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"
37 class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> {
38 typedef InstVisitor<CallAnalyzer, bool> Base;
39 friend class InstVisitor<CallAnalyzer, bool>;
41 // TargetData if available, or null.
42 const TargetData *const TD;
44 // The called function.
49 const bool AlwaysInline;
52 bool ExposesReturnsTwice;
53 bool HasDynamicAlloca;
54 unsigned NumInstructions, NumVectorInstructions;
55 int FiftyPercentVectorBonus, TenPercentVectorBonus;
58 // While we walk the potentially-inlined instructions, we build up and
59 // maintain a mapping of simplified values specific to this callsite. The
60 // idea is to propagate any special information we have about arguments to
61 // this call through the inlinable section of the function, and account for
62 // likely simplifications post-inlining. The most important aspect we track
63 // is CFG altering simplifications -- when we prove a basic block dead, that
64 // can cause dramatic shifts in the cost of inlining a function.
65 DenseMap<Value *, Constant *> SimplifiedValues;
67 // Keep track of the values which map back (through function arguments) to
68 // allocas on the caller stack which could be simplified through SROA.
69 DenseMap<Value *, Value *> SROAArgValues;
71 // The mapping of caller Alloca values to their accumulated cost savings. If
72 // we have to disable SROA for one of the allocas, this tells us how much
73 // cost must be added.
74 DenseMap<Value *, int> SROAArgCosts;
76 // Keep track of values which map to a pointer base and constant offset.
77 DenseMap<Value *, std::pair<Value *, APInt> > ConstantOffsetPtrs;
79 // Custom simplification helper routines.
80 bool isAllocaDerivedArg(Value *V);
81 bool lookupSROAArgAndCost(Value *V, Value *&Arg,
82 DenseMap<Value *, int>::iterator &CostIt);
83 void disableSROA(DenseMap<Value *, int>::iterator CostIt);
84 void disableSROA(Value *V);
85 void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
87 bool handleSROACandidate(bool IsSROAValid,
88 DenseMap<Value *, int>::iterator CostIt,
90 bool isGEPOffsetConstant(GetElementPtrInst &GEP);
91 bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset);
92 ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V);
94 // Custom analysis routines.
95 bool analyzeBlock(BasicBlock *BB);
97 // Disable several entry points to the visitor so we don't accidentally use
98 // them by declaring but not defining them here.
99 void visit(Module *); void visit(Module &);
100 void visit(Function *); void visit(Function &);
101 void visit(BasicBlock *); void visit(BasicBlock &);
103 // Provide base case for our instruction visit.
104 bool visitInstruction(Instruction &I);
106 // Our visit overrides.
107 bool visitAlloca(AllocaInst &I);
108 bool visitPHI(PHINode &I);
109 bool visitGetElementPtr(GetElementPtrInst &I);
110 bool visitBitCast(BitCastInst &I);
111 bool visitPtrToInt(PtrToIntInst &I);
112 bool visitIntToPtr(IntToPtrInst &I);
113 bool visitCastInst(CastInst &I);
114 bool visitUnaryInstruction(UnaryInstruction &I);
115 bool visitICmp(ICmpInst &I);
116 bool visitSub(BinaryOperator &I);
117 bool visitBinaryOperator(BinaryOperator &I);
118 bool visitLoad(LoadInst &I);
119 bool visitStore(StoreInst &I);
120 bool visitCallSite(CallSite CS);
123 CallAnalyzer(const TargetData *TD, Function &Callee, int Threshold)
124 : TD(TD), F(Callee), Threshold(Threshold), Cost(0),
125 AlwaysInline(F.hasFnAttr(Attribute::AlwaysInline)),
126 IsRecursive(false), ExposesReturnsTwice(false), HasDynamicAlloca(false),
127 NumInstructions(0), NumVectorInstructions(0),
128 FiftyPercentVectorBonus(0), TenPercentVectorBonus(0), VectorBonus(0),
129 NumConstantArgs(0), NumConstantOffsetPtrArgs(0), NumAllocaArgs(0),
130 NumConstantPtrCmps(0), NumConstantPtrDiffs(0),
131 NumInstructionsSimplified(0), SROACostSavings(0), SROACostSavingsLost(0) {
134 bool analyzeCall(CallSite CS);
136 int getThreshold() { return Threshold; }
137 int getCost() { return Cost; }
139 // Keep a bunch of stats about the cost savings found so we can print them
140 // out when debugging.
141 unsigned NumConstantArgs;
142 unsigned NumConstantOffsetPtrArgs;
143 unsigned NumAllocaArgs;
144 unsigned NumConstantPtrCmps;
145 unsigned NumConstantPtrDiffs;
146 unsigned NumInstructionsSimplified;
147 unsigned SROACostSavings;
148 unsigned SROACostSavingsLost;
155 /// \brief Test whether the given value is an Alloca-derived function argument.
156 bool CallAnalyzer::isAllocaDerivedArg(Value *V) {
157 return SROAArgValues.count(V);
160 /// \brief Lookup the SROA-candidate argument and cost iterator which V maps to.
161 /// Returns false if V does not map to a SROA-candidate.
162 bool CallAnalyzer::lookupSROAArgAndCost(
163 Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) {
164 if (SROAArgValues.empty() || SROAArgCosts.empty())
167 DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V);
168 if (ArgIt == SROAArgValues.end())
172 CostIt = SROAArgCosts.find(Arg);
173 return CostIt != SROAArgCosts.end();
176 /// \brief Disable SROA for the candidate marked by this cost iterator.
178 /// This markes the candidate as no longer viable for SROA, and adds the cost
179 /// savings associated with it back into the inline cost measurement.
180 void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) {
181 // If we're no longer able to perform SROA we need to undo its cost savings
182 // and prevent subsequent analysis.
183 Cost += CostIt->second;
184 SROACostSavings -= CostIt->second;
185 SROACostSavingsLost += CostIt->second;
186 SROAArgCosts.erase(CostIt);
189 /// \brief If 'V' maps to a SROA candidate, disable SROA for it.
190 void CallAnalyzer::disableSROA(Value *V) {
192 DenseMap<Value *, int>::iterator CostIt;
193 if (lookupSROAArgAndCost(V, SROAArg, CostIt))
197 /// \brief Accumulate the given cost for a particular SROA candidate.
198 void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
199 int InstructionCost) {
200 CostIt->second += InstructionCost;
201 SROACostSavings += InstructionCost;
204 /// \brief Helper for the common pattern of handling a SROA candidate.
205 /// Either accumulates the cost savings if the SROA remains valid, or disables
206 /// SROA for the candidate.
207 bool CallAnalyzer::handleSROACandidate(bool IsSROAValid,
208 DenseMap<Value *, int>::iterator CostIt,
209 int InstructionCost) {
211 accumulateSROACost(CostIt, InstructionCost);
219 /// \brief Check whether a GEP's indices are all constant.
221 /// Respects any simplified values known during the analysis of this callsite.
222 bool CallAnalyzer::isGEPOffsetConstant(GetElementPtrInst &GEP) {
223 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
224 if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I))
230 /// \brief Accumulate a constant GEP offset into an APInt if possible.
232 /// Returns false if unable to compute the offset for any reason. Respects any
233 /// simplified values known during the analysis of this callsite.
234 bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) {
238 unsigned IntPtrWidth = TD->getPointerSizeInBits();
239 assert(IntPtrWidth == Offset.getBitWidth());
241 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
243 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
245 if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand()))
246 OpC = dyn_cast<ConstantInt>(SimpleOp);
249 if (OpC->isZero()) continue;
251 // Handle a struct index, which adds its field offset to the pointer.
252 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
253 unsigned ElementIdx = OpC->getZExtValue();
254 const StructLayout *SL = TD->getStructLayout(STy);
255 Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx));
259 APInt TypeSize(IntPtrWidth, TD->getTypeAllocSize(GTI.getIndexedType()));
260 Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize;
265 bool CallAnalyzer::visitAlloca(AllocaInst &I) {
266 // FIXME: Check whether inlining will turn a dynamic alloca into a static
267 // alloca, and handle that case.
269 // We will happily inline static alloca instructions or dynamic alloca
270 // instructions in always-inline situations.
271 if (AlwaysInline || I.isStaticAlloca())
272 return Base::visitAlloca(I);
274 // FIXME: This is overly conservative. Dynamic allocas are inefficient for
275 // a variety of reasons, and so we would like to not inline them into
276 // functions which don't currently have a dynamic alloca. This simply
277 // disables inlining altogether in the presence of a dynamic alloca.
278 HasDynamicAlloca = true;
282 bool CallAnalyzer::visitPHI(PHINode &I) {
283 // FIXME: We should potentially be tracking values through phi nodes,
284 // especially when they collapse to a single value due to deleted CFG edges
287 // FIXME: We need to propagate SROA *disabling* through phi nodes, even
288 // though we don't want to propagate it's bonuses. The idea is to disable
289 // SROA if it *might* be used in an inappropriate manner.
291 // Phi nodes are always zero-cost.
295 bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) {
297 DenseMap<Value *, int>::iterator CostIt;
298 bool SROACandidate = lookupSROAArgAndCost(I.getPointerOperand(),
301 // Try to fold GEPs of constant-offset call site argument pointers. This
302 // requires target data and inbounds GEPs.
303 if (TD && I.isInBounds()) {
304 // Check if we have a base + offset for the pointer.
305 Value *Ptr = I.getPointerOperand();
306 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr);
307 if (BaseAndOffset.first) {
308 // Check if the offset of this GEP is constant, and if so accumulate it
310 if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) {
311 // Non-constant GEPs aren't folded, and disable SROA.
317 // Add the result as a new mapping to Base + Offset.
318 ConstantOffsetPtrs[&I] = BaseAndOffset;
320 // Also handle SROA candidates here, we already know that the GEP is
321 // all-constant indexed.
323 SROAArgValues[&I] = SROAArg;
329 if (isGEPOffsetConstant(I)) {
331 SROAArgValues[&I] = SROAArg;
333 // Constant GEPs are modeled as free.
337 // Variable GEPs will require math and will disable SROA.
343 bool CallAnalyzer::visitBitCast(BitCastInst &I) {
344 // Propagate constants through bitcasts.
345 if (Constant *COp = dyn_cast<Constant>(I.getOperand(0)))
346 if (Constant *C = ConstantExpr::getBitCast(COp, I.getType())) {
347 SimplifiedValues[&I] = C;
351 // Track base/offsets through casts
352 std::pair<Value *, APInt> BaseAndOffset
353 = ConstantOffsetPtrs.lookup(I.getOperand(0));
354 // Casts don't change the offset, just wrap it up.
355 if (BaseAndOffset.first)
356 ConstantOffsetPtrs[&I] = BaseAndOffset;
358 // Also look for SROA candidates here.
360 DenseMap<Value *, int>::iterator CostIt;
361 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
362 SROAArgValues[&I] = SROAArg;
364 // Bitcasts are always zero cost.
368 bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) {
369 // Propagate constants through ptrtoint.
370 if (Constant *COp = dyn_cast<Constant>(I.getOperand(0)))
371 if (Constant *C = ConstantExpr::getPtrToInt(COp, I.getType())) {
372 SimplifiedValues[&I] = C;
376 // Track base/offset pairs when converted to a plain integer provided the
377 // integer is large enough to represent the pointer.
378 unsigned IntegerSize = I.getType()->getScalarSizeInBits();
379 if (TD && IntegerSize >= TD->getPointerSizeInBits()) {
380 std::pair<Value *, APInt> BaseAndOffset
381 = ConstantOffsetPtrs.lookup(I.getOperand(0));
382 if (BaseAndOffset.first)
383 ConstantOffsetPtrs[&I] = BaseAndOffset;
386 // This is really weird. Technically, ptrtoint will disable SROA. However,
387 // unless that ptrtoint is *used* somewhere in the live basic blocks after
388 // inlining, it will be nuked, and SROA should proceed. All of the uses which
389 // would block SROA would also block SROA if applied directly to a pointer,
390 // and so we can just add the integer in here. The only places where SROA is
391 // preserved either cannot fire on an integer, or won't in-and-of themselves
392 // disable SROA (ext) w/o some later use that we would see and disable.
394 DenseMap<Value *, int>::iterator CostIt;
395 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
396 SROAArgValues[&I] = SROAArg;
398 // A ptrtoint cast is free so long as the result is large enough to store the
399 // pointer, and a legal integer type.
400 return TD && TD->isLegalInteger(IntegerSize) &&
401 IntegerSize >= TD->getPointerSizeInBits();
404 bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) {
405 // Propagate constants through ptrtoint.
406 if (Constant *COp = dyn_cast<Constant>(I.getOperand(0)))
407 if (Constant *C = ConstantExpr::getIntToPtr(COp, I.getType())) {
408 SimplifiedValues[&I] = C;
412 // Track base/offset pairs when round-tripped through a pointer without
413 // modifications provided the integer is not too large.
414 Value *Op = I.getOperand(0);
415 unsigned IntegerSize = Op->getType()->getScalarSizeInBits();
416 if (TD && IntegerSize <= TD->getPointerSizeInBits()) {
417 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op);
418 if (BaseAndOffset.first)
419 ConstantOffsetPtrs[&I] = BaseAndOffset;
422 // "Propagate" SROA here in the same manner as we do for ptrtoint above.
424 DenseMap<Value *, int>::iterator CostIt;
425 if (lookupSROAArgAndCost(Op, SROAArg, CostIt))
426 SROAArgValues[&I] = SROAArg;
428 // An inttoptr cast is free so long as the input is a legal integer type
429 // which doesn't contain values outside the range of a pointer.
430 return TD && TD->isLegalInteger(IntegerSize) &&
431 IntegerSize <= TD->getPointerSizeInBits();
434 bool CallAnalyzer::visitCastInst(CastInst &I) {
435 // Propagate constants through ptrtoint.
436 if (Constant *COp = dyn_cast<Constant>(I.getOperand(0)))
437 if (Constant *C = ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) {
438 SimplifiedValues[&I] = C;
442 // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere.
443 disableSROA(I.getOperand(0));
445 // No-op casts don't have any cost.
446 if (I.isLosslessCast())
449 // trunc to a native type is free (assuming the target has compare and
450 // shift-right of the same width).
451 if (TD && isa<TruncInst>(I) &&
452 TD->isLegalInteger(TD->getTypeSizeInBits(I.getType())))
455 // Result of a cmp instruction is often extended (to be used by other
456 // cmp instructions, logical or return instructions). These are usually
457 // no-ops on most sane targets.
458 if (isa<CmpInst>(I.getOperand(0)))
461 // Assume the rest of the casts require work.
465 bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) {
466 Value *Operand = I.getOperand(0);
467 Constant *Ops[1] = { dyn_cast<Constant>(Operand) };
468 if (Ops[0] || (Ops[0] = SimplifiedValues.lookup(Operand)))
469 if (Constant *C = ConstantFoldInstOperands(I.getOpcode(), I.getType(),
471 SimplifiedValues[&I] = C;
475 // Disable any SROA on the argument to arbitrary unary operators.
476 disableSROA(Operand);
481 bool CallAnalyzer::visitICmp(ICmpInst &I) {
482 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
483 // First try to handle simplified comparisons.
484 if (!isa<Constant>(LHS))
485 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
487 if (!isa<Constant>(RHS))
488 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
490 if (Constant *CLHS = dyn_cast<Constant>(LHS))
491 if (Constant *CRHS = dyn_cast<Constant>(RHS))
492 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
493 SimplifiedValues[&I] = C;
497 // Otherwise look for a comparison between constant offset pointers with
499 Value *LHSBase, *RHSBase;
500 APInt LHSOffset, RHSOffset;
501 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
503 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
504 if (RHSBase && LHSBase == RHSBase) {
505 // We have common bases, fold the icmp to a constant based on the
507 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
508 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
509 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
510 SimplifiedValues[&I] = C;
511 ++NumConstantPtrCmps;
517 // If the comparison is an equality comparison with null, we can simplify it
518 // for any alloca-derived argument.
519 if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1)))
520 if (isAllocaDerivedArg(I.getOperand(0))) {
521 // We can actually predict the result of comparisons between an
522 // alloca-derived value and null. Note that this fires regardless of
524 bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE;
525 SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType())
526 : ConstantInt::getFalse(I.getType());
530 // Finally check for SROA candidates in comparisons.
532 DenseMap<Value *, int>::iterator CostIt;
533 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
534 if (isa<ConstantPointerNull>(I.getOperand(1))) {
535 accumulateSROACost(CostIt, InlineConstants::InstrCost);
545 bool CallAnalyzer::visitSub(BinaryOperator &I) {
546 // Try to handle a special case: we can fold computing the difference of two
547 // constant-related pointers.
548 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
549 Value *LHSBase, *RHSBase;
550 APInt LHSOffset, RHSOffset;
551 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
553 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
554 if (RHSBase && LHSBase == RHSBase) {
555 // We have common bases, fold the subtract to a constant based on the
557 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
558 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
559 if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) {
560 SimplifiedValues[&I] = C;
561 ++NumConstantPtrDiffs;
567 // Otherwise, fall back to the generic logic for simplifying and handling
569 return Base::visitSub(I);
572 bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
573 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
574 if (!isa<Constant>(LHS))
575 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
577 if (!isa<Constant>(RHS))
578 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
580 Value *SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, TD);
581 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
582 SimplifiedValues[&I] = C;
586 // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
593 bool CallAnalyzer::visitLoad(LoadInst &I) {
595 DenseMap<Value *, int>::iterator CostIt;
596 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
598 accumulateSROACost(CostIt, InlineConstants::InstrCost);
608 bool CallAnalyzer::visitStore(StoreInst &I) {
610 DenseMap<Value *, int>::iterator CostIt;
611 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
613 accumulateSROACost(CostIt, InlineConstants::InstrCost);
623 bool CallAnalyzer::visitCallSite(CallSite CS) {
624 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->canReturnTwice() &&
625 !F.hasFnAttr(Attribute::ReturnsTwice)) {
626 // This aborts the entire analysis.
627 ExposesReturnsTwice = true;
631 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
632 switch (II->getIntrinsicID()) {
634 return Base::visitCallSite(CS);
636 case Intrinsic::dbg_declare:
637 case Intrinsic::dbg_value:
638 case Intrinsic::invariant_start:
639 case Intrinsic::invariant_end:
640 case Intrinsic::lifetime_start:
641 case Intrinsic::lifetime_end:
642 case Intrinsic::memset:
643 case Intrinsic::memcpy:
644 case Intrinsic::memmove:
645 case Intrinsic::objectsize:
646 case Intrinsic::ptr_annotation:
647 case Intrinsic::var_annotation:
648 // SROA can usually chew through these intrinsics and they have no cost
649 // so don't pay the price of analyzing them in detail.
654 if (Function *F = CS.getCalledFunction()) {
655 if (F == CS.getInstruction()->getParent()->getParent()) {
656 // This flag will fully abort the analysis, so don't bother with anything
662 if (!callIsSmall(F)) {
663 // We account for the average 1 instruction per call argument setup
665 Cost += CS.arg_size() * InlineConstants::InstrCost;
667 // Everything other than inline ASM will also have a significant cost
668 // merely from making the call.
669 if (!isa<InlineAsm>(CS.getCalledValue()))
670 Cost += InlineConstants::CallPenalty;
673 return Base::visitCallSite(CS);
676 // Otherwise we're in a very special case -- an indirect function call. See
677 // if we can be particularly clever about this.
678 Value *Callee = CS.getCalledValue();
680 // First, pay the price of the argument setup. We account for the average
681 // 1 instruction per call argument setup here.
682 Cost += CS.arg_size() * InlineConstants::InstrCost;
684 // Next, check if this happens to be an indirect function call to a known
685 // function in this inline context. If not, we've done all we can.
686 Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee));
688 return Base::visitCallSite(CS);
690 // If we have a constant that we are calling as a function, we can peer
691 // through it and see the function target. This happens not infrequently
692 // during devirtualization and so we want to give it a hefty bonus for
693 // inlining, but cap that bonus in the event that inlining wouldn't pan
694 // out. Pretend to inline the function, with a custom threshold.
695 CallAnalyzer CA(TD, *F, InlineConstants::IndirectCallThreshold);
696 if (CA.analyzeCall(CS)) {
697 // We were able to inline the indirect call! Subtract the cost from the
698 // bonus we want to apply, but don't go below zero.
699 Cost -= std::max(0, InlineConstants::IndirectCallThreshold - CA.getCost());
702 return Base::visitCallSite(CS);
705 bool CallAnalyzer::visitInstruction(Instruction &I) {
706 // We found something we don't understand or can't handle. Mark any SROA-able
707 // values in the operand list as no longer viable.
708 for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI)
715 /// \brief Analyze a basic block for its contribution to the inline cost.
717 /// This method walks the analyzer over every instruction in the given basic
718 /// block and accounts for their cost during inlining at this callsite. It
719 /// aborts early if the threshold has been exceeded or an impossible to inline
720 /// construct has been detected. It returns false if inlining is no longer
721 /// viable, and true if inlining remains viable.
722 bool CallAnalyzer::analyzeBlock(BasicBlock *BB) {
723 for (BasicBlock::iterator I = BB->begin(), E = llvm::prior(BB->end());
726 if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy())
727 ++NumVectorInstructions;
729 // If the instruction simplified to a constant, there is no cost to this
730 // instruction. Visit the instructions using our InstVisitor to account for
731 // all of the per-instruction logic. The visit tree returns true if we
732 // consumed the instruction in any way, and false if the instruction's base
733 // cost should count against inlining.
735 ++NumInstructionsSimplified;
737 Cost += InlineConstants::InstrCost;
739 // If the visit this instruction detected an uninlinable pattern, abort.
740 if (IsRecursive || ExposesReturnsTwice || HasDynamicAlloca)
743 if (NumVectorInstructions > NumInstructions/2)
744 VectorBonus = FiftyPercentVectorBonus;
745 else if (NumVectorInstructions > NumInstructions/10)
746 VectorBonus = TenPercentVectorBonus;
750 // Check if we've past the threshold so we don't spin in huge basic
751 // blocks that will never inline.
752 if (!AlwaysInline && Cost > (Threshold + VectorBonus))
759 /// \brief Compute the base pointer and cumulative constant offsets for V.
761 /// This strips all constant offsets off of V, leaving it the base pointer, and
762 /// accumulates the total constant offset applied in the returned constant. It
763 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
764 /// no constant offsets applied.
765 ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) {
766 if (!TD || !V->getType()->isPointerTy())
769 unsigned IntPtrWidth = TD->getPointerSizeInBits();
770 APInt Offset = APInt::getNullValue(IntPtrWidth);
772 // Even though we don't look through PHI nodes, we could be called on an
773 // instruction in an unreachable block, which may be on a cycle.
774 SmallPtrSet<Value *, 4> Visited;
777 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
778 if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset))
780 V = GEP->getPointerOperand();
781 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
782 V = cast<Operator>(V)->getOperand(0);
783 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
784 if (GA->mayBeOverridden())
786 V = GA->getAliasee();
790 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
791 } while (Visited.insert(V));
793 Type *IntPtrTy = TD->getIntPtrType(V->getContext());
794 return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset));
797 /// \brief Analyze a call site for potential inlining.
799 /// Returns true if inlining this call is viable, and false if it is not
800 /// viable. It computes the cost and adjusts the threshold based on numerous
801 /// factors and heuristics. If this method returns false but the computed cost
802 /// is below the computed threshold, then inlining was forcibly disabled by
803 /// some artifact of the rountine.
804 bool CallAnalyzer::analyzeCall(CallSite CS) {
805 // Track whether the post-inlining function would have more than one basic
806 // block. A single basic block is often intended for inlining. Balloon the
807 // threshold by 50% until we pass the single-BB phase.
808 bool SingleBB = true;
809 int SingleBBBonus = Threshold / 2;
810 Threshold += SingleBBBonus;
812 // Unless we are always-inlining, perform some tweaks to the cost and
813 // threshold based on the direct callsite information.
815 // We want to more aggressively inline vector-dense kernels, so up the
816 // threshold, and we'll lower it if the % of vector instructions gets too
818 assert(NumInstructions == 0);
819 assert(NumVectorInstructions == 0);
820 FiftyPercentVectorBonus = Threshold;
821 TenPercentVectorBonus = Threshold / 2;
823 // Subtract off one instruction per call argument as those will be free after
825 Cost -= CS.arg_size() * InlineConstants::InstrCost;
827 // If there is only one call of the function, and it has internal linkage,
828 // the cost of inlining it drops dramatically.
829 if (F.hasLocalLinkage() && F.hasOneUse() && &F == CS.getCalledFunction())
830 Cost += InlineConstants::LastCallToStaticBonus;
832 // If the instruction after the call, or if the normal destination of the
833 // invoke is an unreachable instruction, the function is noreturn. As such,
834 // there is little point in inlining this unless there is literally zero cost.
835 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
836 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
838 } else if (isa<UnreachableInst>(++BasicBlock::iterator(CS.getInstruction())))
841 // If this function uses the coldcc calling convention, prefer not to inline
843 if (F.getCallingConv() == CallingConv::Cold)
844 Cost += InlineConstants::ColdccPenalty;
846 // Check if we're done. This can happen due to bonuses and penalties.
847 if (Cost > Threshold)
854 // Track whether we've seen a return instruction. The first return
855 // instruction is free, as at least one will usually disappear in inlining.
856 bool HasReturn = false;
858 // Populate our simplified values by mapping from function arguments to call
859 // arguments with known important simplifications.
860 CallSite::arg_iterator CAI = CS.arg_begin();
861 for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end();
862 FAI != FAE; ++FAI, ++CAI) {
863 assert(CAI != CS.arg_end());
864 if (Constant *C = dyn_cast<Constant>(CAI))
865 SimplifiedValues[FAI] = C;
867 Value *PtrArg = *CAI;
868 if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) {
869 ConstantOffsetPtrs[FAI] = std::make_pair(PtrArg, C->getValue());
871 // We can SROA any pointer arguments derived from alloca instructions.
872 if (isa<AllocaInst>(PtrArg)) {
873 SROAArgValues[FAI] = PtrArg;
874 SROAArgCosts[PtrArg] = 0;
878 NumConstantArgs = SimplifiedValues.size();
879 NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size();
880 NumAllocaArgs = SROAArgValues.size();
882 // The worklist of live basic blocks in the callee *after* inlining. We avoid
883 // adding basic blocks of the callee which can be proven to be dead for this
884 // particular call site in order to get more accurate cost estimates. This
885 // requires a somewhat heavyweight iteration pattern: we need to walk the
886 // basic blocks in a breadth-first order as we insert live successors. To
887 // accomplish this, prioritizing for small iterations because we exit after
888 // crossing our threshold, we use a small-size optimized SetVector.
889 typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>,
890 SmallPtrSet<BasicBlock *, 16> > BBSetVector;
891 BBSetVector BBWorklist;
892 BBWorklist.insert(&F.getEntryBlock());
893 // Note that we *must not* cache the size, this loop grows the worklist.
894 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
895 // Bail out the moment we cross the threshold. This means we'll under-count
896 // the cost, but only when undercounting doesn't matter.
897 if (!AlwaysInline && Cost > (Threshold + VectorBonus))
900 BasicBlock *BB = BBWorklist[Idx];
904 // Handle the terminator cost here where we can track returns and other
905 // function-wide constructs.
906 TerminatorInst *TI = BB->getTerminator();
908 // We never want to inline functions that contain an indirectbr. This is
909 // incorrect because all the blockaddress's (in static global initializers
910 // for example) would be referring to the original function, and this indirect
911 // jump would jump from the inlined copy of the function into the original
912 // function which is extremely undefined behavior.
913 // FIXME: This logic isn't really right; we can safely inline functions
914 // with indirectbr's as long as no other function or global references the
915 // blockaddress of a block within the current function. And as a QOI issue,
916 // if someone is using a blockaddress without an indirectbr, and that
917 // reference somehow ends up in another function or global, we probably
918 // don't want to inline this function.
919 if (isa<IndirectBrInst>(TI))
922 if (!HasReturn && isa<ReturnInst>(TI))
925 Cost += InlineConstants::InstrCost;
927 // Analyze the cost of this block. If we blow through the threshold, this
928 // returns false, and we can bail on out.
929 if (!analyzeBlock(BB)) {
930 if (IsRecursive || ExposesReturnsTwice || HasDynamicAlloca)
935 // Add in the live successors by first checking whether we have terminator
936 // that may be simplified based on the values simplified by this call.
937 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
938 if (BI->isConditional()) {
939 Value *Cond = BI->getCondition();
940 if (ConstantInt *SimpleCond
941 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
942 BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0));
946 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
947 Value *Cond = SI->getCondition();
948 if (ConstantInt *SimpleCond
949 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
950 BBWorklist.insert(SI->findCaseValue(SimpleCond).getCaseSuccessor());
955 // If we're unable to select a particular successor, just count all of
957 for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize; ++TIdx)
958 BBWorklist.insert(TI->getSuccessor(TIdx));
960 // If we had any successors at this point, than post-inlining is likely to
961 // have them as well. Note that we assume any basic blocks which existed
962 // due to branches or switches which folded above will also fold after
964 if (SingleBB && TI->getNumSuccessors() > 1) {
965 // Take off the bonus we applied to the threshold.
966 Threshold -= SingleBBBonus;
971 Threshold += VectorBonus;
973 return AlwaysInline || Cost < Threshold;
976 /// \brief Dump stats about this call's analysis.
977 void CallAnalyzer::dump() {
978 #define DEBUG_PRINT_STAT(x) llvm::dbgs() << " " #x ": " << x << "\n"
979 DEBUG_PRINT_STAT(NumConstantArgs);
980 DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs);
981 DEBUG_PRINT_STAT(NumAllocaArgs);
982 DEBUG_PRINT_STAT(NumConstantPtrCmps);
983 DEBUG_PRINT_STAT(NumConstantPtrDiffs);
984 DEBUG_PRINT_STAT(NumInstructionsSimplified);
985 DEBUG_PRINT_STAT(SROACostSavings);
986 DEBUG_PRINT_STAT(SROACostSavingsLost);
987 #undef DEBUG_PRINT_STAT
990 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS, int Threshold) {
991 return getInlineCost(CS, CS.getCalledFunction(), Threshold);
994 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS, Function *Callee,
996 // Don't inline functions which can be redefined at link-time to mean
997 // something else. Don't inline functions marked noinline or call sites
999 if (!Callee || Callee->mayBeOverridden() ||
1000 Callee->hasFnAttr(Attribute::NoInline) || CS.isNoInline())
1001 return llvm::InlineCost::getNever();
1003 DEBUG(llvm::dbgs() << " Analyzing call of " << Callee->getName() << "...\n");
1005 CallAnalyzer CA(TD, *Callee, Threshold);
1006 bool ShouldInline = CA.analyzeCall(CS);
1010 // Check if there was a reason to force inlining or no inlining.
1011 if (!ShouldInline && CA.getCost() < CA.getThreshold())
1012 return InlineCost::getNever();
1013 if (ShouldInline && CA.getCost() >= CA.getThreshold())
1014 return InlineCost::getAlways();
1016 return llvm::InlineCost::get(CA.getCost(), CA.getThreshold());