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 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/ADT/DenseSet.h"
59 #include "llvm/ADT/Hashing.h"
60 #include "llvm/ADT/STLExtras.h"
61 #include "llvm/ADT/SetVector.h"
62 #include "llvm/ADT/SmallBitVector.h"
63 #include "llvm/Analysis/IVUsers.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DerivedTypes.h"
69 #include "llvm/IR/Dominators.h"
70 #include "llvm/IR/Instructions.h"
71 #include "llvm/IR/IntrinsicInst.h"
72 #include "llvm/IR/ValueHandle.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/Debug.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
81 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
82 /// bail out. This threshold is far beyond the number of users that LSR can
83 /// conceivably solve, so it should not affect generated code, but catches the
84 /// worst cases before LSR burns too much compile time and stack space.
85 static const unsigned MaxIVUsers = 200;
87 // Temporary flag to cleanup congruent phis after LSR phi expansion.
88 // It's currently disabled until we can determine whether it's truly useful or
89 // not. The flag should be removed after the v3.0 release.
90 // This is now needed for ivchains.
91 static cl::opt<bool> EnablePhiElim(
92 "enable-lsr-phielim", cl::Hidden, cl::init(true),
93 cl::desc("Enable LSR phi elimination"));
96 // Stress test IV chain generation.
97 static cl::opt<bool> StressIVChain(
98 "stress-ivchain", cl::Hidden, cl::init(false),
99 cl::desc("Stress test LSR IV chains"));
101 static bool StressIVChain = false;
106 /// RegSortData - This class holds data which is used to order reuse candidates.
109 /// UsedByIndices - This represents the set of LSRUse indices which reference
110 /// a particular register.
111 SmallBitVector UsedByIndices;
115 void print(raw_ostream &OS) const;
121 void RegSortData::print(raw_ostream &OS) const {
122 OS << "[NumUses=" << UsedByIndices.count() << ']';
125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
126 void RegSortData::dump() const {
127 print(errs()); errs() << '\n';
133 /// RegUseTracker - Map register candidates to information about how they are
135 class RegUseTracker {
136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
138 RegUsesTy RegUsesMap;
139 SmallVector<const SCEV *, 16> RegSequence;
142 void CountRegister(const SCEV *Reg, size_t LUIdx);
143 void DropRegister(const SCEV *Reg, size_t LUIdx);
144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
152 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
154 iterator begin() { return RegSequence.begin(); }
155 iterator end() { return RegSequence.end(); }
156 const_iterator begin() const { return RegSequence.begin(); }
157 const_iterator end() const { return RegSequence.end(); }
163 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
164 std::pair<RegUsesTy::iterator, bool> Pair =
165 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
166 RegSortData &RSD = Pair.first->second;
168 RegSequence.push_back(Reg);
169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
170 RSD.UsedByIndices.set(LUIdx);
174 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
175 RegUsesTy::iterator It = RegUsesMap.find(Reg);
176 assert(It != RegUsesMap.end());
177 RegSortData &RSD = It->second;
178 assert(RSD.UsedByIndices.size() > LUIdx);
179 RSD.UsedByIndices.reset(LUIdx);
183 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
184 assert(LUIdx <= LastLUIdx);
186 // Update RegUses. The data structure is not optimized for this purpose;
187 // we must iterate through it and update each of the bit vectors.
188 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
190 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
191 if (LUIdx < UsedByIndices.size())
192 UsedByIndices[LUIdx] =
193 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
194 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
199 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
200 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
201 if (I == RegUsesMap.end())
203 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
204 int i = UsedByIndices.find_first();
205 if (i == -1) return false;
206 if ((size_t)i != LUIdx) return true;
207 return UsedByIndices.find_next(i) != -1;
210 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
211 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
212 assert(I != RegUsesMap.end() && "Unknown register!");
213 return I->second.UsedByIndices;
216 void RegUseTracker::clear() {
223 /// Formula - This class holds information that describes a formula for
224 /// computing satisfying a use. It may include broken-out immediates and scaled
227 /// Global base address used for complex addressing.
230 /// Base offset for complex addressing.
233 /// Whether any complex addressing has a base register.
236 /// The scale of any complex addressing.
239 /// BaseRegs - The list of "base" registers for this use. When this is
241 SmallVector<const SCEV *, 4> BaseRegs;
243 /// ScaledReg - The 'scaled' register for this use. This should be non-null
244 /// when Scale is not zero.
245 const SCEV *ScaledReg;
247 /// UnfoldedOffset - An additional constant offset which added near the
248 /// use. This requires a temporary register, but the offset itself can
249 /// live in an add immediate field rather than a register.
250 int64_t UnfoldedOffset;
253 : BaseGV(0), BaseOffset(0), HasBaseReg(false), Scale(0), ScaledReg(0),
256 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
258 unsigned getNumRegs() const;
259 Type *getType() const;
261 void DeleteBaseReg(const SCEV *&S);
263 bool referencesReg(const SCEV *S) const;
264 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
265 const RegUseTracker &RegUses) const;
267 void print(raw_ostream &OS) const;
273 /// DoInitialMatch - Recursion helper for InitialMatch.
274 static void DoInitialMatch(const SCEV *S, Loop *L,
275 SmallVectorImpl<const SCEV *> &Good,
276 SmallVectorImpl<const SCEV *> &Bad,
277 ScalarEvolution &SE) {
278 // Collect expressions which properly dominate the loop header.
279 if (SE.properlyDominates(S, L->getHeader())) {
284 // Look at add operands.
285 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
286 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
288 DoInitialMatch(*I, L, Good, Bad, SE);
292 // Look at addrec operands.
293 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
294 if (!AR->getStart()->isZero()) {
295 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
296 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
297 AR->getStepRecurrence(SE),
298 // FIXME: AR->getNoWrapFlags()
299 AR->getLoop(), SCEV::FlagAnyWrap),
304 // Handle a multiplication by -1 (negation) if it didn't fold.
305 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
306 if (Mul->getOperand(0)->isAllOnesValue()) {
307 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
308 const SCEV *NewMul = SE.getMulExpr(Ops);
310 SmallVector<const SCEV *, 4> MyGood;
311 SmallVector<const SCEV *, 4> MyBad;
312 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
313 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
314 SE.getEffectiveSCEVType(NewMul->getType())));
315 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
316 E = MyGood.end(); I != E; ++I)
317 Good.push_back(SE.getMulExpr(NegOne, *I));
318 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
319 E = MyBad.end(); I != E; ++I)
320 Bad.push_back(SE.getMulExpr(NegOne, *I));
324 // Ok, we can't do anything interesting. Just stuff the whole thing into a
325 // register and hope for the best.
329 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
330 /// attempting to keep all loop-invariant and loop-computable values in a
331 /// single base register.
332 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
333 SmallVector<const SCEV *, 4> Good;
334 SmallVector<const SCEV *, 4> Bad;
335 DoInitialMatch(S, L, Good, Bad, SE);
337 const SCEV *Sum = SE.getAddExpr(Good);
339 BaseRegs.push_back(Sum);
343 const SCEV *Sum = SE.getAddExpr(Bad);
345 BaseRegs.push_back(Sum);
350 /// getNumRegs - Return the total number of register operands used by this
351 /// formula. This does not include register uses implied by non-constant
353 unsigned Formula::getNumRegs() const {
354 return !!ScaledReg + BaseRegs.size();
357 /// getType - Return the type of this formula, if it has one, or null
358 /// otherwise. This type is meaningless except for the bit size.
359 Type *Formula::getType() const {
360 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
361 ScaledReg ? ScaledReg->getType() :
362 BaseGV ? BaseGV->getType() :
366 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
367 void Formula::DeleteBaseReg(const SCEV *&S) {
368 if (&S != &BaseRegs.back())
369 std::swap(S, BaseRegs.back());
373 /// referencesReg - Test if this formula references the given register.
374 bool Formula::referencesReg(const SCEV *S) const {
375 return S == ScaledReg ||
376 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
379 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
380 /// which are used by uses other than the use with the given index.
381 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
382 const RegUseTracker &RegUses) const {
384 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
386 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
387 E = BaseRegs.end(); I != E; ++I)
388 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
393 void Formula::print(raw_ostream &OS) const {
396 if (!First) OS << " + "; else First = false;
397 BaseGV->printAsOperand(OS, /*PrintType=*/false);
399 if (BaseOffset != 0) {
400 if (!First) OS << " + "; else First = false;
403 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
404 E = BaseRegs.end(); I != E; ++I) {
405 if (!First) OS << " + "; else First = false;
406 OS << "reg(" << **I << ')';
408 if (HasBaseReg && BaseRegs.empty()) {
409 if (!First) OS << " + "; else First = false;
410 OS << "**error: HasBaseReg**";
411 } else if (!HasBaseReg && !BaseRegs.empty()) {
412 if (!First) OS << " + "; else First = false;
413 OS << "**error: !HasBaseReg**";
416 if (!First) OS << " + "; else First = false;
417 OS << Scale << "*reg(";
424 if (UnfoldedOffset != 0) {
425 if (!First) OS << " + ";
426 OS << "imm(" << UnfoldedOffset << ')';
430 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
431 void Formula::dump() const {
432 print(errs()); errs() << '\n';
436 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
437 /// without changing its value.
438 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
440 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
441 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
444 /// isAddSExtable - Return true if the given add can be sign-extended
445 /// without changing its value.
446 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
448 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
449 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
452 /// isMulSExtable - Return true if the given mul can be sign-extended
453 /// without changing its value.
454 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
456 IntegerType::get(SE.getContext(),
457 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
458 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
461 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
462 /// and if the remainder is known to be zero, or null otherwise. If
463 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
464 /// to Y, ignoring that the multiplication may overflow, which is useful when
465 /// the result will be used in a context where the most significant bits are
467 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
469 bool IgnoreSignificantBits = false) {
470 // Handle the trivial case, which works for any SCEV type.
472 return SE.getConstant(LHS->getType(), 1);
474 // Handle a few RHS special cases.
475 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
477 const APInt &RA = RC->getValue()->getValue();
478 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
480 if (RA.isAllOnesValue())
481 return SE.getMulExpr(LHS, RC);
482 // Handle x /s 1 as x.
487 // Check for a division of a constant by a constant.
488 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
491 const APInt &LA = C->getValue()->getValue();
492 const APInt &RA = RC->getValue()->getValue();
493 if (LA.srem(RA) != 0)
495 return SE.getConstant(LA.sdiv(RA));
498 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
499 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
500 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
501 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
502 IgnoreSignificantBits);
504 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
505 IgnoreSignificantBits);
506 if (!Start) return 0;
507 // FlagNW is independent of the start value, step direction, and is
508 // preserved with smaller magnitude steps.
509 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
510 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
515 // Distribute the sdiv over add operands, if the add doesn't overflow.
516 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
517 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
518 SmallVector<const SCEV *, 8> Ops;
519 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
521 const SCEV *Op = getExactSDiv(*I, RHS, SE,
522 IgnoreSignificantBits);
526 return SE.getAddExpr(Ops);
531 // Check for a multiply operand that we can pull RHS out of.
532 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
533 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
534 SmallVector<const SCEV *, 4> Ops;
536 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
540 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
541 IgnoreSignificantBits)) {
547 return Found ? SE.getMulExpr(Ops) : 0;
552 // Otherwise we don't know.
556 /// ExtractImmediate - If S involves the addition of a constant integer value,
557 /// return that integer value, and mutate S to point to a new SCEV with that
559 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
560 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
561 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
562 S = SE.getConstant(C->getType(), 0);
563 return C->getValue()->getSExtValue();
565 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
566 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
567 int64_t Result = ExtractImmediate(NewOps.front(), SE);
569 S = SE.getAddExpr(NewOps);
571 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
572 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
573 int64_t Result = ExtractImmediate(NewOps.front(), SE);
575 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
576 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
583 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
584 /// return that symbol, and mutate S to point to a new SCEV with that
586 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
587 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
588 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
589 S = SE.getConstant(GV->getType(), 0);
592 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
593 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
594 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
596 S = SE.getAddExpr(NewOps);
598 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
599 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
600 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
602 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
603 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
610 /// isAddressUse - Returns true if the specified instruction is using the
611 /// specified value as an address.
612 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
613 bool isAddress = isa<LoadInst>(Inst);
614 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
615 if (SI->getOperand(1) == OperandVal)
617 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
618 // Addressing modes can also be folded into prefetches and a variety
620 switch (II->getIntrinsicID()) {
622 case Intrinsic::prefetch:
623 case Intrinsic::x86_sse_storeu_ps:
624 case Intrinsic::x86_sse2_storeu_pd:
625 case Intrinsic::x86_sse2_storeu_dq:
626 case Intrinsic::x86_sse2_storel_dq:
627 if (II->getArgOperand(0) == OperandVal)
635 /// getAccessType - Return the type of the memory being accessed.
636 static Type *getAccessType(const Instruction *Inst) {
637 Type *AccessTy = Inst->getType();
638 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
639 AccessTy = SI->getOperand(0)->getType();
640 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
641 // Addressing modes can also be folded into prefetches and a variety
643 switch (II->getIntrinsicID()) {
645 case Intrinsic::x86_sse_storeu_ps:
646 case Intrinsic::x86_sse2_storeu_pd:
647 case Intrinsic::x86_sse2_storeu_dq:
648 case Intrinsic::x86_sse2_storel_dq:
649 AccessTy = II->getArgOperand(0)->getType();
654 // All pointers have the same requirements, so canonicalize them to an
655 // arbitrary pointer type to minimize variation.
656 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
657 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
658 PTy->getAddressSpace());
663 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
664 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
665 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
666 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
667 if (SE.isSCEVable(PN->getType()) &&
668 (SE.getEffectiveSCEVType(PN->getType()) ==
669 SE.getEffectiveSCEVType(AR->getType())) &&
670 SE.getSCEV(PN) == AR)
676 /// Check if expanding this expression is likely to incur significant cost. This
677 /// is tricky because SCEV doesn't track which expressions are actually computed
678 /// by the current IR.
680 /// We currently allow expansion of IV increments that involve adds,
681 /// multiplication by constants, and AddRecs from existing phis.
683 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
684 /// obvious multiple of the UDivExpr.
685 static bool isHighCostExpansion(const SCEV *S,
686 SmallPtrSet<const SCEV*, 8> &Processed,
687 ScalarEvolution &SE) {
688 // Zero/One operand expressions
689 switch (S->getSCEVType()) {
694 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
697 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
700 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
704 if (!Processed.insert(S))
707 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
708 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
710 if (isHighCostExpansion(*I, Processed, SE))
716 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
717 if (Mul->getNumOperands() == 2) {
718 // Multiplication by a constant is ok
719 if (isa<SCEVConstant>(Mul->getOperand(0)))
720 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
722 // If we have the value of one operand, check if an existing
723 // multiplication already generates this expression.
724 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
725 Value *UVal = U->getValue();
726 for (User *UR : UVal->users()) {
727 // If U is a constant, it may be used by a ConstantExpr.
728 Instruction *UI = dyn_cast<Instruction>(UR);
729 if (UI && UI->getOpcode() == Instruction::Mul &&
730 SE.isSCEVable(UI->getType())) {
731 return SE.getSCEV(UI) == Mul;
738 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
739 if (isExistingPhi(AR, SE))
743 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
747 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
748 /// specified set are trivially dead, delete them and see if this makes any of
749 /// their operands subsequently dead.
751 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
752 bool Changed = false;
754 while (!DeadInsts.empty()) {
755 Value *V = DeadInsts.pop_back_val();
756 Instruction *I = dyn_cast_or_null<Instruction>(V);
758 if (I == 0 || !isInstructionTriviallyDead(I))
761 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
762 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
765 DeadInsts.push_back(U);
768 I->eraseFromParent();
778 // Check if it is legal to fold 2 base registers.
779 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
781 // Get the cost of the scaling factor used in F for LU.
782 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
783 const LSRUse &LU, const Formula &F);
787 /// Cost - This class is used to measure and compare candidate formulae.
789 /// TODO: Some of these could be merged. Also, a lexical ordering
790 /// isn't always optimal.
794 unsigned NumBaseAdds;
801 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
802 SetupCost(0), ScaleCost(0) {}
804 bool operator<(const Cost &Other) const;
809 // Once any of the metrics loses, they must all remain losers.
811 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
812 | ImmCost | SetupCost | ScaleCost) != ~0u)
813 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
814 & ImmCost & SetupCost & ScaleCost) == ~0u);
819 assert(isValid() && "invalid cost");
820 return NumRegs == ~0u;
823 void RateFormula(const TargetTransformInfo &TTI,
825 SmallPtrSet<const SCEV *, 16> &Regs,
826 const DenseSet<const SCEV *> &VisitedRegs,
828 const SmallVectorImpl<int64_t> &Offsets,
829 ScalarEvolution &SE, DominatorTree &DT,
831 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
833 void print(raw_ostream &OS) const;
837 void RateRegister(const SCEV *Reg,
838 SmallPtrSet<const SCEV *, 16> &Regs,
840 ScalarEvolution &SE, DominatorTree &DT);
841 void RatePrimaryRegister(const SCEV *Reg,
842 SmallPtrSet<const SCEV *, 16> &Regs,
844 ScalarEvolution &SE, DominatorTree &DT,
845 SmallPtrSet<const SCEV *, 16> *LoserRegs);
850 /// RateRegister - Tally up interesting quantities from the given register.
851 void Cost::RateRegister(const SCEV *Reg,
852 SmallPtrSet<const SCEV *, 16> &Regs,
854 ScalarEvolution &SE, DominatorTree &DT) {
855 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
856 // If this is an addrec for another loop, don't second-guess its addrec phi
857 // nodes. LSR isn't currently smart enough to reason about more than one
858 // loop at a time. LSR has already run on inner loops, will not run on outer
859 // loops, and cannot be expected to change sibling loops.
860 if (AR->getLoop() != L) {
861 // If the AddRec exists, consider it's register free and leave it alone.
862 if (isExistingPhi(AR, SE))
865 // Otherwise, do not consider this formula at all.
869 AddRecCost += 1; /// TODO: This should be a function of the stride.
871 // Add the step value register, if it needs one.
872 // TODO: The non-affine case isn't precisely modeled here.
873 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
874 if (!Regs.count(AR->getOperand(1))) {
875 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
883 // Rough heuristic; favor registers which don't require extra setup
884 // instructions in the preheader.
885 if (!isa<SCEVUnknown>(Reg) &&
886 !isa<SCEVConstant>(Reg) &&
887 !(isa<SCEVAddRecExpr>(Reg) &&
888 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
889 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
892 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
893 SE.hasComputableLoopEvolution(Reg, L);
896 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
897 /// before, rate it. Optional LoserRegs provides a way to declare any formula
898 /// that refers to one of those regs an instant loser.
899 void Cost::RatePrimaryRegister(const SCEV *Reg,
900 SmallPtrSet<const SCEV *, 16> &Regs,
902 ScalarEvolution &SE, DominatorTree &DT,
903 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
904 if (LoserRegs && LoserRegs->count(Reg)) {
908 if (Regs.insert(Reg)) {
909 RateRegister(Reg, Regs, L, SE, DT);
910 if (LoserRegs && isLoser())
911 LoserRegs->insert(Reg);
915 void Cost::RateFormula(const TargetTransformInfo &TTI,
917 SmallPtrSet<const SCEV *, 16> &Regs,
918 const DenseSet<const SCEV *> &VisitedRegs,
920 const SmallVectorImpl<int64_t> &Offsets,
921 ScalarEvolution &SE, DominatorTree &DT,
923 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
924 // Tally up the registers.
925 if (const SCEV *ScaledReg = F.ScaledReg) {
926 if (VisitedRegs.count(ScaledReg)) {
930 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
934 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
935 E = F.BaseRegs.end(); I != E; ++I) {
936 const SCEV *BaseReg = *I;
937 if (VisitedRegs.count(BaseReg)) {
941 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
946 // Determine how many (unfolded) adds we'll need inside the loop.
947 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
948 if (NumBaseParts > 1)
949 // Do not count the base and a possible second register if the target
950 // allows to fold 2 registers.
951 NumBaseAdds += NumBaseParts - (1 + isLegal2RegAMUse(TTI, LU, F));
953 // Accumulate non-free scaling amounts.
954 ScaleCost += getScalingFactorCost(TTI, LU, F);
956 // Tally up the non-zero immediates.
957 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
958 E = Offsets.end(); I != E; ++I) {
959 int64_t Offset = (uint64_t)*I + F.BaseOffset;
961 ImmCost += 64; // Handle symbolic values conservatively.
962 // TODO: This should probably be the pointer size.
963 else if (Offset != 0)
964 ImmCost += APInt(64, Offset, true).getMinSignedBits();
966 assert(isValid() && "invalid cost");
969 /// Lose - Set this cost to a losing value.
980 /// operator< - Choose the lower cost.
981 bool Cost::operator<(const Cost &Other) const {
982 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
983 ImmCost, SetupCost) <
984 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
985 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
989 void Cost::print(raw_ostream &OS) const {
990 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
992 OS << ", with addrec cost " << AddRecCost;
994 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
995 if (NumBaseAdds != 0)
996 OS << ", plus " << NumBaseAdds << " base add"
997 << (NumBaseAdds == 1 ? "" : "s");
999 OS << ", plus " << ScaleCost << " scale cost";
1001 OS << ", plus " << ImmCost << " imm cost";
1003 OS << ", plus " << SetupCost << " setup cost";
1006 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1007 void Cost::dump() const {
1008 print(errs()); errs() << '\n';
1014 /// LSRFixup - An operand value in an instruction which is to be replaced
1015 /// with some equivalent, possibly strength-reduced, replacement.
1017 /// UserInst - The instruction which will be updated.
1018 Instruction *UserInst;
1020 /// OperandValToReplace - The operand of the instruction which will
1021 /// be replaced. The operand may be used more than once; every instance
1022 /// will be replaced.
1023 Value *OperandValToReplace;
1025 /// PostIncLoops - If this user is to use the post-incremented value of an
1026 /// induction variable, this variable is non-null and holds the loop
1027 /// associated with the induction variable.
1028 PostIncLoopSet PostIncLoops;
1030 /// LUIdx - The index of the LSRUse describing the expression which
1031 /// this fixup needs, minus an offset (below).
1034 /// Offset - A constant offset to be added to the LSRUse expression.
1035 /// This allows multiple fixups to share the same LSRUse with different
1036 /// offsets, for example in an unrolled loop.
1039 bool isUseFullyOutsideLoop(const Loop *L) const;
1043 void print(raw_ostream &OS) const;
1049 LSRFixup::LSRFixup()
1050 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1052 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1053 /// value outside of the given loop.
1054 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1055 // PHI nodes use their value in their incoming blocks.
1056 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1057 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1058 if (PN->getIncomingValue(i) == OperandValToReplace &&
1059 L->contains(PN->getIncomingBlock(i)))
1064 return !L->contains(UserInst);
1067 void LSRFixup::print(raw_ostream &OS) const {
1069 // Store is common and interesting enough to be worth special-casing.
1070 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1072 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1073 } else if (UserInst->getType()->isVoidTy())
1074 OS << UserInst->getOpcodeName();
1076 UserInst->printAsOperand(OS, /*PrintType=*/false);
1078 OS << ", OperandValToReplace=";
1079 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1081 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1082 E = PostIncLoops.end(); I != E; ++I) {
1083 OS << ", PostIncLoop=";
1084 (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1087 if (LUIdx != ~size_t(0))
1088 OS << ", LUIdx=" << LUIdx;
1091 OS << ", Offset=" << Offset;
1094 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1095 void LSRFixup::dump() const {
1096 print(errs()); errs() << '\n';
1102 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1103 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1104 struct UniquifierDenseMapInfo {
1105 static SmallVector<const SCEV *, 4> getEmptyKey() {
1106 SmallVector<const SCEV *, 4> V;
1107 V.push_back(reinterpret_cast<const SCEV *>(-1));
1111 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1112 SmallVector<const SCEV *, 4> V;
1113 V.push_back(reinterpret_cast<const SCEV *>(-2));
1117 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1118 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1121 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1122 const SmallVector<const SCEV *, 4> &RHS) {
1127 /// LSRUse - This class holds the state that LSR keeps for each use in
1128 /// IVUsers, as well as uses invented by LSR itself. It includes information
1129 /// about what kinds of things can be folded into the user, information about
1130 /// the user itself, and information about how the use may be satisfied.
1131 /// TODO: Represent multiple users of the same expression in common?
1133 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1136 /// KindType - An enum for a kind of use, indicating what types of
1137 /// scaled and immediate operands it might support.
1139 Basic, ///< A normal use, with no folding.
1140 Special, ///< A special case of basic, allowing -1 scales.
1141 Address, ///< An address use; folding according to TargetLowering
1142 ICmpZero ///< An equality icmp with both operands folded into one.
1143 // TODO: Add a generic icmp too?
1146 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1151 SmallVector<int64_t, 8> Offsets;
1155 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1156 /// LSRUse are outside of the loop, in which case some special-case heuristics
1158 bool AllFixupsOutsideLoop;
1160 /// RigidFormula is set to true to guarantee that this use will be associated
1161 /// with a single formula--the one that initially matched. Some SCEV
1162 /// expressions cannot be expanded. This allows LSR to consider the registers
1163 /// used by those expressions without the need to expand them later after
1164 /// changing the formula.
1167 /// WidestFixupType - This records the widest use type for any fixup using
1168 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1169 /// max fixup widths to be equivalent, because the narrower one may be relying
1170 /// on the implicit truncation to truncate away bogus bits.
1171 Type *WidestFixupType;
1173 /// Formulae - A list of ways to build a value that can satisfy this user.
1174 /// After the list is populated, one of these is selected heuristically and
1175 /// used to formulate a replacement for OperandValToReplace in UserInst.
1176 SmallVector<Formula, 12> Formulae;
1178 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1179 SmallPtrSet<const SCEV *, 4> Regs;
1181 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1182 MinOffset(INT64_MAX),
1183 MaxOffset(INT64_MIN),
1184 AllFixupsOutsideLoop(true),
1185 RigidFormula(false),
1186 WidestFixupType(0) {}
1188 bool HasFormulaWithSameRegs(const Formula &F) const;
1189 bool InsertFormula(const Formula &F);
1190 void DeleteFormula(Formula &F);
1191 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1193 void print(raw_ostream &OS) const;
1199 /// HasFormula - Test whether this use as a formula which has the same
1200 /// registers as the given formula.
1201 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1202 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1203 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1204 // Unstable sort by host order ok, because this is only used for uniquifying.
1205 std::sort(Key.begin(), Key.end());
1206 return Uniquifier.count(Key);
1209 /// InsertFormula - If the given formula has not yet been inserted, add it to
1210 /// the list, and return true. Return false otherwise.
1211 bool LSRUse::InsertFormula(const Formula &F) {
1212 if (!Formulae.empty() && RigidFormula)
1215 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1216 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1217 // Unstable sort by host order ok, because this is only used for uniquifying.
1218 std::sort(Key.begin(), Key.end());
1220 if (!Uniquifier.insert(Key).second)
1223 // Using a register to hold the value of 0 is not profitable.
1224 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1225 "Zero allocated in a scaled register!");
1227 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1228 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1229 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1232 // Add the formula to the list.
1233 Formulae.push_back(F);
1235 // Record registers now being used by this use.
1236 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1241 /// DeleteFormula - Remove the given formula from this use's list.
1242 void LSRUse::DeleteFormula(Formula &F) {
1243 if (&F != &Formulae.back())
1244 std::swap(F, Formulae.back());
1245 Formulae.pop_back();
1248 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1249 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1250 // Now that we've filtered out some formulae, recompute the Regs set.
1251 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1253 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1254 E = Formulae.end(); I != E; ++I) {
1255 const Formula &F = *I;
1256 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1257 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1260 // Update the RegTracker.
1261 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1262 E = OldRegs.end(); I != E; ++I)
1263 if (!Regs.count(*I))
1264 RegUses.DropRegister(*I, LUIdx);
1267 void LSRUse::print(raw_ostream &OS) const {
1268 OS << "LSR Use: Kind=";
1270 case Basic: OS << "Basic"; break;
1271 case Special: OS << "Special"; break;
1272 case ICmpZero: OS << "ICmpZero"; break;
1274 OS << "Address of ";
1275 if (AccessTy->isPointerTy())
1276 OS << "pointer"; // the full pointer type could be really verbose
1281 OS << ", Offsets={";
1282 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1283 E = Offsets.end(); I != E; ++I) {
1285 if (std::next(I) != E)
1290 if (AllFixupsOutsideLoop)
1291 OS << ", all-fixups-outside-loop";
1293 if (WidestFixupType)
1294 OS << ", widest fixup type: " << *WidestFixupType;
1297 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1298 void LSRUse::dump() const {
1299 print(errs()); errs() << '\n';
1303 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1304 /// be completely folded into the user instruction at isel time. This includes
1305 /// address-mode folding and special icmp tricks.
1306 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1307 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1308 bool HasBaseReg, int64_t Scale) {
1310 case LSRUse::Address:
1311 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1313 // Otherwise, just guess that reg+reg addressing is legal.
1316 case LSRUse::ICmpZero:
1317 // There's not even a target hook for querying whether it would be legal to
1318 // fold a GV into an ICmp.
1322 // ICmp only has two operands; don't allow more than two non-trivial parts.
1323 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1326 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1327 // putting the scaled register in the other operand of the icmp.
1328 if (Scale != 0 && Scale != -1)
1331 // If we have low-level target information, ask the target if it can fold an
1332 // integer immediate on an icmp.
1333 if (BaseOffset != 0) {
1335 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1336 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1337 // Offs is the ICmp immediate.
1339 // The cast does the right thing with INT64_MIN.
1340 BaseOffset = -(uint64_t)BaseOffset;
1341 return TTI.isLegalICmpImmediate(BaseOffset);
1344 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1348 // Only handle single-register values.
1349 return !BaseGV && Scale == 0 && BaseOffset == 0;
1351 case LSRUse::Special:
1352 // Special case Basic to handle -1 scales.
1353 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1356 llvm_unreachable("Invalid LSRUse Kind!");
1359 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1360 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1361 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1363 // Check for overflow.
1364 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1367 MinOffset = (uint64_t)BaseOffset + MinOffset;
1368 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1371 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1373 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1375 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1378 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1379 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1381 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1382 F.BaseOffset, F.HasBaseReg, F.Scale);
1385 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
1387 // If F is used as an Addressing Mode, it may fold one Base plus one
1388 // scaled register. If the scaled register is nil, do as if another
1389 // element of the base regs is a 1-scaled register.
1390 // This is possible if BaseRegs has at least 2 registers.
1392 // If this is not an address calculation, this is not an addressing mode
1394 if (LU.Kind != LSRUse::Address)
1397 // F is already scaled.
1401 // We need to keep one register for the base and one to scale.
1402 if (F.BaseRegs.size() < 2)
1405 return isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
1406 F.BaseGV, F.BaseOffset, F.HasBaseReg, 1);
1409 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1410 const LSRUse &LU, const Formula &F) {
1413 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1414 LU.AccessTy, F) && "Illegal formula in use.");
1417 case LSRUse::Address: {
1418 // Check the scaling factor cost with both the min and max offsets.
1419 int ScaleCostMinOffset =
1420 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1421 F.BaseOffset + LU.MinOffset,
1422 F.HasBaseReg, F.Scale);
1423 int ScaleCostMaxOffset =
1424 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1425 F.BaseOffset + LU.MaxOffset,
1426 F.HasBaseReg, F.Scale);
1428 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1429 "Legal addressing mode has an illegal cost!");
1430 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1432 case LSRUse::ICmpZero:
1433 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
1434 // Therefore, return 0 in case F.Scale == -1.
1435 return F.Scale != -1;
1438 case LSRUse::Special:
1442 llvm_unreachable("Invalid LSRUse Kind!");
1445 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1446 LSRUse::KindType Kind, Type *AccessTy,
1447 GlobalValue *BaseGV, int64_t BaseOffset,
1449 // Fast-path: zero is always foldable.
1450 if (BaseOffset == 0 && !BaseGV) return true;
1452 // Conservatively, create an address with an immediate and a
1453 // base and a scale.
1454 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1456 // Canonicalize a scale of 1 to a base register if the formula doesn't
1457 // already have a base register.
1458 if (!HasBaseReg && Scale == 1) {
1463 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1466 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1467 ScalarEvolution &SE, int64_t MinOffset,
1468 int64_t MaxOffset, LSRUse::KindType Kind,
1469 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1470 // Fast-path: zero is always foldable.
1471 if (S->isZero()) return true;
1473 // Conservatively, create an address with an immediate and a
1474 // base and a scale.
1475 int64_t BaseOffset = ExtractImmediate(S, SE);
1476 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1478 // If there's anything else involved, it's not foldable.
1479 if (!S->isZero()) return false;
1481 // Fast-path: zero is always foldable.
1482 if (BaseOffset == 0 && !BaseGV) return true;
1484 // Conservatively, create an address with an immediate and a
1485 // base and a scale.
1486 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1488 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1489 BaseOffset, HasBaseReg, Scale);
1494 /// IVInc - An individual increment in a Chain of IV increments.
1495 /// Relate an IV user to an expression that computes the IV it uses from the IV
1496 /// used by the previous link in the Chain.
1498 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1499 /// original IVOperand. The head of the chain's IVOperand is only valid during
1500 /// chain collection, before LSR replaces IV users. During chain generation,
1501 /// IncExpr can be used to find the new IVOperand that computes the same
1504 Instruction *UserInst;
1506 const SCEV *IncExpr;
1508 IVInc(Instruction *U, Value *O, const SCEV *E):
1509 UserInst(U), IVOperand(O), IncExpr(E) {}
1512 // IVChain - The list of IV increments in program order.
1513 // We typically add the head of a chain without finding subsequent links.
1515 SmallVector<IVInc,1> Incs;
1516 const SCEV *ExprBase;
1518 IVChain() : ExprBase(0) {}
1520 IVChain(const IVInc &Head, const SCEV *Base)
1521 : Incs(1, Head), ExprBase(Base) {}
1523 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1525 // begin - return the first increment in the chain.
1526 const_iterator begin() const {
1527 assert(!Incs.empty());
1528 return std::next(Incs.begin());
1530 const_iterator end() const {
1534 // hasIncs - Returns true if this chain contains any increments.
1535 bool hasIncs() const { return Incs.size() >= 2; }
1537 // add - Add an IVInc to the end of this chain.
1538 void add(const IVInc &X) { Incs.push_back(X); }
1540 // tailUserInst - Returns the last UserInst in the chain.
1541 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1543 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1545 bool isProfitableIncrement(const SCEV *OperExpr,
1546 const SCEV *IncExpr,
1550 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1551 /// Distinguish between FarUsers that definitely cross IV increments and
1552 /// NearUsers that may be used between IV increments.
1554 SmallPtrSet<Instruction*, 4> FarUsers;
1555 SmallPtrSet<Instruction*, 4> NearUsers;
1558 /// LSRInstance - This class holds state for the main loop strength reduction
1562 ScalarEvolution &SE;
1565 const TargetTransformInfo &TTI;
1569 /// IVIncInsertPos - This is the insert position that the current loop's
1570 /// induction variable increment should be placed. In simple loops, this is
1571 /// the latch block's terminator. But in more complicated cases, this is a
1572 /// position which will dominate all the in-loop post-increment users.
1573 Instruction *IVIncInsertPos;
1575 /// Factors - Interesting factors between use strides.
1576 SmallSetVector<int64_t, 8> Factors;
1578 /// Types - Interesting use types, to facilitate truncation reuse.
1579 SmallSetVector<Type *, 4> Types;
1581 /// Fixups - The list of operands which are to be replaced.
1582 SmallVector<LSRFixup, 16> Fixups;
1584 /// Uses - The list of interesting uses.
1585 SmallVector<LSRUse, 16> Uses;
1587 /// RegUses - Track which uses use which register candidates.
1588 RegUseTracker RegUses;
1590 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1591 // have more than a few IV increment chains in a loop. Missing a Chain falls
1592 // back to normal LSR behavior for those uses.
1593 static const unsigned MaxChains = 8;
1595 /// IVChainVec - IV users can form a chain of IV increments.
1596 SmallVector<IVChain, MaxChains> IVChainVec;
1598 /// IVIncSet - IV users that belong to profitable IVChains.
1599 SmallPtrSet<Use*, MaxChains> IVIncSet;
1601 void OptimizeShadowIV();
1602 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1603 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1604 void OptimizeLoopTermCond();
1606 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1607 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1608 void FinalizeChain(IVChain &Chain);
1609 void CollectChains();
1610 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1611 SmallVectorImpl<WeakVH> &DeadInsts);
1613 void CollectInterestingTypesAndFactors();
1614 void CollectFixupsAndInitialFormulae();
1616 LSRFixup &getNewFixup() {
1617 Fixups.push_back(LSRFixup());
1618 return Fixups.back();
1621 // Support for sharing of LSRUses between LSRFixups.
1622 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
1625 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1626 LSRUse::KindType Kind, Type *AccessTy);
1628 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1629 LSRUse::KindType Kind,
1632 void DeleteUse(LSRUse &LU, size_t LUIdx);
1634 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1636 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1637 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1638 void CountRegisters(const Formula &F, size_t LUIdx);
1639 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1641 void CollectLoopInvariantFixupsAndFormulae();
1643 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1644 unsigned Depth = 0);
1645 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1646 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1647 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1648 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1649 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1650 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1651 void GenerateCrossUseConstantOffsets();
1652 void GenerateAllReuseFormulae();
1654 void FilterOutUndesirableDedicatedRegisters();
1656 size_t EstimateSearchSpaceComplexity() const;
1657 void NarrowSearchSpaceByDetectingSupersets();
1658 void NarrowSearchSpaceByCollapsingUnrolledCode();
1659 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1660 void NarrowSearchSpaceByPickingWinnerRegs();
1661 void NarrowSearchSpaceUsingHeuristics();
1663 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1665 SmallVectorImpl<const Formula *> &Workspace,
1666 const Cost &CurCost,
1667 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1668 DenseSet<const SCEV *> &VisitedRegs) const;
1669 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1671 BasicBlock::iterator
1672 HoistInsertPosition(BasicBlock::iterator IP,
1673 const SmallVectorImpl<Instruction *> &Inputs) const;
1674 BasicBlock::iterator
1675 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1678 SCEVExpander &Rewriter) const;
1680 Value *Expand(const LSRFixup &LF,
1682 BasicBlock::iterator IP,
1683 SCEVExpander &Rewriter,
1684 SmallVectorImpl<WeakVH> &DeadInsts) const;
1685 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1687 SCEVExpander &Rewriter,
1688 SmallVectorImpl<WeakVH> &DeadInsts,
1690 void Rewrite(const LSRFixup &LF,
1692 SCEVExpander &Rewriter,
1693 SmallVectorImpl<WeakVH> &DeadInsts,
1695 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1699 LSRInstance(Loop *L, Pass *P);
1701 bool getChanged() const { return Changed; }
1703 void print_factors_and_types(raw_ostream &OS) const;
1704 void print_fixups(raw_ostream &OS) const;
1705 void print_uses(raw_ostream &OS) const;
1706 void print(raw_ostream &OS) const;
1712 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1713 /// inside the loop then try to eliminate the cast operation.
1714 void LSRInstance::OptimizeShadowIV() {
1715 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1716 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1719 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1720 UI != E; /* empty */) {
1721 IVUsers::const_iterator CandidateUI = UI;
1723 Instruction *ShadowUse = CandidateUI->getUser();
1725 bool IsSigned = false;
1727 /* If shadow use is a int->float cast then insert a second IV
1728 to eliminate this cast.
1730 for (unsigned i = 0; i < n; ++i)
1736 for (unsigned i = 0; i < n; ++i, ++d)
1739 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1741 DestTy = UCast->getDestTy();
1743 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1745 DestTy = SCast->getDestTy();
1747 if (!DestTy) continue;
1749 // If target does not support DestTy natively then do not apply
1750 // this transformation.
1751 if (!TTI.isTypeLegal(DestTy)) continue;
1753 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1755 if (PH->getNumIncomingValues() != 2) continue;
1757 Type *SrcTy = PH->getType();
1758 int Mantissa = DestTy->getFPMantissaWidth();
1759 if (Mantissa == -1) continue;
1760 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1763 unsigned Entry, Latch;
1764 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1772 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1773 if (!Init) continue;
1774 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1775 (double)Init->getSExtValue() :
1776 (double)Init->getZExtValue());
1778 BinaryOperator *Incr =
1779 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1780 if (!Incr) continue;
1781 if (Incr->getOpcode() != Instruction::Add
1782 && Incr->getOpcode() != Instruction::Sub)
1785 /* Initialize new IV, double d = 0.0 in above example. */
1787 if (Incr->getOperand(0) == PH)
1788 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1789 else if (Incr->getOperand(1) == PH)
1790 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1796 // Ignore negative constants, as the code below doesn't handle them
1797 // correctly. TODO: Remove this restriction.
1798 if (!C->getValue().isStrictlyPositive()) continue;
1800 /* Add new PHINode. */
1801 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1803 /* create new increment. '++d' in above example. */
1804 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1805 BinaryOperator *NewIncr =
1806 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1807 Instruction::FAdd : Instruction::FSub,
1808 NewPH, CFP, "IV.S.next.", Incr);
1810 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1811 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1813 /* Remove cast operation */
1814 ShadowUse->replaceAllUsesWith(NewPH);
1815 ShadowUse->eraseFromParent();
1821 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1822 /// set the IV user and stride information and return true, otherwise return
1824 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1825 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1826 if (UI->getUser() == Cond) {
1827 // NOTE: we could handle setcc instructions with multiple uses here, but
1828 // InstCombine does it as well for simple uses, it's not clear that it
1829 // occurs enough in real life to handle.
1836 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1837 /// a max computation.
1839 /// This is a narrow solution to a specific, but acute, problem. For loops
1845 /// } while (++i < n);
1847 /// the trip count isn't just 'n', because 'n' might not be positive. And
1848 /// unfortunately this can come up even for loops where the user didn't use
1849 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1850 /// will commonly be lowered like this:
1856 /// } while (++i < n);
1859 /// and then it's possible for subsequent optimization to obscure the if
1860 /// test in such a way that indvars can't find it.
1862 /// When indvars can't find the if test in loops like this, it creates a
1863 /// max expression, which allows it to give the loop a canonical
1864 /// induction variable:
1867 /// max = n < 1 ? 1 : n;
1870 /// } while (++i != max);
1872 /// Canonical induction variables are necessary because the loop passes
1873 /// are designed around them. The most obvious example of this is the
1874 /// LoopInfo analysis, which doesn't remember trip count values. It
1875 /// expects to be able to rediscover the trip count each time it is
1876 /// needed, and it does this using a simple analysis that only succeeds if
1877 /// the loop has a canonical induction variable.
1879 /// However, when it comes time to generate code, the maximum operation
1880 /// can be quite costly, especially if it's inside of an outer loop.
1882 /// This function solves this problem by detecting this type of loop and
1883 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1884 /// the instructions for the maximum computation.
1886 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1887 // Check that the loop matches the pattern we're looking for.
1888 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1889 Cond->getPredicate() != CmpInst::ICMP_NE)
1892 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1893 if (!Sel || !Sel->hasOneUse()) return Cond;
1895 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1896 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1898 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1900 // Add one to the backedge-taken count to get the trip count.
1901 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1902 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1904 // Check for a max calculation that matches the pattern. There's no check
1905 // for ICMP_ULE here because the comparison would be with zero, which
1906 // isn't interesting.
1907 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1908 const SCEVNAryExpr *Max = 0;
1909 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1910 Pred = ICmpInst::ICMP_SLE;
1912 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1913 Pred = ICmpInst::ICMP_SLT;
1915 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1916 Pred = ICmpInst::ICMP_ULT;
1923 // To handle a max with more than two operands, this optimization would
1924 // require additional checking and setup.
1925 if (Max->getNumOperands() != 2)
1928 const SCEV *MaxLHS = Max->getOperand(0);
1929 const SCEV *MaxRHS = Max->getOperand(1);
1931 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1932 // for a comparison with 1. For <= and >=, a comparison with zero.
1934 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1937 // Check the relevant induction variable for conformance to
1939 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1940 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1941 if (!AR || !AR->isAffine() ||
1942 AR->getStart() != One ||
1943 AR->getStepRecurrence(SE) != One)
1946 assert(AR->getLoop() == L &&
1947 "Loop condition operand is an addrec in a different loop!");
1949 // Check the right operand of the select, and remember it, as it will
1950 // be used in the new comparison instruction.
1952 if (ICmpInst::isTrueWhenEqual(Pred)) {
1953 // Look for n+1, and grab n.
1954 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1955 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1956 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1957 NewRHS = BO->getOperand(0);
1958 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1959 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1960 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1961 NewRHS = BO->getOperand(0);
1964 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1965 NewRHS = Sel->getOperand(1);
1966 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1967 NewRHS = Sel->getOperand(2);
1968 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1969 NewRHS = SU->getValue();
1971 // Max doesn't match expected pattern.
1974 // Determine the new comparison opcode. It may be signed or unsigned,
1975 // and the original comparison may be either equality or inequality.
1976 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1977 Pred = CmpInst::getInversePredicate(Pred);
1979 // Ok, everything looks ok to change the condition into an SLT or SGE and
1980 // delete the max calculation.
1982 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1984 // Delete the max calculation instructions.
1985 Cond->replaceAllUsesWith(NewCond);
1986 CondUse->setUser(NewCond);
1987 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1988 Cond->eraseFromParent();
1989 Sel->eraseFromParent();
1990 if (Cmp->use_empty())
1991 Cmp->eraseFromParent();
1995 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1996 /// postinc iv when possible.
1998 LSRInstance::OptimizeLoopTermCond() {
1999 SmallPtrSet<Instruction *, 4> PostIncs;
2001 BasicBlock *LatchBlock = L->getLoopLatch();
2002 SmallVector<BasicBlock*, 8> ExitingBlocks;
2003 L->getExitingBlocks(ExitingBlocks);
2005 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2006 BasicBlock *ExitingBlock = ExitingBlocks[i];
2008 // Get the terminating condition for the loop if possible. If we
2009 // can, we want to change it to use a post-incremented version of its
2010 // induction variable, to allow coalescing the live ranges for the IV into
2011 // one register value.
2013 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2016 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2017 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2020 // Search IVUsesByStride to find Cond's IVUse if there is one.
2021 IVStrideUse *CondUse = 0;
2022 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2023 if (!FindIVUserForCond(Cond, CondUse))
2026 // If the trip count is computed in terms of a max (due to ScalarEvolution
2027 // being unable to find a sufficient guard, for example), change the loop
2028 // comparison to use SLT or ULT instead of NE.
2029 // One consequence of doing this now is that it disrupts the count-down
2030 // optimization. That's not always a bad thing though, because in such
2031 // cases it may still be worthwhile to avoid a max.
2032 Cond = OptimizeMax(Cond, CondUse);
2034 // If this exiting block dominates the latch block, it may also use
2035 // the post-inc value if it won't be shared with other uses.
2036 // Check for dominance.
2037 if (!DT.dominates(ExitingBlock, LatchBlock))
2040 // Conservatively avoid trying to use the post-inc value in non-latch
2041 // exits if there may be pre-inc users in intervening blocks.
2042 if (LatchBlock != ExitingBlock)
2043 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2044 // Test if the use is reachable from the exiting block. This dominator
2045 // query is a conservative approximation of reachability.
2046 if (&*UI != CondUse &&
2047 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2048 // Conservatively assume there may be reuse if the quotient of their
2049 // strides could be a legal scale.
2050 const SCEV *A = IU.getStride(*CondUse, L);
2051 const SCEV *B = IU.getStride(*UI, L);
2052 if (!A || !B) continue;
2053 if (SE.getTypeSizeInBits(A->getType()) !=
2054 SE.getTypeSizeInBits(B->getType())) {
2055 if (SE.getTypeSizeInBits(A->getType()) >
2056 SE.getTypeSizeInBits(B->getType()))
2057 B = SE.getSignExtendExpr(B, A->getType());
2059 A = SE.getSignExtendExpr(A, B->getType());
2061 if (const SCEVConstant *D =
2062 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2063 const ConstantInt *C = D->getValue();
2064 // Stride of one or negative one can have reuse with non-addresses.
2065 if (C->isOne() || C->isAllOnesValue())
2066 goto decline_post_inc;
2067 // Avoid weird situations.
2068 if (C->getValue().getMinSignedBits() >= 64 ||
2069 C->getValue().isMinSignedValue())
2070 goto decline_post_inc;
2071 // Check for possible scaled-address reuse.
2072 Type *AccessTy = getAccessType(UI->getUser());
2073 int64_t Scale = C->getSExtValue();
2074 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2076 /*HasBaseReg=*/ false, Scale))
2077 goto decline_post_inc;
2079 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2081 /*HasBaseReg=*/ false, Scale))
2082 goto decline_post_inc;
2086 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2089 // It's possible for the setcc instruction to be anywhere in the loop, and
2090 // possible for it to have multiple users. If it is not immediately before
2091 // the exiting block branch, move it.
2092 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2093 if (Cond->hasOneUse()) {
2094 Cond->moveBefore(TermBr);
2096 // Clone the terminating condition and insert into the loopend.
2097 ICmpInst *OldCond = Cond;
2098 Cond = cast<ICmpInst>(Cond->clone());
2099 Cond->setName(L->getHeader()->getName() + ".termcond");
2100 ExitingBlock->getInstList().insert(TermBr, Cond);
2102 // Clone the IVUse, as the old use still exists!
2103 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2104 TermBr->replaceUsesOfWith(OldCond, Cond);
2108 // If we get to here, we know that we can transform the setcc instruction to
2109 // use the post-incremented version of the IV, allowing us to coalesce the
2110 // live ranges for the IV correctly.
2111 CondUse->transformToPostInc(L);
2114 PostIncs.insert(Cond);
2118 // Determine an insertion point for the loop induction variable increment. It
2119 // must dominate all the post-inc comparisons we just set up, and it must
2120 // dominate the loop latch edge.
2121 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2122 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2123 E = PostIncs.end(); I != E; ++I) {
2125 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2127 if (BB == (*I)->getParent())
2128 IVIncInsertPos = *I;
2129 else if (BB != IVIncInsertPos->getParent())
2130 IVIncInsertPos = BB->getTerminator();
2134 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2135 /// at the given offset and other details. If so, update the use and
2138 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2139 LSRUse::KindType Kind, Type *AccessTy) {
2140 int64_t NewMinOffset = LU.MinOffset;
2141 int64_t NewMaxOffset = LU.MaxOffset;
2142 Type *NewAccessTy = AccessTy;
2144 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2145 // something conservative, however this can pessimize in the case that one of
2146 // the uses will have all its uses outside the loop, for example.
2147 if (LU.Kind != Kind)
2149 // Conservatively assume HasBaseReg is true for now.
2150 if (NewOffset < LU.MinOffset) {
2151 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2152 LU.MaxOffset - NewOffset, HasBaseReg))
2154 NewMinOffset = NewOffset;
2155 } else if (NewOffset > LU.MaxOffset) {
2156 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2157 NewOffset - LU.MinOffset, HasBaseReg))
2159 NewMaxOffset = NewOffset;
2161 // Check for a mismatched access type, and fall back conservatively as needed.
2162 // TODO: Be less conservative when the type is similar and can use the same
2163 // addressing modes.
2164 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2165 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2168 LU.MinOffset = NewMinOffset;
2169 LU.MaxOffset = NewMaxOffset;
2170 LU.AccessTy = NewAccessTy;
2171 if (NewOffset != LU.Offsets.back())
2172 LU.Offsets.push_back(NewOffset);
2176 /// getUse - Return an LSRUse index and an offset value for a fixup which
2177 /// needs the given expression, with the given kind and optional access type.
2178 /// Either reuse an existing use or create a new one, as needed.
2179 std::pair<size_t, int64_t>
2180 LSRInstance::getUse(const SCEV *&Expr,
2181 LSRUse::KindType Kind, Type *AccessTy) {
2182 const SCEV *Copy = Expr;
2183 int64_t Offset = ExtractImmediate(Expr, SE);
2185 // Basic uses can't accept any offset, for example.
2186 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2187 Offset, /*HasBaseReg=*/ true)) {
2192 std::pair<UseMapTy::iterator, bool> P =
2193 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2195 // A use already existed with this base.
2196 size_t LUIdx = P.first->second;
2197 LSRUse &LU = Uses[LUIdx];
2198 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2200 return std::make_pair(LUIdx, Offset);
2203 // Create a new use.
2204 size_t LUIdx = Uses.size();
2205 P.first->second = LUIdx;
2206 Uses.push_back(LSRUse(Kind, AccessTy));
2207 LSRUse &LU = Uses[LUIdx];
2209 // We don't need to track redundant offsets, but we don't need to go out
2210 // of our way here to avoid them.
2211 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2212 LU.Offsets.push_back(Offset);
2214 LU.MinOffset = Offset;
2215 LU.MaxOffset = Offset;
2216 return std::make_pair(LUIdx, Offset);
2219 /// DeleteUse - Delete the given use from the Uses list.
2220 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2221 if (&LU != &Uses.back())
2222 std::swap(LU, Uses.back());
2226 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2229 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2230 /// a formula that has the same registers as the given formula.
2232 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2233 const LSRUse &OrigLU) {
2234 // Search all uses for the formula. This could be more clever.
2235 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2236 LSRUse &LU = Uses[LUIdx];
2237 // Check whether this use is close enough to OrigLU, to see whether it's
2238 // worthwhile looking through its formulae.
2239 // Ignore ICmpZero uses because they may contain formulae generated by
2240 // GenerateICmpZeroScales, in which case adding fixup offsets may
2242 if (&LU != &OrigLU &&
2243 LU.Kind != LSRUse::ICmpZero &&
2244 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2245 LU.WidestFixupType == OrigLU.WidestFixupType &&
2246 LU.HasFormulaWithSameRegs(OrigF)) {
2247 // Scan through this use's formulae.
2248 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2249 E = LU.Formulae.end(); I != E; ++I) {
2250 const Formula &F = *I;
2251 // Check to see if this formula has the same registers and symbols
2253 if (F.BaseRegs == OrigF.BaseRegs &&
2254 F.ScaledReg == OrigF.ScaledReg &&
2255 F.BaseGV == OrigF.BaseGV &&
2256 F.Scale == OrigF.Scale &&
2257 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2258 if (F.BaseOffset == 0)
2260 // This is the formula where all the registers and symbols matched;
2261 // there aren't going to be any others. Since we declined it, we
2262 // can skip the rest of the formulae and proceed to the next LSRUse.
2269 // Nothing looked good.
2273 void LSRInstance::CollectInterestingTypesAndFactors() {
2274 SmallSetVector<const SCEV *, 4> Strides;
2276 // Collect interesting types and strides.
2277 SmallVector<const SCEV *, 4> Worklist;
2278 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2279 const SCEV *Expr = IU.getExpr(*UI);
2281 // Collect interesting types.
2282 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2284 // Add strides for mentioned loops.
2285 Worklist.push_back(Expr);
2287 const SCEV *S = Worklist.pop_back_val();
2288 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2289 if (AR->getLoop() == L)
2290 Strides.insert(AR->getStepRecurrence(SE));
2291 Worklist.push_back(AR->getStart());
2292 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2293 Worklist.append(Add->op_begin(), Add->op_end());
2295 } while (!Worklist.empty());
2298 // Compute interesting factors from the set of interesting strides.
2299 for (SmallSetVector<const SCEV *, 4>::const_iterator
2300 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2301 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2302 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2303 const SCEV *OldStride = *I;
2304 const SCEV *NewStride = *NewStrideIter;
2306 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2307 SE.getTypeSizeInBits(NewStride->getType())) {
2308 if (SE.getTypeSizeInBits(OldStride->getType()) >
2309 SE.getTypeSizeInBits(NewStride->getType()))
2310 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2312 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2314 if (const SCEVConstant *Factor =
2315 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2317 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2318 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2319 } else if (const SCEVConstant *Factor =
2320 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2323 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2324 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2328 // If all uses use the same type, don't bother looking for truncation-based
2330 if (Types.size() == 1)
2333 DEBUG(print_factors_and_types(dbgs()));
2336 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2337 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2338 /// Instructions to IVStrideUses, we could partially skip this.
2339 static User::op_iterator
2340 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2341 Loop *L, ScalarEvolution &SE) {
2342 for(; OI != OE; ++OI) {
2343 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2344 if (!SE.isSCEVable(Oper->getType()))
2347 if (const SCEVAddRecExpr *AR =
2348 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2349 if (AR->getLoop() == L)
2357 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2358 /// operands, so wrap it in a convenient helper.
2359 static Value *getWideOperand(Value *Oper) {
2360 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2361 return Trunc->getOperand(0);
2365 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2367 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2368 Type *LType = LVal->getType();
2369 Type *RType = RVal->getType();
2370 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2373 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2374 /// NULL for any constant. Returning the expression itself is
2375 /// conservative. Returning a deeper subexpression is more precise and valid as
2376 /// long as it isn't less complex than another subexpression. For expressions
2377 /// involving multiple unscaled values, we need to return the pointer-type
2378 /// SCEVUnknown. This avoids forming chains across objects, such as:
2379 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2381 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2382 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2383 static const SCEV *getExprBase(const SCEV *S) {
2384 switch (S->getSCEVType()) {
2385 default: // uncluding scUnknown.
2390 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2392 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2394 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2396 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2397 // there's nothing more complex.
2398 // FIXME: not sure if we want to recognize negation.
2399 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2400 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2401 E(Add->op_begin()); I != E; ++I) {
2402 const SCEV *SubExpr = *I;
2403 if (SubExpr->getSCEVType() == scAddExpr)
2404 return getExprBase(SubExpr);
2406 if (SubExpr->getSCEVType() != scMulExpr)
2409 return S; // all operands are scaled, be conservative.
2412 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2416 /// Return true if the chain increment is profitable to expand into a loop
2417 /// invariant value, which may require its own register. A profitable chain
2418 /// increment will be an offset relative to the same base. We allow such offsets
2419 /// to potentially be used as chain increment as long as it's not obviously
2420 /// expensive to expand using real instructions.
2421 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2422 const SCEV *IncExpr,
2423 ScalarEvolution &SE) {
2424 // Aggressively form chains when -stress-ivchain.
2428 // Do not replace a constant offset from IV head with a nonconstant IV
2430 if (!isa<SCEVConstant>(IncExpr)) {
2431 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2432 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2436 SmallPtrSet<const SCEV*, 8> Processed;
2437 return !isHighCostExpansion(IncExpr, Processed, SE);
2440 /// Return true if the number of registers needed for the chain is estimated to
2441 /// be less than the number required for the individual IV users. First prohibit
2442 /// any IV users that keep the IV live across increments (the Users set should
2443 /// be empty). Next count the number and type of increments in the chain.
2445 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2446 /// effectively use postinc addressing modes. Only consider it profitable it the
2447 /// increments can be computed in fewer registers when chained.
2449 /// TODO: Consider IVInc free if it's already used in another chains.
2451 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2452 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2456 if (!Chain.hasIncs())
2459 if (!Users.empty()) {
2460 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2461 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2462 E = Users.end(); I != E; ++I) {
2463 dbgs() << " " << **I << "\n";
2467 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2469 // The chain itself may require a register, so intialize cost to 1.
2472 // A complete chain likely eliminates the need for keeping the original IV in
2473 // a register. LSR does not currently know how to form a complete chain unless
2474 // the header phi already exists.
2475 if (isa<PHINode>(Chain.tailUserInst())
2476 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2479 const SCEV *LastIncExpr = 0;
2480 unsigned NumConstIncrements = 0;
2481 unsigned NumVarIncrements = 0;
2482 unsigned NumReusedIncrements = 0;
2483 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2486 if (I->IncExpr->isZero())
2489 // Incrementing by zero or some constant is neutral. We assume constants can
2490 // be folded into an addressing mode or an add's immediate operand.
2491 if (isa<SCEVConstant>(I->IncExpr)) {
2492 ++NumConstIncrements;
2496 if (I->IncExpr == LastIncExpr)
2497 ++NumReusedIncrements;
2501 LastIncExpr = I->IncExpr;
2503 // An IV chain with a single increment is handled by LSR's postinc
2504 // uses. However, a chain with multiple increments requires keeping the IV's
2505 // value live longer than it needs to be if chained.
2506 if (NumConstIncrements > 1)
2509 // Materializing increment expressions in the preheader that didn't exist in
2510 // the original code may cost a register. For example, sign-extended array
2511 // indices can produce ridiculous increments like this:
2512 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2513 cost += NumVarIncrements;
2515 // Reusing variable increments likely saves a register to hold the multiple of
2517 cost -= NumReusedIncrements;
2519 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2525 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2527 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2528 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2529 // When IVs are used as types of varying widths, they are generally converted
2530 // to a wider type with some uses remaining narrow under a (free) trunc.
2531 Value *const NextIV = getWideOperand(IVOper);
2532 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2533 const SCEV *const OperExprBase = getExprBase(OperExpr);
2535 // Visit all existing chains. Check if its IVOper can be computed as a
2536 // profitable loop invariant increment from the last link in the Chain.
2537 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2538 const SCEV *LastIncExpr = 0;
2539 for (; ChainIdx < NChains; ++ChainIdx) {
2540 IVChain &Chain = IVChainVec[ChainIdx];
2542 // Prune the solution space aggressively by checking that both IV operands
2543 // are expressions that operate on the same unscaled SCEVUnknown. This
2544 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2545 // first avoids creating extra SCEV expressions.
2546 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2549 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2550 if (!isCompatibleIVType(PrevIV, NextIV))
2553 // A phi node terminates a chain.
2554 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2557 // The increment must be loop-invariant so it can be kept in a register.
2558 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2559 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2560 if (!SE.isLoopInvariant(IncExpr, L))
2563 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2564 LastIncExpr = IncExpr;
2568 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2569 // bother for phi nodes, because they must be last in the chain.
2570 if (ChainIdx == NChains) {
2571 if (isa<PHINode>(UserInst))
2573 if (NChains >= MaxChains && !StressIVChain) {
2574 DEBUG(dbgs() << "IV Chain Limit\n");
2577 LastIncExpr = OperExpr;
2578 // IVUsers may have skipped over sign/zero extensions. We don't currently
2579 // attempt to form chains involving extensions unless they can be hoisted
2580 // into this loop's AddRec.
2581 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2584 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2586 ChainUsersVec.resize(NChains);
2587 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2588 << ") IV=" << *LastIncExpr << "\n");
2590 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2591 << ") IV+" << *LastIncExpr << "\n");
2592 // Add this IV user to the end of the chain.
2593 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2595 IVChain &Chain = IVChainVec[ChainIdx];
2597 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2598 // This chain's NearUsers become FarUsers.
2599 if (!LastIncExpr->isZero()) {
2600 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2605 // All other uses of IVOperand become near uses of the chain.
2606 // We currently ignore intermediate values within SCEV expressions, assuming
2607 // they will eventually be used be the current chain, or can be computed
2608 // from one of the chain increments. To be more precise we could
2609 // transitively follow its user and only add leaf IV users to the set.
2610 for (User *U : IVOper->users()) {
2611 Instruction *OtherUse = dyn_cast<Instruction>(U);
2614 // Uses in the chain will no longer be uses if the chain is formed.
2615 // Include the head of the chain in this iteration (not Chain.begin()).
2616 IVChain::const_iterator IncIter = Chain.Incs.begin();
2617 IVChain::const_iterator IncEnd = Chain.Incs.end();
2618 for( ; IncIter != IncEnd; ++IncIter) {
2619 if (IncIter->UserInst == OtherUse)
2622 if (IncIter != IncEnd)
2625 if (SE.isSCEVable(OtherUse->getType())
2626 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2627 && IU.isIVUserOrOperand(OtherUse)) {
2630 NearUsers.insert(OtherUse);
2633 // Since this user is part of the chain, it's no longer considered a use
2635 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2638 /// CollectChains - Populate the vector of Chains.
2640 /// This decreases ILP at the architecture level. Targets with ample registers,
2641 /// multiple memory ports, and no register renaming probably don't want
2642 /// this. However, such targets should probably disable LSR altogether.
2644 /// The job of LSR is to make a reasonable choice of induction variables across
2645 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2646 /// ILP *within the loop* if the target wants it.
2648 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2649 /// will not reorder memory operations, it will recognize this as a chain, but
2650 /// will generate redundant IV increments. Ideally this would be corrected later
2651 /// by a smart scheduler:
2657 /// TODO: Walk the entire domtree within this loop, not just the path to the
2658 /// loop latch. This will discover chains on side paths, but requires
2659 /// maintaining multiple copies of the Chains state.
2660 void LSRInstance::CollectChains() {
2661 DEBUG(dbgs() << "Collecting IV Chains.\n");
2662 SmallVector<ChainUsers, 8> ChainUsersVec;
2664 SmallVector<BasicBlock *,8> LatchPath;
2665 BasicBlock *LoopHeader = L->getHeader();
2666 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2667 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2668 LatchPath.push_back(Rung->getBlock());
2670 LatchPath.push_back(LoopHeader);
2672 // Walk the instruction stream from the loop header to the loop latch.
2673 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2674 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2675 BBIter != BBEnd; ++BBIter) {
2676 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2678 // Skip instructions that weren't seen by IVUsers analysis.
2679 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2682 // Ignore users that are part of a SCEV expression. This way we only
2683 // consider leaf IV Users. This effectively rediscovers a portion of
2684 // IVUsers analysis but in program order this time.
2685 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2688 // Remove this instruction from any NearUsers set it may be in.
2689 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2690 ChainIdx < NChains; ++ChainIdx) {
2691 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2693 // Search for operands that can be chained.
2694 SmallPtrSet<Instruction*, 4> UniqueOperands;
2695 User::op_iterator IVOpEnd = I->op_end();
2696 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2697 while (IVOpIter != IVOpEnd) {
2698 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2699 if (UniqueOperands.insert(IVOpInst))
2700 ChainInstruction(I, IVOpInst, ChainUsersVec);
2701 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2703 } // Continue walking down the instructions.
2704 } // Continue walking down the domtree.
2705 // Visit phi backedges to determine if the chain can generate the IV postinc.
2706 for (BasicBlock::iterator I = L->getHeader()->begin();
2707 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2708 if (!SE.isSCEVable(PN->getType()))
2712 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2714 ChainInstruction(PN, IncV, ChainUsersVec);
2716 // Remove any unprofitable chains.
2717 unsigned ChainIdx = 0;
2718 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2719 UsersIdx < NChains; ++UsersIdx) {
2720 if (!isProfitableChain(IVChainVec[UsersIdx],
2721 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2723 // Preserve the chain at UsesIdx.
2724 if (ChainIdx != UsersIdx)
2725 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2726 FinalizeChain(IVChainVec[ChainIdx]);
2729 IVChainVec.resize(ChainIdx);
2732 void LSRInstance::FinalizeChain(IVChain &Chain) {
2733 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2734 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2736 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2738 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2739 User::op_iterator UseI =
2740 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2741 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2742 IVIncSet.insert(UseI);
2746 /// Return true if the IVInc can be folded into an addressing mode.
2747 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2748 Value *Operand, const TargetTransformInfo &TTI) {
2749 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2750 if (!IncConst || !isAddressUse(UserInst, Operand))
2753 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2756 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2757 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2758 getAccessType(UserInst), /*BaseGV=*/ 0,
2759 IncOffset, /*HaseBaseReg=*/ false))
2765 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2766 /// materialize the IV user's operand from the previous IV user's operand.
2767 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2768 SmallVectorImpl<WeakVH> &DeadInsts) {
2769 // Find the new IVOperand for the head of the chain. It may have been replaced
2771 const IVInc &Head = Chain.Incs[0];
2772 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2773 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2774 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2777 while (IVOpIter != IVOpEnd) {
2778 IVSrc = getWideOperand(*IVOpIter);
2780 // If this operand computes the expression that the chain needs, we may use
2781 // it. (Check this after setting IVSrc which is used below.)
2783 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2784 // narrow for the chain, so we can no longer use it. We do allow using a
2785 // wider phi, assuming the LSR checked for free truncation. In that case we
2786 // should already have a truncate on this operand such that
2787 // getSCEV(IVSrc) == IncExpr.
2788 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2789 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2792 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2794 if (IVOpIter == IVOpEnd) {
2795 // Gracefully give up on this chain.
2796 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2800 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2801 Type *IVTy = IVSrc->getType();
2802 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2803 const SCEV *LeftOverExpr = 0;
2804 for (IVChain::const_iterator IncI = Chain.begin(),
2805 IncE = Chain.end(); IncI != IncE; ++IncI) {
2807 Instruction *InsertPt = IncI->UserInst;
2808 if (isa<PHINode>(InsertPt))
2809 InsertPt = L->getLoopLatch()->getTerminator();
2811 // IVOper will replace the current IV User's operand. IVSrc is the IV
2812 // value currently held in a register.
2813 Value *IVOper = IVSrc;
2814 if (!IncI->IncExpr->isZero()) {
2815 // IncExpr was the result of subtraction of two narrow values, so must
2817 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2818 LeftOverExpr = LeftOverExpr ?
2819 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2821 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2822 // Expand the IV increment.
2823 Rewriter.clearPostInc();
2824 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2825 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2826 SE.getUnknown(IncV));
2827 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2829 // If an IV increment can't be folded, use it as the next IV value.
2830 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2832 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2837 Type *OperTy = IncI->IVOperand->getType();
2838 if (IVTy != OperTy) {
2839 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2840 "cannot extend a chained IV");
2841 IRBuilder<> Builder(InsertPt);
2842 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2844 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2845 DeadInsts.push_back(IncI->IVOperand);
2847 // If LSR created a new, wider phi, we may also replace its postinc. We only
2848 // do this if we also found a wide value for the head of the chain.
2849 if (isa<PHINode>(Chain.tailUserInst())) {
2850 for (BasicBlock::iterator I = L->getHeader()->begin();
2851 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2852 if (!isCompatibleIVType(Phi, IVSrc))
2854 Instruction *PostIncV = dyn_cast<Instruction>(
2855 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2856 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2858 Value *IVOper = IVSrc;
2859 Type *PostIncTy = PostIncV->getType();
2860 if (IVTy != PostIncTy) {
2861 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2862 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2863 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2864 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2866 Phi->replaceUsesOfWith(PostIncV, IVOper);
2867 DeadInsts.push_back(PostIncV);
2872 void LSRInstance::CollectFixupsAndInitialFormulae() {
2873 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2874 Instruction *UserInst = UI->getUser();
2875 // Skip IV users that are part of profitable IV Chains.
2876 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2877 UI->getOperandValToReplace());
2878 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2879 if (IVIncSet.count(UseI))
2883 LSRFixup &LF = getNewFixup();
2884 LF.UserInst = UserInst;
2885 LF.OperandValToReplace = UI->getOperandValToReplace();
2886 LF.PostIncLoops = UI->getPostIncLoops();
2888 LSRUse::KindType Kind = LSRUse::Basic;
2890 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2891 Kind = LSRUse::Address;
2892 AccessTy = getAccessType(LF.UserInst);
2895 const SCEV *S = IU.getExpr(*UI);
2897 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2898 // (N - i == 0), and this allows (N - i) to be the expression that we work
2899 // with rather than just N or i, so we can consider the register
2900 // requirements for both N and i at the same time. Limiting this code to
2901 // equality icmps is not a problem because all interesting loops use
2902 // equality icmps, thanks to IndVarSimplify.
2903 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2904 if (CI->isEquality()) {
2905 // Swap the operands if needed to put the OperandValToReplace on the
2906 // left, for consistency.
2907 Value *NV = CI->getOperand(1);
2908 if (NV == LF.OperandValToReplace) {
2909 CI->setOperand(1, CI->getOperand(0));
2910 CI->setOperand(0, NV);
2911 NV = CI->getOperand(1);
2915 // x == y --> x - y == 0
2916 const SCEV *N = SE.getSCEV(NV);
2917 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2918 // S is normalized, so normalize N before folding it into S
2919 // to keep the result normalized.
2920 N = TransformForPostIncUse(Normalize, N, CI, 0,
2921 LF.PostIncLoops, SE, DT);
2922 Kind = LSRUse::ICmpZero;
2923 S = SE.getMinusSCEV(N, S);
2926 // -1 and the negations of all interesting strides (except the negation
2927 // of -1) are now also interesting.
2928 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2929 if (Factors[i] != -1)
2930 Factors.insert(-(uint64_t)Factors[i]);
2934 // Set up the initial formula for this use.
2935 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2937 LF.Offset = P.second;
2938 LSRUse &LU = Uses[LF.LUIdx];
2939 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2940 if (!LU.WidestFixupType ||
2941 SE.getTypeSizeInBits(LU.WidestFixupType) <
2942 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2943 LU.WidestFixupType = LF.OperandValToReplace->getType();
2945 // If this is the first use of this LSRUse, give it a formula.
2946 if (LU.Formulae.empty()) {
2947 InsertInitialFormula(S, LU, LF.LUIdx);
2948 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2952 DEBUG(print_fixups(dbgs()));
2955 /// InsertInitialFormula - Insert a formula for the given expression into
2956 /// the given use, separating out loop-variant portions from loop-invariant
2957 /// and loop-computable portions.
2959 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2960 // Mark uses whose expressions cannot be expanded.
2961 if (!isSafeToExpand(S, SE))
2962 LU.RigidFormula = true;
2965 F.InitialMatch(S, L, SE);
2966 bool Inserted = InsertFormula(LU, LUIdx, F);
2967 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2970 /// InsertSupplementalFormula - Insert a simple single-register formula for
2971 /// the given expression into the given use.
2973 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2974 LSRUse &LU, size_t LUIdx) {
2976 F.BaseRegs.push_back(S);
2977 F.HasBaseReg = true;
2978 bool Inserted = InsertFormula(LU, LUIdx, F);
2979 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2982 /// CountRegisters - Note which registers are used by the given formula,
2983 /// updating RegUses.
2984 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2986 RegUses.CountRegister(F.ScaledReg, LUIdx);
2987 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2988 E = F.BaseRegs.end(); I != E; ++I)
2989 RegUses.CountRegister(*I, LUIdx);
2992 /// InsertFormula - If the given formula has not yet been inserted, add it to
2993 /// the list, and return true. Return false otherwise.
2994 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2995 if (!LU.InsertFormula(F))
2998 CountRegisters(F, LUIdx);
3002 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3003 /// loop-invariant values which we're tracking. These other uses will pin these
3004 /// values in registers, making them less profitable for elimination.
3005 /// TODO: This currently misses non-constant addrec step registers.
3006 /// TODO: Should this give more weight to users inside the loop?
3008 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3009 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3010 SmallPtrSet<const SCEV *, 8> Inserted;
3012 while (!Worklist.empty()) {
3013 const SCEV *S = Worklist.pop_back_val();
3015 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3016 Worklist.append(N->op_begin(), N->op_end());
3017 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3018 Worklist.push_back(C->getOperand());
3019 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3020 Worklist.push_back(D->getLHS());
3021 Worklist.push_back(D->getRHS());
3022 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3023 if (!Inserted.insert(US)) continue;
3024 const Value *V = US->getValue();
3025 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3026 // Look for instructions defined outside the loop.
3027 if (L->contains(Inst)) continue;
3028 } else if (isa<UndefValue>(V))
3029 // Undef doesn't have a live range, so it doesn't matter.
3031 for (const Use &U : V->uses()) {
3032 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3033 // Ignore non-instructions.
3036 // Ignore instructions in other functions (as can happen with
3038 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3040 // Ignore instructions not dominated by the loop.
3041 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3042 UserInst->getParent() :
3043 cast<PHINode>(UserInst)->getIncomingBlock(
3044 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3045 if (!DT.dominates(L->getHeader(), UseBB))
3047 // Ignore uses which are part of other SCEV expressions, to avoid
3048 // analyzing them multiple times.
3049 if (SE.isSCEVable(UserInst->getType())) {
3050 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3051 // If the user is a no-op, look through to its uses.
3052 if (!isa<SCEVUnknown>(UserS))
3056 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3060 // Ignore icmp instructions which are already being analyzed.
3061 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3062 unsigned OtherIdx = !U.getOperandNo();
3063 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3064 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3068 LSRFixup &LF = getNewFixup();
3069 LF.UserInst = const_cast<Instruction *>(UserInst);
3070 LF.OperandValToReplace = U;
3071 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
3073 LF.Offset = P.second;
3074 LSRUse &LU = Uses[LF.LUIdx];
3075 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3076 if (!LU.WidestFixupType ||
3077 SE.getTypeSizeInBits(LU.WidestFixupType) <
3078 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3079 LU.WidestFixupType = LF.OperandValToReplace->getType();
3080 InsertSupplementalFormula(US, LU, LF.LUIdx);
3081 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3088 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3089 /// separate registers. If C is non-null, multiply each subexpression by C.
3091 /// Return remainder expression after factoring the subexpressions captured by
3092 /// Ops. If Ops is complete, return NULL.
3093 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3094 SmallVectorImpl<const SCEV *> &Ops,
3096 ScalarEvolution &SE,
3097 unsigned Depth = 0) {
3098 // Arbitrarily cap recursion to protect compile time.
3102 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3103 // Break out add operands.
3104 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3106 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3108 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3111 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3112 // Split a non-zero base out of an addrec.
3113 if (AR->getStart()->isZero())
3116 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3117 C, Ops, L, SE, Depth+1);
3118 // Split the non-zero AddRec unless it is part of a nested recurrence that
3119 // does not pertain to this loop.
3120 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3121 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3124 if (Remainder != AR->getStart()) {
3126 Remainder = SE.getConstant(AR->getType(), 0);
3127 return SE.getAddRecExpr(Remainder,
3128 AR->getStepRecurrence(SE),
3130 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3133 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3134 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3135 if (Mul->getNumOperands() != 2)
3137 if (const SCEVConstant *Op0 =
3138 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3139 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3140 const SCEV *Remainder =
3141 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3143 Ops.push_back(SE.getMulExpr(C, Remainder));
3150 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3152 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3155 // Arbitrarily cap recursion to protect compile time.
3156 if (Depth >= 3) return;
3158 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3159 const SCEV *BaseReg = Base.BaseRegs[i];
3161 SmallVector<const SCEV *, 8> AddOps;
3162 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3164 AddOps.push_back(Remainder);
3166 if (AddOps.size() == 1) continue;
3168 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3169 JE = AddOps.end(); J != JE; ++J) {
3171 // Loop-variant "unknown" values are uninteresting; we won't be able to
3172 // do anything meaningful with them.
3173 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3176 // Don't pull a constant into a register if the constant could be folded
3177 // into an immediate field.
3178 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3179 LU.AccessTy, *J, Base.getNumRegs() > 1))
3182 // Collect all operands except *J.
3183 SmallVector<const SCEV *, 8> InnerAddOps(
3184 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3185 InnerAddOps.append(std::next(J),
3186 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3188 // Don't leave just a constant behind in a register if the constant could
3189 // be folded into an immediate field.
3190 if (InnerAddOps.size() == 1 &&
3191 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3192 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3195 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3196 if (InnerSum->isZero())
3200 // Add the remaining pieces of the add back into the new formula.
3201 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3203 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3204 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3205 InnerSumSC->getValue()->getZExtValue())) {
3206 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3207 InnerSumSC->getValue()->getZExtValue();
3208 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3210 F.BaseRegs[i] = InnerSum;
3212 // Add J as its own register, or an unfolded immediate.
3213 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3214 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3215 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3216 SC->getValue()->getZExtValue()))
3217 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3218 SC->getValue()->getZExtValue();
3220 F.BaseRegs.push_back(*J);
3222 if (InsertFormula(LU, LUIdx, F))
3223 // If that formula hadn't been seen before, recurse to find more like
3225 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3230 /// GenerateCombinations - Generate a formula consisting of all of the
3231 /// loop-dominating registers added into a single register.
3232 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3234 // This method is only interesting on a plurality of registers.
3235 if (Base.BaseRegs.size() <= 1) return;
3239 SmallVector<const SCEV *, 4> Ops;
3240 for (SmallVectorImpl<const SCEV *>::const_iterator
3241 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3242 const SCEV *BaseReg = *I;
3243 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3244 !SE.hasComputableLoopEvolution(BaseReg, L))
3245 Ops.push_back(BaseReg);
3247 F.BaseRegs.push_back(BaseReg);
3249 if (Ops.size() > 1) {
3250 const SCEV *Sum = SE.getAddExpr(Ops);
3251 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3252 // opportunity to fold something. For now, just ignore such cases
3253 // rather than proceed with zero in a register.
3254 if (!Sum->isZero()) {
3255 F.BaseRegs.push_back(Sum);
3256 (void)InsertFormula(LU, LUIdx, F);
3261 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3262 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3264 // We can't add a symbolic offset if the address already contains one.
3265 if (Base.BaseGV) return;
3267 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3268 const SCEV *G = Base.BaseRegs[i];
3269 GlobalValue *GV = ExtractSymbol(G, SE);
3270 if (G->isZero() || !GV)
3274 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3277 (void)InsertFormula(LU, LUIdx, F);
3281 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3282 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3284 // TODO: For now, just add the min and max offset, because it usually isn't
3285 // worthwhile looking at everything inbetween.
3286 SmallVector<int64_t, 2> Worklist;
3287 Worklist.push_back(LU.MinOffset);
3288 if (LU.MaxOffset != LU.MinOffset)
3289 Worklist.push_back(LU.MaxOffset);
3291 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3292 const SCEV *G = Base.BaseRegs[i];
3294 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3295 E = Worklist.end(); I != E; ++I) {
3297 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3298 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3300 // Add the offset to the base register.
3301 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3302 // If it cancelled out, drop the base register, otherwise update it.
3303 if (NewG->isZero()) {
3304 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3305 F.BaseRegs.pop_back();
3307 F.BaseRegs[i] = NewG;
3309 (void)InsertFormula(LU, LUIdx, F);
3313 int64_t Imm = ExtractImmediate(G, SE);
3314 if (G->isZero() || Imm == 0)
3317 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3318 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3321 (void)InsertFormula(LU, LUIdx, F);
3325 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3326 /// the comparison. For example, x == y -> x*c == y*c.
3327 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3329 if (LU.Kind != LSRUse::ICmpZero) return;
3331 // Determine the integer type for the base formula.
3332 Type *IntTy = Base.getType();
3334 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3336 // Don't do this if there is more than one offset.
3337 if (LU.MinOffset != LU.MaxOffset) return;
3339 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3341 // Check each interesting stride.
3342 for (SmallSetVector<int64_t, 8>::const_iterator
3343 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3344 int64_t Factor = *I;
3346 // Check that the multiplication doesn't overflow.
3347 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3349 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3350 if (NewBaseOffset / Factor != Base.BaseOffset)
3352 // If the offset will be truncated at this use, check that it is in bounds.
3353 if (!IntTy->isPointerTy() &&
3354 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3357 // Check that multiplying with the use offset doesn't overflow.
3358 int64_t Offset = LU.MinOffset;
3359 if (Offset == INT64_MIN && Factor == -1)
3361 Offset = (uint64_t)Offset * Factor;
3362 if (Offset / Factor != LU.MinOffset)
3364 // If the offset will be truncated at this use, check that it is in bounds.
3365 if (!IntTy->isPointerTy() &&
3366 !ConstantInt::isValueValidForType(IntTy, Offset))
3370 F.BaseOffset = NewBaseOffset;
3372 // Check that this scale is legal.
3373 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3376 // Compensate for the use having MinOffset built into it.
3377 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3379 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3381 // Check that multiplying with each base register doesn't overflow.
3382 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3383 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3384 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3388 // Check that multiplying with the scaled register doesn't overflow.
3390 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3391 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3395 // Check that multiplying with the unfolded offset doesn't overflow.
3396 if (F.UnfoldedOffset != 0) {
3397 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3399 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3400 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3402 // If the offset will be truncated, check that it is in bounds.
3403 if (!IntTy->isPointerTy() &&
3404 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3408 // If we make it here and it's legal, add it.
3409 (void)InsertFormula(LU, LUIdx, F);
3414 /// GenerateScales - Generate stride factor reuse formulae by making use of
3415 /// scaled-offset address modes, for example.
3416 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3417 // Determine the integer type for the base formula.
3418 Type *IntTy = Base.getType();
3421 // If this Formula already has a scaled register, we can't add another one.
3422 if (Base.Scale != 0) return;
3424 // Check each interesting stride.
3425 for (SmallSetVector<int64_t, 8>::const_iterator
3426 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3427 int64_t Factor = *I;
3429 Base.Scale = Factor;
3430 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3431 // Check whether this scale is going to be legal.
3432 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3434 // As a special-case, handle special out-of-loop Basic users specially.
3435 // TODO: Reconsider this special case.
3436 if (LU.Kind == LSRUse::Basic &&
3437 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3438 LU.AccessTy, Base) &&
3439 LU.AllFixupsOutsideLoop)
3440 LU.Kind = LSRUse::Special;
3444 // For an ICmpZero, negating a solitary base register won't lead to
3446 if (LU.Kind == LSRUse::ICmpZero &&
3447 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3449 // For each addrec base reg, apply the scale, if possible.
3450 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3451 if (const SCEVAddRecExpr *AR =
3452 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3453 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3454 if (FactorS->isZero())
3456 // Divide out the factor, ignoring high bits, since we'll be
3457 // scaling the value back up in the end.
3458 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3459 // TODO: This could be optimized to avoid all the copying.
3461 F.ScaledReg = Quotient;
3462 F.DeleteBaseReg(F.BaseRegs[i]);
3463 (void)InsertFormula(LU, LUIdx, F);
3469 /// GenerateTruncates - Generate reuse formulae from different IV types.
3470 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3471 // Don't bother truncating symbolic values.
3472 if (Base.BaseGV) return;
3474 // Determine the integer type for the base formula.
3475 Type *DstTy = Base.getType();
3477 DstTy = SE.getEffectiveSCEVType(DstTy);
3479 for (SmallSetVector<Type *, 4>::const_iterator
3480 I = Types.begin(), E = Types.end(); I != E; ++I) {
3482 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3485 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3486 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3487 JE = F.BaseRegs.end(); J != JE; ++J)
3488 *J = SE.getAnyExtendExpr(*J, SrcTy);
3490 // TODO: This assumes we've done basic processing on all uses and
3491 // have an idea what the register usage is.
3492 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3495 (void)InsertFormula(LU, LUIdx, F);
3502 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3503 /// defer modifications so that the search phase doesn't have to worry about
3504 /// the data structures moving underneath it.
3508 const SCEV *OrigReg;
3510 WorkItem(size_t LI, int64_t I, const SCEV *R)
3511 : LUIdx(LI), Imm(I), OrigReg(R) {}
3513 void print(raw_ostream &OS) const;
3519 void WorkItem::print(raw_ostream &OS) const {
3520 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3521 << " , add offset " << Imm;
3524 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3525 void WorkItem::dump() const {
3526 print(errs()); errs() << '\n';
3530 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3531 /// distance apart and try to form reuse opportunities between them.
3532 void LSRInstance::GenerateCrossUseConstantOffsets() {
3533 // Group the registers by their value without any added constant offset.
3534 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3535 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3537 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3538 SmallVector<const SCEV *, 8> Sequence;
3539 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3541 const SCEV *Reg = *I;
3542 int64_t Imm = ExtractImmediate(Reg, SE);
3543 std::pair<RegMapTy::iterator, bool> Pair =
3544 Map.insert(std::make_pair(Reg, ImmMapTy()));
3546 Sequence.push_back(Reg);
3547 Pair.first->second.insert(std::make_pair(Imm, *I));
3548 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3551 // Now examine each set of registers with the same base value. Build up
3552 // a list of work to do and do the work in a separate step so that we're
3553 // not adding formulae and register counts while we're searching.
3554 SmallVector<WorkItem, 32> WorkItems;
3555 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3556 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3557 E = Sequence.end(); I != E; ++I) {
3558 const SCEV *Reg = *I;
3559 const ImmMapTy &Imms = Map.find(Reg)->second;
3561 // It's not worthwhile looking for reuse if there's only one offset.
3562 if (Imms.size() == 1)
3565 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3566 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3568 dbgs() << ' ' << J->first;
3571 // Examine each offset.
3572 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3574 const SCEV *OrigReg = J->second;
3576 int64_t JImm = J->first;
3577 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3579 if (!isa<SCEVConstant>(OrigReg) &&
3580 UsedByIndicesMap[Reg].count() == 1) {
3581 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3585 // Conservatively examine offsets between this orig reg a few selected
3587 ImmMapTy::const_iterator OtherImms[] = {
3588 Imms.begin(), std::prev(Imms.end()),
3589 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3592 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3593 ImmMapTy::const_iterator M = OtherImms[i];
3594 if (M == J || M == JE) continue;
3596 // Compute the difference between the two.
3597 int64_t Imm = (uint64_t)JImm - M->first;
3598 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3599 LUIdx = UsedByIndices.find_next(LUIdx))
3600 // Make a memo of this use, offset, and register tuple.
3601 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3602 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3609 UsedByIndicesMap.clear();
3610 UniqueItems.clear();
3612 // Now iterate through the worklist and add new formulae.
3613 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3614 E = WorkItems.end(); I != E; ++I) {
3615 const WorkItem &WI = *I;
3616 size_t LUIdx = WI.LUIdx;
3617 LSRUse &LU = Uses[LUIdx];
3618 int64_t Imm = WI.Imm;
3619 const SCEV *OrigReg = WI.OrigReg;
3621 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3622 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3623 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3625 // TODO: Use a more targeted data structure.
3626 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3627 const Formula &F = LU.Formulae[L];
3628 // Use the immediate in the scaled register.
3629 if (F.ScaledReg == OrigReg) {
3630 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3631 // Don't create 50 + reg(-50).
3632 if (F.referencesReg(SE.getSCEV(
3633 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3636 NewF.BaseOffset = Offset;
3637 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3640 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3642 // If the new scale is a constant in a register, and adding the constant
3643 // value to the immediate would produce a value closer to zero than the
3644 // immediate itself, then the formula isn't worthwhile.
3645 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3646 if (C->getValue()->isNegative() !=
3647 (NewF.BaseOffset < 0) &&
3648 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3649 .ule(abs64(NewF.BaseOffset)))
3653 (void)InsertFormula(LU, LUIdx, NewF);
3655 // Use the immediate in a base register.
3656 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3657 const SCEV *BaseReg = F.BaseRegs[N];
3658 if (BaseReg != OrigReg)
3661 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3662 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3663 LU.Kind, LU.AccessTy, NewF)) {
3664 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3667 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3669 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3671 // If the new formula has a constant in a register, and adding the
3672 // constant value to the immediate would produce a value closer to
3673 // zero than the immediate itself, then the formula isn't worthwhile.
3674 for (SmallVectorImpl<const SCEV *>::const_iterator
3675 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3677 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3678 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3679 abs64(NewF.BaseOffset)) &&
3680 (C->getValue()->getValue() +
3681 NewF.BaseOffset).countTrailingZeros() >=
3682 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3686 (void)InsertFormula(LU, LUIdx, NewF);
3695 /// GenerateAllReuseFormulae - Generate formulae for each use.
3697 LSRInstance::GenerateAllReuseFormulae() {
3698 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3699 // queries are more precise.
3700 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3701 LSRUse &LU = Uses[LUIdx];
3702 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3703 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3704 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3705 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3707 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3708 LSRUse &LU = Uses[LUIdx];
3709 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3710 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3711 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3712 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3713 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3714 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3715 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3716 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3718 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3719 LSRUse &LU = Uses[LUIdx];
3720 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3721 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3724 GenerateCrossUseConstantOffsets();
3726 DEBUG(dbgs() << "\n"
3727 "After generating reuse formulae:\n";
3728 print_uses(dbgs()));
3731 /// If there are multiple formulae with the same set of registers used
3732 /// by other uses, pick the best one and delete the others.
3733 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3734 DenseSet<const SCEV *> VisitedRegs;
3735 SmallPtrSet<const SCEV *, 16> Regs;
3736 SmallPtrSet<const SCEV *, 16> LoserRegs;
3738 bool ChangedFormulae = false;
3741 // Collect the best formula for each unique set of shared registers. This
3742 // is reset for each use.
3743 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3745 BestFormulaeTy BestFormulae;
3747 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3748 LSRUse &LU = Uses[LUIdx];
3749 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3752 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3753 FIdx != NumForms; ++FIdx) {
3754 Formula &F = LU.Formulae[FIdx];
3756 // Some formulas are instant losers. For example, they may depend on
3757 // nonexistent AddRecs from other loops. These need to be filtered
3758 // immediately, otherwise heuristics could choose them over others leading
3759 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3760 // avoids the need to recompute this information across formulae using the
3761 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3762 // the corresponding bad register from the Regs set.
3765 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3767 if (CostF.isLoser()) {
3768 // During initial formula generation, undesirable formulae are generated
3769 // by uses within other loops that have some non-trivial address mode or
3770 // use the postinc form of the IV. LSR needs to provide these formulae
3771 // as the basis of rediscovering the desired formula that uses an AddRec
3772 // corresponding to the existing phi. Once all formulae have been
3773 // generated, these initial losers may be pruned.
3774 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3778 SmallVector<const SCEV *, 4> Key;
3779 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3780 JE = F.BaseRegs.end(); J != JE; ++J) {
3781 const SCEV *Reg = *J;
3782 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3786 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3787 Key.push_back(F.ScaledReg);
3788 // Unstable sort by host order ok, because this is only used for
3790 std::sort(Key.begin(), Key.end());
3792 std::pair<BestFormulaeTy::const_iterator, bool> P =
3793 BestFormulae.insert(std::make_pair(Key, FIdx));
3797 Formula &Best = LU.Formulae[P.first->second];
3801 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3803 if (CostF < CostBest)
3805 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3807 " in favor of formula "; Best.print(dbgs());
3811 ChangedFormulae = true;
3813 LU.DeleteFormula(F);
3819 // Now that we've filtered out some formulae, recompute the Regs set.
3821 LU.RecomputeRegs(LUIdx, RegUses);
3823 // Reset this to prepare for the next use.
3824 BestFormulae.clear();
3827 DEBUG(if (ChangedFormulae) {
3829 "After filtering out undesirable candidates:\n";
3834 // This is a rough guess that seems to work fairly well.
3835 static const size_t ComplexityLimit = UINT16_MAX;
3837 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3838 /// solutions the solver might have to consider. It almost never considers
3839 /// this many solutions because it prune the search space, but the pruning
3840 /// isn't always sufficient.
3841 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3843 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3844 E = Uses.end(); I != E; ++I) {
3845 size_t FSize = I->Formulae.size();
3846 if (FSize >= ComplexityLimit) {
3847 Power = ComplexityLimit;
3851 if (Power >= ComplexityLimit)
3857 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3858 /// of the registers of another formula, it won't help reduce register
3859 /// pressure (though it may not necessarily hurt register pressure); remove
3860 /// it to simplify the system.
3861 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3862 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3863 DEBUG(dbgs() << "The search space is too complex.\n");
3865 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3866 "which use a superset of registers used by other "
3869 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3870 LSRUse &LU = Uses[LUIdx];
3872 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3873 Formula &F = LU.Formulae[i];
3874 // Look for a formula with a constant or GV in a register. If the use
3875 // also has a formula with that same value in an immediate field,
3876 // delete the one that uses a register.
3877 for (SmallVectorImpl<const SCEV *>::const_iterator
3878 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3879 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3881 NewF.BaseOffset += C->getValue()->getSExtValue();
3882 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3883 (I - F.BaseRegs.begin()));
3884 if (LU.HasFormulaWithSameRegs(NewF)) {
3885 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3886 LU.DeleteFormula(F);
3892 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3893 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3897 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3898 (I - F.BaseRegs.begin()));
3899 if (LU.HasFormulaWithSameRegs(NewF)) {
3900 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3902 LU.DeleteFormula(F);
3913 LU.RecomputeRegs(LUIdx, RegUses);
3916 DEBUG(dbgs() << "After pre-selection:\n";
3917 print_uses(dbgs()));
3921 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3922 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3924 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3925 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
3928 DEBUG(dbgs() << "The search space is too complex.\n"
3929 "Narrowing the search space by assuming that uses separated "
3930 "by a constant offset will use the same registers.\n");
3932 // This is especially useful for unrolled loops.
3934 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3935 LSRUse &LU = Uses[LUIdx];
3936 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3937 E = LU.Formulae.end(); I != E; ++I) {
3938 const Formula &F = *I;
3939 if (F.BaseOffset == 0 || F.Scale != 0)
3942 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
3946 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
3947 LU.Kind, LU.AccessTy))
3950 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
3952 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3954 // Update the relocs to reference the new use.
3955 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3956 E = Fixups.end(); I != E; ++I) {
3957 LSRFixup &Fixup = *I;
3958 if (Fixup.LUIdx == LUIdx) {
3959 Fixup.LUIdx = LUThatHas - &Uses.front();
3960 Fixup.Offset += F.BaseOffset;
3961 // Add the new offset to LUThatHas' offset list.
3962 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3963 LUThatHas->Offsets.push_back(Fixup.Offset);
3964 if (Fixup.Offset > LUThatHas->MaxOffset)
3965 LUThatHas->MaxOffset = Fixup.Offset;
3966 if (Fixup.Offset < LUThatHas->MinOffset)
3967 LUThatHas->MinOffset = Fixup.Offset;
3969 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
3971 if (Fixup.LUIdx == NumUses-1)
3972 Fixup.LUIdx = LUIdx;
3975 // Delete formulae from the new use which are no longer legal.
3977 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3978 Formula &F = LUThatHas->Formulae[i];
3979 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
3980 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
3981 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3983 LUThatHas->DeleteFormula(F);
3991 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3993 // Delete the old use.
3994 DeleteUse(LU, LUIdx);
4001 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4004 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4005 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4006 /// we've done more filtering, as it may be able to find more formulae to
4008 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4009 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4010 DEBUG(dbgs() << "The search space is too complex.\n");
4012 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4013 "undesirable dedicated registers.\n");
4015 FilterOutUndesirableDedicatedRegisters();
4017 DEBUG(dbgs() << "After pre-selection:\n";
4018 print_uses(dbgs()));
4022 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4023 /// to be profitable, and then in any use which has any reference to that
4024 /// register, delete all formulae which do not reference that register.
4025 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4026 // With all other options exhausted, loop until the system is simple
4027 // enough to handle.
4028 SmallPtrSet<const SCEV *, 4> Taken;
4029 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4030 // Ok, we have too many of formulae on our hands to conveniently handle.
4031 // Use a rough heuristic to thin out the list.
4032 DEBUG(dbgs() << "The search space is too complex.\n");
4034 // Pick the register which is used by the most LSRUses, which is likely
4035 // to be a good reuse register candidate.
4036 const SCEV *Best = 0;
4037 unsigned BestNum = 0;
4038 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4040 const SCEV *Reg = *I;
4041 if (Taken.count(Reg))
4046 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4047 if (Count > BestNum) {
4054 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4055 << " will yield profitable reuse.\n");
4058 // In any use with formulae which references this register, delete formulae
4059 // which don't reference it.
4060 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4061 LSRUse &LU = Uses[LUIdx];
4062 if (!LU.Regs.count(Best)) continue;
4065 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4066 Formula &F = LU.Formulae[i];
4067 if (!F.referencesReg(Best)) {
4068 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4069 LU.DeleteFormula(F);
4073 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4079 LU.RecomputeRegs(LUIdx, RegUses);
4082 DEBUG(dbgs() << "After pre-selection:\n";
4083 print_uses(dbgs()));
4087 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4088 /// formulae to choose from, use some rough heuristics to prune down the number
4089 /// of formulae. This keeps the main solver from taking an extraordinary amount
4090 /// of time in some worst-case scenarios.
4091 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4092 NarrowSearchSpaceByDetectingSupersets();
4093 NarrowSearchSpaceByCollapsingUnrolledCode();
4094 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4095 NarrowSearchSpaceByPickingWinnerRegs();
4098 /// SolveRecurse - This is the recursive solver.
4099 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4101 SmallVectorImpl<const Formula *> &Workspace,
4102 const Cost &CurCost,
4103 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4104 DenseSet<const SCEV *> &VisitedRegs) const {
4107 // - use more aggressive filtering
4108 // - sort the formula so that the most profitable solutions are found first
4109 // - sort the uses too
4111 // - don't compute a cost, and then compare. compare while computing a cost
4113 // - track register sets with SmallBitVector
4115 const LSRUse &LU = Uses[Workspace.size()];
4117 // If this use references any register that's already a part of the
4118 // in-progress solution, consider it a requirement that a formula must
4119 // reference that register in order to be considered. This prunes out
4120 // unprofitable searching.
4121 SmallSetVector<const SCEV *, 4> ReqRegs;
4122 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4123 E = CurRegs.end(); I != E; ++I)
4124 if (LU.Regs.count(*I))
4127 SmallPtrSet<const SCEV *, 16> NewRegs;
4129 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4130 E = LU.Formulae.end(); I != E; ++I) {
4131 const Formula &F = *I;
4133 // Ignore formulae which do not use any of the required registers.
4134 bool SatisfiedReqReg = true;
4135 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4136 JE = ReqRegs.end(); J != JE; ++J) {
4137 const SCEV *Reg = *J;
4138 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4139 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4141 SatisfiedReqReg = false;
4145 if (!SatisfiedReqReg) {
4146 // If none of the formulae satisfied the required registers, then we could
4147 // clear ReqRegs and try again. Currently, we simply give up in this case.
4151 // Evaluate the cost of the current formula. If it's already worse than
4152 // the current best, prune the search at that point.
4155 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4157 if (NewCost < SolutionCost) {
4158 Workspace.push_back(&F);
4159 if (Workspace.size() != Uses.size()) {
4160 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4161 NewRegs, VisitedRegs);
4162 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4163 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4165 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4166 dbgs() << ".\n Regs:";
4167 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4168 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4169 dbgs() << ' ' << **I;
4172 SolutionCost = NewCost;
4173 Solution = Workspace;
4175 Workspace.pop_back();
4180 /// Solve - Choose one formula from each use. Return the results in the given
4181 /// Solution vector.
4182 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4183 SmallVector<const Formula *, 8> Workspace;
4185 SolutionCost.Lose();
4187 SmallPtrSet<const SCEV *, 16> CurRegs;
4188 DenseSet<const SCEV *> VisitedRegs;
4189 Workspace.reserve(Uses.size());
4191 // SolveRecurse does all the work.
4192 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4193 CurRegs, VisitedRegs);
4194 if (Solution.empty()) {
4195 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4199 // Ok, we've now made all our decisions.
4200 DEBUG(dbgs() << "\n"
4201 "The chosen solution requires "; SolutionCost.print(dbgs());
4203 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4205 Uses[i].print(dbgs());
4208 Solution[i]->print(dbgs());
4212 assert(Solution.size() == Uses.size() && "Malformed solution!");
4215 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4216 /// the dominator tree far as we can go while still being dominated by the
4217 /// input positions. This helps canonicalize the insert position, which
4218 /// encourages sharing.
4219 BasicBlock::iterator
4220 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4221 const SmallVectorImpl<Instruction *> &Inputs)
4224 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4225 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4228 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4229 if (!Rung) return IP;
4230 Rung = Rung->getIDom();
4231 if (!Rung) return IP;
4232 IDom = Rung->getBlock();
4234 // Don't climb into a loop though.
4235 const Loop *IDomLoop = LI.getLoopFor(IDom);
4236 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4237 if (IDomDepth <= IPLoopDepth &&
4238 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4242 bool AllDominate = true;
4243 Instruction *BetterPos = 0;
4244 Instruction *Tentative = IDom->getTerminator();
4245 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4246 E = Inputs.end(); I != E; ++I) {
4247 Instruction *Inst = *I;
4248 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4249 AllDominate = false;
4252 // Attempt to find an insert position in the middle of the block,
4253 // instead of at the end, so that it can be used for other expansions.
4254 if (IDom == Inst->getParent() &&
4255 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4256 BetterPos = std::next(BasicBlock::iterator(Inst));
4269 /// AdjustInsertPositionForExpand - Determine an input position which will be
4270 /// dominated by the operands and which will dominate the result.
4271 BasicBlock::iterator
4272 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4275 SCEVExpander &Rewriter) const {
4276 // Collect some instructions which must be dominated by the
4277 // expanding replacement. These must be dominated by any operands that
4278 // will be required in the expansion.
4279 SmallVector<Instruction *, 4> Inputs;
4280 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4281 Inputs.push_back(I);
4282 if (LU.Kind == LSRUse::ICmpZero)
4283 if (Instruction *I =
4284 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4285 Inputs.push_back(I);
4286 if (LF.PostIncLoops.count(L)) {
4287 if (LF.isUseFullyOutsideLoop(L))
4288 Inputs.push_back(L->getLoopLatch()->getTerminator());
4290 Inputs.push_back(IVIncInsertPos);
4292 // The expansion must also be dominated by the increment positions of any
4293 // loops it for which it is using post-inc mode.
4294 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4295 E = LF.PostIncLoops.end(); I != E; ++I) {
4296 const Loop *PIL = *I;
4297 if (PIL == L) continue;
4299 // Be dominated by the loop exit.
4300 SmallVector<BasicBlock *, 4> ExitingBlocks;
4301 PIL->getExitingBlocks(ExitingBlocks);
4302 if (!ExitingBlocks.empty()) {
4303 BasicBlock *BB = ExitingBlocks[0];
4304 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4305 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4306 Inputs.push_back(BB->getTerminator());
4310 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4311 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4312 "Insertion point must be a normal instruction");
4314 // Then, climb up the immediate dominator tree as far as we can go while
4315 // still being dominated by the input positions.
4316 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4318 // Don't insert instructions before PHI nodes.
4319 while (isa<PHINode>(IP)) ++IP;
4321 // Ignore landingpad instructions.
4322 while (isa<LandingPadInst>(IP)) ++IP;
4324 // Ignore debug intrinsics.
4325 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4327 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4328 // IP consistent across expansions and allows the previously inserted
4329 // instructions to be reused by subsequent expansion.
4330 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4335 /// Expand - Emit instructions for the leading candidate expression for this
4336 /// LSRUse (this is called "expanding").
4337 Value *LSRInstance::Expand(const LSRFixup &LF,
4339 BasicBlock::iterator IP,
4340 SCEVExpander &Rewriter,
4341 SmallVectorImpl<WeakVH> &DeadInsts) const {
4342 const LSRUse &LU = Uses[LF.LUIdx];
4343 if (LU.RigidFormula)
4344 return LF.OperandValToReplace;
4346 // Determine an input position which will be dominated by the operands and
4347 // which will dominate the result.
4348 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4350 // Inform the Rewriter if we have a post-increment use, so that it can
4351 // perform an advantageous expansion.
4352 Rewriter.setPostInc(LF.PostIncLoops);
4354 // This is the type that the user actually needs.
4355 Type *OpTy = LF.OperandValToReplace->getType();
4356 // This will be the type that we'll initially expand to.
4357 Type *Ty = F.getType();
4359 // No type known; just expand directly to the ultimate type.
4361 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4362 // Expand directly to the ultimate type if it's the right size.
4364 // This is the type to do integer arithmetic in.
4365 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4367 // Build up a list of operands to add together to form the full base.
4368 SmallVector<const SCEV *, 8> Ops;
4370 // Expand the BaseRegs portion.
4371 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4372 E = F.BaseRegs.end(); I != E; ++I) {
4373 const SCEV *Reg = *I;
4374 assert(!Reg->isZero() && "Zero allocated in a base register!");
4376 // If we're expanding for a post-inc user, make the post-inc adjustment.
4377 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4378 Reg = TransformForPostIncUse(Denormalize, Reg,
4379 LF.UserInst, LF.OperandValToReplace,
4382 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4385 // Expand the ScaledReg portion.
4386 Value *ICmpScaledV = 0;
4388 const SCEV *ScaledS = F.ScaledReg;
4390 // If we're expanding for a post-inc user, make the post-inc adjustment.
4391 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4392 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4393 LF.UserInst, LF.OperandValToReplace,
4396 if (LU.Kind == LSRUse::ICmpZero) {
4397 // An interesting way of "folding" with an icmp is to use a negated
4398 // scale, which we'll implement by inserting it into the other operand
4400 assert(F.Scale == -1 &&
4401 "The only scale supported by ICmpZero uses is -1!");
4402 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4404 // Otherwise just expand the scaled register and an explicit scale,
4405 // which is expected to be matched as part of the address.
4407 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4408 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4409 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4411 Ops.push_back(SE.getUnknown(FullV));
4413 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4414 ScaledS = SE.getMulExpr(ScaledS,
4415 SE.getConstant(ScaledS->getType(), F.Scale));
4416 Ops.push_back(ScaledS);
4420 // Expand the GV portion.
4422 // Flush the operand list to suppress SCEVExpander hoisting.
4424 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4426 Ops.push_back(SE.getUnknown(FullV));
4428 Ops.push_back(SE.getUnknown(F.BaseGV));
4431 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4432 // unfolded offsets. LSR assumes they both live next to their uses.
4434 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4436 Ops.push_back(SE.getUnknown(FullV));
4439 // Expand the immediate portion.
4440 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4442 if (LU.Kind == LSRUse::ICmpZero) {
4443 // The other interesting way of "folding" with an ICmpZero is to use a
4444 // negated immediate.
4446 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4448 Ops.push_back(SE.getUnknown(ICmpScaledV));
4449 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4452 // Just add the immediate values. These again are expected to be matched
4453 // as part of the address.
4454 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4458 // Expand the unfolded offset portion.
4459 int64_t UnfoldedOffset = F.UnfoldedOffset;
4460 if (UnfoldedOffset != 0) {
4461 // Just add the immediate values.
4462 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4466 // Emit instructions summing all the operands.
4467 const SCEV *FullS = Ops.empty() ?
4468 SE.getConstant(IntTy, 0) :
4470 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4472 // We're done expanding now, so reset the rewriter.
4473 Rewriter.clearPostInc();
4475 // An ICmpZero Formula represents an ICmp which we're handling as a
4476 // comparison against zero. Now that we've expanded an expression for that
4477 // form, update the ICmp's other operand.
4478 if (LU.Kind == LSRUse::ICmpZero) {
4479 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4480 DeadInsts.push_back(CI->getOperand(1));
4481 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4482 "a scale at the same time!");
4483 if (F.Scale == -1) {
4484 if (ICmpScaledV->getType() != OpTy) {
4486 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4488 ICmpScaledV, OpTy, "tmp", CI);
4491 CI->setOperand(1, ICmpScaledV);
4493 assert(F.Scale == 0 &&
4494 "ICmp does not support folding a global value and "
4495 "a scale at the same time!");
4496 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4498 if (C->getType() != OpTy)
4499 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4503 CI->setOperand(1, C);
4510 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4511 /// of their operands effectively happens in their predecessor blocks, so the
4512 /// expression may need to be expanded in multiple places.
4513 void LSRInstance::RewriteForPHI(PHINode *PN,
4516 SCEVExpander &Rewriter,
4517 SmallVectorImpl<WeakVH> &DeadInsts,
4519 DenseMap<BasicBlock *, Value *> Inserted;
4520 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4521 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4522 BasicBlock *BB = PN->getIncomingBlock(i);
4524 // If this is a critical edge, split the edge so that we do not insert
4525 // the code on all predecessor/successor paths. We do this unless this
4526 // is the canonical backedge for this loop, which complicates post-inc
4528 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4529 !isa<IndirectBrInst>(BB->getTerminator())) {
4530 BasicBlock *Parent = PN->getParent();
4531 Loop *PNLoop = LI.getLoopFor(Parent);
4532 if (!PNLoop || Parent != PNLoop->getHeader()) {
4533 // Split the critical edge.
4534 BasicBlock *NewBB = 0;
4535 if (!Parent->isLandingPad()) {
4536 NewBB = SplitCriticalEdge(BB, Parent, P,
4537 /*MergeIdenticalEdges=*/true,
4538 /*DontDeleteUselessPhis=*/true);
4540 SmallVector<BasicBlock*, 2> NewBBs;
4541 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4544 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4545 // phi predecessors are identical. The simple thing to do is skip
4546 // splitting in this case rather than complicate the API.
4548 // If PN is outside of the loop and BB is in the loop, we want to
4549 // move the block to be immediately before the PHI block, not
4550 // immediately after BB.
4551 if (L->contains(BB) && !L->contains(PN))
4552 NewBB->moveBefore(PN->getParent());
4554 // Splitting the edge can reduce the number of PHI entries we have.
4555 e = PN->getNumIncomingValues();
4557 i = PN->getBasicBlockIndex(BB);
4562 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4563 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4565 PN->setIncomingValue(i, Pair.first->second);
4567 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4569 // If this is reuse-by-noop-cast, insert the noop cast.
4570 Type *OpTy = LF.OperandValToReplace->getType();
4571 if (FullV->getType() != OpTy)
4573 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4575 FullV, LF.OperandValToReplace->getType(),
4576 "tmp", BB->getTerminator());
4578 PN->setIncomingValue(i, FullV);
4579 Pair.first->second = FullV;
4584 /// Rewrite - Emit instructions for the leading candidate expression for this
4585 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4586 /// the newly expanded value.
4587 void LSRInstance::Rewrite(const LSRFixup &LF,
4589 SCEVExpander &Rewriter,
4590 SmallVectorImpl<WeakVH> &DeadInsts,
4592 // First, find an insertion point that dominates UserInst. For PHI nodes,
4593 // find the nearest block which dominates all the relevant uses.
4594 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4595 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4597 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4599 // If this is reuse-by-noop-cast, insert the noop cast.
4600 Type *OpTy = LF.OperandValToReplace->getType();
4601 if (FullV->getType() != OpTy) {
4603 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4604 FullV, OpTy, "tmp", LF.UserInst);
4608 // Update the user. ICmpZero is handled specially here (for now) because
4609 // Expand may have updated one of the operands of the icmp already, and
4610 // its new value may happen to be equal to LF.OperandValToReplace, in
4611 // which case doing replaceUsesOfWith leads to replacing both operands
4612 // with the same value. TODO: Reorganize this.
4613 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4614 LF.UserInst->setOperand(0, FullV);
4616 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4619 DeadInsts.push_back(LF.OperandValToReplace);
4622 /// ImplementSolution - Rewrite all the fixup locations with new values,
4623 /// following the chosen solution.
4625 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4627 // Keep track of instructions we may have made dead, so that
4628 // we can remove them after we are done working.
4629 SmallVector<WeakVH, 16> DeadInsts;
4631 SCEVExpander Rewriter(SE, "lsr");
4633 Rewriter.setDebugType(DEBUG_TYPE);
4635 Rewriter.disableCanonicalMode();
4636 Rewriter.enableLSRMode();
4637 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4639 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4640 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4641 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4642 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4643 Rewriter.setChainedPhi(PN);
4646 // Expand the new value definitions and update the users.
4647 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4648 E = Fixups.end(); I != E; ++I) {
4649 const LSRFixup &Fixup = *I;
4651 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4656 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4657 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4658 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4661 // Clean up after ourselves. This must be done before deleting any
4665 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4668 LSRInstance::LSRInstance(Loop *L, Pass *P)
4669 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4670 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
4671 LI(P->getAnalysis<LoopInfo>()),
4672 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4674 // If LoopSimplify form is not available, stay out of trouble.
4675 if (!L->isLoopSimplifyForm())
4678 // If there's no interesting work to be done, bail early.
4679 if (IU.empty()) return;
4681 // If there's too much analysis to be done, bail early. We won't be able to
4682 // model the problem anyway.
4683 unsigned NumUsers = 0;
4684 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4685 if (++NumUsers > MaxIVUsers) {
4686 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4693 // All dominating loops must have preheaders, or SCEVExpander may not be able
4694 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4696 // IVUsers analysis should only create users that are dominated by simple loop
4697 // headers. Since this loop should dominate all of its users, its user list
4698 // should be empty if this loop itself is not within a simple loop nest.
4699 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4700 Rung; Rung = Rung->getIDom()) {
4701 BasicBlock *BB = Rung->getBlock();
4702 const Loop *DomLoop = LI.getLoopFor(BB);
4703 if (DomLoop && DomLoop->getHeader() == BB) {
4704 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4709 DEBUG(dbgs() << "\nLSR on loop ";
4710 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
4713 // First, perform some low-level loop optimizations.
4715 OptimizeLoopTermCond();
4717 // If loop preparation eliminates all interesting IV users, bail.
4718 if (IU.empty()) return;
4720 // Skip nested loops until we can model them better with formulae.
4722 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4726 // Start collecting data and preparing for the solver.
4728 CollectInterestingTypesAndFactors();
4729 CollectFixupsAndInitialFormulae();
4730 CollectLoopInvariantFixupsAndFormulae();
4732 assert(!Uses.empty() && "IVUsers reported at least one use");
4733 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4734 print_uses(dbgs()));
4736 // Now use the reuse data to generate a bunch of interesting ways
4737 // to formulate the values needed for the uses.
4738 GenerateAllReuseFormulae();
4740 FilterOutUndesirableDedicatedRegisters();
4741 NarrowSearchSpaceUsingHeuristics();
4743 SmallVector<const Formula *, 8> Solution;
4746 // Release memory that is no longer needed.
4751 if (Solution.empty())
4755 // Formulae should be legal.
4756 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4758 const LSRUse &LU = *I;
4759 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4760 JE = LU.Formulae.end();
4762 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4763 *J) && "Illegal formula generated!");
4767 // Now that we've decided what we want, make it so.
4768 ImplementSolution(Solution, P);
4771 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4772 if (Factors.empty() && Types.empty()) return;
4774 OS << "LSR has identified the following interesting factors and types: ";
4777 for (SmallSetVector<int64_t, 8>::const_iterator
4778 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4779 if (!First) OS << ", ";
4784 for (SmallSetVector<Type *, 4>::const_iterator
4785 I = Types.begin(), E = Types.end(); I != E; ++I) {
4786 if (!First) OS << ", ";
4788 OS << '(' << **I << ')';
4793 void LSRInstance::print_fixups(raw_ostream &OS) const {
4794 OS << "LSR is examining the following fixup sites:\n";
4795 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4796 E = Fixups.end(); I != E; ++I) {
4803 void LSRInstance::print_uses(raw_ostream &OS) const {
4804 OS << "LSR is examining the following uses:\n";
4805 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4806 E = Uses.end(); I != E; ++I) {
4807 const LSRUse &LU = *I;
4811 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4812 JE = LU.Formulae.end(); J != JE; ++J) {
4820 void LSRInstance::print(raw_ostream &OS) const {
4821 print_factors_and_types(OS);
4826 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4827 void LSRInstance::dump() const {
4828 print(errs()); errs() << '\n';
4834 class LoopStrengthReduce : public LoopPass {
4836 static char ID; // Pass ID, replacement for typeid
4837 LoopStrengthReduce();
4840 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
4841 void getAnalysisUsage(AnalysisUsage &AU) const override;
4846 char LoopStrengthReduce::ID = 0;
4847 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4848 "Loop Strength Reduction", false, false)
4849 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4850 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4851 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4852 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4853 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4854 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4855 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4856 "Loop Strength Reduction", false, false)
4859 Pass *llvm::createLoopStrengthReducePass() {
4860 return new LoopStrengthReduce();
4863 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4864 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4867 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4868 // We split critical edges, so we change the CFG. However, we do update
4869 // many analyses if they are around.
4870 AU.addPreservedID(LoopSimplifyID);
4872 AU.addRequired<LoopInfo>();
4873 AU.addPreserved<LoopInfo>();
4874 AU.addRequiredID(LoopSimplifyID);
4875 AU.addRequired<DominatorTreeWrapperPass>();
4876 AU.addPreserved<DominatorTreeWrapperPass>();
4877 AU.addRequired<ScalarEvolution>();
4878 AU.addPreserved<ScalarEvolution>();
4879 // Requiring LoopSimplify a second time here prevents IVUsers from running
4880 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4881 AU.addRequiredID(LoopSimplifyID);
4882 AU.addRequired<IVUsers>();
4883 AU.addPreserved<IVUsers>();
4884 AU.addRequired<TargetTransformInfo>();
4887 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4888 if (skipOptnoneFunction(L))
4891 bool Changed = false;
4893 // Run the main LSR transformation.
4894 Changed |= LSRInstance(L, this).getChanged();
4896 // Remove any extra phis created by processing inner loops.
4897 Changed |= DeleteDeadPHIs(L->getHeader());
4898 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4899 SmallVector<WeakVH, 16> DeadInsts;
4900 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4902 Rewriter.setDebugType(DEBUG_TYPE);
4904 unsigned numFolded = Rewriter.replaceCongruentIVs(
4905 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
4906 &getAnalysis<TargetTransformInfo>());
4909 DeleteTriviallyDeadInstructions(DeadInsts);
4910 DeleteDeadPHIs(L->getHeader());