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/Analysis/ConstantFolding.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Support/CallSite.h"
19 #include "llvm/Support/Debug.h"
20 #include "llvm/Support/InstVisitor.h"
21 #include "llvm/Support/GetElementPtrTypeIterator.h"
22 #include "llvm/Support/raw_ostream.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/IntrinsicInst.h"
25 #include "llvm/Operator.h"
26 #include "llvm/GlobalAlias.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SetVector.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/Statistic.h"
36 STATISTIC(NumCallsAnalyzed, "Number of call sites analyzed");
40 class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> {
41 typedef InstVisitor<CallAnalyzer, bool> Base;
42 friend class InstVisitor<CallAnalyzer, bool>;
44 // TargetData if available, or null.
45 const TargetData *const TD;
47 // The called function.
52 const bool AlwaysInline;
55 bool ExposesReturnsTwice;
56 bool HasDynamicAlloca;
57 unsigned NumInstructions, NumVectorInstructions;
58 int FiftyPercentVectorBonus, TenPercentVectorBonus;
61 // While we walk the potentially-inlined instructions, we build up and
62 // maintain a mapping of simplified values specific to this callsite. The
63 // idea is to propagate any special information we have about arguments to
64 // this call through the inlinable section of the function, and account for
65 // likely simplifications post-inlining. The most important aspect we track
66 // is CFG altering simplifications -- when we prove a basic block dead, that
67 // can cause dramatic shifts in the cost of inlining a function.
68 DenseMap<Value *, Constant *> SimplifiedValues;
70 // Keep track of the values which map back (through function arguments) to
71 // allocas on the caller stack which could be simplified through SROA.
72 DenseMap<Value *, Value *> SROAArgValues;
74 // The mapping of caller Alloca values to their accumulated cost savings. If
75 // we have to disable SROA for one of the allocas, this tells us how much
76 // cost must be added.
77 DenseMap<Value *, int> SROAArgCosts;
79 // Keep track of values which map to a pointer base and constant offset.
80 DenseMap<Value *, std::pair<Value *, APInt> > ConstantOffsetPtrs;
82 // Custom simplification helper routines.
83 bool isAllocaDerivedArg(Value *V);
84 bool lookupSROAArgAndCost(Value *V, Value *&Arg,
85 DenseMap<Value *, int>::iterator &CostIt);
86 void disableSROA(DenseMap<Value *, int>::iterator CostIt);
87 void disableSROA(Value *V);
88 void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
90 bool handleSROACandidate(bool IsSROAValid,
91 DenseMap<Value *, int>::iterator CostIt,
93 bool isGEPOffsetConstant(GetElementPtrInst &GEP);
94 bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset);
95 ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V);
97 // Custom analysis routines.
98 bool analyzeBlock(BasicBlock *BB);
100 // Disable several entry points to the visitor so we don't accidentally use
101 // them by declaring but not defining them here.
102 void visit(Module *); void visit(Module &);
103 void visit(Function *); void visit(Function &);
104 void visit(BasicBlock *); void visit(BasicBlock &);
106 // Provide base case for our instruction visit.
107 bool visitInstruction(Instruction &I);
109 // Our visit overrides.
110 bool visitAlloca(AllocaInst &I);
111 bool visitPHI(PHINode &I);
112 bool visitGetElementPtr(GetElementPtrInst &I);
113 bool visitBitCast(BitCastInst &I);
114 bool visitPtrToInt(PtrToIntInst &I);
115 bool visitIntToPtr(IntToPtrInst &I);
116 bool visitCastInst(CastInst &I);
117 bool visitUnaryInstruction(UnaryInstruction &I);
118 bool visitICmp(ICmpInst &I);
119 bool visitSub(BinaryOperator &I);
120 bool visitBinaryOperator(BinaryOperator &I);
121 bool visitLoad(LoadInst &I);
122 bool visitStore(StoreInst &I);
123 bool visitCallSite(CallSite CS);
126 CallAnalyzer(const TargetData *TD, Function &Callee, int Threshold)
127 : TD(TD), F(Callee), Threshold(Threshold), Cost(0),
128 AlwaysInline(F.hasFnAttr(Attribute::AlwaysInline)),
129 IsRecursive(false), ExposesReturnsTwice(false), HasDynamicAlloca(false),
130 NumInstructions(0), NumVectorInstructions(0),
131 FiftyPercentVectorBonus(0), TenPercentVectorBonus(0), VectorBonus(0),
132 NumConstantArgs(0), NumConstantOffsetPtrArgs(0), NumAllocaArgs(0),
133 NumConstantPtrCmps(0), NumConstantPtrDiffs(0),
134 NumInstructionsSimplified(0), SROACostSavings(0), SROACostSavingsLost(0) {
137 bool analyzeCall(CallSite CS);
139 int getThreshold() { return Threshold; }
140 int getCost() { return Cost; }
142 // Keep a bunch of stats about the cost savings found so we can print them
143 // out when debugging.
144 unsigned NumConstantArgs;
145 unsigned NumConstantOffsetPtrArgs;
146 unsigned NumAllocaArgs;
147 unsigned NumConstantPtrCmps;
148 unsigned NumConstantPtrDiffs;
149 unsigned NumInstructionsSimplified;
150 unsigned SROACostSavings;
151 unsigned SROACostSavingsLost;
158 /// \brief Test whether the given value is an Alloca-derived function argument.
159 bool CallAnalyzer::isAllocaDerivedArg(Value *V) {
160 return SROAArgValues.count(V);
163 /// \brief Lookup the SROA-candidate argument and cost iterator which V maps to.
164 /// Returns false if V does not map to a SROA-candidate.
165 bool CallAnalyzer::lookupSROAArgAndCost(
166 Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) {
167 if (SROAArgValues.empty() || SROAArgCosts.empty())
170 DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V);
171 if (ArgIt == SROAArgValues.end())
175 CostIt = SROAArgCosts.find(Arg);
176 return CostIt != SROAArgCosts.end();
179 /// \brief Disable SROA for the candidate marked by this cost iterator.
181 /// This markes the candidate as no longer viable for SROA, and adds the cost
182 /// savings associated with it back into the inline cost measurement.
183 void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) {
184 // If we're no longer able to perform SROA we need to undo its cost savings
185 // and prevent subsequent analysis.
186 Cost += CostIt->second;
187 SROACostSavings -= CostIt->second;
188 SROACostSavingsLost += CostIt->second;
189 SROAArgCosts.erase(CostIt);
192 /// \brief If 'V' maps to a SROA candidate, disable SROA for it.
193 void CallAnalyzer::disableSROA(Value *V) {
195 DenseMap<Value *, int>::iterator CostIt;
196 if (lookupSROAArgAndCost(V, SROAArg, CostIt))
200 /// \brief Accumulate the given cost for a particular SROA candidate.
201 void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
202 int InstructionCost) {
203 CostIt->second += InstructionCost;
204 SROACostSavings += InstructionCost;
207 /// \brief Helper for the common pattern of handling a SROA candidate.
208 /// Either accumulates the cost savings if the SROA remains valid, or disables
209 /// SROA for the candidate.
210 bool CallAnalyzer::handleSROACandidate(bool IsSROAValid,
211 DenseMap<Value *, int>::iterator CostIt,
212 int InstructionCost) {
214 accumulateSROACost(CostIt, InstructionCost);
222 /// \brief Check whether a GEP's indices are all constant.
224 /// Respects any simplified values known during the analysis of this callsite.
225 bool CallAnalyzer::isGEPOffsetConstant(GetElementPtrInst &GEP) {
226 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
227 if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I))
233 /// \brief Accumulate a constant GEP offset into an APInt if possible.
235 /// Returns false if unable to compute the offset for any reason. Respects any
236 /// simplified values known during the analysis of this callsite.
237 bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) {
241 unsigned IntPtrWidth = TD->getPointerSizeInBits();
242 assert(IntPtrWidth == Offset.getBitWidth());
244 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
246 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
248 if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand()))
249 OpC = dyn_cast<ConstantInt>(SimpleOp);
252 if (OpC->isZero()) continue;
254 // Handle a struct index, which adds its field offset to the pointer.
255 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
256 unsigned ElementIdx = OpC->getZExtValue();
257 const StructLayout *SL = TD->getStructLayout(STy);
258 Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx));
262 APInt TypeSize(IntPtrWidth, TD->getTypeAllocSize(GTI.getIndexedType()));
263 Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize;
268 bool CallAnalyzer::visitAlloca(AllocaInst &I) {
269 // FIXME: Check whether inlining will turn a dynamic alloca into a static
270 // alloca, and handle that case.
272 // We will happily inline static alloca instructions or dynamic alloca
273 // instructions in always-inline situations.
274 if (AlwaysInline || I.isStaticAlloca())
275 return Base::visitAlloca(I);
277 // FIXME: This is overly conservative. Dynamic allocas are inefficient for
278 // a variety of reasons, and so we would like to not inline them into
279 // functions which don't currently have a dynamic alloca. This simply
280 // disables inlining altogether in the presence of a dynamic alloca.
281 HasDynamicAlloca = true;
285 bool CallAnalyzer::visitPHI(PHINode &I) {
286 // FIXME: We should potentially be tracking values through phi nodes,
287 // especially when they collapse to a single value due to deleted CFG edges
290 // FIXME: We need to propagate SROA *disabling* through phi nodes, even
291 // though we don't want to propagate it's bonuses. The idea is to disable
292 // SROA if it *might* be used in an inappropriate manner.
294 // Phi nodes are always zero-cost.
298 bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) {
300 DenseMap<Value *, int>::iterator CostIt;
301 bool SROACandidate = lookupSROAArgAndCost(I.getPointerOperand(),
304 // Try to fold GEPs of constant-offset call site argument pointers. This
305 // requires target data and inbounds GEPs.
306 if (TD && I.isInBounds()) {
307 // Check if we have a base + offset for the pointer.
308 Value *Ptr = I.getPointerOperand();
309 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr);
310 if (BaseAndOffset.first) {
311 // Check if the offset of this GEP is constant, and if so accumulate it
313 if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) {
314 // Non-constant GEPs aren't folded, and disable SROA.
320 // Add the result as a new mapping to Base + Offset.
321 ConstantOffsetPtrs[&I] = BaseAndOffset;
323 // Also handle SROA candidates here, we already know that the GEP is
324 // all-constant indexed.
326 SROAArgValues[&I] = SROAArg;
332 if (isGEPOffsetConstant(I)) {
334 SROAArgValues[&I] = SROAArg;
336 // Constant GEPs are modeled as free.
340 // Variable GEPs will require math and will disable SROA.
346 bool CallAnalyzer::visitBitCast(BitCastInst &I) {
347 // Propagate constants through bitcasts.
348 if (Constant *COp = dyn_cast<Constant>(I.getOperand(0)))
349 if (Constant *C = ConstantExpr::getBitCast(COp, I.getType())) {
350 SimplifiedValues[&I] = C;
354 // Track base/offsets through casts
355 std::pair<Value *, APInt> BaseAndOffset
356 = ConstantOffsetPtrs.lookup(I.getOperand(0));
357 // Casts don't change the offset, just wrap it up.
358 if (BaseAndOffset.first)
359 ConstantOffsetPtrs[&I] = BaseAndOffset;
361 // Also look for SROA candidates here.
363 DenseMap<Value *, int>::iterator CostIt;
364 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
365 SROAArgValues[&I] = SROAArg;
367 // Bitcasts are always zero cost.
371 bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) {
372 // Propagate constants through ptrtoint.
373 if (Constant *COp = dyn_cast<Constant>(I.getOperand(0)))
374 if (Constant *C = ConstantExpr::getPtrToInt(COp, I.getType())) {
375 SimplifiedValues[&I] = C;
379 // Track base/offset pairs when converted to a plain integer provided the
380 // integer is large enough to represent the pointer.
381 unsigned IntegerSize = I.getType()->getScalarSizeInBits();
382 if (TD && IntegerSize >= TD->getPointerSizeInBits()) {
383 std::pair<Value *, APInt> BaseAndOffset
384 = ConstantOffsetPtrs.lookup(I.getOperand(0));
385 if (BaseAndOffset.first)
386 ConstantOffsetPtrs[&I] = BaseAndOffset;
389 // This is really weird. Technically, ptrtoint will disable SROA. However,
390 // unless that ptrtoint is *used* somewhere in the live basic blocks after
391 // inlining, it will be nuked, and SROA should proceed. All of the uses which
392 // would block SROA would also block SROA if applied directly to a pointer,
393 // and so we can just add the integer in here. The only places where SROA is
394 // preserved either cannot fire on an integer, or won't in-and-of themselves
395 // disable SROA (ext) w/o some later use that we would see and disable.
397 DenseMap<Value *, int>::iterator CostIt;
398 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
399 SROAArgValues[&I] = SROAArg;
401 // A ptrtoint cast is free so long as the result is large enough to store the
402 // pointer, and a legal integer type.
403 return TD && TD->isLegalInteger(IntegerSize) &&
404 IntegerSize >= TD->getPointerSizeInBits();
407 bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) {
408 // Propagate constants through ptrtoint.
409 if (Constant *COp = dyn_cast<Constant>(I.getOperand(0)))
410 if (Constant *C = ConstantExpr::getIntToPtr(COp, I.getType())) {
411 SimplifiedValues[&I] = C;
415 // Track base/offset pairs when round-tripped through a pointer without
416 // modifications provided the integer is not too large.
417 Value *Op = I.getOperand(0);
418 unsigned IntegerSize = Op->getType()->getScalarSizeInBits();
419 if (TD && IntegerSize <= TD->getPointerSizeInBits()) {
420 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op);
421 if (BaseAndOffset.first)
422 ConstantOffsetPtrs[&I] = BaseAndOffset;
425 // "Propagate" SROA here in the same manner as we do for ptrtoint above.
427 DenseMap<Value *, int>::iterator CostIt;
428 if (lookupSROAArgAndCost(Op, SROAArg, CostIt))
429 SROAArgValues[&I] = SROAArg;
431 // An inttoptr cast is free so long as the input is a legal integer type
432 // which doesn't contain values outside the range of a pointer.
433 return TD && TD->isLegalInteger(IntegerSize) &&
434 IntegerSize <= TD->getPointerSizeInBits();
437 bool CallAnalyzer::visitCastInst(CastInst &I) {
438 // Propagate constants through ptrtoint.
439 if (Constant *COp = dyn_cast<Constant>(I.getOperand(0)))
440 if (Constant *C = ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) {
441 SimplifiedValues[&I] = C;
445 // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere.
446 disableSROA(I.getOperand(0));
448 // No-op casts don't have any cost.
449 if (I.isLosslessCast())
452 // trunc to a native type is free (assuming the target has compare and
453 // shift-right of the same width).
454 if (TD && isa<TruncInst>(I) &&
455 TD->isLegalInteger(TD->getTypeSizeInBits(I.getType())))
458 // Result of a cmp instruction is often extended (to be used by other
459 // cmp instructions, logical or return instructions). These are usually
460 // no-ops on most sane targets.
461 if (isa<CmpInst>(I.getOperand(0)))
464 // Assume the rest of the casts require work.
468 bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) {
469 Value *Operand = I.getOperand(0);
470 Constant *Ops[1] = { dyn_cast<Constant>(Operand) };
471 if (Ops[0] || (Ops[0] = SimplifiedValues.lookup(Operand)))
472 if (Constant *C = ConstantFoldInstOperands(I.getOpcode(), I.getType(),
474 SimplifiedValues[&I] = C;
478 // Disable any SROA on the argument to arbitrary unary operators.
479 disableSROA(Operand);
484 bool CallAnalyzer::visitICmp(ICmpInst &I) {
485 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
486 // First try to handle simplified comparisons.
487 if (!isa<Constant>(LHS))
488 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
490 if (!isa<Constant>(RHS))
491 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
493 if (Constant *CLHS = dyn_cast<Constant>(LHS))
494 if (Constant *CRHS = dyn_cast<Constant>(RHS))
495 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
496 SimplifiedValues[&I] = C;
500 // Otherwise look for a comparison between constant offset pointers with
502 Value *LHSBase, *RHSBase;
503 APInt LHSOffset, RHSOffset;
504 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
506 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
507 if (RHSBase && LHSBase == RHSBase) {
508 // We have common bases, fold the icmp to a constant based on the
510 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
511 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
512 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
513 SimplifiedValues[&I] = C;
514 ++NumConstantPtrCmps;
520 // If the comparison is an equality comparison with null, we can simplify it
521 // for any alloca-derived argument.
522 if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1)))
523 if (isAllocaDerivedArg(I.getOperand(0))) {
524 // We can actually predict the result of comparisons between an
525 // alloca-derived value and null. Note that this fires regardless of
527 bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE;
528 SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType())
529 : ConstantInt::getFalse(I.getType());
533 // Finally check for SROA candidates in comparisons.
535 DenseMap<Value *, int>::iterator CostIt;
536 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
537 if (isa<ConstantPointerNull>(I.getOperand(1))) {
538 accumulateSROACost(CostIt, InlineConstants::InstrCost);
548 bool CallAnalyzer::visitSub(BinaryOperator &I) {
549 // Try to handle a special case: we can fold computing the difference of two
550 // constant-related pointers.
551 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
552 Value *LHSBase, *RHSBase;
553 APInt LHSOffset, RHSOffset;
554 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
556 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
557 if (RHSBase && LHSBase == RHSBase) {
558 // We have common bases, fold the subtract to a constant based on the
560 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
561 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
562 if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) {
563 SimplifiedValues[&I] = C;
564 ++NumConstantPtrDiffs;
570 // Otherwise, fall back to the generic logic for simplifying and handling
572 return Base::visitSub(I);
575 bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
576 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
577 if (!isa<Constant>(LHS))
578 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
580 if (!isa<Constant>(RHS))
581 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
583 Value *SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, TD);
584 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
585 SimplifiedValues[&I] = C;
589 // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
596 bool CallAnalyzer::visitLoad(LoadInst &I) {
598 DenseMap<Value *, int>::iterator CostIt;
599 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
601 accumulateSROACost(CostIt, InlineConstants::InstrCost);
611 bool CallAnalyzer::visitStore(StoreInst &I) {
613 DenseMap<Value *, int>::iterator CostIt;
614 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
616 accumulateSROACost(CostIt, InlineConstants::InstrCost);
626 bool CallAnalyzer::visitCallSite(CallSite CS) {
627 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->canReturnTwice() &&
628 !F.hasFnAttr(Attribute::ReturnsTwice)) {
629 // This aborts the entire analysis.
630 ExposesReturnsTwice = true;
634 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
635 switch (II->getIntrinsicID()) {
637 return Base::visitCallSite(CS);
639 case Intrinsic::dbg_declare:
640 case Intrinsic::dbg_value:
641 case Intrinsic::invariant_start:
642 case Intrinsic::invariant_end:
643 case Intrinsic::lifetime_start:
644 case Intrinsic::lifetime_end:
645 case Intrinsic::memset:
646 case Intrinsic::memcpy:
647 case Intrinsic::memmove:
648 case Intrinsic::objectsize:
649 case Intrinsic::ptr_annotation:
650 case Intrinsic::var_annotation:
651 // SROA can usually chew through these intrinsics and they have no cost
652 // so don't pay the price of analyzing them in detail.
657 if (Function *F = CS.getCalledFunction()) {
658 if (F == CS.getInstruction()->getParent()->getParent()) {
659 // This flag will fully abort the analysis, so don't bother with anything
665 if (!callIsSmall(F)) {
666 // We account for the average 1 instruction per call argument setup
668 Cost += CS.arg_size() * InlineConstants::InstrCost;
670 // Everything other than inline ASM will also have a significant cost
671 // merely from making the call.
672 if (!isa<InlineAsm>(CS.getCalledValue()))
673 Cost += InlineConstants::CallPenalty;
676 return Base::visitCallSite(CS);
679 // Otherwise we're in a very special case -- an indirect function call. See
680 // if we can be particularly clever about this.
681 Value *Callee = CS.getCalledValue();
683 // First, pay the price of the argument setup. We account for the average
684 // 1 instruction per call argument setup here.
685 Cost += CS.arg_size() * InlineConstants::InstrCost;
687 // Next, check if this happens to be an indirect function call to a known
688 // function in this inline context. If not, we've done all we can.
689 Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee));
691 return Base::visitCallSite(CS);
693 // If we have a constant that we are calling as a function, we can peer
694 // through it and see the function target. This happens not infrequently
695 // during devirtualization and so we want to give it a hefty bonus for
696 // inlining, but cap that bonus in the event that inlining wouldn't pan
697 // out. Pretend to inline the function, with a custom threshold.
698 CallAnalyzer CA(TD, *F, InlineConstants::IndirectCallThreshold);
699 if (CA.analyzeCall(CS)) {
700 // We were able to inline the indirect call! Subtract the cost from the
701 // bonus we want to apply, but don't go below zero.
702 Cost -= std::max(0, InlineConstants::IndirectCallThreshold - CA.getCost());
705 return Base::visitCallSite(CS);
708 bool CallAnalyzer::visitInstruction(Instruction &I) {
709 // We found something we don't understand or can't handle. Mark any SROA-able
710 // values in the operand list as no longer viable.
711 for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI)
718 /// \brief Analyze a basic block for its contribution to the inline cost.
720 /// This method walks the analyzer over every instruction in the given basic
721 /// block and accounts for their cost during inlining at this callsite. It
722 /// aborts early if the threshold has been exceeded or an impossible to inline
723 /// construct has been detected. It returns false if inlining is no longer
724 /// viable, and true if inlining remains viable.
725 bool CallAnalyzer::analyzeBlock(BasicBlock *BB) {
726 for (BasicBlock::iterator I = BB->begin(), E = llvm::prior(BB->end());
729 if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy())
730 ++NumVectorInstructions;
732 // If the instruction simplified to a constant, there is no cost to this
733 // instruction. Visit the instructions using our InstVisitor to account for
734 // all of the per-instruction logic. The visit tree returns true if we
735 // consumed the instruction in any way, and false if the instruction's base
736 // cost should count against inlining.
738 ++NumInstructionsSimplified;
740 Cost += InlineConstants::InstrCost;
742 // If the visit this instruction detected an uninlinable pattern, abort.
743 if (IsRecursive || ExposesReturnsTwice || HasDynamicAlloca)
746 if (NumVectorInstructions > NumInstructions/2)
747 VectorBonus = FiftyPercentVectorBonus;
748 else if (NumVectorInstructions > NumInstructions/10)
749 VectorBonus = TenPercentVectorBonus;
753 // Check if we've past the threshold so we don't spin in huge basic
754 // blocks that will never inline.
755 if (!AlwaysInline && Cost > (Threshold + VectorBonus))
762 /// \brief Compute the base pointer and cumulative constant offsets for V.
764 /// This strips all constant offsets off of V, leaving it the base pointer, and
765 /// accumulates the total constant offset applied in the returned constant. It
766 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
767 /// no constant offsets applied.
768 ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) {
769 if (!TD || !V->getType()->isPointerTy())
772 unsigned IntPtrWidth = TD->getPointerSizeInBits();
773 APInt Offset = APInt::getNullValue(IntPtrWidth);
775 // Even though we don't look through PHI nodes, we could be called on an
776 // instruction in an unreachable block, which may be on a cycle.
777 SmallPtrSet<Value *, 4> Visited;
780 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
781 if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset))
783 V = GEP->getPointerOperand();
784 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
785 V = cast<Operator>(V)->getOperand(0);
786 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
787 if (GA->mayBeOverridden())
789 V = GA->getAliasee();
793 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
794 } while (Visited.insert(V));
796 Type *IntPtrTy = TD->getIntPtrType(V->getContext());
797 return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset));
800 /// \brief Analyze a call site for potential inlining.
802 /// Returns true if inlining this call is viable, and false if it is not
803 /// viable. It computes the cost and adjusts the threshold based on numerous
804 /// factors and heuristics. If this method returns false but the computed cost
805 /// is below the computed threshold, then inlining was forcibly disabled by
806 /// some artifact of the rountine.
807 bool CallAnalyzer::analyzeCall(CallSite CS) {
810 // Track whether the post-inlining function would have more than one basic
811 // block. A single basic block is often intended for inlining. Balloon the
812 // threshold by 50% until we pass the single-BB phase.
813 bool SingleBB = true;
814 int SingleBBBonus = Threshold / 2;
815 Threshold += SingleBBBonus;
817 // Unless we are always-inlining, perform some tweaks to the cost and
818 // threshold based on the direct callsite information.
820 // We want to more aggressively inline vector-dense kernels, so up the
821 // threshold, and we'll lower it if the % of vector instructions gets too
823 assert(NumInstructions == 0);
824 assert(NumVectorInstructions == 0);
825 FiftyPercentVectorBonus = Threshold;
826 TenPercentVectorBonus = Threshold / 2;
828 // Subtract off one instruction per call argument as those will be free after
830 Cost -= CS.arg_size() * InlineConstants::InstrCost;
832 // If there is only one call of the function, and it has internal linkage,
833 // the cost of inlining it drops dramatically.
834 if (F.hasLocalLinkage() && F.hasOneUse() && &F == CS.getCalledFunction())
835 Cost += InlineConstants::LastCallToStaticBonus;
837 // If the instruction after the call, or if the normal destination of the
838 // invoke is an unreachable instruction, the function is noreturn. As such,
839 // there is little point in inlining this unless there is literally zero cost.
840 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
841 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
843 } else if (isa<UnreachableInst>(++BasicBlock::iterator(CS.getInstruction())))
846 // If this function uses the coldcc calling convention, prefer not to inline
848 if (F.getCallingConv() == CallingConv::Cold)
849 Cost += InlineConstants::ColdccPenalty;
851 // Check if we're done. This can happen due to bonuses and penalties.
852 if (Cost > Threshold)
859 // Track whether we've seen a return instruction. The first return
860 // instruction is free, as at least one will usually disappear in inlining.
861 bool HasReturn = false;
863 // Populate our simplified values by mapping from function arguments to call
864 // arguments with known important simplifications.
865 CallSite::arg_iterator CAI = CS.arg_begin();
866 for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end();
867 FAI != FAE; ++FAI, ++CAI) {
868 assert(CAI != CS.arg_end());
869 if (Constant *C = dyn_cast<Constant>(CAI))
870 SimplifiedValues[FAI] = C;
872 Value *PtrArg = *CAI;
873 if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) {
874 ConstantOffsetPtrs[FAI] = std::make_pair(PtrArg, C->getValue());
876 // We can SROA any pointer arguments derived from alloca instructions.
877 if (isa<AllocaInst>(PtrArg)) {
878 SROAArgValues[FAI] = PtrArg;
879 SROAArgCosts[PtrArg] = 0;
883 NumConstantArgs = SimplifiedValues.size();
884 NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size();
885 NumAllocaArgs = SROAArgValues.size();
887 // The worklist of live basic blocks in the callee *after* inlining. We avoid
888 // adding basic blocks of the callee which can be proven to be dead for this
889 // particular call site in order to get more accurate cost estimates. This
890 // requires a somewhat heavyweight iteration pattern: we need to walk the
891 // basic blocks in a breadth-first order as we insert live successors. To
892 // accomplish this, prioritizing for small iterations because we exit after
893 // crossing our threshold, we use a small-size optimized SetVector.
894 typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>,
895 SmallPtrSet<BasicBlock *, 16> > BBSetVector;
896 BBSetVector BBWorklist;
897 BBWorklist.insert(&F.getEntryBlock());
898 // Note that we *must not* cache the size, this loop grows the worklist.
899 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
900 // Bail out the moment we cross the threshold. This means we'll under-count
901 // the cost, but only when undercounting doesn't matter.
902 if (!AlwaysInline && Cost > (Threshold + VectorBonus))
905 BasicBlock *BB = BBWorklist[Idx];
909 // Handle the terminator cost here where we can track returns and other
910 // function-wide constructs.
911 TerminatorInst *TI = BB->getTerminator();
913 // We never want to inline functions that contain an indirectbr. This is
914 // incorrect because all the blockaddress's (in static global initializers
915 // for example) would be referring to the original function, and this indirect
916 // jump would jump from the inlined copy of the function into the original
917 // function which is extremely undefined behavior.
918 // FIXME: This logic isn't really right; we can safely inline functions
919 // with indirectbr's as long as no other function or global references the
920 // blockaddress of a block within the current function. And as a QOI issue,
921 // if someone is using a blockaddress without an indirectbr, and that
922 // reference somehow ends up in another function or global, we probably
923 // don't want to inline this function.
924 if (isa<IndirectBrInst>(TI))
927 if (!HasReturn && isa<ReturnInst>(TI))
930 Cost += InlineConstants::InstrCost;
932 // Analyze the cost of this block. If we blow through the threshold, this
933 // returns false, and we can bail on out.
934 if (!analyzeBlock(BB)) {
935 if (IsRecursive || ExposesReturnsTwice || HasDynamicAlloca)
940 // Add in the live successors by first checking whether we have terminator
941 // that may be simplified based on the values simplified by this call.
942 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
943 if (BI->isConditional()) {
944 Value *Cond = BI->getCondition();
945 if (ConstantInt *SimpleCond
946 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
947 BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0));
951 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
952 Value *Cond = SI->getCondition();
953 if (ConstantInt *SimpleCond
954 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
955 BBWorklist.insert(SI->findCaseValue(SimpleCond).getCaseSuccessor());
960 // If we're unable to select a particular successor, just count all of
962 for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize; ++TIdx)
963 BBWorklist.insert(TI->getSuccessor(TIdx));
965 // If we had any successors at this point, than post-inlining is likely to
966 // have them as well. Note that we assume any basic blocks which existed
967 // due to branches or switches which folded above will also fold after
969 if (SingleBB && TI->getNumSuccessors() > 1) {
970 // Take off the bonus we applied to the threshold.
971 Threshold -= SingleBBBonus;
976 Threshold += VectorBonus;
978 return AlwaysInline || Cost < Threshold;
981 /// \brief Dump stats about this call's analysis.
982 void CallAnalyzer::dump() {
983 #define DEBUG_PRINT_STAT(x) llvm::dbgs() << " " #x ": " << x << "\n"
984 DEBUG_PRINT_STAT(NumConstantArgs);
985 DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs);
986 DEBUG_PRINT_STAT(NumAllocaArgs);
987 DEBUG_PRINT_STAT(NumConstantPtrCmps);
988 DEBUG_PRINT_STAT(NumConstantPtrDiffs);
989 DEBUG_PRINT_STAT(NumInstructionsSimplified);
990 DEBUG_PRINT_STAT(SROACostSavings);
991 DEBUG_PRINT_STAT(SROACostSavingsLost);
992 #undef DEBUG_PRINT_STAT
995 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS, int Threshold) {
996 return getInlineCost(CS, CS.getCalledFunction(), Threshold);
999 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS, Function *Callee,
1001 // Don't inline functions which can be redefined at link-time to mean
1002 // something else. Don't inline functions marked noinline or call sites
1004 if (!Callee || Callee->mayBeOverridden() ||
1005 Callee->hasFnAttr(Attribute::NoInline) || CS.isNoInline())
1006 return llvm::InlineCost::getNever();
1008 DEBUG(llvm::dbgs() << " Analyzing call of " << Callee->getName() << "...\n");
1010 CallAnalyzer CA(TD, *Callee, Threshold);
1011 bool ShouldInline = CA.analyzeCall(CS);
1015 // Check if there was a reason to force inlining or no inlining.
1016 if (!ShouldInline && CA.getCost() < CA.getThreshold())
1017 return InlineCost::getNever();
1018 if (ShouldInline && CA.getCost() >= CA.getThreshold())
1019 return InlineCost::getAlways();
1021 return llvm::InlineCost::get(CA.getCost(), CA.getThreshold());