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/CallingConv.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/Operator.h"
29 #include "llvm/InstVisitor.h"
30 #include "llvm/Support/CallSite.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.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 TD;
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;
63 /// Number of bytes allocated statically by the callee.
64 uint64_t AllocatedSize;
65 unsigned NumInstructions, NumVectorInstructions;
66 int FiftyPercentVectorBonus, TenPercentVectorBonus;
69 // While we walk the potentially-inlined instructions, we build up and
70 // maintain a mapping of simplified values specific to this callsite. The
71 // idea is to propagate any special information we have about arguments to
72 // this call through the inlinable section of the function, and account for
73 // likely simplifications post-inlining. The most important aspect we track
74 // is CFG altering simplifications -- when we prove a basic block dead, that
75 // can cause dramatic shifts in the cost of inlining a function.
76 DenseMap<Value *, Constant *> SimplifiedValues;
78 // Keep track of the values which map back (through function arguments) to
79 // allocas on the caller stack which could be simplified through SROA.
80 DenseMap<Value *, Value *> SROAArgValues;
82 // The mapping of caller Alloca values to their accumulated cost savings. If
83 // we have to disable SROA for one of the allocas, this tells us how much
84 // cost must be added.
85 DenseMap<Value *, int> SROAArgCosts;
87 // Keep track of values which map to a pointer base and constant offset.
88 DenseMap<Value *, std::pair<Value *, APInt> > ConstantOffsetPtrs;
90 // Custom simplification helper routines.
91 bool isAllocaDerivedArg(Value *V);
92 bool lookupSROAArgAndCost(Value *V, Value *&Arg,
93 DenseMap<Value *, int>::iterator &CostIt);
94 void disableSROA(DenseMap<Value *, int>::iterator CostIt);
95 void disableSROA(Value *V);
96 void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
98 bool handleSROACandidate(bool IsSROAValid,
99 DenseMap<Value *, int>::iterator CostIt,
100 int InstructionCost);
101 bool isGEPOffsetConstant(GetElementPtrInst &GEP);
102 bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset);
103 bool simplifyCallSite(Function *F, CallSite CS);
104 ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V);
106 // Custom analysis routines.
107 bool analyzeBlock(BasicBlock *BB);
109 // Disable several entry points to the visitor so we don't accidentally use
110 // them by declaring but not defining them here.
111 void visit(Module *); void visit(Module &);
112 void visit(Function *); void visit(Function &);
113 void visit(BasicBlock *); void visit(BasicBlock &);
115 // Provide base case for our instruction visit.
116 bool visitInstruction(Instruction &I);
118 // Our visit overrides.
119 bool visitAlloca(AllocaInst &I);
120 bool visitPHI(PHINode &I);
121 bool visitGetElementPtr(GetElementPtrInst &I);
122 bool visitBitCast(BitCastInst &I);
123 bool visitPtrToInt(PtrToIntInst &I);
124 bool visitIntToPtr(IntToPtrInst &I);
125 bool visitCastInst(CastInst &I);
126 bool visitUnaryInstruction(UnaryInstruction &I);
127 bool visitICmp(ICmpInst &I);
128 bool visitSub(BinaryOperator &I);
129 bool visitBinaryOperator(BinaryOperator &I);
130 bool visitLoad(LoadInst &I);
131 bool visitStore(StoreInst &I);
132 bool visitExtractValue(ExtractValueInst &I);
133 bool visitInsertValue(InsertValueInst &I);
134 bool visitCallSite(CallSite CS);
137 CallAnalyzer(const DataLayout *TD, const TargetTransformInfo &TTI,
138 Function &Callee, int Threshold)
139 : TD(TD), TTI(TTI), F(Callee), Threshold(Threshold), Cost(0),
140 IsCallerRecursive(false), IsRecursiveCall(false),
141 ExposesReturnsTwice(false), HasDynamicAlloca(false),
142 ContainsNoDuplicateCall(false), AllocatedSize(0), NumInstructions(0),
143 NumVectorInstructions(0), FiftyPercentVectorBonus(0),
144 TenPercentVectorBonus(0), VectorBonus(0), NumConstantArgs(0),
145 NumConstantOffsetPtrArgs(0), NumAllocaArgs(0), NumConstantPtrCmps(0),
146 NumConstantPtrDiffs(0), NumInstructionsSimplified(0),
147 SROACostSavings(0), SROACostSavingsLost(0) {}
149 bool analyzeCall(CallSite CS);
151 int getThreshold() { return Threshold; }
152 int getCost() { return Cost; }
154 // Keep a bunch of stats about the cost savings found so we can print them
155 // out when debugging.
156 unsigned NumConstantArgs;
157 unsigned NumConstantOffsetPtrArgs;
158 unsigned NumAllocaArgs;
159 unsigned NumConstantPtrCmps;
160 unsigned NumConstantPtrDiffs;
161 unsigned NumInstructionsSimplified;
162 unsigned SROACostSavings;
163 unsigned SROACostSavingsLost;
170 /// \brief Test whether the given value is an Alloca-derived function argument.
171 bool CallAnalyzer::isAllocaDerivedArg(Value *V) {
172 return SROAArgValues.count(V);
175 /// \brief Lookup the SROA-candidate argument and cost iterator which V maps to.
176 /// Returns false if V does not map to a SROA-candidate.
177 bool CallAnalyzer::lookupSROAArgAndCost(
178 Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) {
179 if (SROAArgValues.empty() || SROAArgCosts.empty())
182 DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V);
183 if (ArgIt == SROAArgValues.end())
187 CostIt = SROAArgCosts.find(Arg);
188 return CostIt != SROAArgCosts.end();
191 /// \brief Disable SROA for the candidate marked by this cost iterator.
193 /// This marks the candidate as no longer viable for SROA, and adds the cost
194 /// savings associated with it back into the inline cost measurement.
195 void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) {
196 // If we're no longer able to perform SROA we need to undo its cost savings
197 // and prevent subsequent analysis.
198 Cost += CostIt->second;
199 SROACostSavings -= CostIt->second;
200 SROACostSavingsLost += CostIt->second;
201 SROAArgCosts.erase(CostIt);
204 /// \brief If 'V' maps to a SROA candidate, disable SROA for it.
205 void CallAnalyzer::disableSROA(Value *V) {
207 DenseMap<Value *, int>::iterator CostIt;
208 if (lookupSROAArgAndCost(V, SROAArg, CostIt))
212 /// \brief Accumulate the given cost for a particular SROA candidate.
213 void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
214 int InstructionCost) {
215 CostIt->second += InstructionCost;
216 SROACostSavings += InstructionCost;
219 /// \brief Helper for the common pattern of handling a SROA candidate.
220 /// Either accumulates the cost savings if the SROA remains valid, or disables
221 /// SROA for the candidate.
222 bool CallAnalyzer::handleSROACandidate(bool IsSROAValid,
223 DenseMap<Value *, int>::iterator CostIt,
224 int InstructionCost) {
226 accumulateSROACost(CostIt, InstructionCost);
234 /// \brief Check whether a GEP's indices are all constant.
236 /// Respects any simplified values known during the analysis of this callsite.
237 bool CallAnalyzer::isGEPOffsetConstant(GetElementPtrInst &GEP) {
238 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
239 if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I))
245 /// \brief Accumulate a constant GEP offset into an APInt if possible.
247 /// Returns false if unable to compute the offset for any reason. Respects any
248 /// simplified values known during the analysis of this callsite.
249 bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) {
253 unsigned IntPtrWidth = TD->getPointerSizeInBits();
254 assert(IntPtrWidth == Offset.getBitWidth());
256 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
258 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
260 if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand()))
261 OpC = dyn_cast<ConstantInt>(SimpleOp);
264 if (OpC->isZero()) continue;
266 // Handle a struct index, which adds its field offset to the pointer.
267 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
268 unsigned ElementIdx = OpC->getZExtValue();
269 const StructLayout *SL = TD->getStructLayout(STy);
270 Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx));
274 APInt TypeSize(IntPtrWidth, TD->getTypeAllocSize(GTI.getIndexedType()));
275 Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize;
280 bool CallAnalyzer::visitAlloca(AllocaInst &I) {
281 // FIXME: Check whether inlining will turn a dynamic alloca into a static
282 // alloca, and handle that case.
284 // Accumulate the allocated size.
285 if (I.isStaticAlloca()) {
286 Type *Ty = I.getAllocatedType();
287 AllocatedSize += (TD ? TD->getTypeAllocSize(Ty) :
288 Ty->getPrimitiveSizeInBits());
291 // We will happily inline static alloca instructions.
292 if (I.isStaticAlloca())
293 return Base::visitAlloca(I);
295 // FIXME: This is overly conservative. Dynamic allocas are inefficient for
296 // a variety of reasons, and so we would like to not inline them into
297 // functions which don't currently have a dynamic alloca. This simply
298 // disables inlining altogether in the presence of a dynamic alloca.
299 HasDynamicAlloca = true;
303 bool CallAnalyzer::visitPHI(PHINode &I) {
304 // FIXME: We should potentially be tracking values through phi nodes,
305 // especially when they collapse to a single value due to deleted CFG edges
308 // FIXME: We need to propagate SROA *disabling* through phi nodes, even
309 // though we don't want to propagate it's bonuses. The idea is to disable
310 // SROA if it *might* be used in an inappropriate manner.
312 // Phi nodes are always zero-cost.
316 bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) {
318 DenseMap<Value *, int>::iterator CostIt;
319 bool SROACandidate = lookupSROAArgAndCost(I.getPointerOperand(),
322 // Try to fold GEPs of constant-offset call site argument pointers. This
323 // requires target data and inbounds GEPs.
324 if (TD && I.isInBounds()) {
325 // Check if we have a base + offset for the pointer.
326 Value *Ptr = I.getPointerOperand();
327 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr);
328 if (BaseAndOffset.first) {
329 // Check if the offset of this GEP is constant, and if so accumulate it
331 if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) {
332 // Non-constant GEPs aren't folded, and disable SROA.
338 // Add the result as a new mapping to Base + Offset.
339 ConstantOffsetPtrs[&I] = BaseAndOffset;
341 // Also handle SROA candidates here, we already know that the GEP is
342 // all-constant indexed.
344 SROAArgValues[&I] = SROAArg;
350 if (isGEPOffsetConstant(I)) {
352 SROAArgValues[&I] = SROAArg;
354 // Constant GEPs are modeled as free.
358 // Variable GEPs will require math and will disable SROA.
364 bool CallAnalyzer::visitBitCast(BitCastInst &I) {
365 // Propagate constants through bitcasts.
366 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
368 COp = SimplifiedValues.lookup(I.getOperand(0));
370 if (Constant *C = ConstantExpr::getBitCast(COp, I.getType())) {
371 SimplifiedValues[&I] = C;
375 // Track base/offsets through casts
376 std::pair<Value *, APInt> BaseAndOffset
377 = ConstantOffsetPtrs.lookup(I.getOperand(0));
378 // Casts don't change the offset, just wrap it up.
379 if (BaseAndOffset.first)
380 ConstantOffsetPtrs[&I] = BaseAndOffset;
382 // Also look for SROA candidates here.
384 DenseMap<Value *, int>::iterator CostIt;
385 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
386 SROAArgValues[&I] = SROAArg;
388 // Bitcasts are always zero cost.
392 bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) {
393 // Propagate constants through ptrtoint.
394 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
396 COp = SimplifiedValues.lookup(I.getOperand(0));
398 if (Constant *C = ConstantExpr::getPtrToInt(COp, I.getType())) {
399 SimplifiedValues[&I] = C;
403 // Track base/offset pairs when converted to a plain integer provided the
404 // integer is large enough to represent the pointer.
405 unsigned IntegerSize = I.getType()->getScalarSizeInBits();
406 if (TD && IntegerSize >= TD->getPointerSizeInBits()) {
407 std::pair<Value *, APInt> BaseAndOffset
408 = ConstantOffsetPtrs.lookup(I.getOperand(0));
409 if (BaseAndOffset.first)
410 ConstantOffsetPtrs[&I] = BaseAndOffset;
413 // This is really weird. Technically, ptrtoint will disable SROA. However,
414 // unless that ptrtoint is *used* somewhere in the live basic blocks after
415 // inlining, it will be nuked, and SROA should proceed. All of the uses which
416 // would block SROA would also block SROA if applied directly to a pointer,
417 // and so we can just add the integer in here. The only places where SROA is
418 // preserved either cannot fire on an integer, or won't in-and-of themselves
419 // disable SROA (ext) w/o some later use that we would see and disable.
421 DenseMap<Value *, int>::iterator CostIt;
422 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
423 SROAArgValues[&I] = SROAArg;
425 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
428 bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) {
429 // Propagate constants through ptrtoint.
430 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
432 COp = SimplifiedValues.lookup(I.getOperand(0));
434 if (Constant *C = ConstantExpr::getIntToPtr(COp, I.getType())) {
435 SimplifiedValues[&I] = C;
439 // Track base/offset pairs when round-tripped through a pointer without
440 // modifications provided the integer is not too large.
441 Value *Op = I.getOperand(0);
442 unsigned IntegerSize = Op->getType()->getScalarSizeInBits();
443 if (TD && IntegerSize <= TD->getPointerSizeInBits()) {
444 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op);
445 if (BaseAndOffset.first)
446 ConstantOffsetPtrs[&I] = BaseAndOffset;
449 // "Propagate" SROA here in the same manner as we do for ptrtoint above.
451 DenseMap<Value *, int>::iterator CostIt;
452 if (lookupSROAArgAndCost(Op, SROAArg, CostIt))
453 SROAArgValues[&I] = SROAArg;
455 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
458 bool CallAnalyzer::visitCastInst(CastInst &I) {
459 // Propagate constants through ptrtoint.
460 Constant *COp = dyn_cast<Constant>(I.getOperand(0));
462 COp = SimplifiedValues.lookup(I.getOperand(0));
464 if (Constant *C = ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) {
465 SimplifiedValues[&I] = C;
469 // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere.
470 disableSROA(I.getOperand(0));
472 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
475 bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) {
476 Value *Operand = I.getOperand(0);
477 Constant *Ops[1] = { dyn_cast<Constant>(Operand) };
478 if (Ops[0] || (Ops[0] = SimplifiedValues.lookup(Operand)))
479 if (Constant *C = ConstantFoldInstOperands(I.getOpcode(), I.getType(),
481 SimplifiedValues[&I] = C;
485 // Disable any SROA on the argument to arbitrary unary operators.
486 disableSROA(Operand);
491 bool CallAnalyzer::visitICmp(ICmpInst &I) {
492 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
493 // First try to handle simplified comparisons.
494 if (!isa<Constant>(LHS))
495 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
497 if (!isa<Constant>(RHS))
498 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
500 if (Constant *CLHS = dyn_cast<Constant>(LHS))
501 if (Constant *CRHS = dyn_cast<Constant>(RHS))
502 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
503 SimplifiedValues[&I] = C;
507 // Otherwise look for a comparison between constant offset pointers with
509 Value *LHSBase, *RHSBase;
510 APInt LHSOffset, RHSOffset;
511 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
513 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
514 if (RHSBase && LHSBase == RHSBase) {
515 // We have common bases, fold the icmp to a constant based on the
517 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
518 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
519 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
520 SimplifiedValues[&I] = C;
521 ++NumConstantPtrCmps;
527 // If the comparison is an equality comparison with null, we can simplify it
528 // for any alloca-derived argument.
529 if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1)))
530 if (isAllocaDerivedArg(I.getOperand(0))) {
531 // We can actually predict the result of comparisons between an
532 // alloca-derived value and null. Note that this fires regardless of
534 bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE;
535 SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType())
536 : ConstantInt::getFalse(I.getType());
540 // Finally check for SROA candidates in comparisons.
542 DenseMap<Value *, int>::iterator CostIt;
543 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
544 if (isa<ConstantPointerNull>(I.getOperand(1))) {
545 accumulateSROACost(CostIt, InlineConstants::InstrCost);
555 bool CallAnalyzer::visitSub(BinaryOperator &I) {
556 // Try to handle a special case: we can fold computing the difference of two
557 // constant-related pointers.
558 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
559 Value *LHSBase, *RHSBase;
560 APInt LHSOffset, RHSOffset;
561 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
563 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
564 if (RHSBase && LHSBase == RHSBase) {
565 // We have common bases, fold the subtract to a constant based on the
567 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
568 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
569 if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) {
570 SimplifiedValues[&I] = C;
571 ++NumConstantPtrDiffs;
577 // Otherwise, fall back to the generic logic for simplifying and handling
579 return Base::visitSub(I);
582 bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
583 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
584 if (!isa<Constant>(LHS))
585 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
587 if (!isa<Constant>(RHS))
588 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
590 Value *SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, TD);
591 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
592 SimplifiedValues[&I] = C;
596 // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
603 bool CallAnalyzer::visitLoad(LoadInst &I) {
605 DenseMap<Value *, int>::iterator CostIt;
606 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
608 accumulateSROACost(CostIt, InlineConstants::InstrCost);
618 bool CallAnalyzer::visitStore(StoreInst &I) {
620 DenseMap<Value *, int>::iterator CostIt;
621 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
623 accumulateSROACost(CostIt, InlineConstants::InstrCost);
633 bool CallAnalyzer::visitExtractValue(ExtractValueInst &I) {
634 // Constant folding for extract value is trivial.
635 Constant *C = dyn_cast<Constant>(I.getAggregateOperand());
637 C = SimplifiedValues.lookup(I.getAggregateOperand());
639 SimplifiedValues[&I] = ConstantExpr::getExtractValue(C, I.getIndices());
643 // SROA can look through these but give them a cost.
647 bool CallAnalyzer::visitInsertValue(InsertValueInst &I) {
648 // Constant folding for insert value is trivial.
649 Constant *AggC = dyn_cast<Constant>(I.getAggregateOperand());
651 AggC = SimplifiedValues.lookup(I.getAggregateOperand());
652 Constant *InsertedC = dyn_cast<Constant>(I.getInsertedValueOperand());
654 InsertedC = SimplifiedValues.lookup(I.getInsertedValueOperand());
655 if (AggC && InsertedC) {
656 SimplifiedValues[&I] = ConstantExpr::getInsertValue(AggC, InsertedC,
661 // SROA can look through these but give them a cost.
665 /// \brief Try to simplify a call site.
667 /// Takes a concrete function and callsite and tries to actually simplify it by
668 /// analyzing the arguments and call itself with instsimplify. Returns true if
669 /// it has simplified the callsite to some other entity (a constant), making it
671 bool CallAnalyzer::simplifyCallSite(Function *F, CallSite CS) {
672 // FIXME: Using the instsimplify logic directly for this is inefficient
673 // because we have to continually rebuild the argument list even when no
674 // simplifications can be performed. Until that is fixed with remapping
675 // inside of instsimplify, directly constant fold calls here.
676 if (!canConstantFoldCallTo(F))
679 // Try to re-map the arguments to constants.
680 SmallVector<Constant *, 4> ConstantArgs;
681 ConstantArgs.reserve(CS.arg_size());
682 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
684 Constant *C = dyn_cast<Constant>(*I);
686 C = dyn_cast_or_null<Constant>(SimplifiedValues.lookup(*I));
688 return false; // This argument doesn't map to a constant.
690 ConstantArgs.push_back(C);
692 if (Constant *C = ConstantFoldCall(F, ConstantArgs)) {
693 SimplifiedValues[CS.getInstruction()] = C;
700 bool CallAnalyzer::visitCallSite(CallSite CS) {
701 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->canReturnTwice() &&
702 !F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
703 Attribute::ReturnsTwice)) {
704 // This aborts the entire analysis.
705 ExposesReturnsTwice = true;
709 cast<CallInst>(CS.getInstruction())->hasFnAttr(Attribute::NoDuplicate))
710 ContainsNoDuplicateCall = true;
712 if (Function *F = CS.getCalledFunction()) {
713 // When we have a concrete function, first try to simplify it directly.
714 if (simplifyCallSite(F, CS))
717 // Next check if it is an intrinsic we know about.
718 // FIXME: Lift this into part of the InstVisitor.
719 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
720 switch (II->getIntrinsicID()) {
722 return Base::visitCallSite(CS);
724 case Intrinsic::memset:
725 case Intrinsic::memcpy:
726 case Intrinsic::memmove:
727 // SROA can usually chew through these intrinsics, but they aren't free.
732 if (F == CS.getInstruction()->getParent()->getParent()) {
733 // This flag will fully abort the analysis, so don't bother with anything
735 IsRecursiveCall = true;
739 if (TTI.isLoweredToCall(F)) {
740 // We account for the average 1 instruction per call argument setup
742 Cost += CS.arg_size() * InlineConstants::InstrCost;
744 // Everything other than inline ASM will also have a significant cost
745 // merely from making the call.
746 if (!isa<InlineAsm>(CS.getCalledValue()))
747 Cost += InlineConstants::CallPenalty;
750 return Base::visitCallSite(CS);
753 // Otherwise we're in a very special case -- an indirect function call. See
754 // if we can be particularly clever about this.
755 Value *Callee = CS.getCalledValue();
757 // First, pay the price of the argument setup. We account for the average
758 // 1 instruction per call argument setup here.
759 Cost += CS.arg_size() * InlineConstants::InstrCost;
761 // Next, check if this happens to be an indirect function call to a known
762 // function in this inline context. If not, we've done all we can.
763 Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee));
765 return Base::visitCallSite(CS);
767 // If we have a constant that we are calling as a function, we can peer
768 // through it and see the function target. This happens not infrequently
769 // during devirtualization and so we want to give it a hefty bonus for
770 // inlining, but cap that bonus in the event that inlining wouldn't pan
771 // out. Pretend to inline the function, with a custom threshold.
772 CallAnalyzer CA(TD, TTI, *F, InlineConstants::IndirectCallThreshold);
773 if (CA.analyzeCall(CS)) {
774 // We were able to inline the indirect call! Subtract the cost from the
775 // bonus we want to apply, but don't go below zero.
776 Cost -= std::max(0, InlineConstants::IndirectCallThreshold - CA.getCost());
779 return Base::visitCallSite(CS);
782 bool CallAnalyzer::visitInstruction(Instruction &I) {
783 // Some instructions are free. All of the free intrinsics can also be
784 // handled by SROA, etc.
785 if (TargetTransformInfo::TCC_Free == TTI.getUserCost(&I))
788 // We found something we don't understand or can't handle. Mark any SROA-able
789 // values in the operand list as no longer viable.
790 for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI)
797 /// \brief Analyze a basic block for its contribution to the inline cost.
799 /// This method walks the analyzer over every instruction in the given basic
800 /// block and accounts for their cost during inlining at this callsite. It
801 /// aborts early if the threshold has been exceeded or an impossible to inline
802 /// construct has been detected. It returns false if inlining is no longer
803 /// viable, and true if inlining remains viable.
804 bool CallAnalyzer::analyzeBlock(BasicBlock *BB) {
805 for (BasicBlock::iterator I = BB->begin(), E = llvm::prior(BB->end());
808 if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy())
809 ++NumVectorInstructions;
811 // If the instruction simplified to a constant, there is no cost to this
812 // instruction. Visit the instructions using our InstVisitor to account for
813 // all of the per-instruction logic. The visit tree returns true if we
814 // consumed the instruction in any way, and false if the instruction's base
815 // cost should count against inlining.
817 ++NumInstructionsSimplified;
819 Cost += InlineConstants::InstrCost;
821 // If the visit this instruction detected an uninlinable pattern, abort.
822 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca)
825 // If the caller is a recursive function then we don't want to inline
826 // functions which allocate a lot of stack space because it would increase
827 // the caller stack usage dramatically.
828 if (IsCallerRecursive &&
829 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller)
832 if (NumVectorInstructions > NumInstructions/2)
833 VectorBonus = FiftyPercentVectorBonus;
834 else if (NumVectorInstructions > NumInstructions/10)
835 VectorBonus = TenPercentVectorBonus;
839 // Check if we've past the threshold so we don't spin in huge basic
840 // blocks that will never inline.
841 if (Cost > (Threshold + VectorBonus))
848 /// \brief Compute the base pointer and cumulative constant offsets for V.
850 /// This strips all constant offsets off of V, leaving it the base pointer, and
851 /// accumulates the total constant offset applied in the returned constant. It
852 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
853 /// no constant offsets applied.
854 ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) {
855 if (!TD || !V->getType()->isPointerTy())
858 unsigned IntPtrWidth = TD->getPointerSizeInBits();
859 APInt Offset = APInt::getNullValue(IntPtrWidth);
861 // Even though we don't look through PHI nodes, we could be called on an
862 // instruction in an unreachable block, which may be on a cycle.
863 SmallPtrSet<Value *, 4> Visited;
866 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
867 if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset))
869 V = GEP->getPointerOperand();
870 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
871 V = cast<Operator>(V)->getOperand(0);
872 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
873 if (GA->mayBeOverridden())
875 V = GA->getAliasee();
879 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
880 } while (Visited.insert(V));
882 Type *IntPtrTy = TD->getIntPtrType(V->getContext());
883 return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset));
886 /// \brief Analyze a call site for potential inlining.
888 /// Returns true if inlining this call is viable, and false if it is not
889 /// viable. It computes the cost and adjusts the threshold based on numerous
890 /// factors and heuristics. If this method returns false but the computed cost
891 /// is below the computed threshold, then inlining was forcibly disabled by
892 /// some artifact of the routine.
893 bool CallAnalyzer::analyzeCall(CallSite CS) {
896 // Track whether the post-inlining function would have more than one basic
897 // block. A single basic block is often intended for inlining. Balloon the
898 // threshold by 50% until we pass the single-BB phase.
899 bool SingleBB = true;
900 int SingleBBBonus = Threshold / 2;
901 Threshold += SingleBBBonus;
903 // Perform some tweaks to the cost and threshold based on the direct
904 // callsite information.
906 // We want to more aggressively inline vector-dense kernels, so up the
907 // threshold, and we'll lower it if the % of vector instructions gets too
909 assert(NumInstructions == 0);
910 assert(NumVectorInstructions == 0);
911 FiftyPercentVectorBonus = Threshold;
912 TenPercentVectorBonus = Threshold / 2;
914 // Give out bonuses per argument, as the instructions setting them up will
915 // be gone after inlining.
916 for (unsigned I = 0, E = CS.arg_size(); I != E; ++I) {
917 if (TD && CS.isByValArgument(I)) {
918 // We approximate the number of loads and stores needed by dividing the
919 // size of the byval type by the target's pointer size.
920 PointerType *PTy = cast<PointerType>(CS.getArgument(I)->getType());
921 unsigned TypeSize = TD->getTypeSizeInBits(PTy->getElementType());
922 unsigned PointerSize = TD->getPointerSizeInBits();
924 unsigned NumStores = (TypeSize + PointerSize - 1) / PointerSize;
926 // If it generates more than 8 stores it is likely to be expanded as an
927 // inline memcpy so we take that as an upper bound. Otherwise we assume
928 // one load and one store per word copied.
929 // FIXME: The maxStoresPerMemcpy setting from the target should be used
930 // here instead of a magic number of 8, but it's not available via
932 NumStores = std::min(NumStores, 8U);
934 Cost -= 2 * NumStores * InlineConstants::InstrCost;
936 // For non-byval arguments subtract off one instruction per call
938 Cost -= InlineConstants::InstrCost;
942 // If there is only one call of the function, and it has internal linkage,
943 // the cost of inlining it drops dramatically.
944 bool OnlyOneCallAndLocalLinkage = F.hasLocalLinkage() && F.hasOneUse() &&
945 &F == CS.getCalledFunction();
946 if (OnlyOneCallAndLocalLinkage)
947 Cost += InlineConstants::LastCallToStaticBonus;
949 // If the instruction after the call, or if the normal destination of the
950 // invoke is an unreachable instruction, the function is noreturn. As such,
951 // there is little point in inlining this unless there is literally zero
953 Instruction *Instr = CS.getInstruction();
954 if (InvokeInst *II = dyn_cast<InvokeInst>(Instr)) {
955 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
957 } else if (isa<UnreachableInst>(++BasicBlock::iterator(Instr)))
960 // If this function uses the coldcc calling convention, prefer not to inline
962 if (F.getCallingConv() == CallingConv::Cold)
963 Cost += InlineConstants::ColdccPenalty;
965 // Check if we're done. This can happen due to bonuses and penalties.
966 if (Cost > Threshold)
972 Function *Caller = CS.getInstruction()->getParent()->getParent();
973 // Check if the caller function is recursive itself.
974 for (Value::use_iterator U = Caller->use_begin(), E = Caller->use_end();
976 CallSite Site(cast<Value>(*U));
979 Instruction *I = Site.getInstruction();
980 if (I->getParent()->getParent() == Caller) {
981 IsCallerRecursive = true;
986 // Track whether we've seen a return instruction. The first return
987 // instruction is free, as at least one will usually disappear in inlining.
988 bool HasReturn = false;
990 // Populate our simplified values by mapping from function arguments to call
991 // arguments with known important simplifications.
992 CallSite::arg_iterator CAI = CS.arg_begin();
993 for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end();
994 FAI != FAE; ++FAI, ++CAI) {
995 assert(CAI != CS.arg_end());
996 if (Constant *C = dyn_cast<Constant>(CAI))
997 SimplifiedValues[FAI] = C;
999 Value *PtrArg = *CAI;
1000 if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) {
1001 ConstantOffsetPtrs[FAI] = std::make_pair(PtrArg, C->getValue());
1003 // We can SROA any pointer arguments derived from alloca instructions.
1004 if (isa<AllocaInst>(PtrArg)) {
1005 SROAArgValues[FAI] = PtrArg;
1006 SROAArgCosts[PtrArg] = 0;
1010 NumConstantArgs = SimplifiedValues.size();
1011 NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size();
1012 NumAllocaArgs = SROAArgValues.size();
1014 // The worklist of live basic blocks in the callee *after* inlining. We avoid
1015 // adding basic blocks of the callee which can be proven to be dead for this
1016 // particular call site in order to get more accurate cost estimates. This
1017 // requires a somewhat heavyweight iteration pattern: we need to walk the
1018 // basic blocks in a breadth-first order as we insert live successors. To
1019 // accomplish this, prioritizing for small iterations because we exit after
1020 // crossing our threshold, we use a small-size optimized SetVector.
1021 typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>,
1022 SmallPtrSet<BasicBlock *, 16> > BBSetVector;
1023 BBSetVector BBWorklist;
1024 BBWorklist.insert(&F.getEntryBlock());
1025 // Note that we *must not* cache the size, this loop grows the worklist.
1026 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
1027 // Bail out the moment we cross the threshold. This means we'll under-count
1028 // the cost, but only when undercounting doesn't matter.
1029 if (Cost > (Threshold + VectorBonus))
1032 BasicBlock *BB = BBWorklist[Idx];
1036 // Handle the terminator cost here where we can track returns and other
1037 // function-wide constructs.
1038 TerminatorInst *TI = BB->getTerminator();
1040 // We never want to inline functions that contain an indirectbr. This is
1041 // incorrect because all the blockaddress's (in static global initializers
1042 // for example) would be referring to the original function, and this
1043 // indirect jump would jump from the inlined copy of the function into the
1044 // original function which is extremely undefined behavior.
1045 // FIXME: This logic isn't really right; we can safely inline functions
1046 // with indirectbr's as long as no other function or global references the
1047 // blockaddress of a block within the current function. And as a QOI issue,
1048 // if someone is using a blockaddress without an indirectbr, and that
1049 // reference somehow ends up in another function or global, we probably
1050 // don't want to inline this function.
1051 if (isa<IndirectBrInst>(TI))
1054 if (!HasReturn && isa<ReturnInst>(TI))
1057 Cost += InlineConstants::InstrCost;
1059 // Analyze the cost of this block. If we blow through the threshold, this
1060 // returns false, and we can bail on out.
1061 if (!analyzeBlock(BB)) {
1062 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca)
1065 // If the caller is a recursive function then we don't want to inline
1066 // functions which allocate a lot of stack space because it would increase
1067 // the caller stack usage dramatically.
1068 if (IsCallerRecursive &&
1069 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller)
1075 // Add in the live successors by first checking whether we have terminator
1076 // that may be simplified based on the values simplified by this call.
1077 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1078 if (BI->isConditional()) {
1079 Value *Cond = BI->getCondition();
1080 if (ConstantInt *SimpleCond
1081 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
1082 BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0));
1086 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1087 Value *Cond = SI->getCondition();
1088 if (ConstantInt *SimpleCond
1089 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
1090 BBWorklist.insert(SI->findCaseValue(SimpleCond).getCaseSuccessor());
1095 // If we're unable to select a particular successor, just count all of
1097 for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize;
1099 BBWorklist.insert(TI->getSuccessor(TIdx));
1101 // If we had any successors at this point, than post-inlining is likely to
1102 // have them as well. Note that we assume any basic blocks which existed
1103 // due to branches or switches which folded above will also fold after
1105 if (SingleBB && TI->getNumSuccessors() > 1) {
1106 // Take off the bonus we applied to the threshold.
1107 Threshold -= SingleBBBonus;
1112 // If this is a noduplicate call, we can still inline as long as
1113 // inlining this would cause the removal of the caller (so the instruction
1114 // is not actually duplicated, just moved).
1115 if (!OnlyOneCallAndLocalLinkage && ContainsNoDuplicateCall)
1118 Threshold += VectorBonus;
1120 return Cost < Threshold;
1123 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1124 /// \brief Dump stats about this call's analysis.
1125 void CallAnalyzer::dump() {
1126 #define DEBUG_PRINT_STAT(x) llvm::dbgs() << " " #x ": " << x << "\n"
1127 DEBUG_PRINT_STAT(NumConstantArgs);
1128 DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs);
1129 DEBUG_PRINT_STAT(NumAllocaArgs);
1130 DEBUG_PRINT_STAT(NumConstantPtrCmps);
1131 DEBUG_PRINT_STAT(NumConstantPtrDiffs);
1132 DEBUG_PRINT_STAT(NumInstructionsSimplified);
1133 DEBUG_PRINT_STAT(SROACostSavings);
1134 DEBUG_PRINT_STAT(SROACostSavingsLost);
1135 DEBUG_PRINT_STAT(ContainsNoDuplicateCall);
1136 #undef DEBUG_PRINT_STAT
1140 INITIALIZE_PASS_BEGIN(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis",
1142 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
1143 INITIALIZE_PASS_END(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis",
1146 char InlineCostAnalysis::ID = 0;
1148 InlineCostAnalysis::InlineCostAnalysis() : CallGraphSCCPass(ID), TD(0) {}
1150 InlineCostAnalysis::~InlineCostAnalysis() {}
1152 void InlineCostAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
1153 AU.setPreservesAll();
1154 AU.addRequired<TargetTransformInfo>();
1155 CallGraphSCCPass::getAnalysisUsage(AU);
1158 bool InlineCostAnalysis::runOnSCC(CallGraphSCC &SCC) {
1159 TD = getAnalysisIfAvailable<DataLayout>();
1160 TTI = &getAnalysis<TargetTransformInfo>();
1164 InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, int Threshold) {
1165 return getInlineCost(CS, CS.getCalledFunction(), Threshold);
1168 InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, Function *Callee,
1170 // Cannot inline indirect calls.
1172 return llvm::InlineCost::getNever();
1174 // Calls to functions with always-inline attributes should be inlined
1175 // whenever possible.
1176 if (Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1177 Attribute::AlwaysInline)) {
1178 if (isInlineViable(*Callee))
1179 return llvm::InlineCost::getAlways();
1180 return llvm::InlineCost::getNever();
1183 // Don't inline functions which can be redefined at link-time to mean
1184 // something else. Don't inline functions marked noinline or call sites
1186 if (Callee->mayBeOverridden() ||
1187 Callee->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1188 Attribute::NoInline) ||
1190 return llvm::InlineCost::getNever();
1192 DEBUG(llvm::dbgs() << " Analyzing call of " << Callee->getName()
1195 CallAnalyzer CA(TD, *TTI, *Callee, Threshold);
1196 bool ShouldInline = CA.analyzeCall(CS);
1200 // Check if there was a reason to force inlining or no inlining.
1201 if (!ShouldInline && CA.getCost() < CA.getThreshold())
1202 return InlineCost::getNever();
1203 if (ShouldInline && CA.getCost() >= CA.getThreshold())
1204 return InlineCost::getAlways();
1206 return llvm::InlineCost::get(CA.getCost(), CA.getThreshold());
1209 bool InlineCostAnalysis::isInlineViable(Function &F) {
1211 F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1212 Attribute::ReturnsTwice);
1213 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
1214 // Disallow inlining of functions which contain an indirect branch.
1215 if (isa<IndirectBrInst>(BI->getTerminator()))
1218 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
1224 // Disallow recursive calls.
1225 if (&F == CS.getCalledFunction())
1228 // Disallow calls which expose returns-twice to a function not previously
1229 // attributed as such.
1230 if (!ReturnsTwice && CS.isCall() &&
1231 cast<CallInst>(CS.getInstruction())->canReturnTwice())