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 #include "llvm/Analysis/InlineCost.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/AssumptionCache.h"
21 #include "llvm/Analysis/BlockFrequencyInfo.h"
22 #include "llvm/Analysis/CodeMetrics.h"
23 #include "llvm/Analysis/ConstantFolding.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/ProfileSummaryInfo.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/IR/CallSite.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/GlobalAlias.h"
32 #include "llvm/IR/InstVisitor.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Operator.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
40 #define DEBUG_TYPE "inline-cost"
42 STATISTIC(NumCallsAnalyzed, "Number of call sites analyzed");
44 static cl::opt<int> InlineThreshold(
45 "inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
46 cl::desc("Control the amount of inlining to perform (default = 225)"));
48 static cl::opt<int> HintThreshold(
49 "inlinehint-threshold", cl::Hidden, cl::init(325),
50 cl::desc("Threshold for inlining functions with inline hint"));
53 ColdCallSiteThreshold("inline-cold-callsite-threshold", cl::Hidden,
55 cl::desc("Threshold for inlining cold callsites"));
57 // We introduce this threshold to help performance of instrumentation based
58 // PGO before we actually hook up inliner with analysis passes such as BPI and
60 static cl::opt<int> ColdThreshold(
61 "inlinecold-threshold", cl::Hidden, cl::init(45),
62 cl::desc("Threshold for inlining functions with cold attribute"));
65 HotCallSiteThreshold("hot-callsite-threshold", cl::Hidden, cl::init(3000),
67 cl::desc("Threshold for hot callsites "));
69 static cl::opt<int> ColdCallSiteRelFreq(
70 "cold-callsite-rel-freq", cl::Hidden, cl::init(2), cl::ZeroOrMore,
71 cl::desc("Maxmimum block frequency, expressed as a percentage of caller's "
72 "entry frequency, for a callsite to be cold in the absence of "
73 "profile information."));
77 class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> {
78 typedef InstVisitor<CallAnalyzer, bool> Base;
79 friend class InstVisitor<CallAnalyzer, bool>;
81 /// The TargetTransformInfo available for this compilation.
82 const TargetTransformInfo &TTI;
84 /// Getter for the cache of @llvm.assume intrinsics.
85 std::function<AssumptionCache &(Function &)> &GetAssumptionCache;
87 /// Getter for BlockFrequencyInfo
88 Optional<function_ref<BlockFrequencyInfo &(Function &)>> &GetBFI;
90 /// Profile summary information.
91 ProfileSummaryInfo *PSI;
93 /// The called function.
96 // Cache the DataLayout since we use it a lot.
99 /// The candidate callsite being analyzed. Please do not use this to do
100 /// analysis in the caller function; we want the inline cost query to be
101 /// easily cacheable. Instead, use the cover function paramHasAttr.
102 CallSite CandidateCS;
104 /// Tunable parameters that control the analysis.
105 const InlineParams &Params;
110 bool IsCallerRecursive;
111 bool IsRecursiveCall;
112 bool ExposesReturnsTwice;
113 bool HasDynamicAlloca;
114 bool ContainsNoDuplicateCall;
119 /// Number of bytes allocated statically by the callee.
120 uint64_t AllocatedSize;
121 unsigned NumInstructions, NumVectorInstructions;
122 int FiftyPercentVectorBonus, TenPercentVectorBonus;
125 /// While we walk the potentially-inlined instructions, we build up and
126 /// maintain a mapping of simplified values specific to this callsite. The
127 /// idea is to propagate any special information we have about arguments to
128 /// this call through the inlinable section of the function, and account for
129 /// likely simplifications post-inlining. The most important aspect we track
130 /// is CFG altering simplifications -- when we prove a basic block dead, that
131 /// can cause dramatic shifts in the cost of inlining a function.
132 DenseMap<Value *, Constant *> SimplifiedValues;
134 /// Keep track of the values which map back (through function arguments) to
135 /// allocas on the caller stack which could be simplified through SROA.
136 DenseMap<Value *, Value *> SROAArgValues;
138 /// The mapping of caller Alloca values to their accumulated cost savings. If
139 /// we have to disable SROA for one of the allocas, this tells us how much
140 /// cost must be added.
141 DenseMap<Value *, int> SROAArgCosts;
143 /// Keep track of values which map to a pointer base and constant offset.
144 DenseMap<Value *, std::pair<Value *, APInt>> ConstantOffsetPtrs;
146 // Custom simplification helper routines.
147 bool isAllocaDerivedArg(Value *V);
148 bool lookupSROAArgAndCost(Value *V, Value *&Arg,
149 DenseMap<Value *, int>::iterator &CostIt);
150 void disableSROA(DenseMap<Value *, int>::iterator CostIt);
151 void disableSROA(Value *V);
152 void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
153 int InstructionCost);
154 bool isGEPFree(GetElementPtrInst &GEP);
155 bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset);
156 bool simplifyCallSite(Function *F, CallSite CS);
157 template <typename Callable>
158 bool simplifyInstruction(Instruction &I, Callable Evaluate);
159 ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V);
161 /// Return true if the given argument to the function being considered for
162 /// inlining has the given attribute set either at the call site or the
163 /// function declaration. Primarily used to inspect call site specific
164 /// attributes since these can be more precise than the ones on the callee
166 bool paramHasAttr(Argument *A, Attribute::AttrKind Attr);
168 /// Return true if the given value is known non null within the callee if
169 /// inlined through this particular callsite.
170 bool isKnownNonNullInCallee(Value *V);
172 /// Update Threshold based on callsite properties such as callee
173 /// attributes and callee hotness for PGO builds. The Callee is explicitly
174 /// passed to support analyzing indirect calls whose target is inferred by
176 void updateThreshold(CallSite CS, Function &Callee);
178 /// Return true if size growth is allowed when inlining the callee at CS.
179 bool allowSizeGrowth(CallSite CS);
181 /// Return true if \p CS is a cold callsite.
182 bool isColdCallSite(CallSite CS, BlockFrequencyInfo *CallerBFI);
184 // Custom analysis routines.
185 bool analyzeBlock(BasicBlock *BB, SmallPtrSetImpl<const Value *> &EphValues);
187 // Disable several entry points to the visitor so we don't accidentally use
188 // them by declaring but not defining them here.
189 void visit(Module *);
190 void visit(Module &);
191 void visit(Function *);
192 void visit(Function &);
193 void visit(BasicBlock *);
194 void visit(BasicBlock &);
196 // Provide base case for our instruction visit.
197 bool visitInstruction(Instruction &I);
199 // Our visit overrides.
200 bool visitAlloca(AllocaInst &I);
201 bool visitPHI(PHINode &I);
202 bool visitGetElementPtr(GetElementPtrInst &I);
203 bool visitBitCast(BitCastInst &I);
204 bool visitPtrToInt(PtrToIntInst &I);
205 bool visitIntToPtr(IntToPtrInst &I);
206 bool visitCastInst(CastInst &I);
207 bool visitUnaryInstruction(UnaryInstruction &I);
208 bool visitCmpInst(CmpInst &I);
209 bool visitSub(BinaryOperator &I);
210 bool visitBinaryOperator(BinaryOperator &I);
211 bool visitLoad(LoadInst &I);
212 bool visitStore(StoreInst &I);
213 bool visitExtractValue(ExtractValueInst &I);
214 bool visitInsertValue(InsertValueInst &I);
215 bool visitCallSite(CallSite CS);
216 bool visitReturnInst(ReturnInst &RI);
217 bool visitBranchInst(BranchInst &BI);
218 bool visitSwitchInst(SwitchInst &SI);
219 bool visitIndirectBrInst(IndirectBrInst &IBI);
220 bool visitResumeInst(ResumeInst &RI);
221 bool visitCleanupReturnInst(CleanupReturnInst &RI);
222 bool visitCatchReturnInst(CatchReturnInst &RI);
223 bool visitUnreachableInst(UnreachableInst &I);
226 CallAnalyzer(const TargetTransformInfo &TTI,
227 std::function<AssumptionCache &(Function &)> &GetAssumptionCache,
228 Optional<function_ref<BlockFrequencyInfo &(Function &)>> &GetBFI,
229 ProfileSummaryInfo *PSI, Function &Callee, CallSite CSArg,
230 const InlineParams &Params)
231 : TTI(TTI), GetAssumptionCache(GetAssumptionCache), GetBFI(GetBFI),
232 PSI(PSI), F(Callee), DL(F.getParent()->getDataLayout()),
233 CandidateCS(CSArg), Params(Params), Threshold(Params.DefaultThreshold),
234 Cost(0), IsCallerRecursive(false), IsRecursiveCall(false),
235 ExposesReturnsTwice(false), HasDynamicAlloca(false),
236 ContainsNoDuplicateCall(false), HasReturn(false), HasIndirectBr(false),
237 HasFrameEscape(false), AllocatedSize(0), NumInstructions(0),
238 NumVectorInstructions(0), FiftyPercentVectorBonus(0),
239 TenPercentVectorBonus(0), VectorBonus(0), NumConstantArgs(0),
240 NumConstantOffsetPtrArgs(0), NumAllocaArgs(0), NumConstantPtrCmps(0),
241 NumConstantPtrDiffs(0), NumInstructionsSimplified(0),
242 SROACostSavings(0), SROACostSavingsLost(0) {}
244 bool analyzeCall(CallSite CS);
246 int getThreshold() { return Threshold; }
247 int getCost() { return Cost; }
249 // Keep a bunch of stats about the cost savings found so we can print them
250 // out when debugging.
251 unsigned NumConstantArgs;
252 unsigned NumConstantOffsetPtrArgs;
253 unsigned NumAllocaArgs;
254 unsigned NumConstantPtrCmps;
255 unsigned NumConstantPtrDiffs;
256 unsigned NumInstructionsSimplified;
257 unsigned SROACostSavings;
258 unsigned SROACostSavingsLost;
265 /// \brief Test whether the given value is an Alloca-derived function argument.
266 bool CallAnalyzer::isAllocaDerivedArg(Value *V) {
267 return SROAArgValues.count(V);
270 /// \brief Lookup the SROA-candidate argument and cost iterator which V maps to.
271 /// Returns false if V does not map to a SROA-candidate.
272 bool CallAnalyzer::lookupSROAArgAndCost(
273 Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) {
274 if (SROAArgValues.empty() || SROAArgCosts.empty())
277 DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V);
278 if (ArgIt == SROAArgValues.end())
282 CostIt = SROAArgCosts.find(Arg);
283 return CostIt != SROAArgCosts.end();
286 /// \brief Disable SROA for the candidate marked by this cost iterator.
288 /// This marks the candidate as no longer viable for SROA, and adds the cost
289 /// savings associated with it back into the inline cost measurement.
290 void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) {
291 // If we're no longer able to perform SROA we need to undo its cost savings
292 // and prevent subsequent analysis.
293 Cost += CostIt->second;
294 SROACostSavings -= CostIt->second;
295 SROACostSavingsLost += CostIt->second;
296 SROAArgCosts.erase(CostIt);
299 /// \brief If 'V' maps to a SROA candidate, disable SROA for it.
300 void CallAnalyzer::disableSROA(Value *V) {
302 DenseMap<Value *, int>::iterator CostIt;
303 if (lookupSROAArgAndCost(V, SROAArg, CostIt))
307 /// \brief Accumulate the given cost for a particular SROA candidate.
308 void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
309 int InstructionCost) {
310 CostIt->second += InstructionCost;
311 SROACostSavings += InstructionCost;
314 /// \brief Accumulate a constant GEP offset into an APInt if possible.
316 /// Returns false if unable to compute the offset for any reason. Respects any
317 /// simplified values known during the analysis of this callsite.
318 bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) {
319 unsigned IntPtrWidth = DL.getPointerSizeInBits();
320 assert(IntPtrWidth == Offset.getBitWidth());
322 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
324 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
326 if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand()))
327 OpC = dyn_cast<ConstantInt>(SimpleOp);
333 // Handle a struct index, which adds its field offset to the pointer.
334 if (StructType *STy = GTI.getStructTypeOrNull()) {
335 unsigned ElementIdx = OpC->getZExtValue();
336 const StructLayout *SL = DL.getStructLayout(STy);
337 Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx));
341 APInt TypeSize(IntPtrWidth, DL.getTypeAllocSize(GTI.getIndexedType()));
342 Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize;
347 /// \brief Use TTI to check whether a GEP is free.
349 /// Respects any simplified values known during the analysis of this callsite.
350 bool CallAnalyzer::isGEPFree(GetElementPtrInst &GEP) {
351 SmallVector<Value *, 4> Indices;
352 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
353 if (Constant *SimpleOp = SimplifiedValues.lookup(*I))
354 Indices.push_back(SimpleOp);
356 Indices.push_back(*I);
357 return TargetTransformInfo::TCC_Free ==
358 TTI.getGEPCost(GEP.getSourceElementType(), GEP.getPointerOperand(),
362 bool CallAnalyzer::visitAlloca(AllocaInst &I) {
363 // Check whether inlining will turn a dynamic alloca into a static
364 // alloca and handle that case.
365 if (I.isArrayAllocation()) {
366 Constant *Size = SimplifiedValues.lookup(I.getArraySize());
367 if (auto *AllocSize = dyn_cast_or_null<ConstantInt>(Size)) {
368 Type *Ty = I.getAllocatedType();
369 AllocatedSize = SaturatingMultiplyAdd(
370 AllocSize->getLimitedValue(), DL.getTypeAllocSize(Ty), AllocatedSize);
371 return Base::visitAlloca(I);
375 // Accumulate the allocated size.
376 if (I.isStaticAlloca()) {
377 Type *Ty = I.getAllocatedType();
378 AllocatedSize = SaturatingAdd(DL.getTypeAllocSize(Ty), AllocatedSize);
381 // We will happily inline static alloca instructions.
382 if (I.isStaticAlloca())
383 return Base::visitAlloca(I);
385 // FIXME: This is overly conservative. Dynamic allocas are inefficient for
386 // a variety of reasons, and so we would like to not inline them into
387 // functions which don't currently have a dynamic alloca. This simply
388 // disables inlining altogether in the presence of a dynamic alloca.
389 HasDynamicAlloca = true;
393 bool CallAnalyzer::visitPHI(PHINode &I) {
394 // FIXME: We should potentially be tracking values through phi nodes,
395 // especially when they collapse to a single value due to deleted CFG edges
398 // FIXME: We need to propagate SROA *disabling* through phi nodes, even
399 // though we don't want to propagate it's bonuses. The idea is to disable
400 // SROA if it *might* be used in an inappropriate manner.
402 // Phi nodes are always zero-cost.
406 bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) {
408 DenseMap<Value *, int>::iterator CostIt;
410 lookupSROAArgAndCost(I.getPointerOperand(), SROAArg, CostIt);
412 // Try to fold GEPs of constant-offset call site argument pointers. This
413 // requires target data and inbounds GEPs.
414 if (I.isInBounds()) {
415 // Check if we have a base + offset for the pointer.
416 Value *Ptr = I.getPointerOperand();
417 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr);
418 if (BaseAndOffset.first) {
419 // Check if the offset of this GEP is constant, and if so accumulate it
421 if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) {
422 // Non-constant GEPs aren't folded, and disable SROA.
428 // Add the result as a new mapping to Base + Offset.
429 ConstantOffsetPtrs[&I] = BaseAndOffset;
431 // Also handle SROA candidates here, we already know that the GEP is
432 // all-constant indexed.
434 SROAArgValues[&I] = SROAArg;
440 // Lambda to check whether a GEP's indices are all constant.
441 auto IsGEPOffsetConstant = [&](GetElementPtrInst &GEP) {
442 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
443 if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I))
448 if (IsGEPOffsetConstant(I)) {
450 SROAArgValues[&I] = SROAArg;
452 // Constant GEPs are modeled as free.
456 // Variable GEPs will require math and will disable SROA.
462 /// Simplify \p I if its operands are constants and update SimplifiedValues.
463 /// \p Evaluate is a callable specific to instruction type that evaluates the
464 /// instruction when all the operands are constants.
465 template <typename Callable>
466 bool CallAnalyzer::simplifyInstruction(Instruction &I, Callable Evaluate) {
467 SmallVector<Constant *, 2> COps;
468 for (Value *Op : I.operands()) {
469 Constant *COp = dyn_cast<Constant>(Op);
471 COp = SimplifiedValues.lookup(Op);
476 auto *C = Evaluate(COps);
479 SimplifiedValues[&I] = C;
483 bool CallAnalyzer::visitBitCast(BitCastInst &I) {
484 // Propagate constants through bitcasts.
485 if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
486 return ConstantExpr::getBitCast(COps[0], I.getType());
490 // Track base/offsets through casts
491 std::pair<Value *, APInt> BaseAndOffset =
492 ConstantOffsetPtrs.lookup(I.getOperand(0));
493 // Casts don't change the offset, just wrap it up.
494 if (BaseAndOffset.first)
495 ConstantOffsetPtrs[&I] = BaseAndOffset;
497 // Also look for SROA candidates here.
499 DenseMap<Value *, int>::iterator CostIt;
500 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
501 SROAArgValues[&I] = SROAArg;
503 // Bitcasts are always zero cost.
507 bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) {
508 // Propagate constants through ptrtoint.
509 if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
510 return ConstantExpr::getPtrToInt(COps[0], I.getType());
514 // Track base/offset pairs when converted to a plain integer provided the
515 // integer is large enough to represent the pointer.
516 unsigned IntegerSize = I.getType()->getScalarSizeInBits();
517 if (IntegerSize >= DL.getPointerSizeInBits()) {
518 std::pair<Value *, APInt> BaseAndOffset =
519 ConstantOffsetPtrs.lookup(I.getOperand(0));
520 if (BaseAndOffset.first)
521 ConstantOffsetPtrs[&I] = BaseAndOffset;
524 // This is really weird. Technically, ptrtoint will disable SROA. However,
525 // unless that ptrtoint is *used* somewhere in the live basic blocks after
526 // inlining, it will be nuked, and SROA should proceed. All of the uses which
527 // would block SROA would also block SROA if applied directly to a pointer,
528 // and so we can just add the integer in here. The only places where SROA is
529 // preserved either cannot fire on an integer, or won't in-and-of themselves
530 // disable SROA (ext) w/o some later use that we would see and disable.
532 DenseMap<Value *, int>::iterator CostIt;
533 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
534 SROAArgValues[&I] = SROAArg;
536 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
539 bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) {
540 // Propagate constants through ptrtoint.
541 if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
542 return ConstantExpr::getIntToPtr(COps[0], I.getType());
546 // Track base/offset pairs when round-tripped through a pointer without
547 // modifications provided the integer is not too large.
548 Value *Op = I.getOperand(0);
549 unsigned IntegerSize = Op->getType()->getScalarSizeInBits();
550 if (IntegerSize <= DL.getPointerSizeInBits()) {
551 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op);
552 if (BaseAndOffset.first)
553 ConstantOffsetPtrs[&I] = BaseAndOffset;
556 // "Propagate" SROA here in the same manner as we do for ptrtoint above.
558 DenseMap<Value *, int>::iterator CostIt;
559 if (lookupSROAArgAndCost(Op, SROAArg, CostIt))
560 SROAArgValues[&I] = SROAArg;
562 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
565 bool CallAnalyzer::visitCastInst(CastInst &I) {
566 // Propagate constants through ptrtoint.
567 if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
568 return ConstantExpr::getCast(I.getOpcode(), COps[0], I.getType());
572 // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere.
573 disableSROA(I.getOperand(0));
575 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
578 bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) {
579 Value *Operand = I.getOperand(0);
580 if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
581 return ConstantFoldInstOperands(&I, COps[0], DL);
585 // Disable any SROA on the argument to arbitrary unary operators.
586 disableSROA(Operand);
591 bool CallAnalyzer::paramHasAttr(Argument *A, Attribute::AttrKind Attr) {
592 return CandidateCS.paramHasAttr(A->getArgNo(), Attr);
595 bool CallAnalyzer::isKnownNonNullInCallee(Value *V) {
596 // Does the *call site* have the NonNull attribute set on an argument? We
597 // use the attribute on the call site to memoize any analysis done in the
598 // caller. This will also trip if the callee function has a non-null
599 // parameter attribute, but that's a less interesting case because hopefully
600 // the callee would already have been simplified based on that.
601 if (Argument *A = dyn_cast<Argument>(V))
602 if (paramHasAttr(A, Attribute::NonNull))
605 // Is this an alloca in the caller? This is distinct from the attribute case
606 // above because attributes aren't updated within the inliner itself and we
607 // always want to catch the alloca derived case.
608 if (isAllocaDerivedArg(V))
609 // We can actually predict the result of comparisons between an
610 // alloca-derived value and null. Note that this fires regardless of
617 bool CallAnalyzer::allowSizeGrowth(CallSite CS) {
618 // If the normal destination of the invoke or the parent block of the call
619 // site is unreachable-terminated, there is little point in inlining this
620 // unless there is literally zero cost.
621 // FIXME: Note that it is possible that an unreachable-terminated block has a
622 // hot entry. For example, in below scenario inlining hot_call_X() may be
630 // For now, we are not handling this corner case here as it is rare in real
631 // code. In future, we should elaborate this based on BPI and BFI in more
632 // general threshold adjusting heuristics in updateThreshold().
633 Instruction *Instr = CS.getInstruction();
634 if (InvokeInst *II = dyn_cast<InvokeInst>(Instr)) {
635 if (isa<UnreachableInst>(II->getNormalDest()->getTerminator()))
637 } else if (isa<UnreachableInst>(Instr->getParent()->getTerminator()))
643 bool CallAnalyzer::isColdCallSite(CallSite CS, BlockFrequencyInfo *CallerBFI) {
644 // If global profile summary is available, then callsite's coldness is
645 // determined based on that.
646 if (PSI->hasProfileSummary())
647 return PSI->isColdCallSite(CS, CallerBFI);
651 // In the absence of global profile summary, determine if the callsite is cold
652 // relative to caller's entry. We could potentially cache the computation of
653 // scaled entry frequency, but the added complexity is not worth it unless
654 // this scaling shows up high in the profiles.
655 const BranchProbability ColdProb(ColdCallSiteRelFreq, 100);
656 auto CallSiteBB = CS.getInstruction()->getParent();
657 auto CallSiteFreq = CallerBFI->getBlockFreq(CallSiteBB);
658 auto CallerEntryFreq =
659 CallerBFI->getBlockFreq(&(CS.getCaller()->getEntryBlock()));
660 return CallSiteFreq < CallerEntryFreq * ColdProb;
663 void CallAnalyzer::updateThreshold(CallSite CS, Function &Callee) {
664 // If no size growth is allowed for this inlining, set Threshold to 0.
665 if (!allowSizeGrowth(CS)) {
670 Function *Caller = CS.getCaller();
672 // return min(A, B) if B is valid.
673 auto MinIfValid = [](int A, Optional<int> B) {
674 return B ? std::min(A, B.getValue()) : A;
677 // return max(A, B) if B is valid.
678 auto MaxIfValid = [](int A, Optional<int> B) {
679 return B ? std::max(A, B.getValue()) : A;
682 // Use the OptMinSizeThreshold or OptSizeThreshold knob if they are available
683 // and reduce the threshold if the caller has the necessary attribute.
684 if (Caller->optForMinSize())
685 Threshold = MinIfValid(Threshold, Params.OptMinSizeThreshold);
686 else if (Caller->optForSize())
687 Threshold = MinIfValid(Threshold, Params.OptSizeThreshold);
689 // Adjust the threshold based on inlinehint attribute and profile based
690 // hotness information if the caller does not have MinSize attribute.
691 if (!Caller->optForMinSize()) {
692 if (Callee.hasFnAttribute(Attribute::InlineHint))
693 Threshold = MaxIfValid(Threshold, Params.HintThreshold);
695 BlockFrequencyInfo *CallerBFI = GetBFI ? &((*GetBFI)(*Caller)) : nullptr;
696 // FIXME: After switching to the new passmanager, simplify the logic below
697 // by checking only the callsite hotness/coldness. The check for CallerBFI
698 // exists only because we do not have BFI available with the old PM.
700 // Use callee's hotness information only if we have no way of determining
701 // callsite's hotness information. Callsite hotness can be determined if
702 // sample profile is used (which adds hotness metadata to calls) or if
703 // caller's BlockFrequencyInfo is available.
704 if (CallerBFI || PSI->hasSampleProfile()) {
705 if (PSI->isHotCallSite(CS, CallerBFI)) {
706 DEBUG(dbgs() << "Hot callsite.\n");
707 Threshold = Params.HotCallSiteThreshold.getValue();
708 } else if (isColdCallSite(CS, CallerBFI)) {
709 DEBUG(dbgs() << "Cold callsite.\n");
710 Threshold = MinIfValid(Threshold, Params.ColdCallSiteThreshold);
713 if (PSI->isFunctionEntryHot(&Callee)) {
714 DEBUG(dbgs() << "Hot callee.\n");
715 // If callsite hotness can not be determined, we may still know
716 // that the callee is hot and treat it as a weaker hint for threshold
718 Threshold = MaxIfValid(Threshold, Params.HintThreshold);
719 } else if (PSI->isFunctionEntryCold(&Callee)) {
720 DEBUG(dbgs() << "Cold callee.\n");
721 Threshold = MinIfValid(Threshold, Params.ColdThreshold);
727 // Finally, take the target-specific inlining threshold multiplier into
729 Threshold *= TTI.getInliningThresholdMultiplier();
732 bool CallAnalyzer::visitCmpInst(CmpInst &I) {
733 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
734 // First try to handle simplified comparisons.
735 if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
736 return ConstantExpr::getCompare(I.getPredicate(), COps[0], COps[1]);
740 if (I.getOpcode() == Instruction::FCmp)
743 // Otherwise look for a comparison between constant offset pointers with
745 Value *LHSBase, *RHSBase;
746 APInt LHSOffset, RHSOffset;
747 std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
749 std::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
750 if (RHSBase && LHSBase == RHSBase) {
751 // We have common bases, fold the icmp to a constant based on the
753 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
754 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
755 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
756 SimplifiedValues[&I] = C;
757 ++NumConstantPtrCmps;
763 // If the comparison is an equality comparison with null, we can simplify it
764 // if we know the value (argument) can't be null
765 if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1)) &&
766 isKnownNonNullInCallee(I.getOperand(0))) {
767 bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE;
768 SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType())
769 : ConstantInt::getFalse(I.getType());
772 // Finally check for SROA candidates in comparisons.
774 DenseMap<Value *, int>::iterator CostIt;
775 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
776 if (isa<ConstantPointerNull>(I.getOperand(1))) {
777 accumulateSROACost(CostIt, InlineConstants::InstrCost);
787 bool CallAnalyzer::visitSub(BinaryOperator &I) {
788 // Try to handle a special case: we can fold computing the difference of two
789 // constant-related pointers.
790 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
791 Value *LHSBase, *RHSBase;
792 APInt LHSOffset, RHSOffset;
793 std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
795 std::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
796 if (RHSBase && LHSBase == RHSBase) {
797 // We have common bases, fold the subtract to a constant based on the
799 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
800 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
801 if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) {
802 SimplifiedValues[&I] = C;
803 ++NumConstantPtrDiffs;
809 // Otherwise, fall back to the generic logic for simplifying and handling
811 return Base::visitSub(I);
814 bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
815 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
816 auto Evaluate = [&](SmallVectorImpl<Constant *> &COps) {
817 Value *SimpleV = nullptr;
818 if (auto FI = dyn_cast<FPMathOperator>(&I))
819 SimpleV = SimplifyFPBinOp(I.getOpcode(), COps[0], COps[1],
820 FI->getFastMathFlags(), DL);
822 SimpleV = SimplifyBinOp(I.getOpcode(), COps[0], COps[1], DL);
823 return dyn_cast_or_null<Constant>(SimpleV);
826 if (simplifyInstruction(I, Evaluate))
829 // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
836 bool CallAnalyzer::visitLoad(LoadInst &I) {
838 DenseMap<Value *, int>::iterator CostIt;
839 if (lookupSROAArgAndCost(I.getPointerOperand(), SROAArg, CostIt)) {
841 accumulateSROACost(CostIt, InlineConstants::InstrCost);
851 bool CallAnalyzer::visitStore(StoreInst &I) {
853 DenseMap<Value *, int>::iterator CostIt;
854 if (lookupSROAArgAndCost(I.getPointerOperand(), SROAArg, CostIt)) {
856 accumulateSROACost(CostIt, InlineConstants::InstrCost);
866 bool CallAnalyzer::visitExtractValue(ExtractValueInst &I) {
867 // Constant folding for extract value is trivial.
868 if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
869 return ConstantExpr::getExtractValue(COps[0], I.getIndices());
873 // SROA can look through these but give them a cost.
877 bool CallAnalyzer::visitInsertValue(InsertValueInst &I) {
878 // Constant folding for insert value is trivial.
879 if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
880 return ConstantExpr::getInsertValue(/*AggregateOperand*/ COps[0],
881 /*InsertedValueOperand*/ COps[1],
886 // SROA can look through these but give them a cost.
890 /// \brief Try to simplify a call site.
892 /// Takes a concrete function and callsite and tries to actually simplify it by
893 /// analyzing the arguments and call itself with instsimplify. Returns true if
894 /// it has simplified the callsite to some other entity (a constant), making it
896 bool CallAnalyzer::simplifyCallSite(Function *F, CallSite CS) {
897 // FIXME: Using the instsimplify logic directly for this is inefficient
898 // because we have to continually rebuild the argument list even when no
899 // simplifications can be performed. Until that is fixed with remapping
900 // inside of instsimplify, directly constant fold calls here.
901 if (!canConstantFoldCallTo(CS, F))
904 // Try to re-map the arguments to constants.
905 SmallVector<Constant *, 4> ConstantArgs;
906 ConstantArgs.reserve(CS.arg_size());
907 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E;
909 Constant *C = dyn_cast<Constant>(*I);
911 C = dyn_cast_or_null<Constant>(SimplifiedValues.lookup(*I));
913 return false; // This argument doesn't map to a constant.
915 ConstantArgs.push_back(C);
917 if (Constant *C = ConstantFoldCall(CS, F, ConstantArgs)) {
918 SimplifiedValues[CS.getInstruction()] = C;
925 bool CallAnalyzer::visitCallSite(CallSite CS) {
926 if (CS.hasFnAttr(Attribute::ReturnsTwice) &&
927 !F.hasFnAttribute(Attribute::ReturnsTwice)) {
928 // This aborts the entire analysis.
929 ExposesReturnsTwice = true;
932 if (CS.isCall() && cast<CallInst>(CS.getInstruction())->cannotDuplicate())
933 ContainsNoDuplicateCall = true;
935 if (Function *F = CS.getCalledFunction()) {
936 // When we have a concrete function, first try to simplify it directly.
937 if (simplifyCallSite(F, CS))
940 // Next check if it is an intrinsic we know about.
941 // FIXME: Lift this into part of the InstVisitor.
942 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
943 switch (II->getIntrinsicID()) {
945 return Base::visitCallSite(CS);
947 case Intrinsic::load_relative:
948 // This is normally lowered to 4 LLVM instructions.
949 Cost += 3 * InlineConstants::InstrCost;
952 case Intrinsic::memset:
953 case Intrinsic::memcpy:
954 case Intrinsic::memmove:
955 // SROA can usually chew through these intrinsics, but they aren't free.
957 case Intrinsic::localescape:
958 HasFrameEscape = true;
963 if (F == CS.getInstruction()->getParent()->getParent()) {
964 // This flag will fully abort the analysis, so don't bother with anything
966 IsRecursiveCall = true;
970 if (TTI.isLoweredToCall(F)) {
971 // We account for the average 1 instruction per call argument setup
973 Cost += CS.arg_size() * InlineConstants::InstrCost;
975 // Everything other than inline ASM will also have a significant cost
976 // merely from making the call.
977 if (!isa<InlineAsm>(CS.getCalledValue()))
978 Cost += InlineConstants::CallPenalty;
981 return Base::visitCallSite(CS);
984 // Otherwise we're in a very special case -- an indirect function call. See
985 // if we can be particularly clever about this.
986 Value *Callee = CS.getCalledValue();
988 // First, pay the price of the argument setup. We account for the average
989 // 1 instruction per call argument setup here.
990 Cost += CS.arg_size() * InlineConstants::InstrCost;
992 // Next, check if this happens to be an indirect function call to a known
993 // function in this inline context. If not, we've done all we can.
994 Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee));
996 return Base::visitCallSite(CS);
998 // If we have a constant that we are calling as a function, we can peer
999 // through it and see the function target. This happens not infrequently
1000 // during devirtualization and so we want to give it a hefty bonus for
1001 // inlining, but cap that bonus in the event that inlining wouldn't pan
1002 // out. Pretend to inline the function, with a custom threshold.
1003 auto IndirectCallParams = Params;
1004 IndirectCallParams.DefaultThreshold = InlineConstants::IndirectCallThreshold;
1005 CallAnalyzer CA(TTI, GetAssumptionCache, GetBFI, PSI, *F, CS,
1006 IndirectCallParams);
1007 if (CA.analyzeCall(CS)) {
1008 // We were able to inline the indirect call! Subtract the cost from the
1009 // threshold to get the bonus we want to apply, but don't go below zero.
1010 Cost -= std::max(0, CA.getThreshold() - CA.getCost());
1013 return Base::visitCallSite(CS);
1016 bool CallAnalyzer::visitReturnInst(ReturnInst &RI) {
1017 // At least one return instruction will be free after inlining.
1018 bool Free = !HasReturn;
1023 bool CallAnalyzer::visitBranchInst(BranchInst &BI) {
1024 // We model unconditional branches as essentially free -- they really
1025 // shouldn't exist at all, but handling them makes the behavior of the
1026 // inliner more regular and predictable. Interestingly, conditional branches
1027 // which will fold away are also free.
1028 return BI.isUnconditional() || isa<ConstantInt>(BI.getCondition()) ||
1029 dyn_cast_or_null<ConstantInt>(
1030 SimplifiedValues.lookup(BI.getCondition()));
1033 bool CallAnalyzer::visitSwitchInst(SwitchInst &SI) {
1034 // We model unconditional switches as free, see the comments on handling
1036 if (isa<ConstantInt>(SI.getCondition()))
1038 if (Value *V = SimplifiedValues.lookup(SI.getCondition()))
1039 if (isa<ConstantInt>(V))
1042 // Assume the most general case where the switch is lowered into
1043 // either a jump table, bit test, or a balanced binary tree consisting of
1044 // case clusters without merging adjacent clusters with the same
1045 // destination. We do not consider the switches that are lowered with a mix
1046 // of jump table/bit test/binary search tree. The cost of the switch is
1047 // proportional to the size of the tree or the size of jump table range.
1049 // NB: We convert large switches which are just used to initialize large phi
1050 // nodes to lookup tables instead in simplify-cfg, so this shouldn't prevent
1051 // inlining those. It will prevent inlining in cases where the optimization
1052 // does not (yet) fire.
1054 // Maximum valid cost increased in this function.
1055 int CostUpperBound = INT_MAX - InlineConstants::InstrCost - 1;
1057 // Exit early for a large switch, assuming one case needs at least one
1059 // FIXME: This is not true for a bit test, but ignore such case for now to
1060 // save compile-time.
1061 int64_t CostLowerBound =
1062 std::min((int64_t)CostUpperBound,
1063 (int64_t)SI.getNumCases() * InlineConstants::InstrCost + Cost);
1065 if (CostLowerBound > Threshold) {
1066 Cost = CostLowerBound;
1070 unsigned JumpTableSize = 0;
1071 unsigned NumCaseCluster =
1072 TTI.getEstimatedNumberOfCaseClusters(SI, JumpTableSize);
1074 // If suitable for a jump table, consider the cost for the table size and
1075 // branch to destination.
1076 if (JumpTableSize) {
1077 int64_t JTCost = (int64_t)JumpTableSize * InlineConstants::InstrCost +
1078 4 * InlineConstants::InstrCost;
1080 Cost = std::min((int64_t)CostUpperBound, JTCost + Cost);
1084 // Considering forming a binary search, we should find the number of nodes
1085 // which is same as the number of comparisons when lowered. For a given
1086 // number of clusters, n, we can define a recursive function, f(n), to find
1087 // the number of nodes in the tree. The recursion is :
1088 // f(n) = 1 + f(n/2) + f (n - n/2), when n > 3,
1089 // and f(n) = n, when n <= 3.
1090 // This will lead a binary tree where the leaf should be either f(2) or f(3)
1091 // when n > 3. So, the number of comparisons from leaves should be n, while
1092 // the number of non-leaf should be :
1093 // 2^(log2(n) - 1) - 1
1094 // = 2^log2(n) * 2^-1 - 1
1096 // Considering comparisons from leaf and non-leaf nodes, we can estimate the
1097 // number of comparisons in a simple closed form :
1098 // n + n / 2 - 1 = n * 3 / 2 - 1
1099 if (NumCaseCluster <= 3) {
1100 // Suppose a comparison includes one compare and one conditional branch.
1101 Cost += NumCaseCluster * 2 * InlineConstants::InstrCost;
1105 int64_t ExpectedNumberOfCompare = 3 * (int64_t)NumCaseCluster / 2 - 1;
1106 int64_t SwitchCost =
1107 ExpectedNumberOfCompare * 2 * InlineConstants::InstrCost;
1109 Cost = std::min((int64_t)CostUpperBound, SwitchCost + Cost);
1113 bool CallAnalyzer::visitIndirectBrInst(IndirectBrInst &IBI) {
1114 // We never want to inline functions that contain an indirectbr. This is
1115 // incorrect because all the blockaddress's (in static global initializers
1116 // for example) would be referring to the original function, and this
1117 // indirect jump would jump from the inlined copy of the function into the
1118 // original function which is extremely undefined behavior.
1119 // FIXME: This logic isn't really right; we can safely inline functions with
1120 // indirectbr's as long as no other function or global references the
1121 // blockaddress of a block within the current function.
1122 HasIndirectBr = true;
1126 bool CallAnalyzer::visitResumeInst(ResumeInst &RI) {
1127 // FIXME: It's not clear that a single instruction is an accurate model for
1128 // the inline cost of a resume instruction.
1132 bool CallAnalyzer::visitCleanupReturnInst(CleanupReturnInst &CRI) {
1133 // FIXME: It's not clear that a single instruction is an accurate model for
1134 // the inline cost of a cleanupret instruction.
1138 bool CallAnalyzer::visitCatchReturnInst(CatchReturnInst &CRI) {
1139 // FIXME: It's not clear that a single instruction is an accurate model for
1140 // the inline cost of a catchret instruction.
1144 bool CallAnalyzer::visitUnreachableInst(UnreachableInst &I) {
1145 // FIXME: It might be reasonably to discount the cost of instructions leading
1146 // to unreachable as they have the lowest possible impact on both runtime and
1148 return true; // No actual code is needed for unreachable.
1151 bool CallAnalyzer::visitInstruction(Instruction &I) {
1152 // Some instructions are free. All of the free intrinsics can also be
1153 // handled by SROA, etc.
1154 if (TargetTransformInfo::TCC_Free == TTI.getUserCost(&I))
1157 // We found something we don't understand or can't handle. Mark any SROA-able
1158 // values in the operand list as no longer viable.
1159 for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI)
1165 /// \brief Analyze a basic block for its contribution to the inline cost.
1167 /// This method walks the analyzer over every instruction in the given basic
1168 /// block and accounts for their cost during inlining at this callsite. It
1169 /// aborts early if the threshold has been exceeded or an impossible to inline
1170 /// construct has been detected. It returns false if inlining is no longer
1171 /// viable, and true if inlining remains viable.
1172 bool CallAnalyzer::analyzeBlock(BasicBlock *BB,
1173 SmallPtrSetImpl<const Value *> &EphValues) {
1174 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1175 // FIXME: Currently, the number of instructions in a function regardless of
1176 // our ability to simplify them during inline to constants or dead code,
1177 // are actually used by the vector bonus heuristic. As long as that's true,
1178 // we have to special case debug intrinsics here to prevent differences in
1179 // inlining due to debug symbols. Eventually, the number of unsimplified
1180 // instructions shouldn't factor into the cost computation, but until then,
1181 // hack around it here.
1182 if (isa<DbgInfoIntrinsic>(I))
1185 // Skip ephemeral values.
1186 if (EphValues.count(&*I))
1190 if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy())
1191 ++NumVectorInstructions;
1193 // If the instruction is floating point, and the target says this operation
1194 // is expensive or the function has the "use-soft-float" attribute, this may
1195 // eventually become a library call. Treat the cost as such.
1196 if (I->getType()->isFloatingPointTy()) {
1197 // If the function has the "use-soft-float" attribute, mark it as
1199 if (TTI.getFPOpCost(I->getType()) == TargetTransformInfo::TCC_Expensive ||
1200 (F.getFnAttribute("use-soft-float").getValueAsString() == "true"))
1201 Cost += InlineConstants::CallPenalty;
1204 // If the instruction simplified to a constant, there is no cost to this
1205 // instruction. Visit the instructions using our InstVisitor to account for
1206 // all of the per-instruction logic. The visit tree returns true if we
1207 // consumed the instruction in any way, and false if the instruction's base
1208 // cost should count against inlining.
1209 if (Base::visit(&*I))
1210 ++NumInstructionsSimplified;
1212 Cost += InlineConstants::InstrCost;
1214 // If the visit this instruction detected an uninlinable pattern, abort.
1215 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca ||
1216 HasIndirectBr || HasFrameEscape)
1219 // If the caller is a recursive function then we don't want to inline
1220 // functions which allocate a lot of stack space because it would increase
1221 // the caller stack usage dramatically.
1222 if (IsCallerRecursive &&
1223 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller)
1226 // Check if we've past the maximum possible threshold so we don't spin in
1227 // huge basic blocks that will never inline.
1228 if (Cost > Threshold)
1235 /// \brief Compute the base pointer and cumulative constant offsets for V.
1237 /// This strips all constant offsets off of V, leaving it the base pointer, and
1238 /// accumulates the total constant offset applied in the returned constant. It
1239 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
1240 /// no constant offsets applied.
1241 ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) {
1242 if (!V->getType()->isPointerTy())
1245 unsigned IntPtrWidth = DL.getPointerSizeInBits();
1246 APInt Offset = APInt::getNullValue(IntPtrWidth);
1248 // Even though we don't look through PHI nodes, we could be called on an
1249 // instruction in an unreachable block, which may be on a cycle.
1250 SmallPtrSet<Value *, 4> Visited;
1253 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1254 if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset))
1256 V = GEP->getPointerOperand();
1257 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
1258 V = cast<Operator>(V)->getOperand(0);
1259 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1260 if (GA->isInterposable())
1262 V = GA->getAliasee();
1266 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
1267 } while (Visited.insert(V).second);
1269 Type *IntPtrTy = DL.getIntPtrType(V->getContext());
1270 return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset));
1273 /// \brief Analyze a call site for potential inlining.
1275 /// Returns true if inlining this call is viable, and false if it is not
1276 /// viable. It computes the cost and adjusts the threshold based on numerous
1277 /// factors and heuristics. If this method returns false but the computed cost
1278 /// is below the computed threshold, then inlining was forcibly disabled by
1279 /// some artifact of the routine.
1280 bool CallAnalyzer::analyzeCall(CallSite CS) {
1283 // Perform some tweaks to the cost and threshold based on the direct
1284 // callsite information.
1286 // We want to more aggressively inline vector-dense kernels, so up the
1287 // threshold, and we'll lower it if the % of vector instructions gets too
1288 // low. Note that these bonuses are some what arbitrary and evolved over time
1289 // by accident as much as because they are principled bonuses.
1291 // FIXME: It would be nice to remove all such bonuses. At least it would be
1292 // nice to base the bonus values on something more scientific.
1293 assert(NumInstructions == 0);
1294 assert(NumVectorInstructions == 0);
1296 // Update the threshold based on callsite properties
1297 updateThreshold(CS, F);
1299 FiftyPercentVectorBonus = 3 * Threshold / 2;
1300 TenPercentVectorBonus = 3 * Threshold / 4;
1302 // Track whether the post-inlining function would have more than one basic
1303 // block. A single basic block is often intended for inlining. Balloon the
1304 // threshold by 50% until we pass the single-BB phase.
1305 bool SingleBB = true;
1306 int SingleBBBonus = Threshold / 2;
1308 // Speculatively apply all possible bonuses to Threshold. If cost exceeds
1309 // this Threshold any time, and cost cannot decrease, we can stop processing
1310 // the rest of the function body.
1311 Threshold += (SingleBBBonus + FiftyPercentVectorBonus);
1313 // Give out bonuses for the callsite, as the instructions setting them up
1314 // will be gone after inlining.
1315 Cost -= getCallsiteCost(CS, DL);
1317 // If there is only one call of the function, and it has internal linkage,
1318 // the cost of inlining it drops dramatically.
1319 bool OnlyOneCallAndLocalLinkage =
1320 F.hasLocalLinkage() && F.hasOneUse() && &F == CS.getCalledFunction();
1321 if (OnlyOneCallAndLocalLinkage)
1322 Cost -= InlineConstants::LastCallToStaticBonus;
1324 // If this function uses the coldcc calling convention, prefer not to inline
1326 if (F.getCallingConv() == CallingConv::Cold)
1327 Cost += InlineConstants::ColdccPenalty;
1329 // Check if we're done. This can happen due to bonuses and penalties.
1330 if (Cost > Threshold)
1336 Function *Caller = CS.getInstruction()->getParent()->getParent();
1337 // Check if the caller function is recursive itself.
1338 for (User *U : Caller->users()) {
1342 Instruction *I = Site.getInstruction();
1343 if (I->getParent()->getParent() == Caller) {
1344 IsCallerRecursive = true;
1349 // Populate our simplified values by mapping from function arguments to call
1350 // arguments with known important simplifications.
1351 CallSite::arg_iterator CAI = CS.arg_begin();
1352 for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end();
1353 FAI != FAE; ++FAI, ++CAI) {
1354 assert(CAI != CS.arg_end());
1355 if (Constant *C = dyn_cast<Constant>(CAI))
1356 SimplifiedValues[&*FAI] = C;
1358 Value *PtrArg = *CAI;
1359 if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) {
1360 ConstantOffsetPtrs[&*FAI] = std::make_pair(PtrArg, C->getValue());
1362 // We can SROA any pointer arguments derived from alloca instructions.
1363 if (isa<AllocaInst>(PtrArg)) {
1364 SROAArgValues[&*FAI] = PtrArg;
1365 SROAArgCosts[PtrArg] = 0;
1369 NumConstantArgs = SimplifiedValues.size();
1370 NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size();
1371 NumAllocaArgs = SROAArgValues.size();
1373 // FIXME: If a caller has multiple calls to a callee, we end up recomputing
1374 // the ephemeral values multiple times (and they're completely determined by
1375 // the callee, so this is purely duplicate work).
1376 SmallPtrSet<const Value *, 32> EphValues;
1377 CodeMetrics::collectEphemeralValues(&F, &GetAssumptionCache(F), EphValues);
1379 // The worklist of live basic blocks in the callee *after* inlining. We avoid
1380 // adding basic blocks of the callee which can be proven to be dead for this
1381 // particular call site in order to get more accurate cost estimates. This
1382 // requires a somewhat heavyweight iteration pattern: we need to walk the
1383 // basic blocks in a breadth-first order as we insert live successors. To
1384 // accomplish this, prioritizing for small iterations because we exit after
1385 // crossing our threshold, we use a small-size optimized SetVector.
1386 typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>,
1387 SmallPtrSet<BasicBlock *, 16>>
1389 BBSetVector BBWorklist;
1390 BBWorklist.insert(&F.getEntryBlock());
1391 // Note that we *must not* cache the size, this loop grows the worklist.
1392 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
1393 // Bail out the moment we cross the threshold. This means we'll under-count
1394 // the cost, but only when undercounting doesn't matter.
1395 if (Cost > Threshold)
1398 BasicBlock *BB = BBWorklist[Idx];
1402 // Disallow inlining a blockaddress. A blockaddress only has defined
1403 // behavior for an indirect branch in the same function, and we do not
1404 // currently support inlining indirect branches. But, the inliner may not
1405 // see an indirect branch that ends up being dead code at a particular call
1406 // site. If the blockaddress escapes the function, e.g., via a global
1407 // variable, inlining may lead to an invalid cross-function reference.
1408 if (BB->hasAddressTaken())
1411 // Analyze the cost of this block. If we blow through the threshold, this
1412 // returns false, and we can bail on out.
1413 if (!analyzeBlock(BB, EphValues))
1416 TerminatorInst *TI = BB->getTerminator();
1418 // Add in the live successors by first checking whether we have terminator
1419 // that may be simplified based on the values simplified by this call.
1420 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
1421 if (BI->isConditional()) {
1422 Value *Cond = BI->getCondition();
1423 if (ConstantInt *SimpleCond =
1424 dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
1425 BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0));
1429 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
1430 Value *Cond = SI->getCondition();
1431 if (ConstantInt *SimpleCond =
1432 dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
1433 BBWorklist.insert(SI->findCaseValue(SimpleCond)->getCaseSuccessor());
1438 // If we're unable to select a particular successor, just count all of
1440 for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize;
1442 BBWorklist.insert(TI->getSuccessor(TIdx));
1444 // If we had any successors at this point, than post-inlining is likely to
1445 // have them as well. Note that we assume any basic blocks which existed
1446 // due to branches or switches which folded above will also fold after
1448 if (SingleBB && TI->getNumSuccessors() > 1) {
1449 // Take off the bonus we applied to the threshold.
1450 Threshold -= SingleBBBonus;
1455 // If this is a noduplicate call, we can still inline as long as
1456 // inlining this would cause the removal of the caller (so the instruction
1457 // is not actually duplicated, just moved).
1458 if (!OnlyOneCallAndLocalLinkage && ContainsNoDuplicateCall)
1461 // We applied the maximum possible vector bonus at the beginning. Now,
1462 // subtract the excess bonus, if any, from the Threshold before
1463 // comparing against Cost.
1464 if (NumVectorInstructions <= NumInstructions / 10)
1465 Threshold -= FiftyPercentVectorBonus;
1466 else if (NumVectorInstructions <= NumInstructions / 2)
1467 Threshold -= (FiftyPercentVectorBonus - TenPercentVectorBonus);
1469 return Cost < std::max(1, Threshold);
1472 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1473 /// \brief Dump stats about this call's analysis.
1474 LLVM_DUMP_METHOD void CallAnalyzer::dump() {
1475 #define DEBUG_PRINT_STAT(x) dbgs() << " " #x ": " << x << "\n"
1476 DEBUG_PRINT_STAT(NumConstantArgs);
1477 DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs);
1478 DEBUG_PRINT_STAT(NumAllocaArgs);
1479 DEBUG_PRINT_STAT(NumConstantPtrCmps);
1480 DEBUG_PRINT_STAT(NumConstantPtrDiffs);
1481 DEBUG_PRINT_STAT(NumInstructionsSimplified);
1482 DEBUG_PRINT_STAT(NumInstructions);
1483 DEBUG_PRINT_STAT(SROACostSavings);
1484 DEBUG_PRINT_STAT(SROACostSavingsLost);
1485 DEBUG_PRINT_STAT(ContainsNoDuplicateCall);
1486 DEBUG_PRINT_STAT(Cost);
1487 DEBUG_PRINT_STAT(Threshold);
1488 #undef DEBUG_PRINT_STAT
1492 /// \brief Test that there are no attribute conflicts between Caller and Callee
1493 /// that prevent inlining.
1494 static bool functionsHaveCompatibleAttributes(Function *Caller,
1496 TargetTransformInfo &TTI) {
1497 return TTI.areInlineCompatible(Caller, Callee) &&
1498 AttributeFuncs::areInlineCompatible(*Caller, *Callee);
1501 int llvm::getCallsiteCost(CallSite CS, const DataLayout &DL) {
1503 for (unsigned I = 0, E = CS.arg_size(); I != E; ++I) {
1504 if (CS.isByValArgument(I)) {
1505 // We approximate the number of loads and stores needed by dividing the
1506 // size of the byval type by the target's pointer size.
1507 PointerType *PTy = cast<PointerType>(CS.getArgument(I)->getType());
1508 unsigned TypeSize = DL.getTypeSizeInBits(PTy->getElementType());
1509 unsigned PointerSize = DL.getPointerSizeInBits();
1510 // Ceiling division.
1511 unsigned NumStores = (TypeSize + PointerSize - 1) / PointerSize;
1513 // If it generates more than 8 stores it is likely to be expanded as an
1514 // inline memcpy so we take that as an upper bound. Otherwise we assume
1515 // one load and one store per word copied.
1516 // FIXME: The maxStoresPerMemcpy setting from the target should be used
1517 // here instead of a magic number of 8, but it's not available via
1519 NumStores = std::min(NumStores, 8U);
1521 Cost += 2 * NumStores * InlineConstants::InstrCost;
1523 // For non-byval arguments subtract off one instruction per call
1525 Cost += InlineConstants::InstrCost;
1528 // The call instruction also disappears after inlining.
1529 Cost += InlineConstants::InstrCost + InlineConstants::CallPenalty;
1533 InlineCost llvm::getInlineCost(
1534 CallSite CS, const InlineParams &Params, TargetTransformInfo &CalleeTTI,
1535 std::function<AssumptionCache &(Function &)> &GetAssumptionCache,
1536 Optional<function_ref<BlockFrequencyInfo &(Function &)>> GetBFI,
1537 ProfileSummaryInfo *PSI) {
1538 return getInlineCost(CS, CS.getCalledFunction(), Params, CalleeTTI,
1539 GetAssumptionCache, GetBFI, PSI);
1542 InlineCost llvm::getInlineCost(
1543 CallSite CS, Function *Callee, const InlineParams &Params,
1544 TargetTransformInfo &CalleeTTI,
1545 std::function<AssumptionCache &(Function &)> &GetAssumptionCache,
1546 Optional<function_ref<BlockFrequencyInfo &(Function &)>> GetBFI,
1547 ProfileSummaryInfo *PSI) {
1549 // Cannot inline indirect calls.
1551 return llvm::InlineCost::getNever();
1553 // Calls to functions with always-inline attributes should be inlined
1554 // whenever possible.
1555 if (CS.hasFnAttr(Attribute::AlwaysInline)) {
1556 if (isInlineViable(*Callee))
1557 return llvm::InlineCost::getAlways();
1558 return llvm::InlineCost::getNever();
1561 // Never inline functions with conflicting attributes (unless callee has
1562 // always-inline attribute).
1563 if (!functionsHaveCompatibleAttributes(CS.getCaller(), Callee, CalleeTTI))
1564 return llvm::InlineCost::getNever();
1566 // Don't inline this call if the caller has the optnone attribute.
1567 if (CS.getCaller()->hasFnAttribute(Attribute::OptimizeNone))
1568 return llvm::InlineCost::getNever();
1570 // Don't inline functions which can be interposed at link-time. Don't inline
1571 // functions marked noinline or call sites marked noinline.
1572 // Note: inlining non-exact non-interposable functions is fine, since we know
1573 // we have *a* correct implementation of the source level function.
1574 if (Callee->isInterposable() || Callee->hasFnAttribute(Attribute::NoInline) ||
1576 return llvm::InlineCost::getNever();
1578 DEBUG(llvm::dbgs() << " Analyzing call of " << Callee->getName()
1581 CallAnalyzer CA(CalleeTTI, GetAssumptionCache, GetBFI, PSI, *Callee, CS,
1583 bool ShouldInline = CA.analyzeCall(CS);
1587 // Check if there was a reason to force inlining or no inlining.
1588 if (!ShouldInline && CA.getCost() < CA.getThreshold())
1589 return InlineCost::getNever();
1590 if (ShouldInline && CA.getCost() >= CA.getThreshold())
1591 return InlineCost::getAlways();
1593 return llvm::InlineCost::get(CA.getCost(), CA.getThreshold());
1596 bool llvm::isInlineViable(Function &F) {
1597 bool ReturnsTwice = F.hasFnAttribute(Attribute::ReturnsTwice);
1598 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
1599 // Disallow inlining of functions which contain indirect branches or
1601 if (isa<IndirectBrInst>(BI->getTerminator()) || BI->hasAddressTaken())
1604 for (auto &II : *BI) {
1609 // Disallow recursive calls.
1610 if (&F == CS.getCalledFunction())
1613 // Disallow calls which expose returns-twice to a function not previously
1614 // attributed as such.
1615 if (!ReturnsTwice && CS.isCall() &&
1616 cast<CallInst>(CS.getInstruction())->canReturnTwice())
1619 // Disallow inlining functions that call @llvm.localescape. Doing this
1620 // correctly would require major changes to the inliner.
1621 if (CS.getCalledFunction() &&
1622 CS.getCalledFunction()->getIntrinsicID() ==
1623 llvm::Intrinsic::localescape)
1631 // APIs to create InlineParams based on command line flags and/or other
1634 InlineParams llvm::getInlineParams(int Threshold) {
1635 InlineParams Params;
1637 // This field is the threshold to use for a callee by default. This is
1638 // derived from one or more of:
1639 // * optimization or size-optimization levels,
1640 // * a value passed to createFunctionInliningPass function, or
1641 // * the -inline-threshold flag.
1642 // If the -inline-threshold flag is explicitly specified, that is used
1643 // irrespective of anything else.
1644 if (InlineThreshold.getNumOccurrences() > 0)
1645 Params.DefaultThreshold = InlineThreshold;
1647 Params.DefaultThreshold = Threshold;
1649 // Set the HintThreshold knob from the -inlinehint-threshold.
1650 Params.HintThreshold = HintThreshold;
1652 // Set the HotCallSiteThreshold knob from the -hot-callsite-threshold.
1653 Params.HotCallSiteThreshold = HotCallSiteThreshold;
1655 // Set the ColdCallSiteThreshold knob from the -inline-cold-callsite-threshold.
1656 Params.ColdCallSiteThreshold = ColdCallSiteThreshold;
1658 // Set the OptMinSizeThreshold and OptSizeThreshold params only if the
1659 // -inlinehint-threshold commandline option is not explicitly given. If that
1660 // option is present, then its value applies even for callees with size and
1661 // minsize attributes.
1662 // If the -inline-threshold is not specified, set the ColdThreshold from the
1663 // -inlinecold-threshold even if it is not explicitly passed. If
1664 // -inline-threshold is specified, then -inlinecold-threshold needs to be
1665 // explicitly specified to set the ColdThreshold knob
1666 if (InlineThreshold.getNumOccurrences() == 0) {
1667 Params.OptMinSizeThreshold = InlineConstants::OptMinSizeThreshold;
1668 Params.OptSizeThreshold = InlineConstants::OptSizeThreshold;
1669 Params.ColdThreshold = ColdThreshold;
1670 } else if (ColdThreshold.getNumOccurrences() > 0) {
1671 Params.ColdThreshold = ColdThreshold;
1676 InlineParams llvm::getInlineParams() {
1677 return getInlineParams(InlineThreshold);
1680 // Compute the default threshold for inlining based on the opt level and the
1682 static int computeThresholdFromOptLevels(unsigned OptLevel,
1683 unsigned SizeOptLevel) {
1685 return InlineConstants::OptAggressiveThreshold;
1686 if (SizeOptLevel == 1) // -Os
1687 return InlineConstants::OptSizeThreshold;
1688 if (SizeOptLevel == 2) // -Oz
1689 return InlineConstants::OptMinSizeThreshold;
1690 return InlineThreshold;
1693 InlineParams llvm::getInlineParams(unsigned OptLevel, unsigned SizeOptLevel) {
1694 return getInlineParams(computeThresholdFromOptLevels(OptLevel, SizeOptLevel));