1 //===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
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 file implements inline cost analysis.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "inline-cost"
15 #include "llvm/Analysis/InlineCost.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/CallingConv.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/GetElementPtrTypeIterator.h"
28 #include "llvm/IR/GlobalAlias.h"
29 #include "llvm/IR/InstVisitor.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/Operator.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
37 STATISTIC(NumCallsAnalyzed, "Number of call sites analyzed");
41 class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> {
42 typedef InstVisitor<CallAnalyzer, bool> Base;
43 friend class InstVisitor<CallAnalyzer, bool>;
45 // DataLayout if available, or null.
46 const DataLayout *const DL;
48 /// The TargetTransformInfo available for this compilation.
49 const TargetTransformInfo &TTI;
51 // The called function.
57 bool IsCallerRecursive;
59 bool ExposesReturnsTwice;
60 bool HasDynamicAlloca;
61 bool ContainsNoDuplicateCall;
65 /// Number of bytes allocated statically by the callee.
66 uint64_t AllocatedSize;
67 unsigned NumInstructions, NumVectorInstructions;
68 int FiftyPercentVectorBonus, TenPercentVectorBonus;
71 // While we walk the potentially-inlined instructions, we build up and
72 // maintain a mapping of simplified values specific to this callsite. The
73 // idea is to propagate any special information we have about arguments to
74 // this call through the inlinable section of the function, and account for
75 // likely simplifications post-inlining. The most important aspect we track
76 // is CFG altering simplifications -- when we prove a basic block dead, that
77 // can cause dramatic shifts in the cost of inlining a function.
78 DenseMap<Value *, Constant *> SimplifiedValues;
80 // Keep track of the values which map back (through function arguments) to
81 // allocas on the caller stack which could be simplified through SROA.
82 DenseMap<Value *, Value *> SROAArgValues;
84 // The mapping of caller Alloca values to their accumulated cost savings. If
85 // we have to disable SROA for one of the allocas, this tells us how much
86 // cost must be added.
87 DenseMap<Value *, int> SROAArgCosts;
89 // Keep track of values which map to a pointer base and constant offset.
90 DenseMap<Value *, std::pair<Value *, APInt> > ConstantOffsetPtrs;
92 // Custom simplification helper routines.
93 bool isAllocaDerivedArg(Value *V);
94 bool lookupSROAArgAndCost(Value *V, Value *&Arg,
95 DenseMap<Value *, int>::iterator &CostIt);
96 void disableSROA(DenseMap<Value *, int>::iterator CostIt);
97 void disableSROA(Value *V);
98 void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
100 bool handleSROACandidate(bool IsSROAValid,
101 DenseMap<Value *, int>::iterator CostIt,
102 int InstructionCost);
103 bool isGEPOffsetConstant(GetElementPtrInst &GEP);
104 bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset);
105 bool simplifyCallSite(Function *F, CallSite CS);
106 ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V);
108 // Custom analysis routines.
109 bool analyzeBlock(BasicBlock *BB);
111 // Disable several entry points to the visitor so we don't accidentally use
112 // them by declaring but not defining them here.
113 void visit(Module *); void visit(Module &);
114 void visit(Function *); void visit(Function &);
115 void visit(BasicBlock *); void visit(BasicBlock &);
117 // Provide base case for our instruction visit.
118 bool visitInstruction(Instruction &I);
120 // Our visit overrides.
121 bool visitAlloca(AllocaInst &I);
122 bool visitPHI(PHINode &I);
123 bool visitGetElementPtr(GetElementPtrInst &I);
124 bool visitBitCast(BitCastInst &I);
125 bool visitPtrToInt(PtrToIntInst &I);
126 bool visitIntToPtr(IntToPtrInst &I);
127 bool visitCastInst(CastInst &I);
128 bool visitUnaryInstruction(UnaryInstruction &I);
129 bool visitCmpInst(CmpInst &I);
130 bool visitSub(BinaryOperator &I);
131 bool visitBinaryOperator(BinaryOperator &I);
132 bool visitLoad(LoadInst &I);
133 bool visitStore(StoreInst &I);
134 bool visitExtractValue(ExtractValueInst &I);
135 bool visitInsertValue(InsertValueInst &I);
136 bool visitCallSite(CallSite CS);
137 bool visitReturnInst(ReturnInst &RI);
138 bool visitBranchInst(BranchInst &BI);
139 bool visitSwitchInst(SwitchInst &SI);
140 bool visitIndirectBrInst(IndirectBrInst &IBI);
141 bool visitResumeInst(ResumeInst &RI);
142 bool visitUnreachableInst(UnreachableInst &I);
145 CallAnalyzer(const DataLayout *DL, const TargetTransformInfo &TTI,
146 Function &Callee, int Threshold)
147 : DL(DL), TTI(TTI), F(Callee), Threshold(Threshold), Cost(0),
148 IsCallerRecursive(false), IsRecursiveCall(false),
149 ExposesReturnsTwice(false), HasDynamicAlloca(false),
150 ContainsNoDuplicateCall(false), HasReturn(false), HasIndirectBr(false),
151 AllocatedSize(0), NumInstructions(0), NumVectorInstructions(0),
152 FiftyPercentVectorBonus(0), TenPercentVectorBonus(0), VectorBonus(0),
153 NumConstantArgs(0), NumConstantOffsetPtrArgs(0), NumAllocaArgs(0),
154 NumConstantPtrCmps(0), NumConstantPtrDiffs(0),
155 NumInstructionsSimplified(0), SROACostSavings(0),
156 SROACostSavingsLost(0) {}
158 bool analyzeCall(CallSite CS);
160 int getThreshold() { return Threshold; }
161 int getCost() { return Cost; }
163 // Keep a bunch of stats about the cost savings found so we can print them
164 // out when debugging.
165 unsigned NumConstantArgs;
166 unsigned NumConstantOffsetPtrArgs;
167 unsigned NumAllocaArgs;
168 unsigned NumConstantPtrCmps;
169 unsigned NumConstantPtrDiffs;
170 unsigned NumInstructionsSimplified;
171 unsigned SROACostSavings;
172 unsigned SROACostSavingsLost;
179 /// \brief Test whether the given value is an Alloca-derived function argument.
180 bool CallAnalyzer::isAllocaDerivedArg(Value *V) {
181 return SROAArgValues.count(V);
184 /// \brief Lookup the SROA-candidate argument and cost iterator which V maps to.
185 /// Returns false if V does not map to a SROA-candidate.
186 bool CallAnalyzer::lookupSROAArgAndCost(
187 Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) {
188 if (SROAArgValues.empty() || SROAArgCosts.empty())
191 DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V);
192 if (ArgIt == SROAArgValues.end())
196 CostIt = SROAArgCosts.find(Arg);
197 return CostIt != SROAArgCosts.end();
200 /// \brief Disable SROA for the candidate marked by this cost iterator.
202 /// This marks the candidate as no longer viable for SROA, and adds the cost
203 /// savings associated with it back into the inline cost measurement.
204 void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) {
205 // If we're no longer able to perform SROA we need to undo its cost savings
206 // and prevent subsequent analysis.
207 Cost += CostIt->second;
208 SROACostSavings -= CostIt->second;
209 SROACostSavingsLost += CostIt->second;
210 SROAArgCosts.erase(CostIt);
213 /// \brief If 'V' maps to a SROA candidate, disable SROA for it.
214 void CallAnalyzer::disableSROA(Value *V) {
216 DenseMap<Value *, int>::iterator CostIt;
217 if (lookupSROAArgAndCost(V, SROAArg, CostIt))
221 /// \brief Accumulate the given cost for a particular SROA candidate.
222 void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
223 int InstructionCost) {
224 CostIt->second += InstructionCost;
225 SROACostSavings += InstructionCost;
228 /// \brief Helper for the common pattern of handling a SROA candidate.
229 /// Either accumulates the cost savings if the SROA remains valid, or disables
230 /// SROA for the candidate.
231 bool CallAnalyzer::handleSROACandidate(bool IsSROAValid,
232 DenseMap<Value *, int>::iterator CostIt,
233 int InstructionCost) {
235 accumulateSROACost(CostIt, InstructionCost);
243 /// \brief Check whether a GEP's indices are all constant.
245 /// Respects any simplified values known during the analysis of this callsite.
246 bool CallAnalyzer::isGEPOffsetConstant(GetElementPtrInst &GEP) {
247 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
248 if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I))
254 /// \brief Accumulate a constant GEP offset into an APInt if possible.
256 /// Returns false if unable to compute the offset for any reason. Respects any
257 /// simplified values known during the analysis of this callsite.
258 bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) {
262 unsigned IntPtrWidth = DL->getPointerSizeInBits();
263 assert(IntPtrWidth == Offset.getBitWidth());
265 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
267 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
269 if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand()))
270 OpC = dyn_cast<ConstantInt>(SimpleOp);
273 if (OpC->isZero()) continue;
275 // Handle a struct index, which adds its field offset to the pointer.
276 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
277 unsigned ElementIdx = OpC->getZExtValue();
278 const StructLayout *SL = DL->getStructLayout(STy);
279 Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx));
283 APInt TypeSize(IntPtrWidth, DL->getTypeAllocSize(GTI.getIndexedType()));
284 Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize;
289 bool CallAnalyzer::visitAlloca(AllocaInst &I) {
290 // Check whether inlining will turn a dynamic alloca into a static
291 // alloca, and handle that case.
292 if (I.isArrayAllocation()) {
293 if (Constant *Size = SimplifiedValues.lookup(I.getArraySize())) {
294 ConstantInt *AllocSize = dyn_cast<ConstantInt>(Size);
295 assert(AllocSize && "Allocation size not a constant int?");
296 Type *Ty = I.getAllocatedType();
297 AllocatedSize += Ty->getPrimitiveSizeInBits() * AllocSize->getZExtValue();
298 return Base::visitAlloca(I);
302 // Accumulate the allocated size.
303 if (I.isStaticAlloca()) {
304 Type *Ty = I.getAllocatedType();
305 AllocatedSize += (DL ? DL->getTypeAllocSize(Ty) :
306 Ty->getPrimitiveSizeInBits());
309 // We will happily inline static alloca instructions.
310 if (I.isStaticAlloca())
311 return Base::visitAlloca(I);
313 // FIXME: This is overly conservative. Dynamic allocas are inefficient for
314 // a variety of reasons, and so we would like to not inline them into
315 // functions which don't currently have a dynamic alloca. This simply
316 // disables inlining altogether in the presence of a dynamic alloca.
317 HasDynamicAlloca = true;
321 bool CallAnalyzer::visitPHI(PHINode &I) {
322 // FIXME: We should potentially be tracking values through phi nodes,
323 // especially when they collapse to a single value due to deleted CFG edges
326 // FIXME: We need to propagate SROA *disabling* through phi nodes, even
327 // though we don't want to propagate it's bonuses. The idea is to disable
328 // SROA if it *might* be used in an inappropriate manner.
330 // Phi nodes are always zero-cost.
334 bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) {
336 DenseMap<Value *, int>::iterator CostIt;
337 bool SROACandidate = lookupSROAArgAndCost(I.getPointerOperand(),
340 // Try to fold GEPs of constant-offset call site argument pointers. This
341 // requires target data and inbounds GEPs.
342 if (DL && I.isInBounds()) {
343 // Check if we have a base + offset for the pointer.
344 Value *Ptr = I.getPointerOperand();
345 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr);
346 if (BaseAndOffset.first) {
347 // Check if the offset of this GEP is constant, and if so accumulate it
349 if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) {
350 // Non-constant GEPs aren't folded, and disable SROA.
356 // Add the result as a new mapping to Base + Offset.
357 ConstantOffsetPtrs[&I] = BaseAndOffset;
359 // Also handle SROA candidates here, we already know that the GEP is
360 // all-constant indexed.
362 SROAArgValues[&I] = SROAArg;
368 if (isGEPOffsetConstant(I)) {
370 SROAArgValues[&I] = SROAArg;
372 // Constant GEPs are modeled as free.
376 // Variable GEPs will require math and will disable SROA.
382 bool CallAnalyzer::visitBitCast(BitCastInst &I) {
383 // Propagate constants through bitcasts.
384 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
386 COp = SimplifiedValues.lookup(I.getOperand(0));
388 if (Constant *C = ConstantExpr::getBitCast(COp, I.getType())) {
389 SimplifiedValues[&I] = C;
393 // Track base/offsets through casts
394 std::pair<Value *, APInt> BaseAndOffset
395 = ConstantOffsetPtrs.lookup(I.getOperand(0));
396 // Casts don't change the offset, just wrap it up.
397 if (BaseAndOffset.first)
398 ConstantOffsetPtrs[&I] = BaseAndOffset;
400 // Also look for SROA candidates here.
402 DenseMap<Value *, int>::iterator CostIt;
403 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
404 SROAArgValues[&I] = SROAArg;
406 // Bitcasts are always zero cost.
410 bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) {
411 const DataLayout *DL = I.getDataLayout();
412 // Propagate constants through ptrtoint.
413 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
415 COp = SimplifiedValues.lookup(I.getOperand(0));
417 if (Constant *C = ConstantExpr::getPtrToInt(COp, I.getType())) {
418 SimplifiedValues[&I] = C;
422 // Track base/offset pairs when converted to a plain integer provided the
423 // integer is large enough to represent the pointer.
424 unsigned IntegerSize = I.getType()->getScalarSizeInBits();
425 if (DL && IntegerSize >= DL->getPointerSizeInBits()) {
426 std::pair<Value *, APInt> BaseAndOffset
427 = ConstantOffsetPtrs.lookup(I.getOperand(0));
428 if (BaseAndOffset.first)
429 ConstantOffsetPtrs[&I] = BaseAndOffset;
432 // This is really weird. Technically, ptrtoint will disable SROA. However,
433 // unless that ptrtoint is *used* somewhere in the live basic blocks after
434 // inlining, it will be nuked, and SROA should proceed. All of the uses which
435 // would block SROA would also block SROA if applied directly to a pointer,
436 // and so we can just add the integer in here. The only places where SROA is
437 // preserved either cannot fire on an integer, or won't in-and-of themselves
438 // disable SROA (ext) w/o some later use that we would see and disable.
440 DenseMap<Value *, int>::iterator CostIt;
441 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
442 SROAArgValues[&I] = SROAArg;
444 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
447 bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) {
448 const DataLayout *DL = I.getDataLayout();
449 // Propagate constants through ptrtoint.
450 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
452 COp = SimplifiedValues.lookup(I.getOperand(0));
454 if (Constant *C = ConstantExpr::getIntToPtr(COp, I.getType())) {
455 SimplifiedValues[&I] = C;
459 // Track base/offset pairs when round-tripped through a pointer without
460 // modifications provided the integer is not too large.
461 Value *Op = I.getOperand(0);
462 unsigned IntegerSize = Op->getType()->getScalarSizeInBits();
463 if (DL && IntegerSize <= DL->getPointerSizeInBits()) {
464 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op);
465 if (BaseAndOffset.first)
466 ConstantOffsetPtrs[&I] = BaseAndOffset;
469 // "Propagate" SROA here in the same manner as we do for ptrtoint above.
471 DenseMap<Value *, int>::iterator CostIt;
472 if (lookupSROAArgAndCost(Op, SROAArg, CostIt))
473 SROAArgValues[&I] = SROAArg;
475 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
478 bool CallAnalyzer::visitCastInst(CastInst &I) {
479 // Propagate constants through ptrtoint.
480 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
482 COp = SimplifiedValues.lookup(I.getOperand(0));
484 if (Constant *C = ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) {
485 SimplifiedValues[&I] = C;
489 // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere.
490 disableSROA(I.getOperand(0));
492 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
495 bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) {
496 Value *Operand = I.getOperand(0);
497 Constant *COp = dyn_cast<Constant>(Operand);
499 COp = SimplifiedValues.lookup(Operand);
501 if (Constant *C = ConstantFoldInstOperands(I.getOpcode(), I.getType(),
503 SimplifiedValues[&I] = C;
507 // Disable any SROA on the argument to arbitrary unary operators.
508 disableSROA(Operand);
513 bool CallAnalyzer::visitCmpInst(CmpInst &I) {
514 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
515 // First try to handle simplified comparisons.
516 if (!isa<Constant>(LHS))
517 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
519 if (!isa<Constant>(RHS))
520 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
522 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
523 if (Constant *CRHS = dyn_cast<Constant>(RHS))
524 if (Constant *C = ConstantExpr::getCompare(I.getPredicate(), CLHS, CRHS)) {
525 SimplifiedValues[&I] = C;
530 if (I.getOpcode() == Instruction::FCmp)
533 // Otherwise look for a comparison between constant offset pointers with
535 Value *LHSBase, *RHSBase;
536 APInt LHSOffset, RHSOffset;
537 std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
539 std::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
540 if (RHSBase && LHSBase == RHSBase) {
541 // We have common bases, fold the icmp to a constant based on the
543 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
544 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
545 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
546 SimplifiedValues[&I] = C;
547 ++NumConstantPtrCmps;
553 // If the comparison is an equality comparison with null, we can simplify it
554 // for any alloca-derived argument.
555 if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1)))
556 if (isAllocaDerivedArg(I.getOperand(0))) {
557 // We can actually predict the result of comparisons between an
558 // alloca-derived value and null. Note that this fires regardless of
560 bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE;
561 SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType())
562 : ConstantInt::getFalse(I.getType());
566 // Finally check for SROA candidates in comparisons.
568 DenseMap<Value *, int>::iterator CostIt;
569 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
570 if (isa<ConstantPointerNull>(I.getOperand(1))) {
571 accumulateSROACost(CostIt, InlineConstants::InstrCost);
581 bool CallAnalyzer::visitSub(BinaryOperator &I) {
582 // Try to handle a special case: we can fold computing the difference of two
583 // constant-related pointers.
584 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
585 Value *LHSBase, *RHSBase;
586 APInt LHSOffset, RHSOffset;
587 std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
589 std::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
590 if (RHSBase && LHSBase == RHSBase) {
591 // We have common bases, fold the subtract to a constant based on the
593 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
594 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
595 if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) {
596 SimplifiedValues[&I] = C;
597 ++NumConstantPtrDiffs;
603 // Otherwise, fall back to the generic logic for simplifying and handling
605 return Base::visitSub(I);
608 bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
609 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
610 if (!isa<Constant>(LHS))
611 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
613 if (!isa<Constant>(RHS))
614 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
616 Value *SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, DL);
617 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
618 SimplifiedValues[&I] = C;
622 // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
629 bool CallAnalyzer::visitLoad(LoadInst &I) {
631 DenseMap<Value *, int>::iterator CostIt;
632 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
634 accumulateSROACost(CostIt, InlineConstants::InstrCost);
644 bool CallAnalyzer::visitStore(StoreInst &I) {
646 DenseMap<Value *, int>::iterator CostIt;
647 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
649 accumulateSROACost(CostIt, InlineConstants::InstrCost);
659 bool CallAnalyzer::visitExtractValue(ExtractValueInst &I) {
660 // Constant folding for extract value is trivial.
661 Constant *C = dyn_cast<Constant>(I.getAggregateOperand());
663 C = SimplifiedValues.lookup(I.getAggregateOperand());
665 SimplifiedValues[&I] = ConstantExpr::getExtractValue(C, I.getIndices());
669 // SROA can look through these but give them a cost.
673 bool CallAnalyzer::visitInsertValue(InsertValueInst &I) {
674 // Constant folding for insert value is trivial.
675 Constant *AggC = dyn_cast<Constant>(I.getAggregateOperand());
677 AggC = SimplifiedValues.lookup(I.getAggregateOperand());
678 Constant *InsertedC = dyn_cast<Constant>(I.getInsertedValueOperand());
680 InsertedC = SimplifiedValues.lookup(I.getInsertedValueOperand());
681 if (AggC && InsertedC) {
682 SimplifiedValues[&I] = ConstantExpr::getInsertValue(AggC, InsertedC,
687 // SROA can look through these but give them a cost.
691 /// \brief Try to simplify a call site.
693 /// Takes a concrete function and callsite and tries to actually simplify it by
694 /// analyzing the arguments and call itself with instsimplify. Returns true if
695 /// it has simplified the callsite to some other entity (a constant), making it
697 bool CallAnalyzer::simplifyCallSite(Function *F, CallSite CS) {
698 // FIXME: Using the instsimplify logic directly for this is inefficient
699 // because we have to continually rebuild the argument list even when no
700 // simplifications can be performed. Until that is fixed with remapping
701 // inside of instsimplify, directly constant fold calls here.
702 if (!canConstantFoldCallTo(F))
705 // Try to re-map the arguments to constants.
706 SmallVector<Constant *, 4> ConstantArgs;
707 ConstantArgs.reserve(CS.arg_size());
708 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
710 Constant *C = dyn_cast<Constant>(*I);
712 C = dyn_cast_or_null<Constant>(SimplifiedValues.lookup(*I));
714 return false; // This argument doesn't map to a constant.
716 ConstantArgs.push_back(C);
718 if (Constant *C = ConstantFoldCall(F, ConstantArgs)) {
719 SimplifiedValues[CS.getInstruction()] = C;
726 bool CallAnalyzer::visitCallSite(CallSite CS) {
727 if (CS.hasFnAttr(Attribute::ReturnsTwice) &&
728 !F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
729 Attribute::ReturnsTwice)) {
730 // This aborts the entire analysis.
731 ExposesReturnsTwice = true;
735 cast<CallInst>(CS.getInstruction())->cannotDuplicate())
736 ContainsNoDuplicateCall = true;
738 if (Function *F = CS.getCalledFunction()) {
739 // When we have a concrete function, first try to simplify it directly.
740 if (simplifyCallSite(F, CS))
743 // Next check if it is an intrinsic we know about.
744 // FIXME: Lift this into part of the InstVisitor.
745 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
746 switch (II->getIntrinsicID()) {
748 return Base::visitCallSite(CS);
750 case Intrinsic::memset:
751 case Intrinsic::memcpy:
752 case Intrinsic::memmove:
753 // SROA can usually chew through these intrinsics, but they aren't free.
758 if (F == CS.getInstruction()->getParent()->getParent()) {
759 // This flag will fully abort the analysis, so don't bother with anything
761 IsRecursiveCall = true;
765 if (TTI.isLoweredToCall(F)) {
766 // We account for the average 1 instruction per call argument setup
768 Cost += CS.arg_size() * InlineConstants::InstrCost;
770 // Everything other than inline ASM will also have a significant cost
771 // merely from making the call.
772 if (!isa<InlineAsm>(CS.getCalledValue()))
773 Cost += InlineConstants::CallPenalty;
776 return Base::visitCallSite(CS);
779 // Otherwise we're in a very special case -- an indirect function call. See
780 // if we can be particularly clever about this.
781 Value *Callee = CS.getCalledValue();
783 // First, pay the price of the argument setup. We account for the average
784 // 1 instruction per call argument setup here.
785 Cost += CS.arg_size() * InlineConstants::InstrCost;
787 // Next, check if this happens to be an indirect function call to a known
788 // function in this inline context. If not, we've done all we can.
789 Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee));
791 return Base::visitCallSite(CS);
793 // If we have a constant that we are calling as a function, we can peer
794 // through it and see the function target. This happens not infrequently
795 // during devirtualization and so we want to give it a hefty bonus for
796 // inlining, but cap that bonus in the event that inlining wouldn't pan
797 // out. Pretend to inline the function, with a custom threshold.
798 CallAnalyzer CA(DL, TTI, *F, InlineConstants::IndirectCallThreshold);
799 if (CA.analyzeCall(CS)) {
800 // We were able to inline the indirect call! Subtract the cost from the
801 // bonus we want to apply, but don't go below zero.
802 Cost -= std::max(0, InlineConstants::IndirectCallThreshold - CA.getCost());
805 return Base::visitCallSite(CS);
808 bool CallAnalyzer::visitReturnInst(ReturnInst &RI) {
809 // At least one return instruction will be free after inlining.
810 bool Free = !HasReturn;
815 bool CallAnalyzer::visitBranchInst(BranchInst &BI) {
816 // We model unconditional branches as essentially free -- they really
817 // shouldn't exist at all, but handling them makes the behavior of the
818 // inliner more regular and predictable. Interestingly, conditional branches
819 // which will fold away are also free.
820 return BI.isUnconditional() || isa<ConstantInt>(BI.getCondition()) ||
821 dyn_cast_or_null<ConstantInt>(
822 SimplifiedValues.lookup(BI.getCondition()));
825 bool CallAnalyzer::visitSwitchInst(SwitchInst &SI) {
826 // We model unconditional switches as free, see the comments on handling
828 return isa<ConstantInt>(SI.getCondition()) ||
829 dyn_cast_or_null<ConstantInt>(
830 SimplifiedValues.lookup(SI.getCondition()));
833 bool CallAnalyzer::visitIndirectBrInst(IndirectBrInst &IBI) {
834 // We never want to inline functions that contain an indirectbr. This is
835 // incorrect because all the blockaddress's (in static global initializers
836 // for example) would be referring to the original function, and this
837 // indirect jump would jump from the inlined copy of the function into the
838 // original function which is extremely undefined behavior.
839 // FIXME: This logic isn't really right; we can safely inline functions with
840 // indirectbr's as long as no other function or global references the
841 // blockaddress of a block within the current function. And as a QOI issue,
842 // if someone is using a blockaddress without an indirectbr, and that
843 // reference somehow ends up in another function or global, we probably don't
844 // want to inline this function.
845 HasIndirectBr = true;
849 bool CallAnalyzer::visitResumeInst(ResumeInst &RI) {
850 // FIXME: It's not clear that a single instruction is an accurate model for
851 // the inline cost of a resume instruction.
855 bool CallAnalyzer::visitUnreachableInst(UnreachableInst &I) {
856 // FIXME: It might be reasonably to discount the cost of instructions leading
857 // to unreachable as they have the lowest possible impact on both runtime and
859 return true; // No actual code is needed for unreachable.
862 bool CallAnalyzer::visitInstruction(Instruction &I) {
863 // Some instructions are free. All of the free intrinsics can also be
864 // handled by SROA, etc.
865 if (TargetTransformInfo::TCC_Free == TTI.getUserCost(&I))
868 // We found something we don't understand or can't handle. Mark any SROA-able
869 // values in the operand list as no longer viable.
870 for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI)
877 /// \brief Analyze a basic block for its contribution to the inline cost.
879 /// This method walks the analyzer over every instruction in the given basic
880 /// block and accounts for their cost during inlining at this callsite. It
881 /// aborts early if the threshold has been exceeded or an impossible to inline
882 /// construct has been detected. It returns false if inlining is no longer
883 /// viable, and true if inlining remains viable.
884 bool CallAnalyzer::analyzeBlock(BasicBlock *BB) {
885 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
886 // FIXME: Currently, the number of instructions in a function regardless of
887 // our ability to simplify them during inline to constants or dead code,
888 // are actually used by the vector bonus heuristic. As long as that's true,
889 // we have to special case debug intrinsics here to prevent differences in
890 // inlining due to debug symbols. Eventually, the number of unsimplified
891 // instructions shouldn't factor into the cost computation, but until then,
892 // hack around it here.
893 if (isa<DbgInfoIntrinsic>(I))
897 if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy())
898 ++NumVectorInstructions;
900 // If the instruction simplified to a constant, there is no cost to this
901 // instruction. Visit the instructions using our InstVisitor to account for
902 // all of the per-instruction logic. The visit tree returns true if we
903 // consumed the instruction in any way, and false if the instruction's base
904 // cost should count against inlining.
906 ++NumInstructionsSimplified;
908 Cost += InlineConstants::InstrCost;
910 // If the visit this instruction detected an uninlinable pattern, abort.
911 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca ||
915 // If the caller is a recursive function then we don't want to inline
916 // functions which allocate a lot of stack space because it would increase
917 // the caller stack usage dramatically.
918 if (IsCallerRecursive &&
919 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller)
922 if (NumVectorInstructions > NumInstructions/2)
923 VectorBonus = FiftyPercentVectorBonus;
924 else if (NumVectorInstructions > NumInstructions/10)
925 VectorBonus = TenPercentVectorBonus;
929 // Check if we've past the threshold so we don't spin in huge basic
930 // blocks that will never inline.
931 if (Cost > (Threshold + VectorBonus))
938 /// \brief Compute the base pointer and cumulative constant offsets for V.
940 /// This strips all constant offsets off of V, leaving it the base pointer, and
941 /// accumulates the total constant offset applied in the returned constant. It
942 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
943 /// no constant offsets applied.
944 ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) {
945 if (!DL || !V->getType()->isPointerTy())
948 unsigned IntPtrWidth = DL->getPointerSizeInBits();
949 APInt Offset = APInt::getNullValue(IntPtrWidth);
951 // Even though we don't look through PHI nodes, we could be called on an
952 // instruction in an unreachable block, which may be on a cycle.
953 SmallPtrSet<Value *, 4> Visited;
956 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
957 if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset))
959 V = GEP->getPointerOperand();
960 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
961 V = cast<Operator>(V)->getOperand(0);
962 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
963 if (GA->mayBeOverridden())
965 V = GA->getAliasee();
969 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
970 } while (Visited.insert(V));
972 Type *IntPtrTy = DL->getIntPtrType(V->getContext());
973 return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset));
976 /// \brief Analyze a call site for potential inlining.
978 /// Returns true if inlining this call is viable, and false if it is not
979 /// viable. It computes the cost and adjusts the threshold based on numerous
980 /// factors and heuristics. If this method returns false but the computed cost
981 /// is below the computed threshold, then inlining was forcibly disabled by
982 /// some artifact of the routine.
983 bool CallAnalyzer::analyzeCall(CallSite CS) {
986 // Track whether the post-inlining function would have more than one basic
987 // block. A single basic block is often intended for inlining. Balloon the
988 // threshold by 50% until we pass the single-BB phase.
989 bool SingleBB = true;
990 int SingleBBBonus = Threshold / 2;
991 Threshold += SingleBBBonus;
993 // Perform some tweaks to the cost and threshold based on the direct
994 // callsite information.
996 // We want to more aggressively inline vector-dense kernels, so up the
997 // threshold, and we'll lower it if the % of vector instructions gets too
999 assert(NumInstructions == 0);
1000 assert(NumVectorInstructions == 0);
1001 FiftyPercentVectorBonus = Threshold;
1002 TenPercentVectorBonus = Threshold / 2;
1004 // Give out bonuses per argument, as the instructions setting them up will
1005 // be gone after inlining.
1006 for (unsigned I = 0, E = CS.arg_size(); I != E; ++I) {
1007 if (DL && CS.isByValArgument(I)) {
1008 // We approximate the number of loads and stores needed by dividing the
1009 // size of the byval type by the target's pointer size.
1010 PointerType *PTy = cast<PointerType>(CS.getArgument(I)->getType());
1011 unsigned TypeSize = DL->getTypeSizeInBits(PTy->getElementType());
1012 unsigned PointerSize = DL->getPointerSizeInBits();
1013 // Ceiling division.
1014 unsigned NumStores = (TypeSize + PointerSize - 1) / PointerSize;
1016 // If it generates more than 8 stores it is likely to be expanded as an
1017 // inline memcpy so we take that as an upper bound. Otherwise we assume
1018 // one load and one store per word copied.
1019 // FIXME: The maxStoresPerMemcpy setting from the target should be used
1020 // here instead of a magic number of 8, but it's not available via
1022 NumStores = std::min(NumStores, 8U);
1024 Cost -= 2 * NumStores * InlineConstants::InstrCost;
1026 // For non-byval arguments subtract off one instruction per call
1028 Cost -= InlineConstants::InstrCost;
1032 // If there is only one call of the function, and it has internal linkage,
1033 // the cost of inlining it drops dramatically.
1034 bool OnlyOneCallAndLocalLinkage = F.hasLocalLinkage() && F.hasOneUse() &&
1035 &F == CS.getCalledFunction();
1036 if (OnlyOneCallAndLocalLinkage)
1037 Cost += InlineConstants::LastCallToStaticBonus;
1039 // If the instruction after the call, or if the normal destination of the
1040 // invoke is an unreachable instruction, the function is noreturn. As such,
1041 // there is little point in inlining this unless there is literally zero
1043 Instruction *Instr = CS.getInstruction();
1044 if (InvokeInst *II = dyn_cast<InvokeInst>(Instr)) {
1045 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
1047 } else if (isa<UnreachableInst>(++BasicBlock::iterator(Instr)))
1050 // If this function uses the coldcc calling convention, prefer not to inline
1052 if (F.getCallingConv() == CallingConv::Cold)
1053 Cost += InlineConstants::ColdccPenalty;
1055 // Check if we're done. This can happen due to bonuses and penalties.
1056 if (Cost > Threshold)
1062 Function *Caller = CS.getInstruction()->getParent()->getParent();
1063 // Check if the caller function is recursive itself.
1064 for (User *U : Caller->users()) {
1068 Instruction *I = Site.getInstruction();
1069 if (I->getParent()->getParent() == Caller) {
1070 IsCallerRecursive = true;
1075 // Populate our simplified values by mapping from function arguments to call
1076 // arguments with known important simplifications.
1077 CallSite::arg_iterator CAI = CS.arg_begin();
1078 for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end();
1079 FAI != FAE; ++FAI, ++CAI) {
1080 assert(CAI != CS.arg_end());
1081 if (Constant *C = dyn_cast<Constant>(CAI))
1082 SimplifiedValues[FAI] = C;
1084 Value *PtrArg = *CAI;
1085 if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) {
1086 ConstantOffsetPtrs[FAI] = std::make_pair(PtrArg, C->getValue());
1088 // We can SROA any pointer arguments derived from alloca instructions.
1089 if (isa<AllocaInst>(PtrArg)) {
1090 SROAArgValues[FAI] = PtrArg;
1091 SROAArgCosts[PtrArg] = 0;
1095 NumConstantArgs = SimplifiedValues.size();
1096 NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size();
1097 NumAllocaArgs = SROAArgValues.size();
1099 // The worklist of live basic blocks in the callee *after* inlining. We avoid
1100 // adding basic blocks of the callee which can be proven to be dead for this
1101 // particular call site in order to get more accurate cost estimates. This
1102 // requires a somewhat heavyweight iteration pattern: we need to walk the
1103 // basic blocks in a breadth-first order as we insert live successors. To
1104 // accomplish this, prioritizing for small iterations because we exit after
1105 // crossing our threshold, we use a small-size optimized SetVector.
1106 typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>,
1107 SmallPtrSet<BasicBlock *, 16> > BBSetVector;
1108 BBSetVector BBWorklist;
1109 BBWorklist.insert(&F.getEntryBlock());
1110 // Note that we *must not* cache the size, this loop grows the worklist.
1111 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
1112 // Bail out the moment we cross the threshold. This means we'll under-count
1113 // the cost, but only when undercounting doesn't matter.
1114 if (Cost > (Threshold + VectorBonus))
1117 BasicBlock *BB = BBWorklist[Idx];
1121 // Analyze the cost of this block. If we blow through the threshold, this
1122 // returns false, and we can bail on out.
1123 if (!analyzeBlock(BB)) {
1124 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca ||
1128 // If the caller is a recursive function then we don't want to inline
1129 // functions which allocate a lot of stack space because it would increase
1130 // the caller stack usage dramatically.
1131 if (IsCallerRecursive &&
1132 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller)
1138 TerminatorInst *TI = BB->getTerminator();
1140 // Add in the live successors by first checking whether we have terminator
1141 // that may be simplified based on the values simplified by this call.
1142 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1143 if (BI->isConditional()) {
1144 Value *Cond = BI->getCondition();
1145 if (ConstantInt *SimpleCond
1146 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
1147 BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0));
1151 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1152 Value *Cond = SI->getCondition();
1153 if (ConstantInt *SimpleCond
1154 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
1155 BBWorklist.insert(SI->findCaseValue(SimpleCond).getCaseSuccessor());
1160 // If we're unable to select a particular successor, just count all of
1162 for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize;
1164 BBWorklist.insert(TI->getSuccessor(TIdx));
1166 // If we had any successors at this point, than post-inlining is likely to
1167 // have them as well. Note that we assume any basic blocks which existed
1168 // due to branches or switches which folded above will also fold after
1170 if (SingleBB && TI->getNumSuccessors() > 1) {
1171 // Take off the bonus we applied to the threshold.
1172 Threshold -= SingleBBBonus;
1177 // If this is a noduplicate call, we can still inline as long as
1178 // inlining this would cause the removal of the caller (so the instruction
1179 // is not actually duplicated, just moved).
1180 if (!OnlyOneCallAndLocalLinkage && ContainsNoDuplicateCall)
1183 Threshold += VectorBonus;
1185 return Cost < Threshold;
1188 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1189 /// \brief Dump stats about this call's analysis.
1190 void CallAnalyzer::dump() {
1191 #define DEBUG_PRINT_STAT(x) dbgs() << " " #x ": " << x << "\n"
1192 DEBUG_PRINT_STAT(NumConstantArgs);
1193 DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs);
1194 DEBUG_PRINT_STAT(NumAllocaArgs);
1195 DEBUG_PRINT_STAT(NumConstantPtrCmps);
1196 DEBUG_PRINT_STAT(NumConstantPtrDiffs);
1197 DEBUG_PRINT_STAT(NumInstructionsSimplified);
1198 DEBUG_PRINT_STAT(SROACostSavings);
1199 DEBUG_PRINT_STAT(SROACostSavingsLost);
1200 DEBUG_PRINT_STAT(ContainsNoDuplicateCall);
1201 DEBUG_PRINT_STAT(Cost);
1202 DEBUG_PRINT_STAT(Threshold);
1203 DEBUG_PRINT_STAT(VectorBonus);
1204 #undef DEBUG_PRINT_STAT
1208 INITIALIZE_PASS_BEGIN(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis",
1210 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
1211 INITIALIZE_PASS_END(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis",
1214 char InlineCostAnalysis::ID = 0;
1216 InlineCostAnalysis::InlineCostAnalysis() : CallGraphSCCPass(ID) {}
1218 InlineCostAnalysis::~InlineCostAnalysis() {}
1220 void InlineCostAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1221 AU.setPreservesAll();
1222 AU.addRequired<TargetTransformInfo>();
1223 CallGraphSCCPass::getAnalysisUsage(AU);
1226 bool InlineCostAnalysis::runOnSCC(CallGraphSCC &SCC) {
1227 TTI = &getAnalysis<TargetTransformInfo>();
1231 InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, int Threshold) {
1232 return getInlineCost(CS, CS.getCalledFunction(), Threshold);
1235 /// \brief Test that two functions either have or have not the given attribute
1236 /// at the same time.
1237 static bool attributeMatches(Function *F1, Function *F2,
1238 Attribute::AttrKind Attr) {
1239 return F1->hasFnAttribute(Attr) == F2->hasFnAttribute(Attr);
1242 /// \brief Test that there are no attribute conflicts between Caller and Callee
1243 /// that prevent inlining.
1244 static bool functionsHaveCompatibleAttributes(Function *Caller,
1246 return attributeMatches(Caller, Callee, Attribute::SanitizeAddress) &&
1247 attributeMatches(Caller, Callee, Attribute::SanitizeMemory) &&
1248 attributeMatches(Caller, Callee, Attribute::SanitizeThread);
1251 InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, Function *Callee,
1253 // Cannot inline indirect calls.
1255 return llvm::InlineCost::getNever();
1257 // Calls to functions with always-inline attributes should be inlined
1258 // whenever possible.
1259 if (Callee->hasFnAttribute(Attribute::AlwaysInline)) {
1260 if (isInlineViable(*Callee))
1261 return llvm::InlineCost::getAlways();
1262 return llvm::InlineCost::getNever();
1265 // Never inline functions with conflicting attributes (unless callee has
1266 // always-inline attribute).
1267 if (!functionsHaveCompatibleAttributes(CS.getCaller(), Callee))
1268 return llvm::InlineCost::getNever();
1270 // Don't inline this call if the caller has the optnone attribute.
1271 if (CS.getCaller()->hasFnAttribute(Attribute::OptimizeNone))
1272 return llvm::InlineCost::getNever();
1274 // Don't inline functions which can be redefined at link-time to mean
1275 // something else. Don't inline functions marked noinline or call sites
1277 if (Callee->mayBeOverridden() ||
1278 Callee->hasFnAttribute(Attribute::NoInline) || CS.isNoInline())
1279 return llvm::InlineCost::getNever();
1281 DEBUG(llvm::dbgs() << " Analyzing call of " << Callee->getName()
1284 CallAnalyzer CA(Callee->getDataLayout(), *TTI, *Callee, Threshold);
1285 bool ShouldInline = CA.analyzeCall(CS);
1289 // Check if there was a reason to force inlining or no inlining.
1290 if (!ShouldInline && CA.getCost() < CA.getThreshold())
1291 return InlineCost::getNever();
1292 if (ShouldInline && CA.getCost() >= CA.getThreshold())
1293 return InlineCost::getAlways();
1295 return llvm::InlineCost::get(CA.getCost(), CA.getThreshold());
1298 bool InlineCostAnalysis::isInlineViable(Function &F) {
1300 F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1301 Attribute::ReturnsTwice);
1302 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
1303 // Disallow inlining of functions which contain an indirect branch.
1304 if (isa<IndirectBrInst>(BI->getTerminator()))
1307 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
1313 // Disallow recursive calls.
1314 if (&F == CS.getCalledFunction())
1317 // Disallow calls which expose returns-twice to a function not previously
1318 // attributed as such.
1319 if (!ReturnsTwice && CS.isCall() &&
1320 cast<CallInst>(CS.getInstruction())->canReturnTwice())