1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #include "llvm/Transforms/Scalar.h"
57 #include "llvm/ADT/DenseSet.h"
58 #include "llvm/ADT/Hashing.h"
59 #include "llvm/ADT/STLExtras.h"
60 #include "llvm/ADT/SetVector.h"
61 #include "llvm/ADT/SmallBitVector.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/LoopPass.h"
64 #include "llvm/Analysis/ScalarEvolutionExpander.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/IR/Constants.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/Instructions.h"
70 #include "llvm/IR/IntrinsicInst.h"
71 #include "llvm/IR/ValueHandle.h"
72 #include "llvm/Support/CommandLine.h"
73 #include "llvm/Support/Debug.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
76 #include "llvm/Transforms/Utils/Local.h"
80 #define DEBUG_TYPE "loop-reduce"
82 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
83 /// bail out. This threshold is far beyond the number of users that LSR can
84 /// conceivably solve, so it should not affect generated code, but catches the
85 /// worst cases before LSR burns too much compile time and stack space.
86 static const unsigned MaxIVUsers = 200;
88 // Temporary flag to cleanup congruent phis after LSR phi expansion.
89 // It's currently disabled until we can determine whether it's truly useful or
90 // not. The flag should be removed after the v3.0 release.
91 // This is now needed for ivchains.
92 static cl::opt<bool> EnablePhiElim(
93 "enable-lsr-phielim", cl::Hidden, cl::init(true),
94 cl::desc("Enable LSR phi elimination"));
97 // Stress test IV chain generation.
98 static cl::opt<bool> StressIVChain(
99 "stress-ivchain", cl::Hidden, cl::init(false),
100 cl::desc("Stress test LSR IV chains"));
102 static bool StressIVChain = false;
107 /// RegSortData - This class holds data which is used to order reuse candidates.
110 /// UsedByIndices - This represents the set of LSRUse indices which reference
111 /// a particular register.
112 SmallBitVector UsedByIndices;
116 void print(raw_ostream &OS) const;
122 void RegSortData::print(raw_ostream &OS) const {
123 OS << "[NumUses=" << UsedByIndices.count() << ']';
126 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
127 void RegSortData::dump() const {
128 print(errs()); errs() << '\n';
134 /// RegUseTracker - Map register candidates to information about how they are
136 class RegUseTracker {
137 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
139 RegUsesTy RegUsesMap;
140 SmallVector<const SCEV *, 16> RegSequence;
143 void CountRegister(const SCEV *Reg, size_t LUIdx);
144 void DropRegister(const SCEV *Reg, size_t LUIdx);
145 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
147 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
149 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
153 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
154 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
155 iterator begin() { return RegSequence.begin(); }
156 iterator end() { return RegSequence.end(); }
157 const_iterator begin() const { return RegSequence.begin(); }
158 const_iterator end() const { return RegSequence.end(); }
164 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
165 std::pair<RegUsesTy::iterator, bool> Pair =
166 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
167 RegSortData &RSD = Pair.first->second;
169 RegSequence.push_back(Reg);
170 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
171 RSD.UsedByIndices.set(LUIdx);
175 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
176 RegUsesTy::iterator It = RegUsesMap.find(Reg);
177 assert(It != RegUsesMap.end());
178 RegSortData &RSD = It->second;
179 assert(RSD.UsedByIndices.size() > LUIdx);
180 RSD.UsedByIndices.reset(LUIdx);
184 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
185 assert(LUIdx <= LastLUIdx);
187 // Update RegUses. The data structure is not optimized for this purpose;
188 // we must iterate through it and update each of the bit vectors.
189 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
191 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
192 if (LUIdx < UsedByIndices.size())
193 UsedByIndices[LUIdx] =
194 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
195 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
200 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
201 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
202 if (I == RegUsesMap.end())
204 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
205 int i = UsedByIndices.find_first();
206 if (i == -1) return false;
207 if ((size_t)i != LUIdx) return true;
208 return UsedByIndices.find_next(i) != -1;
211 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
212 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
213 assert(I != RegUsesMap.end() && "Unknown register!");
214 return I->second.UsedByIndices;
217 void RegUseTracker::clear() {
224 /// Formula - This class holds information that describes a formula for
225 /// computing satisfying a use. It may include broken-out immediates and scaled
228 /// Global base address used for complex addressing.
231 /// Base offset for complex addressing.
234 /// Whether any complex addressing has a base register.
237 /// The scale of any complex addressing.
240 /// BaseRegs - The list of "base" registers for this use. When this is
242 SmallVector<const SCEV *, 4> BaseRegs;
244 /// ScaledReg - The 'scaled' register for this use. This should be non-null
245 /// when Scale is not zero.
246 const SCEV *ScaledReg;
248 /// UnfoldedOffset - An additional constant offset which added near the
249 /// use. This requires a temporary register, but the offset itself can
250 /// live in an add immediate field rather than a register.
251 int64_t UnfoldedOffset;
254 : BaseGV(0), BaseOffset(0), HasBaseReg(false), Scale(0), ScaledReg(0),
257 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
259 unsigned getNumRegs() const;
260 Type *getType() const;
262 void DeleteBaseReg(const SCEV *&S);
264 bool referencesReg(const SCEV *S) const;
265 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
266 const RegUseTracker &RegUses) const;
268 void print(raw_ostream &OS) const;
274 /// DoInitialMatch - Recursion helper for InitialMatch.
275 static void DoInitialMatch(const SCEV *S, Loop *L,
276 SmallVectorImpl<const SCEV *> &Good,
277 SmallVectorImpl<const SCEV *> &Bad,
278 ScalarEvolution &SE) {
279 // Collect expressions which properly dominate the loop header.
280 if (SE.properlyDominates(S, L->getHeader())) {
285 // Look at add operands.
286 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
287 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
289 DoInitialMatch(*I, L, Good, Bad, SE);
293 // Look at addrec operands.
294 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
295 if (!AR->getStart()->isZero()) {
296 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
297 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
298 AR->getStepRecurrence(SE),
299 // FIXME: AR->getNoWrapFlags()
300 AR->getLoop(), SCEV::FlagAnyWrap),
305 // Handle a multiplication by -1 (negation) if it didn't fold.
306 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
307 if (Mul->getOperand(0)->isAllOnesValue()) {
308 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
309 const SCEV *NewMul = SE.getMulExpr(Ops);
311 SmallVector<const SCEV *, 4> MyGood;
312 SmallVector<const SCEV *, 4> MyBad;
313 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
314 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
315 SE.getEffectiveSCEVType(NewMul->getType())));
316 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
317 E = MyGood.end(); I != E; ++I)
318 Good.push_back(SE.getMulExpr(NegOne, *I));
319 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
320 E = MyBad.end(); I != E; ++I)
321 Bad.push_back(SE.getMulExpr(NegOne, *I));
325 // Ok, we can't do anything interesting. Just stuff the whole thing into a
326 // register and hope for the best.
330 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
331 /// attempting to keep all loop-invariant and loop-computable values in a
332 /// single base register.
333 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
334 SmallVector<const SCEV *, 4> Good;
335 SmallVector<const SCEV *, 4> Bad;
336 DoInitialMatch(S, L, Good, Bad, SE);
338 const SCEV *Sum = SE.getAddExpr(Good);
340 BaseRegs.push_back(Sum);
344 const SCEV *Sum = SE.getAddExpr(Bad);
346 BaseRegs.push_back(Sum);
351 /// getNumRegs - Return the total number of register operands used by this
352 /// formula. This does not include register uses implied by non-constant
354 unsigned Formula::getNumRegs() const {
355 return !!ScaledReg + BaseRegs.size();
358 /// getType - Return the type of this formula, if it has one, or null
359 /// otherwise. This type is meaningless except for the bit size.
360 Type *Formula::getType() const {
361 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
362 ScaledReg ? ScaledReg->getType() :
363 BaseGV ? BaseGV->getType() :
367 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
368 void Formula::DeleteBaseReg(const SCEV *&S) {
369 if (&S != &BaseRegs.back())
370 std::swap(S, BaseRegs.back());
374 /// referencesReg - Test if this formula references the given register.
375 bool Formula::referencesReg(const SCEV *S) const {
376 return S == ScaledReg ||
377 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
380 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
381 /// which are used by uses other than the use with the given index.
382 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
383 const RegUseTracker &RegUses) const {
385 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
387 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
388 E = BaseRegs.end(); I != E; ++I)
389 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
394 void Formula::print(raw_ostream &OS) const {
397 if (!First) OS << " + "; else First = false;
398 BaseGV->printAsOperand(OS, /*PrintType=*/false);
400 if (BaseOffset != 0) {
401 if (!First) OS << " + "; else First = false;
404 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
405 E = BaseRegs.end(); I != E; ++I) {
406 if (!First) OS << " + "; else First = false;
407 OS << "reg(" << **I << ')';
409 if (HasBaseReg && BaseRegs.empty()) {
410 if (!First) OS << " + "; else First = false;
411 OS << "**error: HasBaseReg**";
412 } else if (!HasBaseReg && !BaseRegs.empty()) {
413 if (!First) OS << " + "; else First = false;
414 OS << "**error: !HasBaseReg**";
417 if (!First) OS << " + "; else First = false;
418 OS << Scale << "*reg(";
425 if (UnfoldedOffset != 0) {
426 if (!First) OS << " + ";
427 OS << "imm(" << UnfoldedOffset << ')';
431 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
432 void Formula::dump() const {
433 print(errs()); errs() << '\n';
437 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
438 /// without changing its value.
439 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
441 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
442 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
445 /// isAddSExtable - Return true if the given add can be sign-extended
446 /// without changing its value.
447 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
449 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
450 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
453 /// isMulSExtable - Return true if the given mul can be sign-extended
454 /// without changing its value.
455 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
457 IntegerType::get(SE.getContext(),
458 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
459 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
462 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
463 /// and if the remainder is known to be zero, or null otherwise. If
464 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
465 /// to Y, ignoring that the multiplication may overflow, which is useful when
466 /// the result will be used in a context where the most significant bits are
468 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
470 bool IgnoreSignificantBits = false) {
471 // Handle the trivial case, which works for any SCEV type.
473 return SE.getConstant(LHS->getType(), 1);
475 // Handle a few RHS special cases.
476 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
478 const APInt &RA = RC->getValue()->getValue();
479 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
481 if (RA.isAllOnesValue())
482 return SE.getMulExpr(LHS, RC);
483 // Handle x /s 1 as x.
488 // Check for a division of a constant by a constant.
489 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
492 const APInt &LA = C->getValue()->getValue();
493 const APInt &RA = RC->getValue()->getValue();
494 if (LA.srem(RA) != 0)
496 return SE.getConstant(LA.sdiv(RA));
499 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
500 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
501 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
502 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
503 IgnoreSignificantBits);
505 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
506 IgnoreSignificantBits);
507 if (!Start) return 0;
508 // FlagNW is independent of the start value, step direction, and is
509 // preserved with smaller magnitude steps.
510 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
511 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
516 // Distribute the sdiv over add operands, if the add doesn't overflow.
517 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
518 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
519 SmallVector<const SCEV *, 8> Ops;
520 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
522 const SCEV *Op = getExactSDiv(*I, RHS, SE,
523 IgnoreSignificantBits);
527 return SE.getAddExpr(Ops);
532 // Check for a multiply operand that we can pull RHS out of.
533 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
534 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
535 SmallVector<const SCEV *, 4> Ops;
537 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
541 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
542 IgnoreSignificantBits)) {
548 return Found ? SE.getMulExpr(Ops) : 0;
553 // Otherwise we don't know.
557 /// ExtractImmediate - If S involves the addition of a constant integer value,
558 /// return that integer value, and mutate S to point to a new SCEV with that
560 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
561 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
562 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
563 S = SE.getConstant(C->getType(), 0);
564 return C->getValue()->getSExtValue();
566 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
567 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
568 int64_t Result = ExtractImmediate(NewOps.front(), SE);
570 S = SE.getAddExpr(NewOps);
572 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
573 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
574 int64_t Result = ExtractImmediate(NewOps.front(), SE);
576 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
577 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
584 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
585 /// return that symbol, and mutate S to point to a new SCEV with that
587 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
588 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
589 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
590 S = SE.getConstant(GV->getType(), 0);
593 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
594 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
595 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
597 S = SE.getAddExpr(NewOps);
599 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
600 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
601 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
603 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
604 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
611 /// isAddressUse - Returns true if the specified instruction is using the
612 /// specified value as an address.
613 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
614 bool isAddress = isa<LoadInst>(Inst);
615 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
616 if (SI->getOperand(1) == OperandVal)
618 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
619 // Addressing modes can also be folded into prefetches and a variety
621 switch (II->getIntrinsicID()) {
623 case Intrinsic::prefetch:
624 case Intrinsic::x86_sse_storeu_ps:
625 case Intrinsic::x86_sse2_storeu_pd:
626 case Intrinsic::x86_sse2_storeu_dq:
627 case Intrinsic::x86_sse2_storel_dq:
628 if (II->getArgOperand(0) == OperandVal)
636 /// getAccessType - Return the type of the memory being accessed.
637 static Type *getAccessType(const Instruction *Inst) {
638 Type *AccessTy = Inst->getType();
639 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
640 AccessTy = SI->getOperand(0)->getType();
641 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
642 // Addressing modes can also be folded into prefetches and a variety
644 switch (II->getIntrinsicID()) {
646 case Intrinsic::x86_sse_storeu_ps:
647 case Intrinsic::x86_sse2_storeu_pd:
648 case Intrinsic::x86_sse2_storeu_dq:
649 case Intrinsic::x86_sse2_storel_dq:
650 AccessTy = II->getArgOperand(0)->getType();
655 // All pointers have the same requirements, so canonicalize them to an
656 // arbitrary pointer type to minimize variation.
657 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
658 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
659 PTy->getAddressSpace());
664 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
665 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
666 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
667 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
668 if (SE.isSCEVable(PN->getType()) &&
669 (SE.getEffectiveSCEVType(PN->getType()) ==
670 SE.getEffectiveSCEVType(AR->getType())) &&
671 SE.getSCEV(PN) == AR)
677 /// Check if expanding this expression is likely to incur significant cost. This
678 /// is tricky because SCEV doesn't track which expressions are actually computed
679 /// by the current IR.
681 /// We currently allow expansion of IV increments that involve adds,
682 /// multiplication by constants, and AddRecs from existing phis.
684 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
685 /// obvious multiple of the UDivExpr.
686 static bool isHighCostExpansion(const SCEV *S,
687 SmallPtrSet<const SCEV*, 8> &Processed,
688 ScalarEvolution &SE) {
689 // Zero/One operand expressions
690 switch (S->getSCEVType()) {
695 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
698 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
701 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
705 if (!Processed.insert(S))
708 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
709 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
711 if (isHighCostExpansion(*I, Processed, SE))
717 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
718 if (Mul->getNumOperands() == 2) {
719 // Multiplication by a constant is ok
720 if (isa<SCEVConstant>(Mul->getOperand(0)))
721 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
723 // If we have the value of one operand, check if an existing
724 // multiplication already generates this expression.
725 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
726 Value *UVal = U->getValue();
727 for (User *UR : UVal->users()) {
728 // If U is a constant, it may be used by a ConstantExpr.
729 Instruction *UI = dyn_cast<Instruction>(UR);
730 if (UI && UI->getOpcode() == Instruction::Mul &&
731 SE.isSCEVable(UI->getType())) {
732 return SE.getSCEV(UI) == Mul;
739 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
740 if (isExistingPhi(AR, SE))
744 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
748 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
749 /// specified set are trivially dead, delete them and see if this makes any of
750 /// their operands subsequently dead.
752 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
753 bool Changed = false;
755 while (!DeadInsts.empty()) {
756 Value *V = DeadInsts.pop_back_val();
757 Instruction *I = dyn_cast_or_null<Instruction>(V);
759 if (I == 0 || !isInstructionTriviallyDead(I))
762 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
763 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
766 DeadInsts.push_back(U);
769 I->eraseFromParent();
779 // Check if it is legal to fold 2 base registers.
780 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
782 // Get the cost of the scaling factor used in F for LU.
783 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
784 const LSRUse &LU, const Formula &F);
788 /// Cost - This class is used to measure and compare candidate formulae.
790 /// TODO: Some of these could be merged. Also, a lexical ordering
791 /// isn't always optimal.
795 unsigned NumBaseAdds;
802 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
803 SetupCost(0), ScaleCost(0) {}
805 bool operator<(const Cost &Other) const;
810 // Once any of the metrics loses, they must all remain losers.
812 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
813 | ImmCost | SetupCost | ScaleCost) != ~0u)
814 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
815 & ImmCost & SetupCost & ScaleCost) == ~0u);
820 assert(isValid() && "invalid cost");
821 return NumRegs == ~0u;
824 void RateFormula(const TargetTransformInfo &TTI,
826 SmallPtrSet<const SCEV *, 16> &Regs,
827 const DenseSet<const SCEV *> &VisitedRegs,
829 const SmallVectorImpl<int64_t> &Offsets,
830 ScalarEvolution &SE, DominatorTree &DT,
832 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
834 void print(raw_ostream &OS) const;
838 void RateRegister(const SCEV *Reg,
839 SmallPtrSet<const SCEV *, 16> &Regs,
841 ScalarEvolution &SE, DominatorTree &DT);
842 void RatePrimaryRegister(const SCEV *Reg,
843 SmallPtrSet<const SCEV *, 16> &Regs,
845 ScalarEvolution &SE, DominatorTree &DT,
846 SmallPtrSet<const SCEV *, 16> *LoserRegs);
851 /// RateRegister - Tally up interesting quantities from the given register.
852 void Cost::RateRegister(const SCEV *Reg,
853 SmallPtrSet<const SCEV *, 16> &Regs,
855 ScalarEvolution &SE, DominatorTree &DT) {
856 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
857 // If this is an addrec for another loop, don't second-guess its addrec phi
858 // nodes. LSR isn't currently smart enough to reason about more than one
859 // loop at a time. LSR has already run on inner loops, will not run on outer
860 // loops, and cannot be expected to change sibling loops.
861 if (AR->getLoop() != L) {
862 // If the AddRec exists, consider it's register free and leave it alone.
863 if (isExistingPhi(AR, SE))
866 // Otherwise, do not consider this formula at all.
870 AddRecCost += 1; /// TODO: This should be a function of the stride.
872 // Add the step value register, if it needs one.
873 // TODO: The non-affine case isn't precisely modeled here.
874 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
875 if (!Regs.count(AR->getOperand(1))) {
876 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
884 // Rough heuristic; favor registers which don't require extra setup
885 // instructions in the preheader.
886 if (!isa<SCEVUnknown>(Reg) &&
887 !isa<SCEVConstant>(Reg) &&
888 !(isa<SCEVAddRecExpr>(Reg) &&
889 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
890 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
893 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
894 SE.hasComputableLoopEvolution(Reg, L);
897 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
898 /// before, rate it. Optional LoserRegs provides a way to declare any formula
899 /// that refers to one of those regs an instant loser.
900 void Cost::RatePrimaryRegister(const SCEV *Reg,
901 SmallPtrSet<const SCEV *, 16> &Regs,
903 ScalarEvolution &SE, DominatorTree &DT,
904 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
905 if (LoserRegs && LoserRegs->count(Reg)) {
909 if (Regs.insert(Reg)) {
910 RateRegister(Reg, Regs, L, SE, DT);
911 if (LoserRegs && isLoser())
912 LoserRegs->insert(Reg);
916 void Cost::RateFormula(const TargetTransformInfo &TTI,
918 SmallPtrSet<const SCEV *, 16> &Regs,
919 const DenseSet<const SCEV *> &VisitedRegs,
921 const SmallVectorImpl<int64_t> &Offsets,
922 ScalarEvolution &SE, DominatorTree &DT,
924 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
925 // Tally up the registers.
926 if (const SCEV *ScaledReg = F.ScaledReg) {
927 if (VisitedRegs.count(ScaledReg)) {
931 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
935 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
936 E = F.BaseRegs.end(); I != E; ++I) {
937 const SCEV *BaseReg = *I;
938 if (VisitedRegs.count(BaseReg)) {
942 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
947 // Determine how many (unfolded) adds we'll need inside the loop.
948 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
949 if (NumBaseParts > 1)
950 // Do not count the base and a possible second register if the target
951 // allows to fold 2 registers.
952 NumBaseAdds += NumBaseParts - (1 + isLegal2RegAMUse(TTI, LU, F));
954 // Accumulate non-free scaling amounts.
955 ScaleCost += getScalingFactorCost(TTI, LU, F);
957 // Tally up the non-zero immediates.
958 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
959 E = Offsets.end(); I != E; ++I) {
960 int64_t Offset = (uint64_t)*I + F.BaseOffset;
962 ImmCost += 64; // Handle symbolic values conservatively.
963 // TODO: This should probably be the pointer size.
964 else if (Offset != 0)
965 ImmCost += APInt(64, Offset, true).getMinSignedBits();
967 assert(isValid() && "invalid cost");
970 /// Lose - Set this cost to a losing value.
981 /// operator< - Choose the lower cost.
982 bool Cost::operator<(const Cost &Other) const {
983 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
984 ImmCost, SetupCost) <
985 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
986 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
990 void Cost::print(raw_ostream &OS) const {
991 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
993 OS << ", with addrec cost " << AddRecCost;
995 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
996 if (NumBaseAdds != 0)
997 OS << ", plus " << NumBaseAdds << " base add"
998 << (NumBaseAdds == 1 ? "" : "s");
1000 OS << ", plus " << ScaleCost << " scale cost";
1002 OS << ", plus " << ImmCost << " imm cost";
1004 OS << ", plus " << SetupCost << " setup cost";
1007 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1008 void Cost::dump() const {
1009 print(errs()); errs() << '\n';
1015 /// LSRFixup - An operand value in an instruction which is to be replaced
1016 /// with some equivalent, possibly strength-reduced, replacement.
1018 /// UserInst - The instruction which will be updated.
1019 Instruction *UserInst;
1021 /// OperandValToReplace - The operand of the instruction which will
1022 /// be replaced. The operand may be used more than once; every instance
1023 /// will be replaced.
1024 Value *OperandValToReplace;
1026 /// PostIncLoops - If this user is to use the post-incremented value of an
1027 /// induction variable, this variable is non-null and holds the loop
1028 /// associated with the induction variable.
1029 PostIncLoopSet PostIncLoops;
1031 /// LUIdx - The index of the LSRUse describing the expression which
1032 /// this fixup needs, minus an offset (below).
1035 /// Offset - A constant offset to be added to the LSRUse expression.
1036 /// This allows multiple fixups to share the same LSRUse with different
1037 /// offsets, for example in an unrolled loop.
1040 bool isUseFullyOutsideLoop(const Loop *L) const;
1044 void print(raw_ostream &OS) const;
1050 LSRFixup::LSRFixup()
1051 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1053 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1054 /// value outside of the given loop.
1055 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1056 // PHI nodes use their value in their incoming blocks.
1057 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1058 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1059 if (PN->getIncomingValue(i) == OperandValToReplace &&
1060 L->contains(PN->getIncomingBlock(i)))
1065 return !L->contains(UserInst);
1068 void LSRFixup::print(raw_ostream &OS) const {
1070 // Store is common and interesting enough to be worth special-casing.
1071 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1073 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1074 } else if (UserInst->getType()->isVoidTy())
1075 OS << UserInst->getOpcodeName();
1077 UserInst->printAsOperand(OS, /*PrintType=*/false);
1079 OS << ", OperandValToReplace=";
1080 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1082 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1083 E = PostIncLoops.end(); I != E; ++I) {
1084 OS << ", PostIncLoop=";
1085 (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1088 if (LUIdx != ~size_t(0))
1089 OS << ", LUIdx=" << LUIdx;
1092 OS << ", Offset=" << Offset;
1095 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1096 void LSRFixup::dump() const {
1097 print(errs()); errs() << '\n';
1103 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1104 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1105 struct UniquifierDenseMapInfo {
1106 static SmallVector<const SCEV *, 4> getEmptyKey() {
1107 SmallVector<const SCEV *, 4> V;
1108 V.push_back(reinterpret_cast<const SCEV *>(-1));
1112 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1113 SmallVector<const SCEV *, 4> V;
1114 V.push_back(reinterpret_cast<const SCEV *>(-2));
1118 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1119 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1122 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1123 const SmallVector<const SCEV *, 4> &RHS) {
1128 /// LSRUse - This class holds the state that LSR keeps for each use in
1129 /// IVUsers, as well as uses invented by LSR itself. It includes information
1130 /// about what kinds of things can be folded into the user, information about
1131 /// the user itself, and information about how the use may be satisfied.
1132 /// TODO: Represent multiple users of the same expression in common?
1134 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1137 /// KindType - An enum for a kind of use, indicating what types of
1138 /// scaled and immediate operands it might support.
1140 Basic, ///< A normal use, with no folding.
1141 Special, ///< A special case of basic, allowing -1 scales.
1142 Address, ///< An address use; folding according to TargetLowering
1143 ICmpZero ///< An equality icmp with both operands folded into one.
1144 // TODO: Add a generic icmp too?
1147 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1152 SmallVector<int64_t, 8> Offsets;
1156 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1157 /// LSRUse are outside of the loop, in which case some special-case heuristics
1159 bool AllFixupsOutsideLoop;
1161 /// RigidFormula is set to true to guarantee that this use will be associated
1162 /// with a single formula--the one that initially matched. Some SCEV
1163 /// expressions cannot be expanded. This allows LSR to consider the registers
1164 /// used by those expressions without the need to expand them later after
1165 /// changing the formula.
1168 /// WidestFixupType - This records the widest use type for any fixup using
1169 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1170 /// max fixup widths to be equivalent, because the narrower one may be relying
1171 /// on the implicit truncation to truncate away bogus bits.
1172 Type *WidestFixupType;
1174 /// Formulae - A list of ways to build a value that can satisfy this user.
1175 /// After the list is populated, one of these is selected heuristically and
1176 /// used to formulate a replacement for OperandValToReplace in UserInst.
1177 SmallVector<Formula, 12> Formulae;
1179 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1180 SmallPtrSet<const SCEV *, 4> Regs;
1182 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1183 MinOffset(INT64_MAX),
1184 MaxOffset(INT64_MIN),
1185 AllFixupsOutsideLoop(true),
1186 RigidFormula(false),
1187 WidestFixupType(0) {}
1189 bool HasFormulaWithSameRegs(const Formula &F) const;
1190 bool InsertFormula(const Formula &F);
1191 void DeleteFormula(Formula &F);
1192 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1194 void print(raw_ostream &OS) const;
1200 /// HasFormula - Test whether this use as a formula which has the same
1201 /// registers as the given formula.
1202 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1203 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1204 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1205 // Unstable sort by host order ok, because this is only used for uniquifying.
1206 std::sort(Key.begin(), Key.end());
1207 return Uniquifier.count(Key);
1210 /// InsertFormula - If the given formula has not yet been inserted, add it to
1211 /// the list, and return true. Return false otherwise.
1212 bool LSRUse::InsertFormula(const Formula &F) {
1213 if (!Formulae.empty() && RigidFormula)
1216 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1217 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1218 // Unstable sort by host order ok, because this is only used for uniquifying.
1219 std::sort(Key.begin(), Key.end());
1221 if (!Uniquifier.insert(Key).second)
1224 // Using a register to hold the value of 0 is not profitable.
1225 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1226 "Zero allocated in a scaled register!");
1228 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1229 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1230 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1233 // Add the formula to the list.
1234 Formulae.push_back(F);
1236 // Record registers now being used by this use.
1237 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1242 /// DeleteFormula - Remove the given formula from this use's list.
1243 void LSRUse::DeleteFormula(Formula &F) {
1244 if (&F != &Formulae.back())
1245 std::swap(F, Formulae.back());
1246 Formulae.pop_back();
1249 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1250 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1251 // Now that we've filtered out some formulae, recompute the Regs set.
1252 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1254 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1255 E = Formulae.end(); I != E; ++I) {
1256 const Formula &F = *I;
1257 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1258 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1261 // Update the RegTracker.
1262 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1263 E = OldRegs.end(); I != E; ++I)
1264 if (!Regs.count(*I))
1265 RegUses.DropRegister(*I, LUIdx);
1268 void LSRUse::print(raw_ostream &OS) const {
1269 OS << "LSR Use: Kind=";
1271 case Basic: OS << "Basic"; break;
1272 case Special: OS << "Special"; break;
1273 case ICmpZero: OS << "ICmpZero"; break;
1275 OS << "Address of ";
1276 if (AccessTy->isPointerTy())
1277 OS << "pointer"; // the full pointer type could be really verbose
1282 OS << ", Offsets={";
1283 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1284 E = Offsets.end(); I != E; ++I) {
1286 if (std::next(I) != E)
1291 if (AllFixupsOutsideLoop)
1292 OS << ", all-fixups-outside-loop";
1294 if (WidestFixupType)
1295 OS << ", widest fixup type: " << *WidestFixupType;
1298 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1299 void LSRUse::dump() const {
1300 print(errs()); errs() << '\n';
1304 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1305 /// be completely folded into the user instruction at isel time. This includes
1306 /// address-mode folding and special icmp tricks.
1307 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1308 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1309 bool HasBaseReg, int64_t Scale) {
1311 case LSRUse::Address:
1312 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1314 // Otherwise, just guess that reg+reg addressing is legal.
1317 case LSRUse::ICmpZero:
1318 // There's not even a target hook for querying whether it would be legal to
1319 // fold a GV into an ICmp.
1323 // ICmp only has two operands; don't allow more than two non-trivial parts.
1324 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1327 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1328 // putting the scaled register in the other operand of the icmp.
1329 if (Scale != 0 && Scale != -1)
1332 // If we have low-level target information, ask the target if it can fold an
1333 // integer immediate on an icmp.
1334 if (BaseOffset != 0) {
1336 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1337 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1338 // Offs is the ICmp immediate.
1340 // The cast does the right thing with INT64_MIN.
1341 BaseOffset = -(uint64_t)BaseOffset;
1342 return TTI.isLegalICmpImmediate(BaseOffset);
1345 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1349 // Only handle single-register values.
1350 return !BaseGV && Scale == 0 && BaseOffset == 0;
1352 case LSRUse::Special:
1353 // Special case Basic to handle -1 scales.
1354 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1357 llvm_unreachable("Invalid LSRUse Kind!");
1360 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1361 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1362 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1364 // Check for overflow.
1365 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1368 MinOffset = (uint64_t)BaseOffset + MinOffset;
1369 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1372 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1374 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1376 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1379 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1380 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1382 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1383 F.BaseOffset, F.HasBaseReg, F.Scale);
1386 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
1388 // If F is used as an Addressing Mode, it may fold one Base plus one
1389 // scaled register. If the scaled register is nil, do as if another
1390 // element of the base regs is a 1-scaled register.
1391 // This is possible if BaseRegs has at least 2 registers.
1393 // If this is not an address calculation, this is not an addressing mode
1395 if (LU.Kind != LSRUse::Address)
1398 // F is already scaled.
1402 // We need to keep one register for the base and one to scale.
1403 if (F.BaseRegs.size() < 2)
1406 return isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
1407 F.BaseGV, F.BaseOffset, F.HasBaseReg, 1);
1410 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1411 const LSRUse &LU, const Formula &F) {
1414 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1415 LU.AccessTy, F) && "Illegal formula in use.");
1418 case LSRUse::Address: {
1419 // Check the scaling factor cost with both the min and max offsets.
1420 int ScaleCostMinOffset =
1421 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1422 F.BaseOffset + LU.MinOffset,
1423 F.HasBaseReg, F.Scale);
1424 int ScaleCostMaxOffset =
1425 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1426 F.BaseOffset + LU.MaxOffset,
1427 F.HasBaseReg, F.Scale);
1429 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1430 "Legal addressing mode has an illegal cost!");
1431 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1433 case LSRUse::ICmpZero:
1434 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
1435 // Therefore, return 0 in case F.Scale == -1.
1436 return F.Scale != -1;
1439 case LSRUse::Special:
1443 llvm_unreachable("Invalid LSRUse Kind!");
1446 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1447 LSRUse::KindType Kind, Type *AccessTy,
1448 GlobalValue *BaseGV, int64_t BaseOffset,
1450 // Fast-path: zero is always foldable.
1451 if (BaseOffset == 0 && !BaseGV) return true;
1453 // Conservatively, create an address with an immediate and a
1454 // base and a scale.
1455 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1457 // Canonicalize a scale of 1 to a base register if the formula doesn't
1458 // already have a base register.
1459 if (!HasBaseReg && Scale == 1) {
1464 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1467 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1468 ScalarEvolution &SE, int64_t MinOffset,
1469 int64_t MaxOffset, LSRUse::KindType Kind,
1470 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1471 // Fast-path: zero is always foldable.
1472 if (S->isZero()) return true;
1474 // Conservatively, create an address with an immediate and a
1475 // base and a scale.
1476 int64_t BaseOffset = ExtractImmediate(S, SE);
1477 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1479 // If there's anything else involved, it's not foldable.
1480 if (!S->isZero()) return false;
1482 // Fast-path: zero is always foldable.
1483 if (BaseOffset == 0 && !BaseGV) return true;
1485 // Conservatively, create an address with an immediate and a
1486 // base and a scale.
1487 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1489 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1490 BaseOffset, HasBaseReg, Scale);
1495 /// IVInc - An individual increment in a Chain of IV increments.
1496 /// Relate an IV user to an expression that computes the IV it uses from the IV
1497 /// used by the previous link in the Chain.
1499 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1500 /// original IVOperand. The head of the chain's IVOperand is only valid during
1501 /// chain collection, before LSR replaces IV users. During chain generation,
1502 /// IncExpr can be used to find the new IVOperand that computes the same
1505 Instruction *UserInst;
1507 const SCEV *IncExpr;
1509 IVInc(Instruction *U, Value *O, const SCEV *E):
1510 UserInst(U), IVOperand(O), IncExpr(E) {}
1513 // IVChain - The list of IV increments in program order.
1514 // We typically add the head of a chain without finding subsequent links.
1516 SmallVector<IVInc,1> Incs;
1517 const SCEV *ExprBase;
1519 IVChain() : ExprBase(0) {}
1521 IVChain(const IVInc &Head, const SCEV *Base)
1522 : Incs(1, Head), ExprBase(Base) {}
1524 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1526 // begin - return the first increment in the chain.
1527 const_iterator begin() const {
1528 assert(!Incs.empty());
1529 return std::next(Incs.begin());
1531 const_iterator end() const {
1535 // hasIncs - Returns true if this chain contains any increments.
1536 bool hasIncs() const { return Incs.size() >= 2; }
1538 // add - Add an IVInc to the end of this chain.
1539 void add(const IVInc &X) { Incs.push_back(X); }
1541 // tailUserInst - Returns the last UserInst in the chain.
1542 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1544 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1546 bool isProfitableIncrement(const SCEV *OperExpr,
1547 const SCEV *IncExpr,
1551 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1552 /// Distinguish between FarUsers that definitely cross IV increments and
1553 /// NearUsers that may be used between IV increments.
1555 SmallPtrSet<Instruction*, 4> FarUsers;
1556 SmallPtrSet<Instruction*, 4> NearUsers;
1559 /// LSRInstance - This class holds state for the main loop strength reduction
1563 ScalarEvolution &SE;
1566 const TargetTransformInfo &TTI;
1570 /// IVIncInsertPos - This is the insert position that the current loop's
1571 /// induction variable increment should be placed. In simple loops, this is
1572 /// the latch block's terminator. But in more complicated cases, this is a
1573 /// position which will dominate all the in-loop post-increment users.
1574 Instruction *IVIncInsertPos;
1576 /// Factors - Interesting factors between use strides.
1577 SmallSetVector<int64_t, 8> Factors;
1579 /// Types - Interesting use types, to facilitate truncation reuse.
1580 SmallSetVector<Type *, 4> Types;
1582 /// Fixups - The list of operands which are to be replaced.
1583 SmallVector<LSRFixup, 16> Fixups;
1585 /// Uses - The list of interesting uses.
1586 SmallVector<LSRUse, 16> Uses;
1588 /// RegUses - Track which uses use which register candidates.
1589 RegUseTracker RegUses;
1591 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1592 // have more than a few IV increment chains in a loop. Missing a Chain falls
1593 // back to normal LSR behavior for those uses.
1594 static const unsigned MaxChains = 8;
1596 /// IVChainVec - IV users can form a chain of IV increments.
1597 SmallVector<IVChain, MaxChains> IVChainVec;
1599 /// IVIncSet - IV users that belong to profitable IVChains.
1600 SmallPtrSet<Use*, MaxChains> IVIncSet;
1602 void OptimizeShadowIV();
1603 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1604 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1605 void OptimizeLoopTermCond();
1607 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1608 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1609 void FinalizeChain(IVChain &Chain);
1610 void CollectChains();
1611 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1612 SmallVectorImpl<WeakVH> &DeadInsts);
1614 void CollectInterestingTypesAndFactors();
1615 void CollectFixupsAndInitialFormulae();
1617 LSRFixup &getNewFixup() {
1618 Fixups.push_back(LSRFixup());
1619 return Fixups.back();
1622 // Support for sharing of LSRUses between LSRFixups.
1623 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1626 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1627 LSRUse::KindType Kind, Type *AccessTy);
1629 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1630 LSRUse::KindType Kind,
1633 void DeleteUse(LSRUse &LU, size_t LUIdx);
1635 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1637 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1638 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1639 void CountRegisters(const Formula &F, size_t LUIdx);
1640 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1642 void CollectLoopInvariantFixupsAndFormulae();
1644 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1645 unsigned Depth = 0);
1646 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1647 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1648 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1649 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1650 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1651 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1652 void GenerateCrossUseConstantOffsets();
1653 void GenerateAllReuseFormulae();
1655 void FilterOutUndesirableDedicatedRegisters();
1657 size_t EstimateSearchSpaceComplexity() const;
1658 void NarrowSearchSpaceByDetectingSupersets();
1659 void NarrowSearchSpaceByCollapsingUnrolledCode();
1660 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1661 void NarrowSearchSpaceByPickingWinnerRegs();
1662 void NarrowSearchSpaceUsingHeuristics();
1664 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1666 SmallVectorImpl<const Formula *> &Workspace,
1667 const Cost &CurCost,
1668 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1669 DenseSet<const SCEV *> &VisitedRegs) const;
1670 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1672 BasicBlock::iterator
1673 HoistInsertPosition(BasicBlock::iterator IP,
1674 const SmallVectorImpl<Instruction *> &Inputs) const;
1675 BasicBlock::iterator
1676 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1679 SCEVExpander &Rewriter) const;
1681 Value *Expand(const LSRFixup &LF,
1683 BasicBlock::iterator IP,
1684 SCEVExpander &Rewriter,
1685 SmallVectorImpl<WeakVH> &DeadInsts) const;
1686 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1688 SCEVExpander &Rewriter,
1689 SmallVectorImpl<WeakVH> &DeadInsts,
1691 void Rewrite(const LSRFixup &LF,
1693 SCEVExpander &Rewriter,
1694 SmallVectorImpl<WeakVH> &DeadInsts,
1696 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1700 LSRInstance(Loop *L, Pass *P);
1702 bool getChanged() const { return Changed; }
1704 void print_factors_and_types(raw_ostream &OS) const;
1705 void print_fixups(raw_ostream &OS) const;
1706 void print_uses(raw_ostream &OS) const;
1707 void print(raw_ostream &OS) const;
1713 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1714 /// inside the loop then try to eliminate the cast operation.
1715 void LSRInstance::OptimizeShadowIV() {
1716 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1717 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1720 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1721 UI != E; /* empty */) {
1722 IVUsers::const_iterator CandidateUI = UI;
1724 Instruction *ShadowUse = CandidateUI->getUser();
1726 bool IsSigned = false;
1728 /* If shadow use is a int->float cast then insert a second IV
1729 to eliminate this cast.
1731 for (unsigned i = 0; i < n; ++i)
1737 for (unsigned i = 0; i < n; ++i, ++d)
1740 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1742 DestTy = UCast->getDestTy();
1744 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1746 DestTy = SCast->getDestTy();
1748 if (!DestTy) continue;
1750 // If target does not support DestTy natively then do not apply
1751 // this transformation.
1752 if (!TTI.isTypeLegal(DestTy)) continue;
1754 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1756 if (PH->getNumIncomingValues() != 2) continue;
1758 Type *SrcTy = PH->getType();
1759 int Mantissa = DestTy->getFPMantissaWidth();
1760 if (Mantissa == -1) continue;
1761 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1764 unsigned Entry, Latch;
1765 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1773 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1774 if (!Init) continue;
1775 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1776 (double)Init->getSExtValue() :
1777 (double)Init->getZExtValue());
1779 BinaryOperator *Incr =
1780 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1781 if (!Incr) continue;
1782 if (Incr->getOpcode() != Instruction::Add
1783 && Incr->getOpcode() != Instruction::Sub)
1786 /* Initialize new IV, double d = 0.0 in above example. */
1788 if (Incr->getOperand(0) == PH)
1789 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1790 else if (Incr->getOperand(1) == PH)
1791 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1797 // Ignore negative constants, as the code below doesn't handle them
1798 // correctly. TODO: Remove this restriction.
1799 if (!C->getValue().isStrictlyPositive()) continue;
1801 /* Add new PHINode. */
1802 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1804 /* create new increment. '++d' in above example. */
1805 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1806 BinaryOperator *NewIncr =
1807 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1808 Instruction::FAdd : Instruction::FSub,
1809 NewPH, CFP, "IV.S.next.", Incr);
1811 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1812 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1814 /* Remove cast operation */
1815 ShadowUse->replaceAllUsesWith(NewPH);
1816 ShadowUse->eraseFromParent();
1822 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1823 /// set the IV user and stride information and return true, otherwise return
1825 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1826 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1827 if (UI->getUser() == Cond) {
1828 // NOTE: we could handle setcc instructions with multiple uses here, but
1829 // InstCombine does it as well for simple uses, it's not clear that it
1830 // occurs enough in real life to handle.
1837 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1838 /// a max computation.
1840 /// This is a narrow solution to a specific, but acute, problem. For loops
1846 /// } while (++i < n);
1848 /// the trip count isn't just 'n', because 'n' might not be positive. And
1849 /// unfortunately this can come up even for loops where the user didn't use
1850 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1851 /// will commonly be lowered like this:
1857 /// } while (++i < n);
1860 /// and then it's possible for subsequent optimization to obscure the if
1861 /// test in such a way that indvars can't find it.
1863 /// When indvars can't find the if test in loops like this, it creates a
1864 /// max expression, which allows it to give the loop a canonical
1865 /// induction variable:
1868 /// max = n < 1 ? 1 : n;
1871 /// } while (++i != max);
1873 /// Canonical induction variables are necessary because the loop passes
1874 /// are designed around them. The most obvious example of this is the
1875 /// LoopInfo analysis, which doesn't remember trip count values. It
1876 /// expects to be able to rediscover the trip count each time it is
1877 /// needed, and it does this using a simple analysis that only succeeds if
1878 /// the loop has a canonical induction variable.
1880 /// However, when it comes time to generate code, the maximum operation
1881 /// can be quite costly, especially if it's inside of an outer loop.
1883 /// This function solves this problem by detecting this type of loop and
1884 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1885 /// the instructions for the maximum computation.
1887 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1888 // Check that the loop matches the pattern we're looking for.
1889 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1890 Cond->getPredicate() != CmpInst::ICMP_NE)
1893 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1894 if (!Sel || !Sel->hasOneUse()) return Cond;
1896 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1897 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1899 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1901 // Add one to the backedge-taken count to get the trip count.
1902 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1903 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1905 // Check for a max calculation that matches the pattern. There's no check
1906 // for ICMP_ULE here because the comparison would be with zero, which
1907 // isn't interesting.
1908 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1909 const SCEVNAryExpr *Max = 0;
1910 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1911 Pred = ICmpInst::ICMP_SLE;
1913 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1914 Pred = ICmpInst::ICMP_SLT;
1916 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1917 Pred = ICmpInst::ICMP_ULT;
1924 // To handle a max with more than two operands, this optimization would
1925 // require additional checking and setup.
1926 if (Max->getNumOperands() != 2)
1929 const SCEV *MaxLHS = Max->getOperand(0);
1930 const SCEV *MaxRHS = Max->getOperand(1);
1932 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1933 // for a comparison with 1. For <= and >=, a comparison with zero.
1935 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1938 // Check the relevant induction variable for conformance to
1940 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1941 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1942 if (!AR || !AR->isAffine() ||
1943 AR->getStart() != One ||
1944 AR->getStepRecurrence(SE) != One)
1947 assert(AR->getLoop() == L &&
1948 "Loop condition operand is an addrec in a different loop!");
1950 // Check the right operand of the select, and remember it, as it will
1951 // be used in the new comparison instruction.
1953 if (ICmpInst::isTrueWhenEqual(Pred)) {
1954 // Look for n+1, and grab n.
1955 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1956 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1957 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1958 NewRHS = BO->getOperand(0);
1959 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1960 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1961 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1962 NewRHS = BO->getOperand(0);
1965 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1966 NewRHS = Sel->getOperand(1);
1967 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1968 NewRHS = Sel->getOperand(2);
1969 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1970 NewRHS = SU->getValue();
1972 // Max doesn't match expected pattern.
1975 // Determine the new comparison opcode. It may be signed or unsigned,
1976 // and the original comparison may be either equality or inequality.
1977 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1978 Pred = CmpInst::getInversePredicate(Pred);
1980 // Ok, everything looks ok to change the condition into an SLT or SGE and
1981 // delete the max calculation.
1983 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1985 // Delete the max calculation instructions.
1986 Cond->replaceAllUsesWith(NewCond);
1987 CondUse->setUser(NewCond);
1988 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1989 Cond->eraseFromParent();
1990 Sel->eraseFromParent();
1991 if (Cmp->use_empty())
1992 Cmp->eraseFromParent();
1996 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1997 /// postinc iv when possible.
1999 LSRInstance::OptimizeLoopTermCond() {
2000 SmallPtrSet<Instruction *, 4> PostIncs;
2002 BasicBlock *LatchBlock = L->getLoopLatch();
2003 SmallVector<BasicBlock*, 8> ExitingBlocks;
2004 L->getExitingBlocks(ExitingBlocks);
2006 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2007 BasicBlock *ExitingBlock = ExitingBlocks[i];
2009 // Get the terminating condition for the loop if possible. If we
2010 // can, we want to change it to use a post-incremented version of its
2011 // induction variable, to allow coalescing the live ranges for the IV into
2012 // one register value.
2014 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2017 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2018 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2021 // Search IVUsesByStride to find Cond's IVUse if there is one.
2022 IVStrideUse *CondUse = 0;
2023 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2024 if (!FindIVUserForCond(Cond, CondUse))
2027 // If the trip count is computed in terms of a max (due to ScalarEvolution
2028 // being unable to find a sufficient guard, for example), change the loop
2029 // comparison to use SLT or ULT instead of NE.
2030 // One consequence of doing this now is that it disrupts the count-down
2031 // optimization. That's not always a bad thing though, because in such
2032 // cases it may still be worthwhile to avoid a max.
2033 Cond = OptimizeMax(Cond, CondUse);
2035 // If this exiting block dominates the latch block, it may also use
2036 // the post-inc value if it won't be shared with other uses.
2037 // Check for dominance.
2038 if (!DT.dominates(ExitingBlock, LatchBlock))
2041 // Conservatively avoid trying to use the post-inc value in non-latch
2042 // exits if there may be pre-inc users in intervening blocks.
2043 if (LatchBlock != ExitingBlock)
2044 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2045 // Test if the use is reachable from the exiting block. This dominator
2046 // query is a conservative approximation of reachability.
2047 if (&*UI != CondUse &&
2048 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2049 // Conservatively assume there may be reuse if the quotient of their
2050 // strides could be a legal scale.
2051 const SCEV *A = IU.getStride(*CondUse, L);
2052 const SCEV *B = IU.getStride(*UI, L);
2053 if (!A || !B) continue;
2054 if (SE.getTypeSizeInBits(A->getType()) !=
2055 SE.getTypeSizeInBits(B->getType())) {
2056 if (SE.getTypeSizeInBits(A->getType()) >
2057 SE.getTypeSizeInBits(B->getType()))
2058 B = SE.getSignExtendExpr(B, A->getType());
2060 A = SE.getSignExtendExpr(A, B->getType());
2062 if (const SCEVConstant *D =
2063 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2064 const ConstantInt *C = D->getValue();
2065 // Stride of one or negative one can have reuse with non-addresses.
2066 if (C->isOne() || C->isAllOnesValue())
2067 goto decline_post_inc;
2068 // Avoid weird situations.
2069 if (C->getValue().getMinSignedBits() >= 64 ||
2070 C->getValue().isMinSignedValue())
2071 goto decline_post_inc;
2072 // Check for possible scaled-address reuse.
2073 Type *AccessTy = getAccessType(UI->getUser());
2074 int64_t Scale = C->getSExtValue();
2075 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2077 /*HasBaseReg=*/ false, Scale))
2078 goto decline_post_inc;
2080 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2082 /*HasBaseReg=*/ false, Scale))
2083 goto decline_post_inc;
2087 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2090 // It's possible for the setcc instruction to be anywhere in the loop, and
2091 // possible for it to have multiple users. If it is not immediately before
2092 // the exiting block branch, move it.
2093 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2094 if (Cond->hasOneUse()) {
2095 Cond->moveBefore(TermBr);
2097 // Clone the terminating condition and insert into the loopend.
2098 ICmpInst *OldCond = Cond;
2099 Cond = cast<ICmpInst>(Cond->clone());
2100 Cond->setName(L->getHeader()->getName() + ".termcond");
2101 ExitingBlock->getInstList().insert(TermBr, Cond);
2103 // Clone the IVUse, as the old use still exists!
2104 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2105 TermBr->replaceUsesOfWith(OldCond, Cond);
2109 // If we get to here, we know that we can transform the setcc instruction to
2110 // use the post-incremented version of the IV, allowing us to coalesce the
2111 // live ranges for the IV correctly.
2112 CondUse->transformToPostInc(L);
2115 PostIncs.insert(Cond);
2119 // Determine an insertion point for the loop induction variable increment. It
2120 // must dominate all the post-inc comparisons we just set up, and it must
2121 // dominate the loop latch edge.
2122 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2123 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2124 E = PostIncs.end(); I != E; ++I) {
2126 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2128 if (BB == (*I)->getParent())
2129 IVIncInsertPos = *I;
2130 else if (BB != IVIncInsertPos->getParent())
2131 IVIncInsertPos = BB->getTerminator();
2135 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2136 /// at the given offset and other details. If so, update the use and
2139 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2140 LSRUse::KindType Kind, Type *AccessTy) {
2141 int64_t NewMinOffset = LU.MinOffset;
2142 int64_t NewMaxOffset = LU.MaxOffset;
2143 Type *NewAccessTy = AccessTy;
2145 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2146 // something conservative, however this can pessimize in the case that one of
2147 // the uses will have all its uses outside the loop, for example.
2148 if (LU.Kind != Kind)
2150 // Conservatively assume HasBaseReg is true for now.
2151 if (NewOffset < LU.MinOffset) {
2152 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2153 LU.MaxOffset - NewOffset, HasBaseReg))
2155 NewMinOffset = NewOffset;
2156 } else if (NewOffset > LU.MaxOffset) {
2157 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2158 NewOffset - LU.MinOffset, HasBaseReg))
2160 NewMaxOffset = NewOffset;
2162 // Check for a mismatched access type, and fall back conservatively as needed.
2163 // TODO: Be less conservative when the type is similar and can use the same
2164 // addressing modes.
2165 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2166 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2169 LU.MinOffset = NewMinOffset;
2170 LU.MaxOffset = NewMaxOffset;
2171 LU.AccessTy = NewAccessTy;
2172 if (NewOffset != LU.Offsets.back())
2173 LU.Offsets.push_back(NewOffset);
2177 /// getUse - Return an LSRUse index and an offset value for a fixup which
2178 /// needs the given expression, with the given kind and optional access type.
2179 /// Either reuse an existing use or create a new one, as needed.
2180 std::pair<size_t, int64_t>
2181 LSRInstance::getUse(const SCEV *&Expr,
2182 LSRUse::KindType Kind, Type *AccessTy) {
2183 const SCEV *Copy = Expr;
2184 int64_t Offset = ExtractImmediate(Expr, SE);
2186 // Basic uses can't accept any offset, for example.
2187 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2188 Offset, /*HasBaseReg=*/ true)) {
2193 std::pair<UseMapTy::iterator, bool> P =
2194 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2196 // A use already existed with this base.
2197 size_t LUIdx = P.first->second;
2198 LSRUse &LU = Uses[LUIdx];
2199 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2201 return std::make_pair(LUIdx, Offset);
2204 // Create a new use.
2205 size_t LUIdx = Uses.size();
2206 P.first->second = LUIdx;
2207 Uses.push_back(LSRUse(Kind, AccessTy));
2208 LSRUse &LU = Uses[LUIdx];
2210 // We don't need to track redundant offsets, but we don't need to go out
2211 // of our way here to avoid them.
2212 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2213 LU.Offsets.push_back(Offset);
2215 LU.MinOffset = Offset;
2216 LU.MaxOffset = Offset;
2217 return std::make_pair(LUIdx, Offset);
2220 /// DeleteUse - Delete the given use from the Uses list.
2221 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2222 if (&LU != &Uses.back())
2223 std::swap(LU, Uses.back());
2227 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2230 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2231 /// a formula that has the same registers as the given formula.
2233 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2234 const LSRUse &OrigLU) {
2235 // Search all uses for the formula. This could be more clever.
2236 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2237 LSRUse &LU = Uses[LUIdx];
2238 // Check whether this use is close enough to OrigLU, to see whether it's
2239 // worthwhile looking through its formulae.
2240 // Ignore ICmpZero uses because they may contain formulae generated by
2241 // GenerateICmpZeroScales, in which case adding fixup offsets may
2243 if (&LU != &OrigLU &&
2244 LU.Kind != LSRUse::ICmpZero &&
2245 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2246 LU.WidestFixupType == OrigLU.WidestFixupType &&
2247 LU.HasFormulaWithSameRegs(OrigF)) {
2248 // Scan through this use's formulae.
2249 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2250 E = LU.Formulae.end(); I != E; ++I) {
2251 const Formula &F = *I;
2252 // Check to see if this formula has the same registers and symbols
2254 if (F.BaseRegs == OrigF.BaseRegs &&
2255 F.ScaledReg == OrigF.ScaledReg &&
2256 F.BaseGV == OrigF.BaseGV &&
2257 F.Scale == OrigF.Scale &&
2258 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2259 if (F.BaseOffset == 0)
2261 // This is the formula where all the registers and symbols matched;
2262 // there aren't going to be any others. Since we declined it, we
2263 // can skip the rest of the formulae and proceed to the next LSRUse.
2270 // Nothing looked good.
2274 void LSRInstance::CollectInterestingTypesAndFactors() {
2275 SmallSetVector<const SCEV *, 4> Strides;
2277 // Collect interesting types and strides.
2278 SmallVector<const SCEV *, 4> Worklist;
2279 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2280 const SCEV *Expr = IU.getExpr(*UI);
2282 // Collect interesting types.
2283 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2285 // Add strides for mentioned loops.
2286 Worklist.push_back(Expr);
2288 const SCEV *S = Worklist.pop_back_val();
2289 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2290 if (AR->getLoop() == L)
2291 Strides.insert(AR->getStepRecurrence(SE));
2292 Worklist.push_back(AR->getStart());
2293 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2294 Worklist.append(Add->op_begin(), Add->op_end());
2296 } while (!Worklist.empty());
2299 // Compute interesting factors from the set of interesting strides.
2300 for (SmallSetVector<const SCEV *, 4>::const_iterator
2301 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2302 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2303 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2304 const SCEV *OldStride = *I;
2305 const SCEV *NewStride = *NewStrideIter;
2307 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2308 SE.getTypeSizeInBits(NewStride->getType())) {
2309 if (SE.getTypeSizeInBits(OldStride->getType()) >
2310 SE.getTypeSizeInBits(NewStride->getType()))
2311 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2313 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2315 if (const SCEVConstant *Factor =
2316 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2318 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2319 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2320 } else if (const SCEVConstant *Factor =
2321 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2324 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2325 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2329 // If all uses use the same type, don't bother looking for truncation-based
2331 if (Types.size() == 1)
2334 DEBUG(print_factors_and_types(dbgs()));
2337 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2338 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2339 /// Instructions to IVStrideUses, we could partially skip this.
2340 static User::op_iterator
2341 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2342 Loop *L, ScalarEvolution &SE) {
2343 for(; OI != OE; ++OI) {
2344 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2345 if (!SE.isSCEVable(Oper->getType()))
2348 if (const SCEVAddRecExpr *AR =
2349 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2350 if (AR->getLoop() == L)
2358 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2359 /// operands, so wrap it in a convenient helper.
2360 static Value *getWideOperand(Value *Oper) {
2361 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2362 return Trunc->getOperand(0);
2366 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2368 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2369 Type *LType = LVal->getType();
2370 Type *RType = RVal->getType();
2371 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2374 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2375 /// NULL for any constant. Returning the expression itself is
2376 /// conservative. Returning a deeper subexpression is more precise and valid as
2377 /// long as it isn't less complex than another subexpression. For expressions
2378 /// involving multiple unscaled values, we need to return the pointer-type
2379 /// SCEVUnknown. This avoids forming chains across objects, such as:
2380 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2382 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2383 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2384 static const SCEV *getExprBase(const SCEV *S) {
2385 switch (S->getSCEVType()) {
2386 default: // uncluding scUnknown.
2391 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2393 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2395 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2397 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2398 // there's nothing more complex.
2399 // FIXME: not sure if we want to recognize negation.
2400 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2401 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2402 E(Add->op_begin()); I != E; ++I) {
2403 const SCEV *SubExpr = *I;
2404 if (SubExpr->getSCEVType() == scAddExpr)
2405 return getExprBase(SubExpr);
2407 if (SubExpr->getSCEVType() != scMulExpr)
2410 return S; // all operands are scaled, be conservative.
2413 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2417 /// Return true if the chain increment is profitable to expand into a loop
2418 /// invariant value, which may require its own register. A profitable chain
2419 /// increment will be an offset relative to the same base. We allow such offsets
2420 /// to potentially be used as chain increment as long as it's not obviously
2421 /// expensive to expand using real instructions.
2422 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2423 const SCEV *IncExpr,
2424 ScalarEvolution &SE) {
2425 // Aggressively form chains when -stress-ivchain.
2429 // Do not replace a constant offset from IV head with a nonconstant IV
2431 if (!isa<SCEVConstant>(IncExpr)) {
2432 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2433 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2437 SmallPtrSet<const SCEV*, 8> Processed;
2438 return !isHighCostExpansion(IncExpr, Processed, SE);
2441 /// Return true if the number of registers needed for the chain is estimated to
2442 /// be less than the number required for the individual IV users. First prohibit
2443 /// any IV users that keep the IV live across increments (the Users set should
2444 /// be empty). Next count the number and type of increments in the chain.
2446 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2447 /// effectively use postinc addressing modes. Only consider it profitable it the
2448 /// increments can be computed in fewer registers when chained.
2450 /// TODO: Consider IVInc free if it's already used in another chains.
2452 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2453 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2457 if (!Chain.hasIncs())
2460 if (!Users.empty()) {
2461 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2462 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2463 E = Users.end(); I != E; ++I) {
2464 dbgs() << " " << **I << "\n";
2468 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2470 // The chain itself may require a register, so intialize cost to 1.
2473 // A complete chain likely eliminates the need for keeping the original IV in
2474 // a register. LSR does not currently know how to form a complete chain unless
2475 // the header phi already exists.
2476 if (isa<PHINode>(Chain.tailUserInst())
2477 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2480 const SCEV *LastIncExpr = 0;
2481 unsigned NumConstIncrements = 0;
2482 unsigned NumVarIncrements = 0;
2483 unsigned NumReusedIncrements = 0;
2484 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2487 if (I->IncExpr->isZero())
2490 // Incrementing by zero or some constant is neutral. We assume constants can
2491 // be folded into an addressing mode or an add's immediate operand.
2492 if (isa<SCEVConstant>(I->IncExpr)) {
2493 ++NumConstIncrements;
2497 if (I->IncExpr == LastIncExpr)
2498 ++NumReusedIncrements;
2502 LastIncExpr = I->IncExpr;
2504 // An IV chain with a single increment is handled by LSR's postinc
2505 // uses. However, a chain with multiple increments requires keeping the IV's
2506 // value live longer than it needs to be if chained.
2507 if (NumConstIncrements > 1)
2510 // Materializing increment expressions in the preheader that didn't exist in
2511 // the original code may cost a register. For example, sign-extended array
2512 // indices can produce ridiculous increments like this:
2513 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2514 cost += NumVarIncrements;
2516 // Reusing variable increments likely saves a register to hold the multiple of
2518 cost -= NumReusedIncrements;
2520 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2526 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2528 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2529 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2530 // When IVs are used as types of varying widths, they are generally converted
2531 // to a wider type with some uses remaining narrow under a (free) trunc.
2532 Value *const NextIV = getWideOperand(IVOper);
2533 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2534 const SCEV *const OperExprBase = getExprBase(OperExpr);
2536 // Visit all existing chains. Check if its IVOper can be computed as a
2537 // profitable loop invariant increment from the last link in the Chain.
2538 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2539 const SCEV *LastIncExpr = 0;
2540 for (; ChainIdx < NChains; ++ChainIdx) {
2541 IVChain &Chain = IVChainVec[ChainIdx];
2543 // Prune the solution space aggressively by checking that both IV operands
2544 // are expressions that operate on the same unscaled SCEVUnknown. This
2545 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2546 // first avoids creating extra SCEV expressions.
2547 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2550 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2551 if (!isCompatibleIVType(PrevIV, NextIV))
2554 // A phi node terminates a chain.
2555 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2558 // The increment must be loop-invariant so it can be kept in a register.
2559 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2560 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2561 if (!SE.isLoopInvariant(IncExpr, L))
2564 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2565 LastIncExpr = IncExpr;
2569 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2570 // bother for phi nodes, because they must be last in the chain.
2571 if (ChainIdx == NChains) {
2572 if (isa<PHINode>(UserInst))
2574 if (NChains >= MaxChains && !StressIVChain) {
2575 DEBUG(dbgs() << "IV Chain Limit\n");
2578 LastIncExpr = OperExpr;
2579 // IVUsers may have skipped over sign/zero extensions. We don't currently
2580 // attempt to form chains involving extensions unless they can be hoisted
2581 // into this loop's AddRec.
2582 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2585 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2587 ChainUsersVec.resize(NChains);
2588 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2589 << ") IV=" << *LastIncExpr << "\n");
2591 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2592 << ") IV+" << *LastIncExpr << "\n");
2593 // Add this IV user to the end of the chain.
2594 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2596 IVChain &Chain = IVChainVec[ChainIdx];
2598 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2599 // This chain's NearUsers become FarUsers.
2600 if (!LastIncExpr->isZero()) {
2601 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2606 // All other uses of IVOperand become near uses of the chain.
2607 // We currently ignore intermediate values within SCEV expressions, assuming
2608 // they will eventually be used be the current chain, or can be computed
2609 // from one of the chain increments. To be more precise we could
2610 // transitively follow its user and only add leaf IV users to the set.
2611 for (User *U : IVOper->users()) {
2612 Instruction *OtherUse = dyn_cast<Instruction>(U);
2615 // Uses in the chain will no longer be uses if the chain is formed.
2616 // Include the head of the chain in this iteration (not Chain.begin()).
2617 IVChain::const_iterator IncIter = Chain.Incs.begin();
2618 IVChain::const_iterator IncEnd = Chain.Incs.end();
2619 for( ; IncIter != IncEnd; ++IncIter) {
2620 if (IncIter->UserInst == OtherUse)
2623 if (IncIter != IncEnd)
2626 if (SE.isSCEVable(OtherUse->getType())
2627 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2628 && IU.isIVUserOrOperand(OtherUse)) {
2631 NearUsers.insert(OtherUse);
2634 // Since this user is part of the chain, it's no longer considered a use
2636 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2639 /// CollectChains - Populate the vector of Chains.
2641 /// This decreases ILP at the architecture level. Targets with ample registers,
2642 /// multiple memory ports, and no register renaming probably don't want
2643 /// this. However, such targets should probably disable LSR altogether.
2645 /// The job of LSR is to make a reasonable choice of induction variables across
2646 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2647 /// ILP *within the loop* if the target wants it.
2649 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2650 /// will not reorder memory operations, it will recognize this as a chain, but
2651 /// will generate redundant IV increments. Ideally this would be corrected later
2652 /// by a smart scheduler:
2658 /// TODO: Walk the entire domtree within this loop, not just the path to the
2659 /// loop latch. This will discover chains on side paths, but requires
2660 /// maintaining multiple copies of the Chains state.
2661 void LSRInstance::CollectChains() {
2662 DEBUG(dbgs() << "Collecting IV Chains.\n");
2663 SmallVector<ChainUsers, 8> ChainUsersVec;
2665 SmallVector<BasicBlock *,8> LatchPath;
2666 BasicBlock *LoopHeader = L->getHeader();
2667 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2668 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2669 LatchPath.push_back(Rung->getBlock());
2671 LatchPath.push_back(LoopHeader);
2673 // Walk the instruction stream from the loop header to the loop latch.
2674 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2675 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2676 BBIter != BBEnd; ++BBIter) {
2677 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2679 // Skip instructions that weren't seen by IVUsers analysis.
2680 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2683 // Ignore users that are part of a SCEV expression. This way we only
2684 // consider leaf IV Users. This effectively rediscovers a portion of
2685 // IVUsers analysis but in program order this time.
2686 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2689 // Remove this instruction from any NearUsers set it may be in.
2690 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2691 ChainIdx < NChains; ++ChainIdx) {
2692 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2694 // Search for operands that can be chained.
2695 SmallPtrSet<Instruction*, 4> UniqueOperands;
2696 User::op_iterator IVOpEnd = I->op_end();
2697 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2698 while (IVOpIter != IVOpEnd) {
2699 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2700 if (UniqueOperands.insert(IVOpInst))
2701 ChainInstruction(I, IVOpInst, ChainUsersVec);
2702 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2704 } // Continue walking down the instructions.
2705 } // Continue walking down the domtree.
2706 // Visit phi backedges to determine if the chain can generate the IV postinc.
2707 for (BasicBlock::iterator I = L->getHeader()->begin();
2708 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2709 if (!SE.isSCEVable(PN->getType()))
2713 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2715 ChainInstruction(PN, IncV, ChainUsersVec);
2717 // Remove any unprofitable chains.
2718 unsigned ChainIdx = 0;
2719 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2720 UsersIdx < NChains; ++UsersIdx) {
2721 if (!isProfitableChain(IVChainVec[UsersIdx],
2722 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2724 // Preserve the chain at UsesIdx.
2725 if (ChainIdx != UsersIdx)
2726 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2727 FinalizeChain(IVChainVec[ChainIdx]);
2730 IVChainVec.resize(ChainIdx);
2733 void LSRInstance::FinalizeChain(IVChain &Chain) {
2734 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2735 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2737 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2739 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2740 User::op_iterator UseI =
2741 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2742 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2743 IVIncSet.insert(UseI);
2747 /// Return true if the IVInc can be folded into an addressing mode.
2748 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2749 Value *Operand, const TargetTransformInfo &TTI) {
2750 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2751 if (!IncConst || !isAddressUse(UserInst, Operand))
2754 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2757 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2758 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2759 getAccessType(UserInst), /*BaseGV=*/ 0,
2760 IncOffset, /*HaseBaseReg=*/ false))
2766 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2767 /// materialize the IV user's operand from the previous IV user's operand.
2768 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2769 SmallVectorImpl<WeakVH> &DeadInsts) {
2770 // Find the new IVOperand for the head of the chain. It may have been replaced
2772 const IVInc &Head = Chain.Incs[0];
2773 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2774 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2775 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2778 while (IVOpIter != IVOpEnd) {
2779 IVSrc = getWideOperand(*IVOpIter);
2781 // If this operand computes the expression that the chain needs, we may use
2782 // it. (Check this after setting IVSrc which is used below.)
2784 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2785 // narrow for the chain, so we can no longer use it. We do allow using a
2786 // wider phi, assuming the LSR checked for free truncation. In that case we
2787 // should already have a truncate on this operand such that
2788 // getSCEV(IVSrc) == IncExpr.
2789 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2790 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2793 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2795 if (IVOpIter == IVOpEnd) {
2796 // Gracefully give up on this chain.
2797 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2801 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2802 Type *IVTy = IVSrc->getType();
2803 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2804 const SCEV *LeftOverExpr = 0;
2805 for (IVChain::const_iterator IncI = Chain.begin(),
2806 IncE = Chain.end(); IncI != IncE; ++IncI) {
2808 Instruction *InsertPt = IncI->UserInst;
2809 if (isa<PHINode>(InsertPt))
2810 InsertPt = L->getLoopLatch()->getTerminator();
2812 // IVOper will replace the current IV User's operand. IVSrc is the IV
2813 // value currently held in a register.
2814 Value *IVOper = IVSrc;
2815 if (!IncI->IncExpr->isZero()) {
2816 // IncExpr was the result of subtraction of two narrow values, so must
2818 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2819 LeftOverExpr = LeftOverExpr ?
2820 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2822 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2823 // Expand the IV increment.
2824 Rewriter.clearPostInc();
2825 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2826 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2827 SE.getUnknown(IncV));
2828 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2830 // If an IV increment can't be folded, use it as the next IV value.
2831 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2833 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2838 Type *OperTy = IncI->IVOperand->getType();
2839 if (IVTy != OperTy) {
2840 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2841 "cannot extend a chained IV");
2842 IRBuilder<> Builder(InsertPt);
2843 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2845 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2846 DeadInsts.push_back(IncI->IVOperand);
2848 // If LSR created a new, wider phi, we may also replace its postinc. We only
2849 // do this if we also found a wide value for the head of the chain.
2850 if (isa<PHINode>(Chain.tailUserInst())) {
2851 for (BasicBlock::iterator I = L->getHeader()->begin();
2852 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2853 if (!isCompatibleIVType(Phi, IVSrc))
2855 Instruction *PostIncV = dyn_cast<Instruction>(
2856 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2857 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2859 Value *IVOper = IVSrc;
2860 Type *PostIncTy = PostIncV->getType();
2861 if (IVTy != PostIncTy) {
2862 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2863 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2864 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2865 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2867 Phi->replaceUsesOfWith(PostIncV, IVOper);
2868 DeadInsts.push_back(PostIncV);
2873 void LSRInstance::CollectFixupsAndInitialFormulae() {
2874 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2875 Instruction *UserInst = UI->getUser();
2876 // Skip IV users that are part of profitable IV Chains.
2877 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2878 UI->getOperandValToReplace());
2879 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2880 if (IVIncSet.count(UseI))
2884 LSRFixup &LF = getNewFixup();
2885 LF.UserInst = UserInst;
2886 LF.OperandValToReplace = UI->getOperandValToReplace();
2887 LF.PostIncLoops = UI->getPostIncLoops();
2889 LSRUse::KindType Kind = LSRUse::Basic;
2891 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2892 Kind = LSRUse::Address;
2893 AccessTy = getAccessType(LF.UserInst);
2896 const SCEV *S = IU.getExpr(*UI);
2898 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2899 // (N - i == 0), and this allows (N - i) to be the expression that we work
2900 // with rather than just N or i, so we can consider the register
2901 // requirements for both N and i at the same time. Limiting this code to
2902 // equality icmps is not a problem because all interesting loops use
2903 // equality icmps, thanks to IndVarSimplify.
2904 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2905 if (CI->isEquality()) {
2906 // Swap the operands if needed to put the OperandValToReplace on the
2907 // left, for consistency.
2908 Value *NV = CI->getOperand(1);
2909 if (NV == LF.OperandValToReplace) {
2910 CI->setOperand(1, CI->getOperand(0));
2911 CI->setOperand(0, NV);
2912 NV = CI->getOperand(1);
2916 // x == y --> x - y == 0
2917 const SCEV *N = SE.getSCEV(NV);
2918 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2919 // S is normalized, so normalize N before folding it into S
2920 // to keep the result normalized.
2921 N = TransformForPostIncUse(Normalize, N, CI, 0,
2922 LF.PostIncLoops, SE, DT);
2923 Kind = LSRUse::ICmpZero;
2924 S = SE.getMinusSCEV(N, S);
2927 // -1 and the negations of all interesting strides (except the negation
2928 // of -1) are now also interesting.
2929 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2930 if (Factors[i] != -1)
2931 Factors.insert(-(uint64_t)Factors[i]);
2935 // Set up the initial formula for this use.
2936 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2938 LF.Offset = P.second;
2939 LSRUse &LU = Uses[LF.LUIdx];
2940 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2941 if (!LU.WidestFixupType ||
2942 SE.getTypeSizeInBits(LU.WidestFixupType) <
2943 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2944 LU.WidestFixupType = LF.OperandValToReplace->getType();
2946 // If this is the first use of this LSRUse, give it a formula.
2947 if (LU.Formulae.empty()) {
2948 InsertInitialFormula(S, LU, LF.LUIdx);
2949 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2953 DEBUG(print_fixups(dbgs()));
2956 /// InsertInitialFormula - Insert a formula for the given expression into
2957 /// the given use, separating out loop-variant portions from loop-invariant
2958 /// and loop-computable portions.
2960 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2961 // Mark uses whose expressions cannot be expanded.
2962 if (!isSafeToExpand(S, SE))
2963 LU.RigidFormula = true;
2966 F.InitialMatch(S, L, SE);
2967 bool Inserted = InsertFormula(LU, LUIdx, F);
2968 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2971 /// InsertSupplementalFormula - Insert a simple single-register formula for
2972 /// the given expression into the given use.
2974 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2975 LSRUse &LU, size_t LUIdx) {
2977 F.BaseRegs.push_back(S);
2978 F.HasBaseReg = true;
2979 bool Inserted = InsertFormula(LU, LUIdx, F);
2980 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2983 /// CountRegisters - Note which registers are used by the given formula,
2984 /// updating RegUses.
2985 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2987 RegUses.CountRegister(F.ScaledReg, LUIdx);
2988 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2989 E = F.BaseRegs.end(); I != E; ++I)
2990 RegUses.CountRegister(*I, LUIdx);
2993 /// InsertFormula - If the given formula has not yet been inserted, add it to
2994 /// the list, and return true. Return false otherwise.
2995 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2996 if (!LU.InsertFormula(F))
2999 CountRegisters(F, LUIdx);
3003 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3004 /// loop-invariant values which we're tracking. These other uses will pin these
3005 /// values in registers, making them less profitable for elimination.
3006 /// TODO: This currently misses non-constant addrec step registers.
3007 /// TODO: Should this give more weight to users inside the loop?
3009 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3010 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3011 SmallPtrSet<const SCEV *, 8> Inserted;
3013 while (!Worklist.empty()) {
3014 const SCEV *S = Worklist.pop_back_val();
3016 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3017 Worklist.append(N->op_begin(), N->op_end());
3018 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3019 Worklist.push_back(C->getOperand());
3020 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3021 Worklist.push_back(D->getLHS());
3022 Worklist.push_back(D->getRHS());
3023 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3024 if (!Inserted.insert(US)) continue;
3025 const Value *V = US->getValue();
3026 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3027 // Look for instructions defined outside the loop.
3028 if (L->contains(Inst)) continue;
3029 } else if (isa<UndefValue>(V))
3030 // Undef doesn't have a live range, so it doesn't matter.
3032 for (const Use &U : V->uses()) {
3033 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3034 // Ignore non-instructions.
3037 // Ignore instructions in other functions (as can happen with
3039 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3041 // Ignore instructions not dominated by the loop.
3042 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3043 UserInst->getParent() :
3044 cast<PHINode>(UserInst)->getIncomingBlock(
3045 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3046 if (!DT.dominates(L->getHeader(), UseBB))
3048 // Ignore uses which are part of other SCEV expressions, to avoid
3049 // analyzing them multiple times.
3050 if (SE.isSCEVable(UserInst->getType())) {
3051 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3052 // If the user is a no-op, look through to its uses.
3053 if (!isa<SCEVUnknown>(UserS))
3057 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3061 // Ignore icmp instructions which are already being analyzed.
3062 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3063 unsigned OtherIdx = !U.getOperandNo();
3064 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3065 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3069 LSRFixup &LF = getNewFixup();
3070 LF.UserInst = const_cast<Instruction *>(UserInst);
3071 LF.OperandValToReplace = U;
3072 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
3074 LF.Offset = P.second;
3075 LSRUse &LU = Uses[LF.LUIdx];
3076 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3077 if (!LU.WidestFixupType ||
3078 SE.getTypeSizeInBits(LU.WidestFixupType) <
3079 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3080 LU.WidestFixupType = LF.OperandValToReplace->getType();
3081 InsertSupplementalFormula(US, LU, LF.LUIdx);
3082 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3089 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3090 /// separate registers. If C is non-null, multiply each subexpression by C.
3092 /// Return remainder expression after factoring the subexpressions captured by
3093 /// Ops. If Ops is complete, return NULL.
3094 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3095 SmallVectorImpl<const SCEV *> &Ops,
3097 ScalarEvolution &SE,
3098 unsigned Depth = 0) {
3099 // Arbitrarily cap recursion to protect compile time.
3103 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3104 // Break out add operands.
3105 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3107 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3109 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3112 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3113 // Split a non-zero base out of an addrec.
3114 if (AR->getStart()->isZero())
3117 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3118 C, Ops, L, SE, Depth+1);
3119 // Split the non-zero AddRec unless it is part of a nested recurrence that
3120 // does not pertain to this loop.
3121 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3122 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3125 if (Remainder != AR->getStart()) {
3127 Remainder = SE.getConstant(AR->getType(), 0);
3128 return SE.getAddRecExpr(Remainder,
3129 AR->getStepRecurrence(SE),
3131 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3134 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3135 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3136 if (Mul->getNumOperands() != 2)
3138 if (const SCEVConstant *Op0 =
3139 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3140 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3141 const SCEV *Remainder =
3142 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3144 Ops.push_back(SE.getMulExpr(C, Remainder));
3151 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3153 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3156 // Arbitrarily cap recursion to protect compile time.
3157 if (Depth >= 3) return;
3159 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3160 const SCEV *BaseReg = Base.BaseRegs[i];
3162 SmallVector<const SCEV *, 8> AddOps;
3163 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3165 AddOps.push_back(Remainder);
3167 if (AddOps.size() == 1) continue;
3169 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3170 JE = AddOps.end(); J != JE; ++J) {
3172 // Loop-variant "unknown" values are uninteresting; we won't be able to
3173 // do anything meaningful with them.
3174 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3177 // Don't pull a constant into a register if the constant could be folded
3178 // into an immediate field.
3179 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3180 LU.AccessTy, *J, Base.getNumRegs() > 1))
3183 // Collect all operands except *J.
3184 SmallVector<const SCEV *, 8> InnerAddOps(
3185 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3186 InnerAddOps.append(std::next(J),
3187 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3189 // Don't leave just a constant behind in a register if the constant could
3190 // be folded into an immediate field.
3191 if (InnerAddOps.size() == 1 &&
3192 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3193 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3196 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3197 if (InnerSum->isZero())
3201 // Add the remaining pieces of the add back into the new formula.
3202 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3204 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3205 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3206 InnerSumSC->getValue()->getZExtValue())) {
3207 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3208 InnerSumSC->getValue()->getZExtValue();
3209 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3211 F.BaseRegs[i] = InnerSum;
3213 // Add J as its own register, or an unfolded immediate.
3214 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3215 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3216 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3217 SC->getValue()->getZExtValue()))
3218 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3219 SC->getValue()->getZExtValue();
3221 F.BaseRegs.push_back(*J);
3223 if (InsertFormula(LU, LUIdx, F))
3224 // If that formula hadn't been seen before, recurse to find more like
3226 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3231 /// GenerateCombinations - Generate a formula consisting of all of the
3232 /// loop-dominating registers added into a single register.
3233 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3235 // This method is only interesting on a plurality of registers.
3236 if (Base.BaseRegs.size() <= 1) return;
3240 SmallVector<const SCEV *, 4> Ops;
3241 for (SmallVectorImpl<const SCEV *>::const_iterator
3242 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3243 const SCEV *BaseReg = *I;
3244 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3245 !SE.hasComputableLoopEvolution(BaseReg, L))
3246 Ops.push_back(BaseReg);
3248 F.BaseRegs.push_back(BaseReg);
3250 if (Ops.size() > 1) {
3251 const SCEV *Sum = SE.getAddExpr(Ops);
3252 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3253 // opportunity to fold something. For now, just ignore such cases
3254 // rather than proceed with zero in a register.
3255 if (!Sum->isZero()) {
3256 F.BaseRegs.push_back(Sum);
3257 (void)InsertFormula(LU, LUIdx, F);
3262 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3263 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3265 // We can't add a symbolic offset if the address already contains one.
3266 if (Base.BaseGV) return;
3268 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3269 const SCEV *G = Base.BaseRegs[i];
3270 GlobalValue *GV = ExtractSymbol(G, SE);
3271 if (G->isZero() || !GV)
3275 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3278 (void)InsertFormula(LU, LUIdx, F);
3282 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3283 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3285 // TODO: For now, just add the min and max offset, because it usually isn't
3286 // worthwhile looking at everything inbetween.
3287 SmallVector<int64_t, 2> Worklist;
3288 Worklist.push_back(LU.MinOffset);
3289 if (LU.MaxOffset != LU.MinOffset)
3290 Worklist.push_back(LU.MaxOffset);
3292 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3293 const SCEV *G = Base.BaseRegs[i];
3295 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3296 E = Worklist.end(); I != E; ++I) {
3298 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3299 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3301 // Add the offset to the base register.
3302 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3303 // If it cancelled out, drop the base register, otherwise update it.
3304 if (NewG->isZero()) {
3305 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3306 F.BaseRegs.pop_back();
3308 F.BaseRegs[i] = NewG;
3310 (void)InsertFormula(LU, LUIdx, F);
3314 int64_t Imm = ExtractImmediate(G, SE);
3315 if (G->isZero() || Imm == 0)
3318 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3319 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3322 (void)InsertFormula(LU, LUIdx, F);
3326 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3327 /// the comparison. For example, x == y -> x*c == y*c.
3328 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3330 if (LU.Kind != LSRUse::ICmpZero) return;
3332 // Determine the integer type for the base formula.
3333 Type *IntTy = Base.getType();
3335 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3337 // Don't do this if there is more than one offset.
3338 if (LU.MinOffset != LU.MaxOffset) return;
3340 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3342 // Check each interesting stride.
3343 for (SmallSetVector<int64_t, 8>::const_iterator
3344 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3345 int64_t Factor = *I;
3347 // Check that the multiplication doesn't overflow.
3348 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3350 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3351 if (NewBaseOffset / Factor != Base.BaseOffset)
3353 // If the offset will be truncated at this use, check that it is in bounds.
3354 if (!IntTy->isPointerTy() &&
3355 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3358 // Check that multiplying with the use offset doesn't overflow.
3359 int64_t Offset = LU.MinOffset;
3360 if (Offset == INT64_MIN && Factor == -1)
3362 Offset = (uint64_t)Offset * Factor;
3363 if (Offset / Factor != LU.MinOffset)
3365 // If the offset will be truncated at this use, check that it is in bounds.
3366 if (!IntTy->isPointerTy() &&
3367 !ConstantInt::isValueValidForType(IntTy, Offset))
3371 F.BaseOffset = NewBaseOffset;
3373 // Check that this scale is legal.
3374 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3377 // Compensate for the use having MinOffset built into it.
3378 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3380 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3382 // Check that multiplying with each base register doesn't overflow.
3383 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3384 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3385 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3389 // Check that multiplying with the scaled register doesn't overflow.
3391 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3392 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3396 // Check that multiplying with the unfolded offset doesn't overflow.
3397 if (F.UnfoldedOffset != 0) {
3398 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3400 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3401 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3403 // If the offset will be truncated, check that it is in bounds.
3404 if (!IntTy->isPointerTy() &&
3405 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3409 // If we make it here and it's legal, add it.
3410 (void)InsertFormula(LU, LUIdx, F);
3415 /// GenerateScales - Generate stride factor reuse formulae by making use of
3416 /// scaled-offset address modes, for example.
3417 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3418 // Determine the integer type for the base formula.
3419 Type *IntTy = Base.getType();
3422 // If this Formula already has a scaled register, we can't add another one.
3423 if (Base.Scale != 0) return;
3425 // Check each interesting stride.
3426 for (SmallSetVector<int64_t, 8>::const_iterator
3427 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3428 int64_t Factor = *I;
3430 Base.Scale = Factor;
3431 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3432 // Check whether this scale is going to be legal.
3433 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3435 // As a special-case, handle special out-of-loop Basic users specially.
3436 // TODO: Reconsider this special case.
3437 if (LU.Kind == LSRUse::Basic &&
3438 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3439 LU.AccessTy, Base) &&
3440 LU.AllFixupsOutsideLoop)
3441 LU.Kind = LSRUse::Special;
3445 // For an ICmpZero, negating a solitary base register won't lead to
3447 if (LU.Kind == LSRUse::ICmpZero &&
3448 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3450 // For each addrec base reg, apply the scale, if possible.
3451 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3452 if (const SCEVAddRecExpr *AR =
3453 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3454 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3455 if (FactorS->isZero())
3457 // Divide out the factor, ignoring high bits, since we'll be
3458 // scaling the value back up in the end.
3459 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3460 // TODO: This could be optimized to avoid all the copying.
3462 F.ScaledReg = Quotient;
3463 F.DeleteBaseReg(F.BaseRegs[i]);
3464 (void)InsertFormula(LU, LUIdx, F);
3470 /// GenerateTruncates - Generate reuse formulae from different IV types.
3471 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3472 // Don't bother truncating symbolic values.
3473 if (Base.BaseGV) return;
3475 // Determine the integer type for the base formula.
3476 Type *DstTy = Base.getType();
3478 DstTy = SE.getEffectiveSCEVType(DstTy);
3480 for (SmallSetVector<Type *, 4>::const_iterator
3481 I = Types.begin(), E = Types.end(); I != E; ++I) {
3483 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3486 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3487 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3488 JE = F.BaseRegs.end(); J != JE; ++J)
3489 *J = SE.getAnyExtendExpr(*J, SrcTy);
3491 // TODO: This assumes we've done basic processing on all uses and
3492 // have an idea what the register usage is.
3493 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3496 (void)InsertFormula(LU, LUIdx, F);
3503 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3504 /// defer modifications so that the search phase doesn't have to worry about
3505 /// the data structures moving underneath it.
3509 const SCEV *OrigReg;
3511 WorkItem(size_t LI, int64_t I, const SCEV *R)
3512 : LUIdx(LI), Imm(I), OrigReg(R) {}
3514 void print(raw_ostream &OS) const;
3520 void WorkItem::print(raw_ostream &OS) const {
3521 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3522 << " , add offset " << Imm;
3525 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3526 void WorkItem::dump() const {
3527 print(errs()); errs() << '\n';
3531 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3532 /// distance apart and try to form reuse opportunities between them.
3533 void LSRInstance::GenerateCrossUseConstantOffsets() {
3534 // Group the registers by their value without any added constant offset.
3535 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3536 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3538 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3539 SmallVector<const SCEV *, 8> Sequence;
3540 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3542 const SCEV *Reg = *I;
3543 int64_t Imm = ExtractImmediate(Reg, SE);
3544 std::pair<RegMapTy::iterator, bool> Pair =
3545 Map.insert(std::make_pair(Reg, ImmMapTy()));
3547 Sequence.push_back(Reg);
3548 Pair.first->second.insert(std::make_pair(Imm, *I));
3549 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3552 // Now examine each set of registers with the same base value. Build up
3553 // a list of work to do and do the work in a separate step so that we're
3554 // not adding formulae and register counts while we're searching.
3555 SmallVector<WorkItem, 32> WorkItems;
3556 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3557 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3558 E = Sequence.end(); I != E; ++I) {
3559 const SCEV *Reg = *I;
3560 const ImmMapTy &Imms = Map.find(Reg)->second;
3562 // It's not worthwhile looking for reuse if there's only one offset.
3563 if (Imms.size() == 1)
3566 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3567 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3569 dbgs() << ' ' << J->first;
3572 // Examine each offset.
3573 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3575 const SCEV *OrigReg = J->second;
3577 int64_t JImm = J->first;
3578 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3580 if (!isa<SCEVConstant>(OrigReg) &&
3581 UsedByIndicesMap[Reg].count() == 1) {
3582 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3586 // Conservatively examine offsets between this orig reg a few selected
3588 ImmMapTy::const_iterator OtherImms[] = {
3589 Imms.begin(), std::prev(Imms.end()),
3590 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3593 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3594 ImmMapTy::const_iterator M = OtherImms[i];
3595 if (M == J || M == JE) continue;
3597 // Compute the difference between the two.
3598 int64_t Imm = (uint64_t)JImm - M->first;
3599 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3600 LUIdx = UsedByIndices.find_next(LUIdx))
3601 // Make a memo of this use, offset, and register tuple.
3602 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3603 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3610 UsedByIndicesMap.clear();
3611 UniqueItems.clear();
3613 // Now iterate through the worklist and add new formulae.
3614 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3615 E = WorkItems.end(); I != E; ++I) {
3616 const WorkItem &WI = *I;
3617 size_t LUIdx = WI.LUIdx;
3618 LSRUse &LU = Uses[LUIdx];
3619 int64_t Imm = WI.Imm;
3620 const SCEV *OrigReg = WI.OrigReg;
3622 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3623 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3624 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3626 // TODO: Use a more targeted data structure.
3627 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3628 const Formula &F = LU.Formulae[L];
3629 // Use the immediate in the scaled register.
3630 if (F.ScaledReg == OrigReg) {
3631 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3632 // Don't create 50 + reg(-50).
3633 if (F.referencesReg(SE.getSCEV(
3634 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3637 NewF.BaseOffset = Offset;
3638 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3641 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3643 // If the new scale is a constant in a register, and adding the constant
3644 // value to the immediate would produce a value closer to zero than the
3645 // immediate itself, then the formula isn't worthwhile.
3646 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3647 if (C->getValue()->isNegative() !=
3648 (NewF.BaseOffset < 0) &&
3649 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3650 .ule(abs64(NewF.BaseOffset)))
3654 (void)InsertFormula(LU, LUIdx, NewF);
3656 // Use the immediate in a base register.
3657 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3658 const SCEV *BaseReg = F.BaseRegs[N];
3659 if (BaseReg != OrigReg)
3662 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3663 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3664 LU.Kind, LU.AccessTy, NewF)) {
3665 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3668 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3670 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3672 // If the new formula has a constant in a register, and adding the
3673 // constant value to the immediate would produce a value closer to
3674 // zero than the immediate itself, then the formula isn't worthwhile.
3675 for (SmallVectorImpl<const SCEV *>::const_iterator
3676 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3678 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3679 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3680 abs64(NewF.BaseOffset)) &&
3681 (C->getValue()->getValue() +
3682 NewF.BaseOffset).countTrailingZeros() >=
3683 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3687 (void)InsertFormula(LU, LUIdx, NewF);
3696 /// GenerateAllReuseFormulae - Generate formulae for each use.
3698 LSRInstance::GenerateAllReuseFormulae() {
3699 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3700 // queries are more precise.
3701 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3702 LSRUse &LU = Uses[LUIdx];
3703 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3704 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3705 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3706 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3708 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3709 LSRUse &LU = Uses[LUIdx];
3710 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3711 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3712 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3713 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3714 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3715 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3716 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3717 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3719 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3720 LSRUse &LU = Uses[LUIdx];
3721 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3722 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3725 GenerateCrossUseConstantOffsets();
3727 DEBUG(dbgs() << "\n"
3728 "After generating reuse formulae:\n";
3729 print_uses(dbgs()));
3732 /// If there are multiple formulae with the same set of registers used
3733 /// by other uses, pick the best one and delete the others.
3734 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3735 DenseSet<const SCEV *> VisitedRegs;
3736 SmallPtrSet<const SCEV *, 16> Regs;
3737 SmallPtrSet<const SCEV *, 16> LoserRegs;
3739 bool ChangedFormulae = false;
3742 // Collect the best formula for each unique set of shared registers. This
3743 // is reset for each use.
3744 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3746 BestFormulaeTy BestFormulae;
3748 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3749 LSRUse &LU = Uses[LUIdx];
3750 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3753 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3754 FIdx != NumForms; ++FIdx) {
3755 Formula &F = LU.Formulae[FIdx];
3757 // Some formulas are instant losers. For example, they may depend on
3758 // nonexistent AddRecs from other loops. These need to be filtered
3759 // immediately, otherwise heuristics could choose them over others leading
3760 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3761 // avoids the need to recompute this information across formulae using the
3762 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3763 // the corresponding bad register from the Regs set.
3766 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3768 if (CostF.isLoser()) {
3769 // During initial formula generation, undesirable formulae are generated
3770 // by uses within other loops that have some non-trivial address mode or
3771 // use the postinc form of the IV. LSR needs to provide these formulae
3772 // as the basis of rediscovering the desired formula that uses an AddRec
3773 // corresponding to the existing phi. Once all formulae have been
3774 // generated, these initial losers may be pruned.
3775 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3779 SmallVector<const SCEV *, 4> Key;
3780 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3781 JE = F.BaseRegs.end(); J != JE; ++J) {
3782 const SCEV *Reg = *J;
3783 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3787 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3788 Key.push_back(F.ScaledReg);
3789 // Unstable sort by host order ok, because this is only used for
3791 std::sort(Key.begin(), Key.end());
3793 std::pair<BestFormulaeTy::const_iterator, bool> P =
3794 BestFormulae.insert(std::make_pair(Key, FIdx));
3798 Formula &Best = LU.Formulae[P.first->second];
3802 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3804 if (CostF < CostBest)
3806 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3808 " in favor of formula "; Best.print(dbgs());
3812 ChangedFormulae = true;
3814 LU.DeleteFormula(F);
3820 // Now that we've filtered out some formulae, recompute the Regs set.
3822 LU.RecomputeRegs(LUIdx, RegUses);
3824 // Reset this to prepare for the next use.
3825 BestFormulae.clear();
3828 DEBUG(if (ChangedFormulae) {
3830 "After filtering out undesirable candidates:\n";
3835 // This is a rough guess that seems to work fairly well.
3836 static const size_t ComplexityLimit = UINT16_MAX;
3838 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3839 /// solutions the solver might have to consider. It almost never considers
3840 /// this many solutions because it prune the search space, but the pruning
3841 /// isn't always sufficient.
3842 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3844 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3845 E = Uses.end(); I != E; ++I) {
3846 size_t FSize = I->Formulae.size();
3847 if (FSize >= ComplexityLimit) {
3848 Power = ComplexityLimit;
3852 if (Power >= ComplexityLimit)
3858 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3859 /// of the registers of another formula, it won't help reduce register
3860 /// pressure (though it may not necessarily hurt register pressure); remove
3861 /// it to simplify the system.
3862 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3863 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3864 DEBUG(dbgs() << "The search space is too complex.\n");
3866 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3867 "which use a superset of registers used by other "
3870 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3871 LSRUse &LU = Uses[LUIdx];
3873 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3874 Formula &F = LU.Formulae[i];
3875 // Look for a formula with a constant or GV in a register. If the use
3876 // also has a formula with that same value in an immediate field,
3877 // delete the one that uses a register.
3878 for (SmallVectorImpl<const SCEV *>::const_iterator
3879 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3880 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3882 NewF.BaseOffset += C->getValue()->getSExtValue();
3883 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3884 (I - F.BaseRegs.begin()));
3885 if (LU.HasFormulaWithSameRegs(NewF)) {
3886 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3887 LU.DeleteFormula(F);
3893 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3894 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3898 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3899 (I - F.BaseRegs.begin()));
3900 if (LU.HasFormulaWithSameRegs(NewF)) {
3901 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3903 LU.DeleteFormula(F);
3914 LU.RecomputeRegs(LUIdx, RegUses);
3917 DEBUG(dbgs() << "After pre-selection:\n";
3918 print_uses(dbgs()));
3922 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3923 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3925 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3926 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
3929 DEBUG(dbgs() << "The search space is too complex.\n"
3930 "Narrowing the search space by assuming that uses separated "
3931 "by a constant offset will use the same registers.\n");
3933 // This is especially useful for unrolled loops.
3935 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3936 LSRUse &LU = Uses[LUIdx];
3937 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3938 E = LU.Formulae.end(); I != E; ++I) {
3939 const Formula &F = *I;
3940 if (F.BaseOffset == 0 || F.Scale != 0)
3943 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
3947 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
3948 LU.Kind, LU.AccessTy))
3951 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
3953 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3955 // Update the relocs to reference the new use.
3956 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3957 E = Fixups.end(); I != E; ++I) {
3958 LSRFixup &Fixup = *I;
3959 if (Fixup.LUIdx == LUIdx) {
3960 Fixup.LUIdx = LUThatHas - &Uses.front();
3961 Fixup.Offset += F.BaseOffset;
3962 // Add the new offset to LUThatHas' offset list.
3963 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3964 LUThatHas->Offsets.push_back(Fixup.Offset);
3965 if (Fixup.Offset > LUThatHas->MaxOffset)
3966 LUThatHas->MaxOffset = Fixup.Offset;
3967 if (Fixup.Offset < LUThatHas->MinOffset)
3968 LUThatHas->MinOffset = Fixup.Offset;
3970 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
3972 if (Fixup.LUIdx == NumUses-1)
3973 Fixup.LUIdx = LUIdx;
3976 // Delete formulae from the new use which are no longer legal.
3978 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3979 Formula &F = LUThatHas->Formulae[i];
3980 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
3981 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
3982 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3984 LUThatHas->DeleteFormula(F);
3992 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3994 // Delete the old use.
3995 DeleteUse(LU, LUIdx);
4002 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4005 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4006 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4007 /// we've done more filtering, as it may be able to find more formulae to
4009 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4010 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4011 DEBUG(dbgs() << "The search space is too complex.\n");
4013 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4014 "undesirable dedicated registers.\n");
4016 FilterOutUndesirableDedicatedRegisters();
4018 DEBUG(dbgs() << "After pre-selection:\n";
4019 print_uses(dbgs()));
4023 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4024 /// to be profitable, and then in any use which has any reference to that
4025 /// register, delete all formulae which do not reference that register.
4026 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4027 // With all other options exhausted, loop until the system is simple
4028 // enough to handle.
4029 SmallPtrSet<const SCEV *, 4> Taken;
4030 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4031 // Ok, we have too many of formulae on our hands to conveniently handle.
4032 // Use a rough heuristic to thin out the list.
4033 DEBUG(dbgs() << "The search space is too complex.\n");
4035 // Pick the register which is used by the most LSRUses, which is likely
4036 // to be a good reuse register candidate.
4037 const SCEV *Best = 0;
4038 unsigned BestNum = 0;
4039 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4041 const SCEV *Reg = *I;
4042 if (Taken.count(Reg))
4047 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4048 if (Count > BestNum) {
4055 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4056 << " will yield profitable reuse.\n");
4059 // In any use with formulae which references this register, delete formulae
4060 // which don't reference it.
4061 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4062 LSRUse &LU = Uses[LUIdx];
4063 if (!LU.Regs.count(Best)) continue;
4066 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4067 Formula &F = LU.Formulae[i];
4068 if (!F.referencesReg(Best)) {
4069 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4070 LU.DeleteFormula(F);
4074 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4080 LU.RecomputeRegs(LUIdx, RegUses);
4083 DEBUG(dbgs() << "After pre-selection:\n";
4084 print_uses(dbgs()));
4088 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4089 /// formulae to choose from, use some rough heuristics to prune down the number
4090 /// of formulae. This keeps the main solver from taking an extraordinary amount
4091 /// of time in some worst-case scenarios.
4092 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4093 NarrowSearchSpaceByDetectingSupersets();
4094 NarrowSearchSpaceByCollapsingUnrolledCode();
4095 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4096 NarrowSearchSpaceByPickingWinnerRegs();
4099 /// SolveRecurse - This is the recursive solver.
4100 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4102 SmallVectorImpl<const Formula *> &Workspace,
4103 const Cost &CurCost,
4104 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4105 DenseSet<const SCEV *> &VisitedRegs) const {
4108 // - use more aggressive filtering
4109 // - sort the formula so that the most profitable solutions are found first
4110 // - sort the uses too
4112 // - don't compute a cost, and then compare. compare while computing a cost
4114 // - track register sets with SmallBitVector
4116 const LSRUse &LU = Uses[Workspace.size()];
4118 // If this use references any register that's already a part of the
4119 // in-progress solution, consider it a requirement that a formula must
4120 // reference that register in order to be considered. This prunes out
4121 // unprofitable searching.
4122 SmallSetVector<const SCEV *, 4> ReqRegs;
4123 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4124 E = CurRegs.end(); I != E; ++I)
4125 if (LU.Regs.count(*I))
4128 SmallPtrSet<const SCEV *, 16> NewRegs;
4130 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4131 E = LU.Formulae.end(); I != E; ++I) {
4132 const Formula &F = *I;
4134 // Ignore formulae which do not use any of the required registers.
4135 bool SatisfiedReqReg = true;
4136 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4137 JE = ReqRegs.end(); J != JE; ++J) {
4138 const SCEV *Reg = *J;
4139 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4140 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4142 SatisfiedReqReg = false;
4146 if (!SatisfiedReqReg) {
4147 // If none of the formulae satisfied the required registers, then we could
4148 // clear ReqRegs and try again. Currently, we simply give up in this case.
4152 // Evaluate the cost of the current formula. If it's already worse than
4153 // the current best, prune the search at that point.
4156 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4158 if (NewCost < SolutionCost) {
4159 Workspace.push_back(&F);
4160 if (Workspace.size() != Uses.size()) {
4161 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4162 NewRegs, VisitedRegs);
4163 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4164 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4166 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4167 dbgs() << ".\n Regs:";
4168 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4169 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4170 dbgs() << ' ' << **I;
4173 SolutionCost = NewCost;
4174 Solution = Workspace;
4176 Workspace.pop_back();
4181 /// Solve - Choose one formula from each use. Return the results in the given
4182 /// Solution vector.
4183 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4184 SmallVector<const Formula *, 8> Workspace;
4186 SolutionCost.Lose();
4188 SmallPtrSet<const SCEV *, 16> CurRegs;
4189 DenseSet<const SCEV *> VisitedRegs;
4190 Workspace.reserve(Uses.size());
4192 // SolveRecurse does all the work.
4193 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4194 CurRegs, VisitedRegs);
4195 if (Solution.empty()) {
4196 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4200 // Ok, we've now made all our decisions.
4201 DEBUG(dbgs() << "\n"
4202 "The chosen solution requires "; SolutionCost.print(dbgs());
4204 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4206 Uses[i].print(dbgs());
4209 Solution[i]->print(dbgs());
4213 assert(Solution.size() == Uses.size() && "Malformed solution!");
4216 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4217 /// the dominator tree far as we can go while still being dominated by the
4218 /// input positions. This helps canonicalize the insert position, which
4219 /// encourages sharing.
4220 BasicBlock::iterator
4221 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4222 const SmallVectorImpl<Instruction *> &Inputs)
4225 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4226 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4229 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4230 if (!Rung) return IP;
4231 Rung = Rung->getIDom();
4232 if (!Rung) return IP;
4233 IDom = Rung->getBlock();
4235 // Don't climb into a loop though.
4236 const Loop *IDomLoop = LI.getLoopFor(IDom);
4237 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4238 if (IDomDepth <= IPLoopDepth &&
4239 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4243 bool AllDominate = true;
4244 Instruction *BetterPos = 0;
4245 Instruction *Tentative = IDom->getTerminator();
4246 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4247 E = Inputs.end(); I != E; ++I) {
4248 Instruction *Inst = *I;
4249 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4250 AllDominate = false;
4253 // Attempt to find an insert position in the middle of the block,
4254 // instead of at the end, so that it can be used for other expansions.
4255 if (IDom == Inst->getParent() &&
4256 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4257 BetterPos = std::next(BasicBlock::iterator(Inst));
4270 /// AdjustInsertPositionForExpand - Determine an input position which will be
4271 /// dominated by the operands and which will dominate the result.
4272 BasicBlock::iterator
4273 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4276 SCEVExpander &Rewriter) const {
4277 // Collect some instructions which must be dominated by the
4278 // expanding replacement. These must be dominated by any operands that
4279 // will be required in the expansion.
4280 SmallVector<Instruction *, 4> Inputs;
4281 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4282 Inputs.push_back(I);
4283 if (LU.Kind == LSRUse::ICmpZero)
4284 if (Instruction *I =
4285 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4286 Inputs.push_back(I);
4287 if (LF.PostIncLoops.count(L)) {
4288 if (LF.isUseFullyOutsideLoop(L))
4289 Inputs.push_back(L->getLoopLatch()->getTerminator());
4291 Inputs.push_back(IVIncInsertPos);
4293 // The expansion must also be dominated by the increment positions of any
4294 // loops it for which it is using post-inc mode.
4295 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4296 E = LF.PostIncLoops.end(); I != E; ++I) {
4297 const Loop *PIL = *I;
4298 if (PIL == L) continue;
4300 // Be dominated by the loop exit.
4301 SmallVector<BasicBlock *, 4> ExitingBlocks;
4302 PIL->getExitingBlocks(ExitingBlocks);
4303 if (!ExitingBlocks.empty()) {
4304 BasicBlock *BB = ExitingBlocks[0];
4305 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4306 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4307 Inputs.push_back(BB->getTerminator());
4311 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4312 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4313 "Insertion point must be a normal instruction");
4315 // Then, climb up the immediate dominator tree as far as we can go while
4316 // still being dominated by the input positions.
4317 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4319 // Don't insert instructions before PHI nodes.
4320 while (isa<PHINode>(IP)) ++IP;
4322 // Ignore landingpad instructions.
4323 while (isa<LandingPadInst>(IP)) ++IP;
4325 // Ignore debug intrinsics.
4326 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4328 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4329 // IP consistent across expansions and allows the previously inserted
4330 // instructions to be reused by subsequent expansion.
4331 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4336 /// Expand - Emit instructions for the leading candidate expression for this
4337 /// LSRUse (this is called "expanding").
4338 Value *LSRInstance::Expand(const LSRFixup &LF,
4340 BasicBlock::iterator IP,
4341 SCEVExpander &Rewriter,
4342 SmallVectorImpl<WeakVH> &DeadInsts) const {
4343 const LSRUse &LU = Uses[LF.LUIdx];
4344 if (LU.RigidFormula)
4345 return LF.OperandValToReplace;
4347 // Determine an input position which will be dominated by the operands and
4348 // which will dominate the result.
4349 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4351 // Inform the Rewriter if we have a post-increment use, so that it can
4352 // perform an advantageous expansion.
4353 Rewriter.setPostInc(LF.PostIncLoops);
4355 // This is the type that the user actually needs.
4356 Type *OpTy = LF.OperandValToReplace->getType();
4357 // This will be the type that we'll initially expand to.
4358 Type *Ty = F.getType();
4360 // No type known; just expand directly to the ultimate type.
4362 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4363 // Expand directly to the ultimate type if it's the right size.
4365 // This is the type to do integer arithmetic in.
4366 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4368 // Build up a list of operands to add together to form the full base.
4369 SmallVector<const SCEV *, 8> Ops;
4371 // Expand the BaseRegs portion.
4372 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4373 E = F.BaseRegs.end(); I != E; ++I) {
4374 const SCEV *Reg = *I;
4375 assert(!Reg->isZero() && "Zero allocated in a base register!");
4377 // If we're expanding for a post-inc user, make the post-inc adjustment.
4378 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4379 Reg = TransformForPostIncUse(Denormalize, Reg,
4380 LF.UserInst, LF.OperandValToReplace,
4383 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4386 // Expand the ScaledReg portion.
4387 Value *ICmpScaledV = 0;
4389 const SCEV *ScaledS = F.ScaledReg;
4391 // If we're expanding for a post-inc user, make the post-inc adjustment.
4392 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4393 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4394 LF.UserInst, LF.OperandValToReplace,
4397 if (LU.Kind == LSRUse::ICmpZero) {
4398 // An interesting way of "folding" with an icmp is to use a negated
4399 // scale, which we'll implement by inserting it into the other operand
4401 assert(F.Scale == -1 &&
4402 "The only scale supported by ICmpZero uses is -1!");
4403 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4405 // Otherwise just expand the scaled register and an explicit scale,
4406 // which is expected to be matched as part of the address.
4408 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4409 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4410 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4412 Ops.push_back(SE.getUnknown(FullV));
4414 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4415 ScaledS = SE.getMulExpr(ScaledS,
4416 SE.getConstant(ScaledS->getType(), F.Scale));
4417 Ops.push_back(ScaledS);
4421 // Expand the GV portion.
4423 // Flush the operand list to suppress SCEVExpander hoisting.
4425 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4427 Ops.push_back(SE.getUnknown(FullV));
4429 Ops.push_back(SE.getUnknown(F.BaseGV));
4432 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4433 // unfolded offsets. LSR assumes they both live next to their uses.
4435 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4437 Ops.push_back(SE.getUnknown(FullV));
4440 // Expand the immediate portion.
4441 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4443 if (LU.Kind == LSRUse::ICmpZero) {
4444 // The other interesting way of "folding" with an ICmpZero is to use a
4445 // negated immediate.
4447 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4449 Ops.push_back(SE.getUnknown(ICmpScaledV));
4450 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4453 // Just add the immediate values. These again are expected to be matched
4454 // as part of the address.
4455 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4459 // Expand the unfolded offset portion.
4460 int64_t UnfoldedOffset = F.UnfoldedOffset;
4461 if (UnfoldedOffset != 0) {
4462 // Just add the immediate values.
4463 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4467 // Emit instructions summing all the operands.
4468 const SCEV *FullS = Ops.empty() ?
4469 SE.getConstant(IntTy, 0) :
4471 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4473 // We're done expanding now, so reset the rewriter.
4474 Rewriter.clearPostInc();
4476 // An ICmpZero Formula represents an ICmp which we're handling as a
4477 // comparison against zero. Now that we've expanded an expression for that
4478 // form, update the ICmp's other operand.
4479 if (LU.Kind == LSRUse::ICmpZero) {
4480 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4481 DeadInsts.push_back(CI->getOperand(1));
4482 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4483 "a scale at the same time!");
4484 if (F.Scale == -1) {
4485 if (ICmpScaledV->getType() != OpTy) {
4487 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4489 ICmpScaledV, OpTy, "tmp", CI);
4492 CI->setOperand(1, ICmpScaledV);
4494 assert(F.Scale == 0 &&
4495 "ICmp does not support folding a global value and "
4496 "a scale at the same time!");
4497 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4499 if (C->getType() != OpTy)
4500 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4504 CI->setOperand(1, C);
4511 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4512 /// of their operands effectively happens in their predecessor blocks, so the
4513 /// expression may need to be expanded in multiple places.
4514 void LSRInstance::RewriteForPHI(PHINode *PN,
4517 SCEVExpander &Rewriter,
4518 SmallVectorImpl<WeakVH> &DeadInsts,
4520 DenseMap<BasicBlock *, Value *> Inserted;
4521 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4522 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4523 BasicBlock *BB = PN->getIncomingBlock(i);
4525 // If this is a critical edge, split the edge so that we do not insert
4526 // the code on all predecessor/successor paths. We do this unless this
4527 // is the canonical backedge for this loop, which complicates post-inc
4529 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4530 !isa<IndirectBrInst>(BB->getTerminator())) {
4531 BasicBlock *Parent = PN->getParent();
4532 Loop *PNLoop = LI.getLoopFor(Parent);
4533 if (!PNLoop || Parent != PNLoop->getHeader()) {
4534 // Split the critical edge.
4535 BasicBlock *NewBB = 0;
4536 if (!Parent->isLandingPad()) {
4537 NewBB = SplitCriticalEdge(BB, Parent, P,
4538 /*MergeIdenticalEdges=*/true,
4539 /*DontDeleteUselessPhis=*/true);
4541 SmallVector<BasicBlock*, 2> NewBBs;
4542 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4545 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4546 // phi predecessors are identical. The simple thing to do is skip
4547 // splitting in this case rather than complicate the API.
4549 // If PN is outside of the loop and BB is in the loop, we want to
4550 // move the block to be immediately before the PHI block, not
4551 // immediately after BB.
4552 if (L->contains(BB) && !L->contains(PN))
4553 NewBB->moveBefore(PN->getParent());
4555 // Splitting the edge can reduce the number of PHI entries we have.
4556 e = PN->getNumIncomingValues();
4558 i = PN->getBasicBlockIndex(BB);
4563 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4564 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4566 PN->setIncomingValue(i, Pair.first->second);
4568 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4570 // If this is reuse-by-noop-cast, insert the noop cast.
4571 Type *OpTy = LF.OperandValToReplace->getType();
4572 if (FullV->getType() != OpTy)
4574 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4576 FullV, LF.OperandValToReplace->getType(),
4577 "tmp", BB->getTerminator());
4579 PN->setIncomingValue(i, FullV);
4580 Pair.first->second = FullV;
4585 /// Rewrite - Emit instructions for the leading candidate expression for this
4586 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4587 /// the newly expanded value.
4588 void LSRInstance::Rewrite(const LSRFixup &LF,
4590 SCEVExpander &Rewriter,
4591 SmallVectorImpl<WeakVH> &DeadInsts,
4593 // First, find an insertion point that dominates UserInst. For PHI nodes,
4594 // find the nearest block which dominates all the relevant uses.
4595 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4596 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4598 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4600 // If this is reuse-by-noop-cast, insert the noop cast.
4601 Type *OpTy = LF.OperandValToReplace->getType();
4602 if (FullV->getType() != OpTy) {
4604 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4605 FullV, OpTy, "tmp", LF.UserInst);
4609 // Update the user. ICmpZero is handled specially here (for now) because
4610 // Expand may have updated one of the operands of the icmp already, and
4611 // its new value may happen to be equal to LF.OperandValToReplace, in
4612 // which case doing replaceUsesOfWith leads to replacing both operands
4613 // with the same value. TODO: Reorganize this.
4614 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4615 LF.UserInst->setOperand(0, FullV);
4617 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4620 DeadInsts.push_back(LF.OperandValToReplace);
4623 /// ImplementSolution - Rewrite all the fixup locations with new values,
4624 /// following the chosen solution.
4626 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4628 // Keep track of instructions we may have made dead, so that
4629 // we can remove them after we are done working.
4630 SmallVector<WeakVH, 16> DeadInsts;
4632 SCEVExpander Rewriter(SE, "lsr");
4634 Rewriter.setDebugType(DEBUG_TYPE);
4636 Rewriter.disableCanonicalMode();
4637 Rewriter.enableLSRMode();
4638 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4640 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4641 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4642 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4643 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4644 Rewriter.setChainedPhi(PN);
4647 // Expand the new value definitions and update the users.
4648 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4649 E = Fixups.end(); I != E; ++I) {
4650 const LSRFixup &Fixup = *I;
4652 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4657 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4658 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4659 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4662 // Clean up after ourselves. This must be done before deleting any
4666 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4669 LSRInstance::LSRInstance(Loop *L, Pass *P)
4670 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4671 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4672 LI(P->getAnalysis<LoopInfo>()),
4673 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4675 // If LoopSimplify form is not available, stay out of trouble.
4676 if (!L->isLoopSimplifyForm())
4679 // If there's no interesting work to be done, bail early.
4680 if (IU.empty()) return;
4682 // If there's too much analysis to be done, bail early. We won't be able to
4683 // model the problem anyway.
4684 unsigned NumUsers = 0;
4685 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4686 if (++NumUsers > MaxIVUsers) {
4687 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4694 // All dominating loops must have preheaders, or SCEVExpander may not be able
4695 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4697 // IVUsers analysis should only create users that are dominated by simple loop
4698 // headers. Since this loop should dominate all of its users, its user list
4699 // should be empty if this loop itself is not within a simple loop nest.
4700 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4701 Rung; Rung = Rung->getIDom()) {
4702 BasicBlock *BB = Rung->getBlock();
4703 const Loop *DomLoop = LI.getLoopFor(BB);
4704 if (DomLoop && DomLoop->getHeader() == BB) {
4705 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4710 DEBUG(dbgs() << "\nLSR on loop ";
4711 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4714 // First, perform some low-level loop optimizations.
4716 OptimizeLoopTermCond();
4718 // If loop preparation eliminates all interesting IV users, bail.
4719 if (IU.empty()) return;
4721 // Skip nested loops until we can model them better with formulae.
4723 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4727 // Start collecting data and preparing for the solver.
4729 CollectInterestingTypesAndFactors();
4730 CollectFixupsAndInitialFormulae();
4731 CollectLoopInvariantFixupsAndFormulae();
4733 assert(!Uses.empty() && "IVUsers reported at least one use");
4734 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4735 print_uses(dbgs()));
4737 // Now use the reuse data to generate a bunch of interesting ways
4738 // to formulate the values needed for the uses.
4739 GenerateAllReuseFormulae();
4741 FilterOutUndesirableDedicatedRegisters();
4742 NarrowSearchSpaceUsingHeuristics();
4744 SmallVector<const Formula *, 8> Solution;
4747 // Release memory that is no longer needed.
4752 if (Solution.empty())
4756 // Formulae should be legal.
4757 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4759 const LSRUse &LU = *I;
4760 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4761 JE = LU.Formulae.end();
4763 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4764 *J) && "Illegal formula generated!");
4768 // Now that we've decided what we want, make it so.
4769 ImplementSolution(Solution, P);
4772 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4773 if (Factors.empty() && Types.empty()) return;
4775 OS << "LSR has identified the following interesting factors and types: ";
4778 for (SmallSetVector<int64_t, 8>::const_iterator
4779 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4780 if (!First) OS << ", ";
4785 for (SmallSetVector<Type *, 4>::const_iterator
4786 I = Types.begin(), E = Types.end(); I != E; ++I) {
4787 if (!First) OS << ", ";
4789 OS << '(' << **I << ')';
4794 void LSRInstance::print_fixups(raw_ostream &OS) const {
4795 OS << "LSR is examining the following fixup sites:\n";
4796 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4797 E = Fixups.end(); I != E; ++I) {
4804 void LSRInstance::print_uses(raw_ostream &OS) const {
4805 OS << "LSR is examining the following uses:\n";
4806 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4807 E = Uses.end(); I != E; ++I) {
4808 const LSRUse &LU = *I;
4812 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4813 JE = LU.Formulae.end(); J != JE; ++J) {
4821 void LSRInstance::print(raw_ostream &OS) const {
4822 print_factors_and_types(OS);
4827 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4828 void LSRInstance::dump() const {
4829 print(errs()); errs() << '\n';
4835 class LoopStrengthReduce : public LoopPass {
4837 static char ID; // Pass ID, replacement for typeid
4838 LoopStrengthReduce();
4841 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
4842 void getAnalysisUsage(AnalysisUsage &AU) const override;
4847 char LoopStrengthReduce::ID = 0;
4848 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4849 "Loop Strength Reduction", false, false)
4850 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4851 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4852 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4853 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4854 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4855 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4856 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4857 "Loop Strength Reduction", false, false)
4860 Pass *llvm::createLoopStrengthReducePass() {
4861 return new LoopStrengthReduce();
4864 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4865 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4868 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4869 // We split critical edges, so we change the CFG. However, we do update
4870 // many analyses if they are around.
4871 AU.addPreservedID(LoopSimplifyID);
4873 AU.addRequired<LoopInfo>();
4874 AU.addPreserved<LoopInfo>();
4875 AU.addRequiredID(LoopSimplifyID);
4876 AU.addRequired<DominatorTreeWrapperPass>();
4877 AU.addPreserved<DominatorTreeWrapperPass>();
4878 AU.addRequired<ScalarEvolution>();
4879 AU.addPreserved<ScalarEvolution>();
4880 // Requiring LoopSimplify a second time here prevents IVUsers from running
4881 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4882 AU.addRequiredID(LoopSimplifyID);
4883 AU.addRequired<IVUsers>();
4884 AU.addPreserved<IVUsers>();
4885 AU.addRequired<TargetTransformInfo>();
4888 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4889 if (skipOptnoneFunction(L))
4892 bool Changed = false;
4894 // Run the main LSR transformation.
4895 Changed |= LSRInstance(L, this).getChanged();
4897 // Remove any extra phis created by processing inner loops.
4898 Changed |= DeleteDeadPHIs(L->getHeader());
4899 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4900 SmallVector<WeakVH, 16> DeadInsts;
4901 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4903 Rewriter.setDebugType(DEBUG_TYPE);
4905 unsigned numFolded = Rewriter.replaceCongruentIVs(
4906 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
4907 &getAnalysis<TargetTransformInfo>());
4910 DeleteTriviallyDeadInstructions(DeadInsts);
4911 DeleteDeadPHIs(L->getHeader());