inc users in intervening blocks.
1988 if (LatchBlock != ExitingBlock)
1989 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1990 // Test if the use is reachable from the exiting block. This dominator
1991 // query is a conservative approximation of reachability.
1992 if (&*UI != CondUse &&
1993 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1994 // Conservatively assume there may be reuse if the quotient of their
1995 // strides could be a legal scale.
1996 const SCEV *A = IU.getStride(*CondUse, L);
1997 const SCEV *B = IU.getStride(*UI, L);
1998 if (!A || !B) continue;
1999 if (SE.getTypeSizeInBits(A->getType()) !=
2000 SE.getTypeSizeInBits(B->getType())) {
2001 if (SE.getTypeSizeInBits(A->getType()) >
2002 SE.getTypeSizeInBits(B->getType()))
2003 B = SE.getSignExtendExpr(B, A->getType());
2005 A = SE.getSignExtendExpr(A, B->getType());
2007 if (const SCEVConstant *D =
2008 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2009 const ConstantInt *C = D->getValue();
2010 // Stride of one or negative one can have reuse with non-addresses.
2011 if (C->isOne() || C->isAllOnesValue())
2012 goto decline_post_inc;
2013 // Avoid weird situations.
2014 if (C->getValue().getMinSignedBits() >= 64 ||
2015 C->getValue().isMinSignedValue())
2016 goto decline_post_inc;
2017 // Without STTI, assume that any stride might be valid, and so any
2018 // use might be shared.
2020 goto decline_post_inc;
2021 // Check for possible scaled-address reuse.
2022 Type *AccessTy = getAccessType(UI->getUser());
2024 AM.Scale = C->getSExtValue();
2025 if (STTI->isLegalAddressingMode(AM, AccessTy))
2026 goto decline_post_inc;
2027 AM.Scale = -AM.Scale;
2028 if (STTI->isLegalAddressingMode(AM, AccessTy))
2029 goto decline_post_inc;
2033 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2036 // It's possible for the setcc instruction to be anywhere in the loop, and
2037 // possible for it to have multiple users. If it is not immediately before
2038 // the exiting block branch, move it.
2039 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2040 if (Cond->hasOneUse()) {
2041 Cond->moveBefore(TermBr);
2043 // Clone the terminating condition and insert into the loopend.
2044 ICmpInst *OldCond = Cond;
2045 Cond = cast<ICmpInst>(Cond->clone());
2046 Cond->setName(L->getHeader()->getName() + ".termcond");
2047 ExitingBlock->getInstList().insert(TermBr, Cond);
2049 // Clone the IVUse, as the old use still exists!
2050 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2051 TermBr->replaceUsesOfWith(OldCond, Cond);
2055 // If we get to here, we know that we can transform the setcc instruction to
2056 // use the post-incremented version of the IV, allowing us to coalesce the
2057 // live ranges for the IV correctly.
2058 CondUse->transformToPostInc(L);
2061 PostIncs.insert(Cond);
2065 // Determine an insertion point for the loop induction variable increment. It
2066 // must dominate all the post-inc comparisons we just set up, and it must
2067 // dominate the loop latch edge.
2068 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2069 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2070 E = PostIncs.end(); I != E; ++I) {
2072 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2074 if (BB == (*I)->getParent())
2075 IVIncInsertPos = *I;
2076 else if (BB != IVIncInsertPos->getParent())
2077 IVIncInsertPos = BB->getTerminator();
2081 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2082 /// at the given offset and other details. If so, update the use and
2085 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2086 LSRUse::KindType Kind, Type *AccessTy) {
2087 int64_t NewMinOffset = LU.MinOffset;
2088 int64_t NewMaxOffset = LU.MaxOffset;
2089 Type *NewAccessTy = AccessTy;
2091 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2092 // something conservative, however this can pessimize in the case that one of
2093 // the uses will have all its uses outside the loop, for example.
2094 if (LU.Kind != Kind)
2096 // Conservatively assume HasBaseReg is true for now.
2097 if (NewOffset < LU.MinOffset) {
2098 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
2099 Kind, AccessTy, STTI))
2101 NewMinOffset = NewOffset;
2102 } else if (NewOffset > LU.MaxOffset) {
2103 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
2104 Kind, AccessTy, STTI))
2106 NewMaxOffset = NewOffset;
2108 // Check for a mismatched access type, and fall back conservatively as needed.
2109 // TODO: Be less conservative when the type is similar and can use the same
2110 // addressing modes.
2111 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2112 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2115 LU.MinOffset = NewMinOffset;
2116 LU.MaxOffset = NewMaxOffset;
2117 LU.AccessTy = NewAccessTy;
2118 if (NewOffset != LU.Offsets.back())
2119 LU.Offsets.push_back(NewOffset);
2123 /// getUse - Return an LSRUse index and an offset value for a fixup which
2124 /// needs the given expression, with the given kind and optional access type.
2125 /// Either reuse an existing use or create a new one, as needed.
2126 std::pair<size_t, int64_t>
2127 LSRInstance::getUse(const SCEV *&Expr,
2128 LSRUse::KindType Kind, Type *AccessTy) {
2129 const SCEV *Copy = Expr;
2130 int64_t Offset = ExtractImmediate(Expr, SE);
2132 // Basic uses can't accept any offset, for example.
2133 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, STTI)) {
2138 std::pair<UseMapTy::iterator, bool> P =
2139 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2141 // A use already existed with this base.
2142 size_t LUIdx = P.first->second;
2143 LSRUse &LU = Uses[LUIdx];
2144 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2146 return std::make_pair(LUIdx, Offset);
2149 // Create a new use.
2150 size_t LUIdx = Uses.size();
2151 P.first->second = LUIdx;
2152 Uses.push_back(LSRUse(Kind, AccessTy));
2153 LSRUse &LU = Uses[LUIdx];
2155 // We don't need to track redundant offsets, but we don't need to go out
2156 // of our way here to avoid them.
2157 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2158 LU.Offsets.push_back(Offset);
2160 LU.MinOffset = Offset;
2161 LU.MaxOffset = Offset;
2162 return std::make_pair(LUIdx, Offset);
2165 /// DeleteUse - Delete the given use from the Uses list.
2166 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2167 if (&LU != &Uses.back())
2168 std::swap(LU, Uses.back());
2172 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2175 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2176 /// a formula that has the same registers as the given formula.
2178 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2179 const LSRUse &OrigLU) {
2180 // Search all uses for the formula. This could be more clever.
2181 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2182 LSRUse &LU = Uses[LUIdx];
2183 // Check whether this use is close enough to OrigLU, to see whether it's
2184 // worthwhile looking through its formulae.
2185 // Ignore ICmpZero uses because they may contain formulae generated by
2186 // GenerateICmpZeroScales, in which case adding fixup offsets may
2188 if (&LU != &OrigLU &&
2189 LU.Kind != LSRUse::ICmpZero &&
2190 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2191 LU.WidestFixupType == OrigLU.WidestFixupType &&
2192 LU.HasFormulaWithSameRegs(OrigF)) {
2193 // Scan through this use's formulae.
2194 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2195 E = LU.Formulae.end(); I != E; ++I) {
2196 const Formula &F = *I;
2197 // Check to see if this formula has the same registers and symbols
2199 if (F.BaseRegs == OrigF.BaseRegs &&
2200 F.ScaledReg == OrigF.ScaledReg &&
2201 F.AM.BaseGV == OrigF.AM.BaseGV &&
2202 F.AM.Scale == OrigF.AM.Scale &&
2203 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2204 if (F.AM.BaseOffs == 0)
2206 // This is the formula where all the registers and symbols matched;
2207 // there aren't going to be any others. Since we declined it, we
2208 // can skip the rest of the formulae and proceed to the next LSRUse.
2215 // Nothing looked good.
2219 void LSRInstance::CollectInterestingTypesAndFactors() {
2220 SmallSetVector<const SCEV *, 4> Strides;
2222 // Collect interesting types and strides.
2223 SmallVector<const SCEV *, 4> Worklist;
2224 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2225 const SCEV *Expr = IU.getExpr(*UI);
2227 // Collect interesting types.
2228 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2230 // Add strides for mentioned loops.
2231 Worklist.push_back(Expr);
2233 const SCEV *S = Worklist.pop_back_val();
2234 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2235 if (AR->getLoop() == L)
2236 Strides.insert(AR->getStepRecurrence(SE));
2237 Worklist.push_back(AR->getStart());
2238 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2239 Worklist.append(Add->op_begin(), Add->op_end());
2241 } while (!Worklist.empty());
2244 // Compute interesting factors from the set of interesting strides.
2245 for (SmallSetVector<const SCEV *, 4>::const_iterator
2246 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2247 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2248 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2249 const SCEV *OldStride = *I;
2250 const SCEV *NewStride = *NewStrideIter;
2252 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2253 SE.getTypeSizeInBits(NewStride->getType())) {
2254 if (SE.getTypeSizeInBits(OldStride->getType()) >
2255 SE.getTypeSizeInBits(NewStride->getType()))
2256 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2258 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2260 if (const SCEVConstant *Factor =
2261 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2263 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2264 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2265 } else if (const SCEVConstant *Factor =
2266 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2269 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2270 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2274 // If all uses use the same type, don't bother looking for truncation-based
2276 if (Types.size() == 1)
2279 DEBUG(print_factors_and_types(dbgs()));
2282 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2283 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2284 /// Instructions to IVStrideUses, we could partially skip this.
2285 static User::op_iterator
2286 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2287 Loop *L, ScalarEvolution &SE) {
2288 for(; OI != OE; ++OI) {
2289 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2290 if (!SE.isSCEVable(Oper->getType()))
2293 if (const SCEVAddRecExpr *AR =
2294 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2295 if (AR->getLoop() == L)
2303 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2304 /// operands, so wrap it in a convenient helper.
2305 static Value *getWideOperand(Value *Oper) {
2306 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2307 return Trunc->getOperand(0);
2311 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2313 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2314 Type *LType = LVal->getType();
2315 Type *RType = RVal->getType();
2316 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2319 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2320 /// NULL for any constant. Returning the expression itself is
2321 /// conservative. Returning a deeper subexpression is more precise and valid as
2322 /// long as it isn't less complex than another subexpression. For expressions
2323 /// involving multiple unscaled values, we need to return the pointer-type
2324 /// SCEVUnknown. This avoids forming chains across objects, such as:
2325 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2327 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2328 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2329 static const SCEV *getExprBase(const SCEV *S) {
2330 switch (S->getSCEVType()) {
2331 default: // uncluding scUnknown.
2336 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2338 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2340 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2342 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2343 // there's nothing more complex.
2344 // FIXME: not sure if we want to recognize negation.
2345 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2346 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2347 E(Add->op_begin()); I != E; ++I) {
2348 const SCEV *SubExpr = *I;
2349 if (SubExpr->getSCEVType() == scAddExpr)
2350 return getExprBase(SubExpr);
2352 if (SubExpr->getSCEVType() != scMulExpr)
2355 return S; // all operands are scaled, be conservative.
2358 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2362 /// Return true if the chain increment is profitable to expand into a loop
2363 /// invariant value, which may require its own register. A profitable chain
2364 /// increment will be an offset relative to the same base. We allow such offsets
2365 /// to potentially be used as chain increment as long as it's not obviously
2366 /// expensive to expand using real instructions.
2367 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2368 const SCEV *IncExpr,
2369 ScalarEvolution &SE) {
2370 // Aggressively form chains when -stress-ivchain.
2374 // Do not replace a constant offset from IV head with a nonconstant IV
2376 if (!isa<SCEVConstant>(IncExpr)) {
2377 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2378 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2382 SmallPtrSet<const SCEV*, 8> Processed;
2383 return !isHighCostExpansion(IncExpr, Processed, SE);
2386 /// Return true if the number of registers needed for the chain is estimated to
2387 /// be less than the number required for the individual IV users. First prohibit
2388 /// any IV users that keep the IV live across increments (the Users set should
2389 /// be empty). Next count the number and type of increments in the chain.
2391 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2392 /// effectively use postinc addressing modes. Only consider it profitable it the
2393 /// increments can be computed in fewer registers when chained.
2395 /// TODO: Consider IVInc free if it's already used in another chains.
2397 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2398 ScalarEvolution &SE, const ScalarTargetTransformInfo *STTI) {
2402 if (!Chain.hasIncs())
2405 if (!Users.empty()) {
2406 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2407 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2408 E = Users.end(); I != E; ++I) {
2409 dbgs() << " " << **I << "\n";
2413 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2415 // The chain itself may require a register, so intialize cost to 1.
2418 // A complete chain likely eliminates the need for keeping the original IV in
2419 // a register. LSR does not currently know how to form a complete chain unless
2420 // the header phi already exists.
2421 if (isa<PHINode>(Chain.tailUserInst())
2422 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2425 const SCEV *LastIncExpr = 0;
2426 unsigned NumConstIncrements = 0;
2427 unsigned NumVarIncrements = 0;
2428 unsigned NumReusedIncrements = 0;
2429 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2432 if (I->IncExpr->isZero())
2435 // Incrementing by zero or some constant is neutral. We assume constants can
2436 // be folded into an addressing mode or an add's immediate operand.
2437 if (isa<SCEVConstant>(I->IncExpr)) {
2438 ++NumConstIncrements;
2442 if (I->IncExpr == LastIncExpr)
2443 ++NumReusedIncrements;
2447 LastIncExpr = I->IncExpr;
2449 // An IV chain with a single increment is handled by LSR's postinc
2450 // uses. However, a chain with multiple increments requires keeping the IV's
2451 // value live longer than it needs to be if chained.
2452 if (NumConstIncrements > 1)
2455 // Materializing increment expressions in the preheader that didn't exist in
2456 // the original code may cost a register. For example, sign-extended array
2457 // indices can produce ridiculous increments like this:
2458 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2459 cost += NumVarIncrements;
2461 // Reusing variable increments likely saves a register to hold the multiple of
2463 cost -= NumReusedIncrements;
2465 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2471 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2473 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2474 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2475 // When IVs are used as types of varying widths, they are generally converted
2476 // to a wider type with some uses remaining narrow under a (free) trunc.
2477 Value *const NextIV = getWideOperand(IVOper);
2478 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2479 const SCEV *const OperExprBase = getExprBase(OperExpr);
2481 // Visit all existing chains. Check if its IVOper can be computed as a
2482 // profitable loop invariant increment from the last link in the Chain.
2483 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2484 const SCEV *LastIncExpr = 0;
2485 for (; ChainIdx < NChains; ++ChainIdx) {
2486 IVChain &Chain = IVChainVec[ChainIdx];
2488 // Prune the solution space aggressively by checking that both IV operands
2489 // are expressions that operate on the same unscaled SCEVUnknown. This
2490 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2491 // first avoids creating extra SCEV expressions.
2492 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2495 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2496 if (!isCompatibleIVType(PrevIV, NextIV))
2499 // A phi node terminates a chain.
2500 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2503 // The increment must be loop-invariant so it can be kept in a register.
2504 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2505 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2506 if (!SE.isLoopInvariant(IncExpr, L))
2509 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2510 LastIncExpr = IncExpr;
2514 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2515 // bother for phi nodes, because they must be last in the chain.
2516 if (ChainIdx == NChains) {
2517 if (isa<PHINode>(UserInst))
2519 if (NChains >= MaxChains && !StressIVChain) {
2520 DEBUG(dbgs() << "IV Chain Limit\n");
2523 LastIncExpr = OperExpr;
2524 // IVUsers may have skipped over sign/zero extensions. We don't currently
2525 // attempt to form chains involving extensions unless they can be hoisted
2526 // into this loop's AddRec.
2527 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2530 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2532 ChainUsersVec.resize(NChains);
2533 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2534 << ") IV=" << *LastIncExpr << "\n");
2536 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2537 << ") IV+" << *LastIncExpr << "\n");
2538 // Add this IV user to the end of the chain.
2539 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2542 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2543 // This chain's NearUsers become FarUsers.
2544 if (!LastIncExpr->isZero()) {
2545 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2550 // All other uses of IVOperand become near uses of the chain.
2551 // We currently ignore intermediate values within SCEV expressions, assuming
2552 // they will eventually be used be the current chain, or can be computed
2553 // from one of the chain increments. To be more precise we could
2554 // transitively follow its user and only add leaf IV users to the set.
2555 for (Value::use_iterator UseIter = IVOper->use_begin(),
2556 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2557 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2558 if (!OtherUse || OtherUse == UserInst)
2560 if (SE.isSCEVable(OtherUse->getType())
2561 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2562 && IU.isIVUserOrOperand(OtherUse)) {
2565 NearUsers.insert(OtherUse);
2568 // Since this user is part of the chain, it's no longer considered a use
2570 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2573 /// CollectChains - Populate the vector of Chains.
2575 /// This decreases ILP at the architecture level. Targets with ample registers,
2576 /// multiple memory ports, and no register renaming probably don't want
2577 /// this. However, such targets should probably disable LSR altogether.
2579 /// The job of LSR is to make a reasonable choice of induction variables across
2580 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2581 /// ILP *within the loop* if the target wants it.
2583 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2584 /// will not reorder memory operations, it will recognize this as a chain, but
2585 /// will generate redundant IV increments. Ideally this would be corrected later
2586 /// by a smart scheduler:
2592 /// TODO: Walk the entire domtree within this loop, not just the path to the
2593 /// loop latch. This will discover chains on side paths, but requires
2594 /// maintaining multiple copies of the Chains state.
2595 void LSRInstance::CollectChains() {
2596 DEBUG(dbgs() << "Collecting IV Chains.\n");
2597 SmallVector<ChainUsers, 8> ChainUsersVec;
2599 SmallVector<BasicBlock *,8> LatchPath;
2600 BasicBlock *LoopHeader = L->getHeader();
2601 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2602 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2603 LatchPath.push_back(Rung->getBlock());
2605 LatchPath.push_back(LoopHeader);
2607 // Walk the instruction stream from the loop header to the loop latch.
2608 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2609 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2610 BBIter != BBEnd; ++BBIter) {
2611 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2613 // Skip instructions that weren't seen by IVUsers analysis.
2614 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2617 // Ignore users that are part of a SCEV expression. This way we only
2618 // consider leaf IV Users. This effectively rediscovers a portion of
2619 // IVUsers analysis but in program order this time.
2620 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2623 // Remove this instruction from any NearUsers set it may be in.
2624 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2625 ChainIdx < NChains; ++ChainIdx) {
2626 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2628 // Search for operands that can be chained.
2629 SmallPtrSet<Instruction*, 4> UniqueOperands;
2630 User::op_iterator IVOpEnd = I->op_end();
2631 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2632 while (IVOpIter != IVOpEnd) {
2633 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2634 if (UniqueOperands.insert(IVOpInst))
2635 ChainInstruction(I, IVOpInst, ChainUsersVec);
2636 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2638 } // Continue walking down the instructions.
2639 } // Continue walking down the domtree.
2640 // Visit phi backedges to determine if the chain can generate the IV postinc.
2641 for (BasicBlock::iterator I = L->getHeader()->begin();
2642 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2643 if (!SE.isSCEVable(PN->getType()))
2647 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2649 ChainInstruction(PN, IncV, ChainUsersVec);
2651 // Remove any unprofitable chains.
2652 unsigned ChainIdx = 0;
2653 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2654 UsersIdx < NChains; ++UsersIdx) {
2655 if (!isProfitableChain(IVChainVec[UsersIdx],
2656 ChainUsersVec[UsersIdx].FarUsers, SE, STTI))
2658 // Preserve the chain at UsesIdx.
2659 if (ChainIdx != UsersIdx)
2660 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2661 FinalizeChain(IVChainVec[ChainIdx]);
2664 IVChainVec.resize(ChainIdx);
2667 void LSRInstance::FinalizeChain(IVChain &Chain) {
2668 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2669 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2671 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2673 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2674 User::op_iterator UseI =
2675 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2676 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2677 IVIncSet.insert(UseI);
2681 /// Return true if the IVInc can be folded into an addressing mode.
2682 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2684 const ScalarTargetTransformInfo *STTI) {
2685 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2686 if (!IncConst || !isAddressUse(UserInst, Operand))
2689 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2692 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2693 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false,
2694 LSRUse::Address, getAccessType(UserInst), STTI))
2700 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2701 /// materialize the IV user's operand from the previous IV user's operand.
2702 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2703 SmallVectorImpl<WeakVH> &DeadInsts) {
2704 // Find the new IVOperand for the head of the chain. It may have been replaced
2706 const IVInc &Head = Chain.Incs[0];
2707 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2708 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2711 while (IVOpIter != IVOpEnd) {
2712 IVSrc = getWideOperand(*IVOpIter);
2714 // If this operand computes the expression that the chain needs, we may use
2715 // it. (Check this after setting IVSrc which is used below.)
2717 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2718 // narrow for the chain, so we can no longer use it. We do allow using a
2719 // wider phi, assuming the LSR checked for free truncation. In that case we
2720 // should already have a truncate on this operand such that
2721 // getSCEV(IVSrc) == IncExpr.
2722 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2723 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2726 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2728 if (IVOpIter == IVOpEnd) {
2729 // Gracefully give up on this chain.
2730 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2734 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2735 Type *IVTy = IVSrc->getType();
2736 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2737 const SCEV *LeftOverExpr = 0;
2738 for (IVChain::const_iterator IncI = Chain.begin(),
2739 IncE = Chain.end(); IncI != IncE; ++IncI) {
2741 Instruction *InsertPt = IncI->UserInst;
2742 if (isa<PHINode>(InsertPt))
2743 InsertPt = L->getLoopLatch()->getTerminator();
2745 // IVOper will replace the current IV User's operand. IVSrc is the IV
2746 // value currently held in a register.
2747 Value *IVOper = IVSrc;
2748 if (!IncI->IncExpr->isZero()) {
2749 // IncExpr was the result of subtraction of two narrow values, so must
2751 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2752 LeftOverExpr = LeftOverExpr ?
2753 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2755 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2756 // Expand the IV increment.
2757 Rewriter.clearPostInc();
2758 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2759 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2760 SE.getUnknown(IncV));
2761 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2763 // If an IV increment can't be folded, use it as the next IV value.
2764 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2766 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2771 Type *OperTy = IncI->IVOperand->getType();
2772 if (IVTy != OperTy) {
2773 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2774 "cannot extend a chained IV");
2775 IRBuilder<> Builder(InsertPt);
2776 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2778 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2779 DeadInsts.push_back(IncI->IVOperand);
2781 // If LSR created a new, wider phi, we may also replace its postinc. We only
2782 // do this if we also found a wide value for the head of the chain.
2783 if (isa<PHINode>(Chain.tailUserInst())) {
2784 for (BasicBlock::iterator I = L->getHeader()->begin();
2785 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2786 if (!isCompatibleIVType(Phi, IVSrc))
2788 Instruction *PostIncV = dyn_cast<Instruction>(
2789 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2790 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2792 Value *IVOper = IVSrc;
2793 Type *PostIncTy = PostIncV->getType();
2794 if (IVTy != PostIncTy) {
2795 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2796 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2797 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2798 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2800 Phi->replaceUsesOfWith(PostIncV, IVOper);
2801 DeadInsts.push_back(PostIncV);
2806 void LSRInstance::CollectFixupsAndInitialFormulae() {
2807 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2808 Instruction *UserInst = UI->getUser();
2809 // Skip IV users that are part of profitable IV Chains.
2810 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2811 UI->getOperandValToReplace());
2812 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2813 if (IVIncSet.count(UseI))
2817 LSRFixup &LF = getNewFixup();
2818 LF.UserInst = UserInst;
2819 LF.OperandValToReplace = UI->getOperandValToReplace();
2820 LF.PostIncLoops = UI->getPostIncLoops();
2822 LSRUse::KindType Kind = LSRUse::Basic;
2824 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2825 Kind = LSRUse::Address;
2826 AccessTy = getAccessType(LF.UserInst);
2829 const SCEV *S = IU.getExpr(*UI);
2831 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2832 // (N - i == 0), and this allows (N - i) to be the expression that we work
2833 // with rather than just N or i, so we can consider the register
2834 // requirements for both N and i at the same time. Limiting this code to
2835 // equality icmps is not a problem because all interesting loops use
2836 // equality icmps, thanks to IndVarSimplify.
2837 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2838 if (CI->isEquality()) {
2839 // Swap the operands if needed to put the OperandValToReplace on the
2840 // left, for consistency.
2841 Value *NV = CI->getOperand(1);
2842 if (NV == LF.OperandValToReplace) {
2843 CI->setOperand(1, CI->getOperand(0));
2844 CI->setOperand(0, NV);
2845 NV = CI->getOperand(1);
2849 // x == y --> x - y == 0
2850 const SCEV *N = SE.getSCEV(NV);
2851 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N)) {
2852 // S is normalized, so normalize N before folding it into S
2853 // to keep the result normalized.
2854 N = TransformForPostIncUse(Normalize, N, CI, 0,
2855 LF.PostIncLoops, SE, DT);
2856 Kind = LSRUse::ICmpZero;
2857 S = SE.getMinusSCEV(N, S);
2860 // -1 and the negations of all interesting strides (except the negation
2861 // of -1) are now also interesting.
2862 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2863 if (Factors[i] != -1)
2864 Factors.insert(-(uint64_t)Factors[i]);
2868 // Set up the initial formula for this use.
2869 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2871 LF.Offset = P.second;
2872 LSRUse &LU = Uses[LF.LUIdx];
2873 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2874 if (!LU.WidestFixupType ||
2875 SE.getTypeSizeInBits(LU.WidestFixupType) <
2876 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2877 LU.WidestFixupType = LF.OperandValToReplace->getType();
2879 // If this is the first use of this LSRUse, give it a formula.
2880 if (LU.Formulae.empty()) {
2881 InsertInitialFormula(S, LU, LF.LUIdx);
2882 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2886 DEBUG(print_fixups(dbgs()));
2889 /// InsertInitialFormula - Insert a formula for the given expression into
2890 /// the given use, separating out loop-variant portions from loop-invariant
2891 /// and loop-computable portions.
2893 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2895 F.InitialMatch(S, L, SE);
2896 bool Inserted = InsertFormula(LU, LUIdx, F);
2897 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2900 /// InsertSupplementalFormula - Insert a simple single-register formula for
2901 /// the given expression into the given use.
2903 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2904 LSRUse &LU, size_t LUIdx) {
2906 F.BaseRegs.push_back(S);
2907 F.AM.HasBaseReg = true;
2908 bool Inserted = InsertFormula(LU, LUIdx, F);
2909 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2912 /// CountRegisters - Note which registers are used by the given formula,
2913 /// updating RegUses.
2914 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2916 RegUses.CountRegister(F.ScaledReg, LUIdx);
2917 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2918 E = F.BaseRegs.end(); I != E; ++I)
2919 RegUses.CountRegister(*I, LUIdx);
2922 /// InsertFormula - If the given formula has not yet been inserted, add it to
2923 /// the list, and return true. Return false otherwise.
2924 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2925 if (!LU.InsertFormula(F))
2928 CountRegisters(F, LUIdx);
2932 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2933 /// loop-invariant values which we're tracking. These other uses will pin these
2934 /// values in registers, making them less profitable for elimination.
2935 /// TODO: This currently misses non-constant addrec step registers.
2936 /// TODO: Should this give more weight to users inside the loop?
2938 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2939 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2940 SmallPtrSet<const SCEV *, 8> Inserted;
2942 while (!Worklist.empty()) {
2943 const SCEV *S = Worklist.pop_back_val();
2945 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2946 Worklist.append(N->op_begin(), N->op_end());
2947 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2948 Worklist.push_back(C->getOperand());
2949 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2950 Worklist.push_back(D->getLHS());
2951 Worklist.push_back(D->getRHS());
2952 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2953 if (!Inserted.insert(U)) continue;
2954 const Value *V = U->getValue();
2955 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2956 // Look for instructions defined outside the loop.
2957 if (L->contains(Inst)) continue;
2958 } else if (isa<UndefValue>(V))
2959 // Undef doesn't have a live range, so it doesn't matter.
2961 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2963 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2964 // Ignore non-instructions.
2967 // Ignore instructions in other functions (as can happen with
2969 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2971 // Ignore instructions not dominated by the loop.
2972 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2973 UserInst->getParent() :
2974 cast<PHINode>(UserInst)->getIncomingBlock(
2975 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2976 if (!DT.dominates(L->getHeader(), UseBB))
2978 // Ignore uses which are part of other SCEV expressions, to avoid
2979 // analyzing them multiple times.
2980 if (SE.isSCEVable(UserInst->getType())) {
2981 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2982 // If the user is a no-op, look through to its uses.
2983 if (!isa<SCEVUnknown>(UserS))
2987 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2991 // Ignore icmp instructions which are already being analyzed.
2992 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2993 unsigned OtherIdx = !UI.getOperandNo();
2994 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2995 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2999 LSRFixup &LF = getNewFixup();
3000 LF.UserInst = const_cast<Instruction *>(UserInst);
3001 LF.OperandValToReplace = UI.getUse();
3002 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
3004 LF.Offset = P.second;
3005 LSRUse &LU = Uses[LF.LUIdx];
3006 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3007 if (!LU.WidestFixupType ||
3008 SE.getTypeSizeInBits(LU.WidestFixupType) <
3009 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3010 LU.WidestFixupType = LF.OperandValToReplace->getType();
3011 InsertSupplementalFormula(U, LU, LF.LUIdx);
3012 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3019 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3020 /// separate registers. If C is non-null, multiply each subexpression by C.
3022 /// Return remainder expression after factoring the subexpressions captured by
3023 /// Ops. If Ops is complete, return NULL.
3024 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3025 SmallVectorImpl<const SCEV *> &Ops,
3027 ScalarEvolution &SE,
3028 unsigned Depth = 0) {
3029 // Arbitrarily cap recursion to protect compile time.
3033 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3034 // Break out add operands.
3035 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3037 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3039 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3042 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3043 // Split a non-zero base out of an addrec.
3044 if (AR->getStart()->isZero())
3047 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3048 C, Ops, L, SE, Depth+1);
3049 // Split the non-zero AddRec unless it is part of a nested recurrence that
3050 // does not pertain to this loop.
3051 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3052 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3055 if (Remainder != AR->getStart()) {
3057 Remainder = SE.getConstant(AR->getType(), 0);
3058 return SE.getAddRecExpr(Remainder,
3059 AR->getStepRecurrence(SE),
3061 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3064 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3065 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3066 if (Mul->getNumOperands() != 2)
3068 if (const SCEVConstant *Op0 =
3069 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3070 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3071 const SCEV *Remainder =
3072 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3074 Ops.push_back(SE.getMulExpr(C, Remainder));
3081 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3083 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3086 // Arbitrarily cap recursion to protect compile time.
3087 if (Depth >= 3) return;
3089 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3090 const SCEV *BaseReg = Base.BaseRegs[i];
3092 SmallVector<const SCEV *, 8> AddOps;
3093 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3095 AddOps.push_back(Remainder);
3097 if (AddOps.size() == 1) continue;
3099 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3100 JE = AddOps.end(); J != JE; ++J) {
3102 // Loop-variant "unknown" values are uninteresting; we won't be able to
3103 // do anything meaningful with them.
3104 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3107 // Don't pull a constant into a register if the constant could be folded
3108 // into an immediate field.
3109 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
3110 Base.getNumRegs() > 1,
3111 LU.Kind, LU.AccessTy, STTI, SE))
3114 // Collect all operands except *J.
3115 SmallVector<const SCEV *, 8> InnerAddOps
3116 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3118 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3120 // Don't leave just a constant behind in a register if the constant could
3121 // be folded into an immediate field.
3122 if (InnerAddOps.size() == 1 &&
3123 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
3124 Base.getNumRegs() > 1,
3125 LU.Kind, LU.AccessTy, STTI, SE))
3128 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3129 if (InnerSum->isZero())
3133 // Add the remaining pieces of the add back into the new formula.
3134 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3135 if (STTI && InnerSumSC &&
3136 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3137 STTI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3138 InnerSumSC->getValue()->getZExtValue())) {
3139 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3140 InnerSumSC->getValue()->getZExtValue();
3141 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3143 F.BaseRegs[i] = InnerSum;
3145 // Add J as its own register, or an unfolded immediate.
3146 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3147 if (STTI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3148 STTI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3149 SC->getValue()->getZExtValue()))
3150 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3151 SC->getValue()->getZExtValue();
3153 F.BaseRegs.push_back(*J);
3155 if (InsertFormula(LU, LUIdx, F))
3156 // If that formula hadn't been seen before, recurse to find more like
3158 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3163 /// GenerateCombinations - Generate a formula consisting of all of the
3164 /// loop-dominating registers added into a single register.
3165 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3167 // This method is only interesting on a plurality of registers.
3168 if (Base.BaseRegs.size() <= 1) return;
3172 SmallVector<const SCEV *, 4> Ops;
3173 for (SmallVectorImpl<const SCEV *>::const_iterator
3174 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3175 const SCEV *BaseReg = *I;
3176 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3177 !SE.hasComputableLoopEvolution(BaseReg, L))
3178 Ops.push_back(BaseReg);
3180 F.BaseRegs.push_back(BaseReg);
3182 if (Ops.size() > 1) {
3183 const SCEV *Sum = SE.getAddExpr(Ops);
3184 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3185 // opportunity to fold something. For now, just ignore such cases
3186 // rather than proceed with zero in a register.
3187 if (!Sum->isZero()) {
3188 F.BaseRegs.push_back(Sum);
3189 (void)InsertFormula(LU, LUIdx, F);
3194 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3195 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3197 // We can't add a symbolic offset if the address already contains one.
3198 if (Base.AM.BaseGV) return;
3200 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3201 const SCEV *G = Base.BaseRegs[i];
3202 GlobalValue *GV = ExtractSymbol(G, SE);
3203 if (G->isZero() || !GV)
3207 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3208 LU.Kind, LU.AccessTy, STTI))
3211 (void)InsertFormula(LU, LUIdx, F);
3215 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3216 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3218 // TODO: For now, just add the min and max offset, because it usually isn't
3219 // worthwhile looking at everything inbetween.
3220 SmallVector<int64_t, 2> Worklist;
3221 Worklist.push_back(LU.MinOffset);
3222 if (LU.MaxOffset != LU.MinOffset)
3223 Worklist.push_back(LU.MaxOffset);
3225 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3226 const SCEV *G = Base.BaseRegs[i];
3228 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3229 E = Worklist.end(); I != E; ++I) {
3231 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
3232 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
3233 LU.Kind, LU.AccessTy, STTI)) {
3234 // Add the offset to the base register.
3235 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3236 // If it cancelled out, drop the base register, otherwise update it.
3237 if (NewG->isZero()) {
3238 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3239 F.BaseRegs.pop_back();
3241 F.BaseRegs[i] = NewG;
3243 (void)InsertFormula(LU, LUIdx, F);
3247 int64_t Imm = ExtractImmediate(G, SE);
3248 if (G->isZero() || Imm == 0)
3251 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
3252 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3253 LU.Kind, LU.AccessTy, STTI))
3256 (void)InsertFormula(LU, LUIdx, F);
3260 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3261 /// the comparison. For example, x == y -> x*c == y*c.
3262 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3264 if (LU.Kind != LSRUse::ICmpZero) return;
3266 // Determine the integer type for the base formula.
3267 Type *IntTy = Base.getType();
3269 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3271 // Don't do this if there is more than one offset.
3272 if (LU.MinOffset != LU.MaxOffset) return;
3274 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
3276 // Check each interesting stride.
3277 for (SmallSetVector<int64_t, 8>::const_iterator
3278 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3279 int64_t Factor = *I;
3281 // Check that the multiplication doesn't overflow.
3282 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
3284 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
3285 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
3288 // Check that multiplying with the use offset doesn't overflow.
3289 int64_t Offset = LU.MinOffset;
3290 if (Offset == INT64_MIN && Factor == -1)
3292 Offset = (uint64_t)Offset * Factor;
3293 if (Offset / Factor != LU.MinOffset)
3297 F.AM.BaseOffs = NewBaseOffs;
3299 // Check that this scale is legal.
3300 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, STTI))
3303 // Compensate for the use having MinOffset built into it.
3304 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
3306 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3308 // Check that multiplying with each base register doesn't overflow.
3309 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3310 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3311 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3315 // Check that multiplying with the scaled register doesn't overflow.
3317 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3318 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3322 // Check that multiplying with the unfolded offset doesn't overflow.
3323 if (F.UnfoldedOffset != 0) {
3324 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3326 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3327 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3331 // If we make it here and it's legal, add it.
3332 (void)InsertFormula(LU, LUIdx, F);
3337 /// GenerateScales - Generate stride factor reuse formulae by making use of
3338 /// scaled-offset address modes, for example.
3339 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3340 // Determine the integer type for the base formula.
3341 Type *IntTy = Base.getType();
3344 // If this Formula already has a scaled register, we can't add another one.
3345 if (Base.AM.Scale != 0) return;
3347 // Check each interesting stride.
3348 for (SmallSetVector<int64_t, 8>::const_iterator
3349 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3350 int64_t Factor = *I;
3352 Base.AM.Scale = Factor;
3353 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
3354 // Check whether this scale is going to be legal.
3355 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3356 LU.Kind, LU.AccessTy, STTI)) {
3357 // As a special-case, handle special out-of-loop Basic users specially.
3358 // TODO: Reconsider this special case.
3359 if (LU.Kind == LSRUse::Basic &&
3360 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3361 LSRUse::Special, LU.AccessTy, STTI) &&
3362 LU.AllFixupsOutsideLoop)
3363 LU.Kind = LSRUse::Special;
3367 // For an ICmpZero, negating a solitary base register won't lead to
3369 if (LU.Kind == LSRUse::ICmpZero &&
3370 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
3372 // For each addrec base reg, apply the scale, if possible.
3373 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3374 if (const SCEVAddRecExpr *AR =
3375 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3376 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3377 if (FactorS->isZero())
3379 // Divide out the factor, ignoring high bits, since we'll be
3380 // scaling the value back up in the end.
3381 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3382 // TODO: This could be optimized to avoid all the copying.
3384 F.ScaledReg = Quotient;
3385 F.DeleteBaseReg(F.BaseRegs[i]);
3386 (void)InsertFormula(LU, LUIdx, F);
3392 /// GenerateTruncates - Generate reuse formulae from different IV types.
3393 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3394 // This requires ScalarTargetTransformInfo to tell us which truncates are free.
3397 // Don't bother truncating symbolic values.
3398 if (Base.AM.BaseGV) return;
3400 // Determine the integer type for the base formula.
3401 Type *DstTy = Base.getType();
3403 DstTy = SE.getEffectiveSCEVType(DstTy);
3405 for (SmallSetVector<Type *, 4>::const_iterator
3406 I = Types.begin(), E = Types.end(); I != E; ++I) {
3408 if (SrcTy != DstTy && STTI->isTruncateFree(SrcTy, DstTy)) {
3411 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3412 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3413 JE = F.BaseRegs.end(); J != JE; ++J)
3414 *J = SE.getAnyExtendExpr(*J, SrcTy);
3416 // TODO: This assumes we've done basic processing on all uses and
3417 // have an idea what the register usage is.
3418 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3421 (void)InsertFormula(LU, LUIdx, F);
3428 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3429 /// defer modifications so that the search phase doesn't have to worry about
3430 /// the data structures moving underneath it.
3434 const SCEV *OrigReg;
3436 WorkItem(size_t LI, int64_t I, const SCEV *R)
3437 : LUIdx(LI), Imm(I), OrigReg(R) {}
3439 void print(raw_ostream &OS) const;
3445 void WorkItem::print(raw_ostream &OS) const {
3446 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3447 << " , add offset " << Imm;
3450 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3451 void WorkItem::dump() const {
3452 print(errs()); errs() << '\n';
3456 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3457 /// distance apart and try to form reuse opportunities between them.
3458 void LSRInstance::GenerateCrossUseConstantOffsets() {
3459 // Group the registers by their value without any added constant offset.
3460 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3461 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3463 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3464 SmallVector<const SCEV *, 8> Sequence;
3465 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3467 const SCEV *Reg = *I;
3468 int64_t Imm = ExtractImmediate(Reg, SE);
3469 std::pair<RegMapTy::iterator, bool> Pair =
3470 Map.insert(std::make_pair(Reg, ImmMapTy()));
3472 Sequence.push_back(Reg);
3473 Pair.first->second.insert(std::make_pair(Imm, *I));
3474 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3477 // Now examine each set of registers with the same base value. Build up
3478 // a list of work to do and do the work in a separate step so that we're
3479 // not adding formulae and register counts while we're searching.
3480 SmallVector<WorkItem, 32> WorkItems;
3481 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3482 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3483 E = Sequence.end(); I != E; ++I) {
3484 const SCEV *Reg = *I;
3485 const ImmMapTy &Imms = Map.find(Reg)->second;
3487 // It's not worthwhile looking for reuse if there's only one offset.
3488 if (Imms.size() == 1)
3491 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3492 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3494 dbgs() << ' ' << J->first;
3497 // Examine each offset.
3498 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3500 const SCEV *OrigReg = J->second;
3502 int64_t JImm = J->first;
3503 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3505 if (!isa<SCEVConstant>(OrigReg) &&
3506 UsedByIndicesMap[Reg].count() == 1) {
3507 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3511 // Conservatively examine offsets between this orig reg a few selected
3513 ImmMapTy::const_iterator OtherImms[] = {
3514 Imms.begin(), prior(Imms.end()),
3515 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3517 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3518 ImmMapTy::const_iterator M = OtherImms[i];
3519 if (M == J || M == JE) continue;
3521 // Compute the difference between the two.
3522 int64_t Imm = (uint64_t)JImm - M->first;
3523 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3524 LUIdx = UsedByIndices.find_next(LUIdx))
3525 // Make a memo of this use, offset, and register tuple.
3526 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3527 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3534 UsedByIndicesMap.clear();
3535 UniqueItems.clear();
3537 // Now iterate through the worklist and add new formulae.
3538 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3539 E = WorkItems.end(); I != E; ++I) {
3540 const WorkItem &WI = *I;
3541 size_t LUIdx = WI.LUIdx;
3542 LSRUse &LU = Uses[LUIdx];
3543 int64_t Imm = WI.Imm;
3544 const SCEV *OrigReg = WI.OrigReg;
3546 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3547 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3548 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3550 // TODO: Use a more targeted data structure.
3551 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3552 const Formula &F = LU.Formulae[L];
3553 // Use the immediate in the scaled register.
3554 if (F.ScaledReg == OrigReg) {
3555 int64_t Offs = (uint64_t)F.AM.BaseOffs +
3556 Imm * (uint64_t)F.AM.Scale;
3557 // Don't create 50 + reg(-50).
3558 if (F.referencesReg(SE.getSCEV(
3559 ConstantInt::get(IntTy, -(uint64_t)Offs))))
3562 NewF.AM.BaseOffs = Offs;
3563 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3564 LU.Kind, LU.AccessTy, STTI))
3566 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3568 // If the new scale is a constant in a register, and adding the constant
3569 // value to the immediate would produce a value closer to zero than the
3570 // immediate itself, then the formula isn't worthwhile.
3571 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3572 if (C->getValue()->isNegative() !=
3573 (NewF.AM.BaseOffs < 0) &&
3574 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
3575 .ule(abs64(NewF.AM.BaseOffs)))
3579 (void)InsertFormula(LU, LUIdx, NewF);
3581 // Use the immediate in a base register.
3582 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3583 const SCEV *BaseReg = F.BaseRegs[N];
3584 if (BaseReg != OrigReg)
3587 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
3588 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3589 LU.Kind, LU.AccessTy, STTI)) {
3591 !STTI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3594 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3596 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3598 // If the new formula has a constant in a register, and adding the
3599 // constant value to the immediate would produce a value closer to
3600 // zero than the immediate itself, then the formula isn't worthwhile.
3601 for (SmallVectorImpl<const SCEV *>::const_iterator
3602 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3604 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3605 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
3606 abs64(NewF.AM.BaseOffs)) &&
3607 (C->getValue()->getValue() +
3608 NewF.AM.BaseOffs).countTrailingZeros() >=
3609 CountTrailingZeros_64(NewF.AM.BaseOffs))
3613 (void)InsertFormula(LU, LUIdx, NewF);
3622 /// GenerateAllReuseFormulae - Generate formulae for each use.
3624 LSRInstance::GenerateAllReuseFormulae() {
3625 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3626 // queries are more precise.
3627 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3628 LSRUse &LU = Uses[LUIdx];
3629 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3630 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3631 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3632 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3634 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3635 LSRUse &LU = Uses[LUIdx];
3636 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3637 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3638 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3639 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3640 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3641 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3642 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3643 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3645 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3646 LSRUse &LU = Uses[LUIdx];
3647 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3648 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3651 GenerateCrossUseConstantOffsets();
3653 DEBUG(dbgs() << "\n"
3654 "After generating reuse formulae:\n";
3655 print_uses(dbgs()));
3658 /// If there are multiple formulae with the same set of registers used
3659 /// by other uses, pick the best one and delete the others.
3660 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3661 DenseSet<const SCEV *> VisitedRegs;
3662 SmallPtrSet<const SCEV *, 16> Regs;
3663 SmallPtrSet<const SCEV *, 16> LoserRegs;
3665 bool ChangedFormulae = false;
3668 // Collect the best formula for each unique set of shared registers. This
3669 // is reset for each use.
3670 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
3672 BestFormulaeTy BestFormulae;
3674 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3675 LSRUse &LU = Uses[LUIdx];
3676 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3679 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3680 FIdx != NumForms; ++FIdx) {
3681 Formula &F = LU.Formulae[FIdx];
3683 // Some formulas are instant losers. For example, they may depend on
3684 // nonexistent AddRecs from other loops. These need to be filtered
3685 // immediately, otherwise heuristics could choose them over others leading
3686 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3687 // avoids the need to recompute this information across formulae using the
3688 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3689 // the corresponding bad register from the Regs set.
3692 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3694 if (CostF.isLoser()) {
3695 // During initial formula generation, undesirable formulae are generated
3696 // by uses within other loops that have some non-trivial address mode or
3697 // use the postinc form of the IV. LSR needs to provide these formulae
3698 // as the basis of rediscovering the desired formula that uses an AddRec
3699 // corresponding to the existing phi. Once all formulae have been
3700 // generated, these initial losers may be pruned.
3701 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3705 SmallVector<const SCEV *, 2> Key;
3706 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3707 JE = F.BaseRegs.end(); J != JE; ++J) {
3708 const SCEV *Reg = *J;
3709 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3713 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3714 Key.push_back(F.ScaledReg);
3715 // Unstable sort by host order ok, because this is only used for
3717 std::sort(Key.begin(), Key.end());
3719 std::pair<BestFormulaeTy::const_iterator, bool> P =
3720 BestFormulae.insert(std::make_pair(Key, FIdx));
3724 Formula &Best = LU.Formulae[P.first->second];
3728 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3729 if (CostF < CostBest)
3731 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3733 " in favor of formula "; Best.print(dbgs());
3737 ChangedFormulae = true;
3739 LU.DeleteFormula(F);
3745 // Now that we've filtered out some formulae, recompute the Regs set.
3747 LU.RecomputeRegs(LUIdx, RegUses);
3749 // Reset this to prepare for the next use.
3750 BestFormulae.clear();
3753 DEBUG(if (ChangedFormulae) {
3755 "After filtering out undesirable candidates:\n";
3760 // This is a rough guess that seems to work fairly well.
3761 static const size_t ComplexityLimit = UINT16_MAX;
3763 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3764 /// solutions the solver might have to consider. It almost never considers
3765 /// this many solutions because it prune the search space, but the pruning
3766 /// isn't always sufficient.
3767 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3769 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3770 E = Uses.end(); I != E; ++I) {
3771 size_t FSize = I->Formulae.size();
3772 if (FSize >= ComplexityLimit) {
3773 Power = ComplexityLimit;
3777 if (Power >= ComplexityLimit)
3783 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3784 /// of the registers of another formula, it won't help reduce register
3785 /// pressure (though it may not necessarily hurt register pressure); remove
3786 /// it to simplify the system.
3787 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3788 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3789 DEBUG(dbgs() << "The search space is too complex.\n");
3791 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3792 "which use a superset of registers used by other "
3795 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3796 LSRUse &LU = Uses[LUIdx];
3798 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3799 Formula &F = LU.Formulae[i];
3800 // Look for a formula with a constant or GV in a register. If the use
3801 // also has a formula with that same value in an immediate field,
3802 // delete the one that uses a register.
3803 for (SmallVectorImpl<const SCEV *>::const_iterator
3804 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3805 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3807 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3808 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3809 (I - F.BaseRegs.begin()));
3810 if (LU.HasFormulaWithSameRegs(NewF)) {
3811 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3812 LU.DeleteFormula(F);
3818 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3819 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3822 NewF.AM.BaseGV = GV;
3823 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3824 (I - F.BaseRegs.begin()));
3825 if (LU.HasFormulaWithSameRegs(NewF)) {
3826 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3828 LU.DeleteFormula(F);
3839 LU.RecomputeRegs(LUIdx, RegUses);
3842 DEBUG(dbgs() << "After pre-selection:\n";
3843 print_uses(dbgs()));
3847 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3848 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3850 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3851 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3852 DEBUG(dbgs() << "The search space is too complex.\n");
3854 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3855 "separated by a constant offset will use the same "
3858 // This is especially useful for unrolled loops.
3860 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3861 LSRUse &LU = Uses[LUIdx];
3862 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3863 E = LU.Formulae.end(); I != E; ++I) {
3864 const Formula &F = *I;
3865 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3866 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3867 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3868 /*HasBaseReg=*/false,
3869 LU.Kind, LU.AccessTy)) {
3870 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3873 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3875 // Update the relocs to reference the new use.
3876 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3877 E = Fixups.end(); I != E; ++I) {
3878 LSRFixup &Fixup = *I;
3879 if (Fixup.LUIdx == LUIdx) {
3880 Fixup.LUIdx = LUThatHas - &Uses.front();
3881 Fixup.Offset += F.AM.BaseOffs;
3882 // Add the new offset to LUThatHas' offset list.
3883 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3884 LUThatHas->Offsets.push_back(Fixup.Offset);
3885 if (Fixup.Offset > LUThatHas->MaxOffset)
3886 LUThatHas->MaxOffset = Fixup.Offset;
3887 if (Fixup.Offset < LUThatHas->MinOffset)
3888 LUThatHas->MinOffset = Fixup.Offset;
3890 DEBUG(dbgs() << "New fixup has offset "
3891 << Fixup.Offset << '\n');
3893 if (Fixup.LUIdx == NumUses-1)
3894 Fixup.LUIdx = LUIdx;
3897 // Delete formulae from the new use which are no longer legal.
3899 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3900 Formula &F = LUThatHas->Formulae[i];
3901 if (!isLegalUse(F.AM,
3902 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3903 LUThatHas->Kind, LUThatHas->AccessTy, STTI)) {
3904 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3906 LUThatHas->DeleteFormula(F);
3913 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3915 // Delete the old use.
3916 DeleteUse(LU, LUIdx);
3926 DEBUG(dbgs() << "After pre-selection:\n";
3927 print_uses(dbgs()));
3931 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3932 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3933 /// we've done more filtering, as it may be able to find more formulae to
3935 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3936 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3937 DEBUG(dbgs() << "The search space is too complex.\n");
3939 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3940 "undesirable dedicated registers.\n");
3942 FilterOutUndesirableDedicatedRegisters();
3944 DEBUG(dbgs() << "After pre-selection:\n";
3945 print_uses(dbgs()));
3949 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3950 /// to be profitable, and then in any use which has any reference to that
3951 /// register, delete all formulae which do not reference that register.
3952 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3953 // With all other options exhausted, loop until the system is simple
3954 // enough to handle.
3955 SmallPtrSet<const SCEV *, 4> Taken;
3956 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3957 // Ok, we have too many of formulae on our hands to conveniently handle.
3958 // Use a rough heuristic to thin out the list.
3959 DEBUG(dbgs() << "The search space is too complex.\n");
3961 // Pick the register which is used by the most LSRUses, which is likely
3962 // to be a good reuse register candidate.
3963 const SCEV *Best = 0;
3964 unsigned BestNum = 0;
3965 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3967 const SCEV *Reg = *I;
3968 if (Taken.count(Reg))
3973 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3974 if (Count > BestNum) {
3981 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3982 << " will yield profitable reuse.\n");
3985 // In any use with formulae which references this register, delete formulae
3986 // which don't reference it.
3987 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3988 LSRUse &LU = Uses[LUIdx];
3989 if (!LU.Regs.count(Best)) continue;
3992 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3993 Formula &F = LU.Formulae[i];
3994 if (!F.referencesReg(Best)) {
3995 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3996 LU.DeleteFormula(F);
4000 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4006 LU.RecomputeRegs(LUIdx, RegUses);
4009 DEBUG(dbgs() << "After pre-selection:\n";
4010 print_uses(dbgs()));
4014 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4015 /// formulae to choose from, use some rough heuristics to prune down the number
4016 /// of formulae. This keeps the main solver from taking an extraordinary amount
4017 /// of time in some worst-case scenarios.
4018 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4019 NarrowSearchSpaceByDetectingSupersets();
4020 NarrowSearchSpaceByCollapsingUnrolledCode();
4021 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4022 NarrowSearchSpaceByPickingWinnerRegs();
4025 /// SolveRecurse - This is the recursive solver.
4026 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4028 SmallVectorImpl<const Formula *> &Workspace,
4029 const Cost &CurCost,
4030 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4031 DenseSet<const SCEV *> &VisitedRegs) const {
4034 // - use more aggressive filtering
4035 // - sort the formula so that the most profitable solutions are found first
4036 // - sort the uses too
4038 // - don't compute a cost, and then compare. compare while computing a cost
4040 // - track register sets with SmallBitVector
4042 const LSRUse &LU = Uses[Workspace.size()];
4044 // If this use references any register that's already a part of the
4045 // in-progress solution, consider it a requirement that a formula must
4046 // reference that register in order to be considered. This prunes out
4047 // unprofitable searching.
4048 SmallSetVector<const SCEV *, 4> ReqRegs;
4049 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4050 E = CurRegs.end(); I != E; ++I)
4051 if (LU.Regs.count(*I))
4054 SmallPtrSet<const SCEV *, 16> NewRegs;
4056 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4057 E = LU.Formulae.end(); I != E; ++I) {
4058 const Formula &F = *I;
4060 // Ignore formulae which do not use any of the required registers.
4061 bool SatisfiedReqReg = true;
4062 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4063 JE = ReqRegs.end(); J != JE; ++J) {
4064 const SCEV *Reg = *J;
4065 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4066 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4068 SatisfiedReqReg = false;
4072 if (!SatisfiedReqReg) {
4073 // If none of the formulae satisfied the required registers, then we could
4074 // clear ReqRegs and try again. Currently, we simply give up in this case.
4078 // Evaluate the cost of the current formula. If it's already worse than
4079 // the current best, prune the search at that point.
4082 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
4083 if (NewCost < SolutionCost) {
4084 Workspace.push_back(&F);
4085 if (Workspace.size() != Uses.size()) {
4086 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4087 NewRegs, VisitedRegs);
4088 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4089 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4091 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4092 dbgs() << ".\n Regs:";
4093 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4094 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4095 dbgs() << ' ' << **I;
4098 SolutionCost = NewCost;
4099 Solution = Workspace;
4101 Workspace.pop_back();
4106 /// Solve - Choose one formula from each use. Return the results in the given
4107 /// Solution vector.
4108 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4109 SmallVector<const Formula *, 8> Workspace;
4111 SolutionCost.Loose();
4113 SmallPtrSet<const SCEV *, 16> CurRegs;
4114 DenseSet<const SCEV *> VisitedRegs;
4115 Workspace.reserve(Uses.size());
4117 // SolveRecurse does all the work.
4118 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4119 CurRegs, VisitedRegs);
4120 if (Solution.empty()) {
4121 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4125 // Ok, we've now made all our decisions.
4126 DEBUG(dbgs() << "\n"
4127 "The chosen solution requires "; SolutionCost.print(dbgs());
4129 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4131 Uses[i].print(dbgs());
4134 Solution[i]->print(dbgs());
4138 assert(Solution.size() == Uses.size() && "Malformed solution!");
4141 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4142 /// the dominator tree far as we can go while still being dominated by the
4143 /// input positions. This helps canonicalize the insert position, which
4144 /// encourages sharing.
4145 BasicBlock::iterator
4146 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4147 const SmallVectorImpl<Instruction *> &Inputs)
4150 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4151 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4154 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4155 if (!Rung) return IP;
4156 Rung = Rung->getIDom();
4157 if (!Rung) return IP;
4158 IDom = Rung->getBlock();
4160 // Don't climb into a loop though.
4161 const Loop *IDomLoop = LI.getLoopFor(IDom);
4162 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4163 if (IDomDepth <= IPLoopDepth &&
4164 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4168 bool AllDominate = true;
4169 Instruction *BetterPos = 0;
4170 Instruction *Tentative = IDom->getTerminator();
4171 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4172 E = Inputs.end(); I != E; ++I) {
4173 Instruction *Inst = *I;
4174 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4175 AllDominate = false;
4178 // Attempt to find an insert position in the middle of the block,
4179 // instead of at the end, so that it can be used for other expansions.
4180 if (IDom == Inst->getParent() &&
4181 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4182 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4195 /// AdjustInsertPositionForExpand - Determine an input position which will be
4196 /// dominated by the operands and which will dominate the result.
4197 BasicBlock::iterator
4198 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4201 SCEVExpander &Rewriter) const {
4202 // Collect some instructions which must be dominated by the
4203 // expanding replacement. These must be dominated by any operands that
4204 // will be required in the expansion.
4205 SmallVector<Instruction *, 4> Inputs;
4206 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4207 Inputs.push_back(I);
4208 if (LU.Kind == LSRUse::ICmpZero)
4209 if (Instruction *I =
4210 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4211 Inputs.push_back(I);
4212 if (LF.PostIncLoops.count(L)) {
4213 if (LF.isUseFullyOutsideLoop(L))
4214 Inputs.push_back(L->getLoopLatch()->getTerminator());
4216 Inputs.push_back(IVIncInsertPos);
4218 // The expansion must also be dominated by the increment positions of any
4219 // loops it for which it is using post-inc mode.
4220 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4221 E = LF.PostIncLoops.end(); I != E; ++I) {
4222 const Loop *PIL = *I;
4223 if (PIL == L) continue;
4225 // Be dominated by the loop exit.
4226 SmallVector<BasicBlock *, 4> ExitingBlocks;
4227 PIL->getExitingBlocks(ExitingBlocks);
4228 if (!ExitingBlocks.empty()) {
4229 BasicBlock *BB = ExitingBlocks[0];
4230 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4231 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4232 Inputs.push_back(BB->getTerminator());
4236 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4237 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4238 "Insertion point must be a normal instruction");
4240 // Then, climb up the immediate dominator tree as far as we can go while
4241 // still being dominated by the input positions.
4242 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4244 // Don't insert instructions before PHI nodes.
4245 while (isa<PHINode>(IP)) ++IP;
4247 // Ignore landingpad instructions.
4248 while (isa<LandingPadInst>(IP)) ++IP;
4250 // Ignore debug intrinsics.
4251 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4253 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4254 // IP consistent across expansions and allows the previously inserted
4255 // instructions to be reused by subsequent expansion.
4256 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4261 /// Expand - Emit instructions for the leading candidate expression for this
4262 /// LSRUse (this is called "expanding").
4263 Value *LSRInstance::Expand(const LSRFixup &LF,
4265 BasicBlock::iterator IP,
4266 SCEVExpander &Rewriter,
4267 SmallVectorImpl<WeakVH> &DeadInsts) const {
4268 const LSRUse &LU = Uses[LF.LUIdx];
4270 // Determine an input position which will be dominated by the operands and
4271 // which will dominate the result.
4272 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4274 // Inform the Rewriter if we have a post-increment use, so that it can
4275 // perform an advantageous expansion.
4276 Rewriter.setPostInc(LF.PostIncLoops);
4278 // This is the type that the user actually needs.
4279 Type *OpTy = LF.OperandValToReplace->getType();
4280 // This will be the type that we'll initially expand to.
4281 Type *Ty = F.getType();
4283 // No type known; just expand directly to the ultimate type.
4285 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4286 // Expand directly to the ultimate type if it's the right size.
4288 // This is the type to do integer arithmetic in.
4289 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4291 // Build up a list of operands to add together to form the full base.
4292 SmallVector<const SCEV *, 8> Ops;
4294 // Expand the BaseRegs portion.
4295 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4296 E = F.BaseRegs.end(); I != E; ++I) {
4297 const SCEV *Reg = *I;
4298 assert(!Reg->isZero() && "Zero allocated in a base register!");
4300 // If we're expanding for a post-inc user, make the post-inc adjustment.
4301 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4302 Reg = TransformForPostIncUse(Denormalize, Reg,
4303 LF.UserInst, LF.OperandValToReplace,
4306 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4309 // Expand the ScaledReg portion.
4310 Value *ICmpScaledV = 0;
4311 if (F.AM.Scale != 0) {
4312 const SCEV *ScaledS = F.ScaledReg;
4314 // If we're expanding for a post-inc user, make the post-inc adjustment.
4315 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4316 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4317 LF.UserInst, LF.OperandValToReplace,
4320 if (LU.Kind == LSRUse::ICmpZero) {
4321 // An interesting way of "folding" with an icmp is to use a negated
4322 // scale, which we'll implement by inserting it into the other operand
4324 assert(F.AM.Scale == -1 &&
4325 "The only scale supported by ICmpZero uses is -1!");
4326 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4328 // Otherwise just expand the scaled register and an explicit scale,
4329 // which is expected to be matched as part of the address.
4331 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4332 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4333 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4335 Ops.push_back(SE.getUnknown(FullV));
4337 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4338 ScaledS = SE.getMulExpr(ScaledS,
4339 SE.getConstant(ScaledS->getType(), F.AM.Scale));
4340 Ops.push_back(ScaledS);
4344 // Expand the GV portion.
4346 // Flush the operand list to suppress SCEVExpander hoisting.
4348 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4350 Ops.push_back(SE.getUnknown(FullV));
4352 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
4355 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4356 // unfolded offsets. LSR assumes they both live next to their uses.
4358 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4360 Ops.push_back(SE.getUnknown(FullV));
4363 // Expand the immediate portion.
4364 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
4366 if (LU.Kind == LSRUse::ICmpZero) {
4367 // The other interesting way of "folding" with an ICmpZero is to use a
4368 // negated immediate.
4370 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4372 Ops.push_back(SE.getUnknown(ICmpScaledV));
4373 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4376 // Just add the immediate values. These again are expected to be matched
4377 // as part of the address.
4378 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4382 // Expand the unfolded offset portion.
4383 int64_t UnfoldedOffset = F.UnfoldedOffset;
4384 if (UnfoldedOffset != 0) {
4385 // Just add the immediate values.
4386 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4390 // Emit instructions summing all the operands.
4391 const SCEV *FullS = Ops.empty() ?
4392 SE.getConstant(IntTy, 0) :
4394 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4396 // We're done expanding now, so reset the rewriter.
4397 Rewriter.clearPostInc();
4399 // An ICmpZero Formula represents an ICmp which we're handling as a
4400 // comparison against zero. Now that we've expanded an expression for that
4401 // form, update the ICmp's other operand.
4402 if (LU.Kind == LSRUse::ICmpZero) {
4403 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4404 DeadInsts.push_back(CI->getOperand(1));
4405 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
4406 "a scale at the same time!");
4407 if (F.AM.Scale == -1) {
4408 if (ICmpScaledV->getType() != OpTy) {
4410 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4412 ICmpScaledV, OpTy, "tmp", CI);
4415 CI->setOperand(1, ICmpScaledV);
4417 assert(F.AM.Scale == 0 &&
4418 "ICmp does not support folding a global value and "
4419 "a scale at the same time!");
4420 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4422 if (C->getType() != OpTy)
4423 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4427 CI->setOperand(1, C);
4434 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4435 /// of their operands effectively happens in their predecessor blocks, so the
4436 /// expression may need to be expanded in multiple places.
4437 void LSRInstance::RewriteForPHI(PHINode *PN,
4440 SCEVExpander &Rewriter,
4441 SmallVectorImpl<WeakVH> &DeadInsts,
4443 DenseMap<BasicBlock *, Value *> Inserted;
4444 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4445 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4446 BasicBlock *BB = PN->getIncomingBlock(i);
4448 // If this is a critical edge, split the edge so that we do not insert
4449 // the code on all predecessor/successor paths. We do this unless this
4450 // is the canonical backedge for this loop, which complicates post-inc
4452 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4453 !isa<IndirectBrInst>(BB->getTerminator())) {
4454 BasicBlock *Parent = PN->getParent();
4455 Loop *PNLoop = LI.getLoopFor(Parent);
4456 if (!PNLoop || Parent != PNLoop->getHeader()) {
4457 // Split the critical edge.
4458 BasicBlock *NewBB = 0;
4459 if (!Parent->isLandingPad()) {
4460 NewBB = SplitCriticalEdge(BB, Parent, P,
4461 /*MergeIdenticalEdges=*/true,
4462 /*DontDeleteUselessPhis=*/true);
4464 SmallVector<BasicBlock*, 2> NewBBs;
4465 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4468 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4469 // phi predecessors are identical. The simple thing to do is skip
4470 // splitting in this case rather than complicate the API.
4472 // If PN is outside of the loop and BB is in the loop, we want to
4473 // move the block to be immediately before the PHI block, not
4474 // immediately after BB.
4475 if (L->contains(BB) && !L->contains(PN))
4476 NewBB->moveBefore(PN->getParent());
4478 // Splitting the edge can reduce the number of PHI entries we have.
4479 e = PN->getNumIncomingValues();
4481 i = PN->getBasicBlockIndex(BB);
4486 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4487 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4489 PN->setIncomingValue(i, Pair.first->second);
4491 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4493 // If this is reuse-by-noop-cast, insert the noop cast.
4494 Type *OpTy = LF.OperandValToReplace->getType();
4495 if (FullV->getType() != OpTy)
4497 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4499 FullV, LF.OperandValToReplace->getType(),
4500 "tmp", BB->getTerminator());
4502 PN->setIncomingValue(i, FullV);
4503 Pair.first->second = FullV;
4508 /// Rewrite - Emit instructions for the leading candidate expression for this
4509 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4510 /// the newly expanded value.
4511 void LSRInstance::Rewrite(const LSRFixup &LF,
4513 SCEVExpander &Rewriter,
4514 SmallVectorImpl<WeakVH> &DeadInsts,
4516 // First, find an insertion point that dominates UserInst. For PHI nodes,
4517 // find the nearest block which dominates all the relevant uses.
4518 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4519 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4521 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4523 // If this is reuse-by-noop-cast, insert the noop cast.
4524 Type *OpTy = LF.OperandValToReplace->getType();
4525 if (FullV->getType() != OpTy) {
4527 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4528 FullV, OpTy, "tmp", LF.UserInst);
4532 // Update the user. ICmpZero is handled specially here (for now) because
4533 // Expand may have updated one of the operands of the icmp already, and
4534 // its new value may happen to be equal to LF.OperandValToReplace, in
4535 // which case doing replaceUsesOfWith leads to replacing both operands
4536 // with the same value. TODO: Reorganize this.
4537 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4538 LF.UserInst->setOperand(0, FullV);
4540 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4543 DeadInsts.push_back(LF.OperandValToReplace);
4546 /// ImplementSolution - Rewrite all the fixup locations with new values,
4547 /// following the chosen solution.
4549 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4551 // Keep track of instructions we may have made dead, so that
4552 // we can remove them after we are done working.
4553 SmallVector<WeakVH, 16> DeadInsts;
4555 SCEVExpander Rewriter(SE, "lsr");
4557 Rewriter.setDebugType(DEBUG_TYPE);
4559 Rewriter.disableCanonicalMode();
4560 Rewriter.enableLSRMode();
4561 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4563 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4564 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4565 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4566 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4567 Rewriter.setChainedPhi(PN);
4570 // Expand the new value definitions and update the users.
4571 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4572 E = Fixups.end(); I != E; ++I) {
4573 const LSRFixup &Fixup = *I;
4575 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4580 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4581 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4582 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4585 // Clean up after ourselves. This must be done before deleting any
4589 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4592 LSRInstance::LSRInstance(const ScalarTargetTransformInfo *stti, Loop *l, Pass *P)
4593 : IU(P->getAnalysis<IVUsers>()),
4594 SE(P->getAnalysis<ScalarEvolution>()),
4595 DT(P->getAnalysis<DominatorTree>()),
4596 LI(P->getAnalysis<LoopInfo>()),
4597 STTI(stti), L(l), Changed(false), IVIncInsertPos(0) {
4599 // If LoopSimplify form is not available, stay out of trouble.
4600 if (!L->isLoopSimplifyForm())
4603 // If there's no interesting work to be done, bail early.
4604 if (IU.empty()) return;
4606 // If there's too much analysis to be done, bail early. We won't be able to
4607 // model the problem anyway.
4608 unsigned NumUsers = 0;
4609 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4610 if (++NumUsers > MaxIVUsers) {
4611 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4618 // All dominating loops must have preheaders, or SCEVExpander may not be able
4619 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4621 // IVUsers analysis should only create users that are dominated by simple loop
4622 // headers. Since this loop should dominate all of its users, its user list
4623 // should be empty if this loop itself is not within a simple loop nest.
4624 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4625 Rung; Rung = Rung->getIDom()) {
4626 BasicBlock *BB = Rung->getBlock();
4627 const Loop *DomLoop = LI.getLoopFor(BB);
4628 if (DomLoop && DomLoop->getHeader() == BB) {
4629 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4634 DEBUG(dbgs() << "\nLSR on loop ";
4635 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4638 // First, perform some low-level loop optimizations.
4640 OptimizeLoopTermCond();
4642 // If loop preparation eliminates all interesting IV users, bail.
4643 if (IU.empty()) return;
4645 // Skip nested loops until we can model them better with formulae.
4647 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4651 // Start collecting data and preparing for the solver.
4653 CollectInterestingTypesAndFactors();
4654 CollectFixupsAndInitialFormulae();
4655 CollectLoopInvariantFixupsAndFormulae();
4657 assert(!Uses.empty() && "IVUsers reported at least one use");
4658 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4659 print_uses(dbgs()));
4661 // Now use the reuse data to generate a bunch of interesting ways
4662 // to formulate the values needed for the uses.
4663 GenerateAllReuseFormulae();
4665 FilterOutUndesirableDedicatedRegisters();
4666 NarrowSearchSpaceUsingHeuristics();
4668 SmallVector<const Formula *, 8> Solution;
4671 // Release memory that is no longer needed.
4676 if (Solution.empty())
4680 // Formulae should be legal.
4681 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4682 E = Uses.end(); I != E; ++I) {
4683 const LSRUse &LU = *I;
4684 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4685 JE = LU.Formulae.end(); J != JE; ++J)
4686 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
4687 LU.Kind, LU.AccessTy, STTI) &&
4688 "Illegal formula generated!");
4692 // Now that we've decided what we want, make it so.
4693 ImplementSolution(Solution, P);
4696 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4697 if (Factors.empty() && Types.empty()) return;
4699 OS << "LSR has identified the following interesting factors and types: ";
4702 for (SmallSetVector<int64_t, 8>::const_iterator
4703 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4704 if (!First) OS << ", ";
4709 for (SmallSetVector<Type *, 4>::const_iterator
4710 I = Types.begin(), E = Types.end(); I != E; ++I) {
4711 if (!First) OS << ", ";
4713 OS << '(' << **I << ')';
4718 void LSRInstance::print_fixups(raw_ostream &OS) const {
4719 OS << "LSR is examining the following fixup sites:\n";
4720 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4721 E = Fixups.end(); I != E; ++I) {
4728 void LSRInstance::print_uses(raw_ostream &OS) const {
4729 OS << "LSR is examining the following uses:\n";
4730 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4731 E = Uses.end(); I != E; ++I) {
4732 const LSRUse &LU = *I;
4736 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4737 JE = LU.Formulae.end(); J != JE; ++J) {
4745 void LSRInstance::print(raw_ostream &OS) const {
4746 print_factors_and_types(OS);
4751 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4752 void LSRInstance::dump() const {
4753 print(errs()); errs() << '\n';
4759 class LoopStrengthReduce : public LoopPass {
4760 /// ScalarTargetTransformInfo provides target information that is needed
4761 /// for strength reducing loops.
4762 const ScalarTargetTransformInfo *STTI;
4765 static char ID; // Pass ID, replacement for typeid
4766 LoopStrengthReduce();
4769 bool runOnLoop(Loop *L, LPPassManager &LPM);
4770 void getAnalysisUsage(AnalysisUsage &AU) const;
4775 char LoopStrengthReduce::ID = 0;
4776 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4777 "Loop Strength Reduction", false, false)
4778 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4779 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4780 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4781 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4782 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4783 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4784 "Loop Strength Reduction", false, false)
4786 Pass *llvm::createLoopStrengthReducePass() {
4787 return new LoopStrengthReduce();
4790 LoopStrengthReduce::LoopStrengthReduce()
4791 : LoopPass(ID), STTI(0) {
4792 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4795 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4796 // We split critical edges, so we change the CFG. However, we do update
4797 // many analyses if they are around.
4798 AU.addPreservedID(LoopSimplifyID);
4800 AU.addRequired<LoopInfo>();
4801 AU.addPreserved<LoopInfo>();
4802 AU.addRequiredID(LoopSimplifyID);
4803 AU.addRequired<DominatorTree>();
4804 AU.addPreserved<DominatorTree>();
4805 AU.addRequired<ScalarEvolution>();
4806 AU.addPreserved<ScalarEvolution>();
4807 // Requiring LoopSimplify a second time here prevents IVUsers from running
4808 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4809 AU.addRequiredID(LoopSimplifyID);
4810 AU.addRequired<IVUsers>();
4811 AU.addPreserved<IVUsers>();
4814 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4815 bool Changed = false;
4817 TargetTransformInfo *TTI = getAnalysisIfAvailable<TargetTransformInfo>();
4820 STTI = TTI->getScalarTargetTransformInfo();
4822 // Run the main LSR transformation.
4823 Changed |= LSRInstance(STTI, L, this).getChanged();
4825 // Remove any extra phis created by processing inner loops.
4826 Changed |= DeleteDeadPHIs(L->getHeader());
4827 if (EnablePhiElim) {
4828 SmallVector<WeakVH, 16> DeadInsts;
4829 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4831 Rewriter.setDebugType(DEBUG_TYPE);
4833 unsigned numFolded = Rewriter.
4834 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, STTI);
4837 DeleteTriviallyDeadInstructions(DeadInsts);
4838 DeleteDeadPHIs(L->getHeader());