LCOV - code coverage report
Current view: top level - lib/Analysis - InlineCost.cpp (source / functions) Hit Total Coverage
Test: llvm-toolchain.info Lines: 607 627 96.8 %
Date: 2017-09-14 15:23:50 Functions: 55 56 98.2 %
Legend: Lines: hit not hit

          Line data    Source code
       1             : //===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
       2             : //
       3             : //                     The LLVM Compiler Infrastructure
       4             : //
       5             : // This file is distributed under the University of Illinois Open Source
       6             : // License. See LICENSE.TXT for details.
       7             : //
       8             : //===----------------------------------------------------------------------===//
       9             : //
      10             : // This file implements inline cost analysis.
      11             : //
      12             : //===----------------------------------------------------------------------===//
      13             : 
      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"
      37             : 
      38             : using namespace llvm;
      39             : 
      40             : #define DEBUG_TYPE "inline-cost"
      41             : 
      42             : STATISTIC(NumCallsAnalyzed, "Number of call sites analyzed");
      43             : 
      44       72306 : static cl::opt<int> InlineThreshold(
      45      216918 :     "inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
      46      289224 :     cl::desc("Control the amount of inlining to perform (default = 225)"));
      47             : 
      48       72306 : static cl::opt<int> HintThreshold(
      49      216918 :     "inlinehint-threshold", cl::Hidden, cl::init(325),
      50      289224 :     cl::desc("Threshold for inlining functions with inline hint"));
      51             : 
      52             : static cl::opt<int>
      53       72306 :     ColdCallSiteThreshold("inline-cold-callsite-threshold", cl::Hidden,
      54      216918 :                           cl::init(45),
      55      289224 :                           cl::desc("Threshold for inlining cold callsites"));
      56             : 
      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
      59             : // BFI.
      60       72306 : static cl::opt<int> ColdThreshold(
      61      216918 :     "inlinecold-threshold", cl::Hidden, cl::init(45),
      62      289224 :     cl::desc("Threshold for inlining functions with cold attribute"));
      63             : 
      64             : static cl::opt<int>
      65      289224 :     HotCallSiteThreshold("hot-callsite-threshold", cl::Hidden, cl::init(3000),
      66             :                          cl::ZeroOrMore,
      67      289224 :                          cl::desc("Threshold for hot callsites "));
      68             : 
      69       72306 : static cl::opt<int> LocallyHotCallSiteThreshold(
      70      216918 :     "locally-hot-callsite-threshold", cl::Hidden, cl::init(525), cl::ZeroOrMore,
      71      289224 :     cl::desc("Threshold for locally hot callsites "));
      72             : 
      73       72306 : static cl::opt<int> ColdCallSiteRelFreq(
      74      216918 :     "cold-callsite-rel-freq", cl::Hidden, cl::init(2), cl::ZeroOrMore,
      75      216918 :     cl::desc("Maxmimum block frequency, expressed as a percentage of caller's "
      76             :              "entry frequency, for a callsite to be cold in the absence of "
      77      216918 :              "profile information."));
      78             : 
      79       72306 : static cl::opt<int> HotCallSiteRelFreq(
      80      216918 :     "hot-callsite-rel-freq", cl::Hidden, cl::init(60), cl::ZeroOrMore,
      81      216918 :     cl::desc("Minimum block frequency, expressed as a multiple of caller's "
      82             :              "entry frequency, for a callsite to be hot in the absence of "
      83      216918 :              "profile information."));
      84             : 
      85       72306 : static cl::opt<bool> ComputeFullInlineCost(
      86      216918 :     "inline-cost-full", cl::Hidden, cl::init(false),
      87      216918 :     cl::desc("Compute the full inline cost of a call site even when the cost "
      88       72306 :              "exceeds the threshold."));
      89             : 
      90             : namespace {
      91             : 
      92     1140745 : class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> {
      93             :   typedef InstVisitor<CallAnalyzer, bool> Base;
      94             :   friend class InstVisitor<CallAnalyzer, bool>;
      95             : 
      96             :   /// The TargetTransformInfo available for this compilation.
      97             :   const TargetTransformInfo &TTI;
      98             : 
      99             :   /// Getter for the cache of @llvm.assume intrinsics.
     100             :   std::function<AssumptionCache &(Function &)> &GetAssumptionCache;
     101             : 
     102             :   /// Getter for BlockFrequencyInfo
     103             :   Optional<function_ref<BlockFrequencyInfo &(Function &)>> &GetBFI;
     104             : 
     105             :   /// Profile summary information.
     106             :   ProfileSummaryInfo *PSI;
     107             : 
     108             :   /// The called function.
     109             :   Function &F;
     110             : 
     111             :   // Cache the DataLayout since we use it a lot.
     112             :   const DataLayout &DL;
     113             : 
     114             :   /// The OptimizationRemarkEmitter available for this compilation.
     115             :   OptimizationRemarkEmitter *ORE;
     116             : 
     117             :   /// The candidate callsite being analyzed. Please do not use this to do
     118             :   /// analysis in the caller function; we want the inline cost query to be
     119             :   /// easily cacheable. Instead, use the cover function paramHasAttr.
     120             :   CallSite CandidateCS;
     121             : 
     122             :   /// Tunable parameters that control the analysis.
     123             :   const InlineParams &Params;
     124             : 
     125             :   int Threshold;
     126             :   int Cost;
     127             : 
     128             :   bool IsCallerRecursive;
     129             :   bool IsRecursiveCall;
     130             :   bool ExposesReturnsTwice;
     131             :   bool HasDynamicAlloca;
     132             :   bool ContainsNoDuplicateCall;
     133             :   bool HasReturn;
     134             :   bool HasIndirectBr;
     135             :   bool HasFrameEscape;
     136             : 
     137             :   /// Number of bytes allocated statically by the callee.
     138             :   uint64_t AllocatedSize;
     139             :   unsigned NumInstructions, NumVectorInstructions;
     140             :   int VectorBonus, TenPercentVectorBonus;
     141             :   // Bonus to be applied when the callee has only one reachable basic block.
     142             :   int SingleBBBonus;
     143             : 
     144             :   /// While we walk the potentially-inlined instructions, we build up and
     145             :   /// maintain a mapping of simplified values specific to this callsite. The
     146             :   /// idea is to propagate any special information we have about arguments to
     147             :   /// this call through the inlinable section of the function, and account for
     148             :   /// likely simplifications post-inlining. The most important aspect we track
     149             :   /// is CFG altering simplifications -- when we prove a basic block dead, that
     150             :   /// can cause dramatic shifts in the cost of inlining a function.
     151             :   DenseMap<Value *, Constant *> SimplifiedValues;
     152             : 
     153             :   /// Keep track of the values which map back (through function arguments) to
     154             :   /// allocas on the caller stack which could be simplified through SROA.
     155             :   DenseMap<Value *, Value *> SROAArgValues;
     156             : 
     157             :   /// The mapping of caller Alloca values to their accumulated cost savings. If
     158             :   /// we have to disable SROA for one of the allocas, this tells us how much
     159             :   /// cost must be added.
     160             :   DenseMap<Value *, int> SROAArgCosts;
     161             : 
     162             :   /// Keep track of values which map to a pointer base and constant offset.
     163             :   DenseMap<Value *, std::pair<Value *, APInt>> ConstantOffsetPtrs;
     164             : 
     165             :   // Custom simplification helper routines.
     166             :   bool isAllocaDerivedArg(Value *V);
     167             :   bool lookupSROAArgAndCost(Value *V, Value *&Arg,
     168             :                             DenseMap<Value *, int>::iterator &CostIt);
     169             :   void disableSROA(DenseMap<Value *, int>::iterator CostIt);
     170             :   void disableSROA(Value *V);
     171             :   void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
     172             :                           int InstructionCost);
     173             :   bool isGEPFree(GetElementPtrInst &GEP);
     174             :   bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset);
     175             :   bool simplifyCallSite(Function *F, CallSite CS);
     176             :   template <typename Callable>
     177             :   bool simplifyInstruction(Instruction &I, Callable Evaluate);
     178             :   ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V);
     179             : 
     180             :   /// Return true if the given argument to the function being considered for
     181             :   /// inlining has the given attribute set either at the call site or the
     182             :   /// function declaration.  Primarily used to inspect call site specific
     183             :   /// attributes since these can be more precise than the ones on the callee
     184             :   /// itself.
     185             :   bool paramHasAttr(Argument *A, Attribute::AttrKind Attr);
     186             : 
     187             :   /// Return true if the given value is known non null within the callee if
     188             :   /// inlined through this particular callsite.
     189             :   bool isKnownNonNullInCallee(Value *V);
     190             : 
     191             :   /// Update Threshold based on callsite properties such as callee
     192             :   /// attributes and callee hotness for PGO builds. The Callee is explicitly
     193             :   /// passed to support analyzing indirect calls whose target is inferred by
     194             :   /// analysis.
     195             :   void updateThreshold(CallSite CS, Function &Callee);
     196             : 
     197             :   /// Return true if size growth is allowed when inlining the callee at CS.
     198             :   bool allowSizeGrowth(CallSite CS);
     199             : 
     200             :   /// Return true if \p CS is a cold callsite.
     201             :   bool isColdCallSite(CallSite CS, BlockFrequencyInfo *CallerBFI);
     202             : 
     203             :   /// Return a higher threshold if \p CS is a hot callsite.
     204             :   Optional<int> getHotCallSiteThreshold(CallSite CS,
     205             :                                         BlockFrequencyInfo *CallerBFI);
     206             : 
     207             :   // Custom analysis routines.
     208             :   bool analyzeBlock(BasicBlock *BB, SmallPtrSetImpl<const Value *> &EphValues);
     209             : 
     210             :   // Disable several entry points to the visitor so we don't accidentally use
     211             :   // them by declaring but not defining them here.
     212             :   void visit(Module *);
     213             :   void visit(Module &);
     214             :   void visit(Function *);
     215             :   void visit(Function &);
     216             :   void visit(BasicBlock *);
     217             :   void visit(BasicBlock &);
     218             : 
     219             :   // Provide base case for our instruction visit.
     220             :   bool visitInstruction(Instruction &I);
     221             : 
     222             :   // Our visit overrides.
     223             :   bool visitAlloca(AllocaInst &I);
     224             :   bool visitPHI(PHINode &I);
     225             :   bool visitGetElementPtr(GetElementPtrInst &I);
     226             :   bool visitBitCast(BitCastInst &I);
     227             :   bool visitPtrToInt(PtrToIntInst &I);
     228             :   bool visitIntToPtr(IntToPtrInst &I);
     229             :   bool visitCastInst(CastInst &I);
     230             :   bool visitUnaryInstruction(UnaryInstruction &I);
     231             :   bool visitCmpInst(CmpInst &I);
     232             :   bool visitAnd(BinaryOperator &I);
     233             :   bool visitOr(BinaryOperator &I);
     234             :   bool visitSub(BinaryOperator &I);
     235             :   bool visitBinaryOperator(BinaryOperator &I);
     236             :   bool visitLoad(LoadInst &I);
     237             :   bool visitStore(StoreInst &I);
     238             :   bool visitExtractValue(ExtractValueInst &I);
     239             :   bool visitInsertValue(InsertValueInst &I);
     240             :   bool visitCallSite(CallSite CS);
     241             :   bool visitReturnInst(ReturnInst &RI);
     242             :   bool visitBranchInst(BranchInst &BI);
     243             :   bool visitSwitchInst(SwitchInst &SI);
     244             :   bool visitIndirectBrInst(IndirectBrInst &IBI);
     245             :   bool visitResumeInst(ResumeInst &RI);
     246             :   bool visitCleanupReturnInst(CleanupReturnInst &RI);
     247             :   bool visitCatchReturnInst(CatchReturnInst &RI);
     248             :   bool visitUnreachableInst(UnreachableInst &I);
     249             : 
     250             : public:
     251      228149 :   CallAnalyzer(const TargetTransformInfo &TTI,
     252             :                std::function<AssumptionCache &(Function &)> &GetAssumptionCache,
     253             :                Optional<function_ref<BlockFrequencyInfo &(Function &)>> &GetBFI,
     254             :                ProfileSummaryInfo *PSI, OptimizationRemarkEmitter *ORE,
     255             :                Function &Callee, CallSite CSArg, const InlineParams &Params)
     256      228149 :       : TTI(TTI), GetAssumptionCache(GetAssumptionCache), GetBFI(GetBFI),
     257      228149 :         PSI(PSI), F(Callee), DL(F.getParent()->getDataLayout()), ORE(ORE),
     258      228149 :         CandidateCS(CSArg), Params(Params), Threshold(Params.DefaultThreshold),
     259             :         Cost(0), IsCallerRecursive(false), IsRecursiveCall(false),
     260             :         ExposesReturnsTwice(false), HasDynamicAlloca(false),
     261             :         ContainsNoDuplicateCall(false), HasReturn(false), HasIndirectBr(false),
     262             :         HasFrameEscape(false), AllocatedSize(0), NumInstructions(0),
     263             :         NumVectorInstructions(0), VectorBonus(0), SingleBBBonus(0),
     264             :         NumConstantArgs(0), NumConstantOffsetPtrArgs(0), NumAllocaArgs(0),
     265             :         NumConstantPtrCmps(0), NumConstantPtrDiffs(0),
     266             :         NumInstructionsSimplified(0), SROACostSavings(0),
     267     1597043 :         SROACostSavingsLost(0) {}
     268             : 
     269             :   bool analyzeCall(CallSite CS);
     270             : 
     271             :   int getThreshold() { return Threshold; }
     272             :   int getCost() { return Cost; }
     273             : 
     274             :   // Keep a bunch of stats about the cost savings found so we can print them
     275             :   // out when debugging.
     276             :   unsigned NumConstantArgs;
     277             :   unsigned NumConstantOffsetPtrArgs;
     278             :   unsigned NumAllocaArgs;
     279             :   unsigned NumConstantPtrCmps;
     280             :   unsigned NumConstantPtrDiffs;
     281             :   unsigned NumInstructionsSimplified;
     282             :   unsigned SROACostSavings;
     283             :   unsigned SROACostSavingsLost;
     284             : 
     285             :   void dump();
     286             : };
     287             : 
     288             : } // namespace
     289             : 
     290             : /// \brief Test whether the given value is an Alloca-derived function argument.
     291             : bool CallAnalyzer::isAllocaDerivedArg(Value *V) {
     292      108120 :   return SROAArgValues.count(V);
     293             : }
     294             : 
     295             : /// \brief Lookup the SROA-candidate argument and cost iterator which V maps to.
     296             : /// Returns false if V does not map to a SROA-candidate.
     297    14789751 : bool CallAnalyzer::lookupSROAArgAndCost(
     298             :     Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) {
     299    37681600 :   if (SROAArgValues.empty() || SROAArgCosts.empty())
     300             :     return false;
     301             : 
     302     5921650 :   DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V);
     303    17764950 :   if (ArgIt == SROAArgValues.end())
     304             :     return false;
     305             : 
     306      391283 :   Arg = ArgIt->second;
     307      391283 :   CostIt = SROAArgCosts.find(Arg);
     308     1173849 :   return CostIt != SROAArgCosts.end();
     309             : }
     310             : 
     311             : /// \brief Disable SROA for the candidate marked by this cost iterator.
     312             : ///
     313             : /// This marks the candidate as no longer viable for SROA, and adds the cost
     314             : /// savings associated with it back into the inline cost measurement.
     315             : void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) {
     316             :   // If we're no longer able to perform SROA we need to undo its cost savings
     317             :   // and prevent subsequent analysis.
     318      105789 :   Cost += CostIt->second;
     319      105789 :   SROACostSavings -= CostIt->second;
     320      105789 :   SROACostSavingsLost += CostIt->second;
     321      211578 :   SROAArgCosts.erase(CostIt);
     322             : }
     323             : 
     324             : /// \brief If 'V' maps to a SROA candidate, disable SROA for it.
     325     8775648 : void CallAnalyzer::disableSROA(Value *V) {
     326             :   Value *SROAArg;
     327     8775648 :   DenseMap<Value *, int>::iterator CostIt;
     328     8775648 :   if (lookupSROAArgAndCost(V, SROAArg, CostIt))
     329      101363 :     disableSROA(CostIt);
     330     8775648 : }
     331             : 
     332             : /// \brief Accumulate the given cost for a particular SROA candidate.
     333             : void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt,
     334             :                                       int InstructionCost) {
     335      133930 :   CostIt->second += InstructionCost;
     336      133930 :   SROACostSavings += InstructionCost;
     337             : }
     338             : 
     339             : /// \brief Accumulate a constant GEP offset into an APInt if possible.
     340             : ///
     341             : /// Returns false if unable to compute the offset for any reason. Respects any
     342             : /// simplified values known during the analysis of this callsite.
     343      370882 : bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) {
     344      741764 :   unsigned IntPtrWidth = DL.getPointerSizeInBits();
     345             :   assert(IntPtrWidth == Offset.getBitWidth());
     346             : 
     347     1231174 :   for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
     348     1231174 :        GTI != GTE; ++GTI) {
     349     1776718 :     ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
     350             :     if (!OpC)
     351       92523 :       if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand()))
     352             :         OpC = dyn_cast<ConstantInt>(SimpleOp);
     353      888359 :     if (!OpC)
     354       28067 :       return false;
     355      860292 :     if (OpC->isZero())
     356     1535251 :       continue;
     357             : 
     358             :     // Handle a struct index, which adds its field offset to the pointer.
     359      145407 :     if (StructType *STy = GTI.getStructTypeOrNull()) {
     360      145407 :       unsigned ElementIdx = OpC->getZExtValue();
     361      145407 :       const StructLayout *SL = DL.getStructLayout(STy);
     362      436221 :       Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx));
     363      145407 :       continue;
     364             :     }
     365             : 
     366       59889 :     APInt TypeSize(IntPtrWidth, DL.getTypeAllocSize(GTI.getIndexedType()));
     367       59889 :     Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize;
     368             :   }
     369      342815 :   return true;
     370             : }
     371             : 
     372             : /// \brief Use TTI to check whether a GEP is free.
     373             : ///
     374             : /// Respects any simplified values known during the analysis of this callsite.
     375      225037 : bool CallAnalyzer::isGEPFree(GetElementPtrInst &GEP) {
     376      450074 :   SmallVector<Value *, 4> Operands;
     377      225037 :   Operands.push_back(GEP.getOperand(0));
     378      911222 :   for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
     379      922296 :     if (Constant *SimpleOp = SimplifiedValues.lookup(*I))
     380         140 :        Operands.push_back(SimpleOp);
     381             :      else
     382      461008 :        Operands.push_back(*I);
     383      675111 :   return TargetTransformInfo::TCC_Free == TTI.getUserCost(&GEP, Operands);
     384             : }
     385             : 
     386     1316896 : bool CallAnalyzer::visitAlloca(AllocaInst &I) {
     387             :   // Check whether inlining will turn a dynamic alloca into a static
     388             :   // alloca and handle that case.
     389     1316896 :   if (I.isArrayAllocation()) {
     390         150 :     Constant *Size = SimplifiedValues.lookup(I.getArraySize());
     391          46 :     if (auto *AllocSize = dyn_cast_or_null<ConstantInt>(Size)) {
     392          41 :       Type *Ty = I.getAllocatedType();
     393         123 :       AllocatedSize = SaturatingMultiplyAdd(
     394          41 :           AllocSize->getLimitedValue(), DL.getTypeAllocSize(Ty), AllocatedSize);
     395          41 :       return Base::visitAlloca(I);
     396             :     }
     397             :   }
     398             : 
     399             :   // Accumulate the allocated size.
     400     1316855 :   if (I.isStaticAlloca()) {
     401     1316848 :     Type *Ty = I.getAllocatedType();
     402     2633696 :     AllocatedSize = SaturatingAdd(DL.getTypeAllocSize(Ty), AllocatedSize);
     403             :   }
     404             : 
     405             :   // We will happily inline static alloca instructions.
     406     1316855 :   if (I.isStaticAlloca())
     407     1316848 :     return Base::visitAlloca(I);
     408             : 
     409             :   // FIXME: This is overly conservative. Dynamic allocas are inefficient for
     410             :   // a variety of reasons, and so we would like to not inline them into
     411             :   // functions which don't currently have a dynamic alloca. This simply
     412             :   // disables inlining altogether in the presence of a dynamic alloca.
     413           7 :   HasDynamicAlloca = true;
     414           7 :   return false;
     415             : }
     416             : 
     417             : bool CallAnalyzer::visitPHI(PHINode &I) {
     418             :   // FIXME: We should potentially be tracking values through phi nodes,
     419             :   // especially when they collapse to a single value due to deleted CFG edges
     420             :   // during inlining.
     421             : 
     422             :   // FIXME: We need to propagate SROA *disabling* through phi nodes, even
     423             :   // though we don't want to propagate it's bonuses. The idea is to disable
     424             :   // SROA if it *might* be used in an inappropriate manner.
     425             : 
     426             :   // Phi nodes are always zero-cost.
     427             :   return true;
     428             : }
     429             : 
     430      710554 : bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) {
     431             :   Value *SROAArg;
     432      710554 :   DenseMap<Value *, int>::iterator CostIt;
     433             :   bool SROACandidate =
     434      710554 :       lookupSROAArgAndCost(I.getPointerOperand(), SROAArg, CostIt);
     435             : 
     436             :   // Try to fold GEPs of constant-offset call site argument pointers. This
     437             :   // requires target data and inbounds GEPs.
     438      710554 :   if (I.isInBounds()) {
     439             :     // Check if we have a base + offset for the pointer.
     440      704882 :     Value *Ptr = I.getPointerOperand();
     441     1108924 :     std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr);
     442      704882 :     if (BaseAndOffset.first) {
     443             :       // Check if the offset of this GEP is constant, and if so accumulate it
     444             :       // into Offset.
     445      300840 :       if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) {
     446             :         // Non-constant GEPs aren't folded, and disable SROA.
     447       24488 :         if (SROACandidate)
     448        3930 :           disableSROA(CostIt);
     449      325328 :         return isGEPFree(I);
     450             :       }
     451             : 
     452             :       // Add the result as a new mapping to Base + Offset.
     453      829056 :       ConstantOffsetPtrs[&I] = BaseAndOffset;
     454             : 
     455             :       // Also handle SROA candidates here, we already know that the GEP is
     456             :       // all-constant indexed.
     457      276352 :       if (SROACandidate)
     458      229318 :         SROAArgValues[&I] = SROAArg;
     459             : 
     460             :       return true;
     461             :     }
     462             :   }
     463             : 
     464             :   // Lambda to check whether a GEP's indices are all constant.
     465      409714 :   auto IsGEPOffsetConstant = [&](GetElementPtrInst &GEP) {
     466     1509229 :     for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I)
     467      402522 :       if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I))
     468             :         return false;
     469             :     return true;
     470      409714 :   };
     471             : 
     472      409714 :   if (IsGEPOffsetConstant(I)) {
     473      209165 :     if (SROACandidate)
     474          36 :       SROAArgValues[&I] = SROAArg;
     475             : 
     476             :     // Constant GEPs are modeled as free.
     477             :     return true;
     478             :   }
     479             : 
     480             :   // Variable GEPs will require math and will disable SROA.
     481      200549 :   if (SROACandidate)
     482           0 :     disableSROA(CostIt);
     483      200549 :   return isGEPFree(I);
     484             : }
     485             : 
     486             : /// Simplify \p I if its operands are constants and update SimplifiedValues.
     487             : /// \p Evaluate is a callable specific to instruction type that evaluates the
     488             : /// instruction when all the operands are constants.
     489             : template <typename Callable>
     490     4155770 : bool CallAnalyzer::simplifyInstruction(Instruction &I, Callable Evaluate) {
     491     8311540 :   SmallVector<Constant *, 2> COps;
     492     9667169 :   for (Value *Op : I.operands()) {
     493     4180623 :     Constant *COp = dyn_cast<Constant>(Op);
     494     4180623 :     if (!COp)
     495     5679746 :       COp = SimplifiedValues.lookup(Op);
     496     4180623 :     if (!COp)
     497     2824994 :       return false;
     498     1355629 :     COps.push_back(COp);
     499             :   }
     500     1330776 :   auto *C = Evaluate(COps);
     501     1330776 :   if (!C)
     502             :     return false;
     503       27774 :   SimplifiedValues[&I] = C;
     504       13887 :   return true;
     505             : }
     506             : 
     507      268715 : bool CallAnalyzer::visitBitCast(BitCastInst &I) {
     508             :   // Propagate constants through bitcasts.
     509      268715 :   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
     510         596 :         return ConstantExpr::getBitCast(COps[0], I.getType());
     511         298 :       }))
     512             :     return true;
     513             : 
     514             :   // Track base/offsets through casts
     515             :   std::pair<Value *, APInt> BaseAndOffset =
     516      536834 :       ConstantOffsetPtrs.lookup(I.getOperand(0));
     517             :   // Casts don't change the offset, just wrap it up.
     518      268417 :   if (BaseAndOffset.first)
     519      199596 :     ConstantOffsetPtrs[&I] = BaseAndOffset;
     520             : 
     521             :   // Also look for SROA candidates here.
     522             :   Value *SROAArg;
     523      268417 :   DenseMap<Value *, int>::iterator CostIt;
     524      536834 :   if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
     525       42598 :     SROAArgValues[&I] = SROAArg;
     526             : 
     527             :   // Bitcasts are always zero cost.
     528      268417 :   return true;
     529             : }
     530             : 
     531       15766 : bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) {
     532             :   // Propagate constants through ptrtoint.
     533       15766 :   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
     534          30 :         return ConstantExpr::getPtrToInt(COps[0], I.getType());
     535          15 :       }))
     536             :     return true;
     537             : 
     538             :   // Track base/offset pairs when converted to a plain integer provided the
     539             :   // integer is large enough to represent the pointer.
     540       15751 :   unsigned IntegerSize = I.getType()->getScalarSizeInBits();
     541       31502 :   if (IntegerSize >= DL.getPointerSizeInBits()) {
     542             :     std::pair<Value *, APInt> BaseAndOffset =
     543       47253 :         ConstantOffsetPtrs.lookup(I.getOperand(0));
     544       15751 :     if (BaseAndOffset.first)
     545       25101 :       ConstantOffsetPtrs[&I] = BaseAndOffset;
     546             :   }
     547             : 
     548             :   // This is really weird. Technically, ptrtoint will disable SROA. However,
     549             :   // unless that ptrtoint is *used* somewhere in the live basic blocks after
     550             :   // inlining, it will be nuked, and SROA should proceed. All of the uses which
     551             :   // would block SROA would also block SROA if applied directly to a pointer,
     552             :   // and so we can just add the integer in here. The only places where SROA is
     553             :   // preserved either cannot fire on an integer, or won't in-and-of themselves
     554             :   // disable SROA (ext) w/o some later use that we would see and disable.
     555             :   Value *SROAArg;
     556       15751 :   DenseMap<Value *, int>::iterator CostIt;
     557       31502 :   if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt))
     558         522 :     SROAArgValues[&I] = SROAArg;
     559             : 
     560       15751 :   return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
     561             : }
     562             : 
     563        6268 : bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) {
     564             :   // Propagate constants through ptrtoint.
     565        6268 :   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
     566         120 :         return ConstantExpr::getIntToPtr(COps[0], I.getType());
     567          60 :       }))
     568             :     return true;
     569             : 
     570             :   // Track base/offset pairs when round-tripped through a pointer without
     571             :   // modifications provided the integer is not too large.
     572       12416 :   Value *Op = I.getOperand(0);
     573        6208 :   unsigned IntegerSize = Op->getType()->getScalarSizeInBits();
     574       12416 :   if (IntegerSize <= DL.getPointerSizeInBits()) {
     575       12416 :     std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op);
     576        6208 :     if (BaseAndOffset.first)
     577           0 :       ConstantOffsetPtrs[&I] = BaseAndOffset;
     578             :   }
     579             : 
     580             :   // "Propagate" SROA here in the same manner as we do for ptrtoint above.
     581             :   Value *SROAArg;
     582        6208 :   DenseMap<Value *, int>::iterator CostIt;
     583        6208 :   if (lookupSROAArgAndCost(Op, SROAArg, CostIt))
     584           0 :     SROAArgValues[&I] = SROAArg;
     585             : 
     586        6208 :   return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
     587             : }
     588             : 
     589       21449 : bool CallAnalyzer::visitCastInst(CastInst &I) {
     590             :   // Propagate constants through ptrtoint.
     591       21449 :   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
     592        5649 :         return ConstantExpr::getCast(I.getOpcode(), COps[0], I.getType());
     593        1883 :       }))
     594             :     return true;
     595             : 
     596             :   // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere.
     597       39132 :   disableSROA(I.getOperand(0));
     598             : 
     599       19566 :   return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I);
     600             : }
     601             : 
     602     1316889 : bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) {
     603     1316889 :   Value *Operand = I.getOperand(0);
     604     1316889 :   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
     605     5267556 :         return ConstantFoldInstOperands(&I, COps[0], DL);
     606     1316889 :       }))
     607             :     return true;
     608             : 
     609             :   // Disable any SROA on the argument to arbitrary unary operators.
     610     1316889 :   disableSROA(Operand);
     611             : 
     612     1316889 :   return false;
     613             : }
     614             : 
     615             : bool CallAnalyzer::paramHasAttr(Argument *A, Attribute::AttrKind Attr) {
     616        4462 :   return CandidateCS.paramHasAttr(A->getArgNo(), Attr);
     617             : }
     618             : 
     619       54675 : bool CallAnalyzer::isKnownNonNullInCallee(Value *V) {
     620             :   // Does the *call site* have the NonNull attribute set on an argument?  We
     621             :   // use the attribute on the call site to memoize any analysis done in the
     622             :   // caller. This will also trip if the callee function has a non-null
     623             :   // parameter attribute, but that's a less interesting case because hopefully
     624             :   // the callee would already have been simplified based on that.
     625        4462 :   if (Argument *A = dyn_cast<Argument>(V))
     626        4462 :     if (paramHasAttr(A, Attribute::NonNull))
     627             :       return true;
     628             : 
     629             :   // Is this an alloca in the caller?  This is distinct from the attribute case
     630             :   // above because attributes aren't updated within the inliner itself and we
     631             :   // always want to catch the alloca derived case.
     632       54060 :   if (isAllocaDerivedArg(V))
     633             :     // We can actually predict the result of comparisons between an
     634             :     // alloca-derived value and null. Note that this fires regardless of
     635             :     // SROA firing.
     636             :     return true;
     637             : 
     638             :   return false;
     639             : }
     640             : 
     641      228149 : bool CallAnalyzer::allowSizeGrowth(CallSite CS) {
     642             :   // If the normal destination of the invoke or the parent block of the call
     643             :   // site is unreachable-terminated, there is little point in inlining this
     644             :   // unless there is literally zero cost.
     645             :   // FIXME: Note that it is possible that an unreachable-terminated block has a
     646             :   // hot entry. For example, in below scenario inlining hot_call_X() may be
     647             :   // beneficial :
     648             :   // main() {
     649             :   //   hot_call_1();
     650             :   //   ...
     651             :   //   hot_call_N()
     652             :   //   exit(0);
     653             :   // }
     654             :   // For now, we are not handling this corner case here as it is rare in real
     655             :   // code. In future, we should elaborate this based on BPI and BFI in more
     656             :   // general threshold adjusting heuristics in updateThreshold().
     657      228149 :   Instruction *Instr = CS.getInstruction();
     658       57018 :   if (InvokeInst *II = dyn_cast<InvokeInst>(Instr)) {
     659      171054 :     if (isa<UnreachableInst>(II->getNormalDest()->getTerminator()))
     660             :       return false;
     661      513393 :   } else if (isa<UnreachableInst>(Instr->getParent()->getTerminator()))
     662             :     return false;
     663             : 
     664             :   return true;
     665             : }
     666             : 
     667      226826 : bool CallAnalyzer::isColdCallSite(CallSite CS, BlockFrequencyInfo *CallerBFI) {
     668             :   // If global profile summary is available, then callsite's coldness is
     669             :   // determined based on that.
     670      453266 :   if (PSI && PSI->hasProfileSummary())
     671          20 :     return PSI->isColdCallSite(CS, CallerBFI);
     672             : 
     673             :   // Otherwise we need BFI to be available.
     674      226806 :   if (!CallerBFI)
     675             :     return false;
     676             : 
     677             :   // Determine if the callsite is cold relative to caller's entry. We could
     678             :   // potentially cache the computation of scaled entry frequency, but the added
     679             :   // complexity is not worth it unless this scaling shows up high in the
     680             :   // profiles.
     681         419 :   const BranchProbability ColdProb(ColdCallSiteRelFreq, 100);
     682         419 :   auto CallSiteBB = CS.getInstruction()->getParent();
     683         419 :   auto CallSiteFreq = CallerBFI->getBlockFreq(CallSiteBB);
     684             :   auto CallerEntryFreq =
     685         838 :       CallerBFI->getBlockFreq(&(CS.getCaller()->getEntryBlock()));
     686         838 :   return CallSiteFreq < CallerEntryFreq * ColdProb;
     687             : }
     688             : 
     689             : Optional<int>
     690      226838 : CallAnalyzer::getHotCallSiteThreshold(CallSite CS,
     691             :                                       BlockFrequencyInfo *CallerBFI) {
     692             : 
     693             :   // If global profile summary is available, then callsite's hotness is
     694             :   // determined based on that.
     695      453290 :   if (PSI && PSI->hasProfileSummary() && PSI->isHotCallSite(CS, CallerBFI))
     696          12 :     return Params.HotCallSiteThreshold;
     697             : 
     698             :   // Otherwise we need BFI to be available and to have a locally hot callsite
     699             :   // threshold.
     700      227253 :   if (!CallerBFI || !Params.LocallyHotCallSiteThreshold)
     701             :     return None;
     702             : 
     703             :   // Determine if the callsite is hot relative to caller's entry. We could
     704             :   // potentially cache the computation of scaled entry frequency, but the added
     705             :   // complexity is not worth it unless this scaling shows up high in the
     706             :   // profiles.
     707           3 :   auto CallSiteBB = CS.getInstruction()->getParent();
     708           6 :   auto CallSiteFreq = CallerBFI->getBlockFreq(CallSiteBB).getFrequency();
     709           3 :   auto CallerEntryFreq = CallerBFI->getEntryFreq();
     710           3 :   if (CallSiteFreq >= CallerEntryFreq * HotCallSiteRelFreq)
     711           0 :     return Params.LocallyHotCallSiteThreshold;
     712             : 
     713             :   // Otherwise treat it normally.
     714             :   return None;
     715             : }
     716             : 
     717      228149 : void CallAnalyzer::updateThreshold(CallSite CS, Function &Callee) {
     718             :   // If no size growth is allowed for this inlining, set Threshold to 0.
     719      228149 :   if (!allowSizeGrowth(CS)) {
     720        1306 :     Threshold = 0;
     721        1306 :     return;
     722             :   }
     723             : 
     724      226843 :   Function *Caller = CS.getCaller();
     725             : 
     726             :   // return min(A, B) if B is valid.
     727             :   auto MinIfValid = [](int A, Optional<int> B) {
     728         139 :     return B ? std::min(A, B.getValue()) : A;
     729             :   };
     730             : 
     731             :   // return max(A, B) if B is valid.
     732             :   auto MaxIfValid = [](int A, Optional<int> B) {
     733      148986 :     return B ? std::max(A, B.getValue()) : A;
     734             :   };
     735             : 
     736             :   // Various bonus percentages. These are multiplied by Threshold to get the
     737             :   // bonus values.
     738             :   // SingleBBBonus: This bonus is applied if the callee has a single reachable
     739             :   // basic block at the given callsite context. This is speculatively applied
     740             :   // and withdrawn if more than one basic block is seen.
     741             :   //
     742             :   // Vector bonuses: We want to more aggressively inline vector-dense kernels
     743             :   // and apply this bonus based on the percentage of vector instructions. A
     744             :   // bonus is applied if the vector instructions exceed 50% and half that amount
     745             :   // is applied if it exceeds 10%. Note that these bonuses are some what
     746             :   // arbitrary and evolved over time by accident as much as because they are
     747             :   // principled bonuses.
     748             :   // FIXME: It would be nice to base the bonus values on something more
     749             :   // scientific.
     750             :   //
     751             :   // LstCallToStaticBonus: This large bonus is applied to ensure the inlining
     752             :   // of the last call to a static function as inlining such functions is
     753             :   // guaranteed to reduce code size.
     754             :   //
     755             :   // These bonus percentages may be set to 0 based on properties of the caller
     756             :   // and the callsite.
     757      226843 :   int SingleBBBonusPercent = 50;
     758      226843 :   int VectorBonusPercent = 150;
     759      226843 :   int LastCallToStaticBonus = InlineConstants::LastCallToStaticBonus;
     760             : 
     761             :   // Lambda to set all the above bonus and bonus percentages to 0.
     762             :   auto DisallowAllBonuses = [&]() {
     763          27 :     SingleBBBonusPercent = 0;
     764          27 :     VectorBonusPercent = 0;
     765          27 :     LastCallToStaticBonus = 0;
     766      226843 :   };
     767             : 
     768             :   // Use the OptMinSizeThreshold or OptSizeThreshold knob if they are available
     769             :   // and reduce the threshold if the caller has the necessary attribute.
     770      226843 :   if (Caller->optForMinSize()) {
     771          20 :     Threshold = MinIfValid(Threshold, Params.OptMinSizeThreshold);
     772             :     // For minsize, we want to disable the single BB bonus and the vector
     773             :     // bonuses, but not the last-call-to-static bonus. Inlining the last call to
     774             :     // a static function will, at the minimum, eliminate the parameter setup and
     775             :     // call/return instructions.
     776           5 :     SingleBBBonusPercent = 0;
     777           5 :     VectorBonusPercent = 0;
     778      226838 :   } else if (Caller->optForSize())
     779          68 :     Threshold = MinIfValid(Threshold, Params.OptSizeThreshold);
     780             : 
     781             :   // Adjust the threshold based on inlinehint attribute and profile based
     782             :   // hotness information if the caller does not have MinSize attribute.
     783      226843 :   if (!Caller->optForMinSize()) {
     784      226838 :     if (Callee.hasFnAttribute(Attribute::InlineHint))
     785      198640 :       Threshold = MaxIfValid(Threshold, Params.HintThreshold);
     786             : 
     787             :     // FIXME: After switching to the new passmanager, simplify the logic below
     788             :     // by checking only the callsite hotness/coldness as we will reliably
     789             :     // have local profile information.
     790             :     //
     791             :     // Callsite hotness and coldness can be determined if sample profile is
     792             :     // used (which adds hotness metadata to calls) or if caller's
     793             :     // BlockFrequencyInfo is available.
     794      227706 :     BlockFrequencyInfo *CallerBFI = GetBFI ? &((*GetBFI)(*Caller)) : nullptr;
     795      453676 :     auto HotCallSiteThreshold = getHotCallSiteThreshold(CS, CallerBFI);
     796      226838 :     if (!Caller->optForSize() && HotCallSiteThreshold) {
     797             :       DEBUG(dbgs() << "Hot callsite.\n");
     798             :       // FIXME: This should update the threshold only if it exceeds the
     799             :       // current threshold, but AutoFDO + ThinLTO currently relies on this
     800             :       // behavior to prevent inlining of hot callsites during ThinLTO
     801             :       // compile phase.
     802          12 :       Threshold = HotCallSiteThreshold.getValue();
     803      226826 :     } else if (isColdCallSite(CS, CallerBFI)) {
     804             :       DEBUG(dbgs() << "Cold callsite.\n");
     805             :       // Do not apply bonuses for a cold callsite including the
     806             :       // LastCallToStatic bonus. While this bonus might result in code size
     807             :       // reduction, it can cause the size of a non-cold caller to increase
     808             :       // preventing it from being inlined.
     809          14 :       DisallowAllBonuses();
     810          56 :       Threshold = MinIfValid(Threshold, Params.ColdCallSiteThreshold);
     811      226812 :     } else if (PSI) {
     812             :       // Use callee's global profile information only if we have no way of
     813             :       // determining this via callsite information.
     814      226430 :       if (PSI->isFunctionEntryHot(&Callee)) {
     815             :         DEBUG(dbgs() << "Hot callee.\n");
     816             :         // If callsite hotness can not be determined, we may still know
     817             :         // that the callee is hot and treat it as a weaker hint for threshold
     818             :         // increase.
     819           8 :         Threshold = MaxIfValid(Threshold, Params.HintThreshold);
     820      226428 :       } else if (PSI->isFunctionEntryCold(&Callee)) {
     821             :         DEBUG(dbgs() << "Cold callee.\n");
     822             :         // Do not apply bonuses for a cold callee including the
     823             :         // LastCallToStatic bonus. While this bonus might result in code size
     824             :         // reduction, it can cause the size of a non-cold caller to increase
     825             :         // preventing it from being inlined.
     826          13 :         DisallowAllBonuses();
     827          52 :         Threshold = MinIfValid(Threshold, Params.ColdThreshold);
     828             :       }
     829             :     }
     830             :   }
     831             : 
     832             :   // Finally, take the target-specific inlining threshold multiplier into
     833             :   // account.
     834      226843 :   Threshold *= TTI.getInliningThresholdMultiplier();
     835             : 
     836      226843 :   SingleBBBonus = Threshold * SingleBBBonusPercent / 100;
     837      226843 :   VectorBonus = Threshold * VectorBonusPercent / 100;
     838             : 
     839             :   bool OnlyOneCallAndLocalLinkage =
     840      259536 :       F.hasLocalLinkage() && F.hasOneUse() && &F == CS.getCalledFunction();
     841             :   // If there is only one call of the function, and it has internal linkage,
     842             :   // the cost of inlining it drops dramatically. It may seem odd to update
     843             :   // Cost in updateThreshold, but the bonus depends on the logic in this method.
     844             :   if (OnlyOneCallAndLocalLinkage)
     845        3918 :     Cost -= LastCallToStaticBonus;
     846             : }
     847             : 
     848      196494 : bool CallAnalyzer::visitCmpInst(CmpInst &I) {
     849      392988 :   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
     850             :   // First try to handle simplified comparisons.
     851      196494 :   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
     852       25872 :         return ConstantExpr::getCompare(I.getPredicate(), COps[0], COps[1]);
     853        6468 :       }))
     854             :     return true;
     855             : 
     856      190026 :   if (I.getOpcode() == Instruction::FCmp)
     857             :     return false;
     858             : 
     859             :   // Otherwise look for a comparison between constant offset pointers with
     860             :   // a common base.
     861             :   Value *LHSBase, *RHSBase;
     862      569619 :   APInt LHSOffset, RHSOffset;
     863      569619 :   std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
     864      189873 :   if (LHSBase) {
     865       20745 :     std::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
     866        6915 :     if (RHSBase && LHSBase == RHSBase) {
     867             :       // We have common bases, fold the icmp to a constant based on the
     868             :       // offsets.
     869          43 :       Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
     870          43 :       Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
     871          43 :       if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) {
     872          86 :         SimplifiedValues[&I] = C;
     873          43 :         ++NumConstantPtrCmps;
     874          43 :         return true;
     875             :       }
     876             :     }
     877             :   }
     878             : 
     879             :   // If the comparison is an equality comparison with null, we can simplify it
     880             :   // if we know the value (argument) can't be null
     881      535293 :   if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1)) &&
     882       54675 :       isKnownNonNullInCallee(I.getOperand(0))) {
     883         619 :     bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE;
     884        1857 :     SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType())
     885          49 :                                       : ConstantInt::getFalse(I.getType());
     886         619 :     return true;
     887             :   }
     888             :   // Finally check for SROA candidates in comparisons.
     889             :   Value *SROAArg;
     890      189211 :   DenseMap<Value *, int>::iterator CostIt;
     891      189211 :   if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
     892         390 :     if (isa<ConstantPointerNull>(I.getOperand(1))) {
     893           0 :       accumulateSROACost(CostIt, InlineConstants::InstrCost);
     894           0 :       return true;
     895             :     }
     896             : 
     897         195 :     disableSROA(CostIt);
     898             :   }
     899             : 
     900             :   return false;
     901             : }
     902             : 
     903        9318 : bool CallAnalyzer::visitOr(BinaryOperator &I) {
     904             :   // This is necessary because the generic simplify instruction only works if
     905             :   // both operands are constants.
     906       27928 :   for (unsigned i = 0; i < 2; ++i) {
     907       55876 :     if (ConstantInt *C = dyn_cast_or_null<ConstantInt>(
     908       37246 :             SimplifiedValues.lookup(I.getOperand(i))))
     909         887 :       if (C->isAllOnesValue()) {
     910          40 :         SimplifiedValues[&I] = C;
     911          20 :         return true;
     912             :       }
     913             :   }
     914        9298 :   return Base::visitOr(I);
     915             : }
     916             : 
     917       18989 : bool CallAnalyzer::visitAnd(BinaryOperator &I) {
     918             :   // This is necessary because the generic simplify instruction only works if
     919             :   // both operands are constants.
     920       55667 :   for (unsigned i = 0; i < 2; ++i) {
     921      112065 :     if (ConstantInt *C = dyn_cast_or_null<ConstantInt>(
     922       74693 :             SimplifiedValues.lookup(I.getOperand(i))))
     923        2259 :       if (C->isZero()) {
     924        1388 :         SimplifiedValues[&I] = C;
     925         694 :         return true;
     926             :       }
     927             :   }
     928       18295 :   return Base::visitAnd(I);
     929             : }
     930             : 
     931       21394 : bool CallAnalyzer::visitSub(BinaryOperator &I) {
     932             :   // Try to handle a special case: we can fold computing the difference of two
     933             :   // constant-related pointers.
     934       42788 :   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
     935             :   Value *LHSBase, *RHSBase;
     936       85576 :   APInt LHSOffset, RHSOffset;
     937       64182 :   std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
     938       21394 :   if (LHSBase) {
     939       11433 :     std::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS);
     940        3811 :     if (RHSBase && LHSBase == RHSBase) {
     941             :       // We have common bases, fold the subtract to a constant based on the
     942             :       // offsets.
     943          11 :       Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset);
     944          11 :       Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset);
     945          11 :       if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) {
     946          22 :         SimplifiedValues[&I] = C;
     947          11 :         ++NumConstantPtrDiffs;
     948          11 :         return true;
     949             :       }
     950             :     }
     951             :   }
     952             : 
     953             :   // Otherwise, fall back to the generic logic for simplifying and handling
     954             :   // instructions.
     955       21383 :   return Base::visitSub(I);
     956             : }
     957             : 
     958     2212238 : bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
     959     4424476 :   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
     960        5161 :   auto Evaluate = [&](SmallVectorImpl<Constant *> &COps) {
     961        5161 :     Value *SimpleV = nullptr;
     962       10324 :     if (auto FI = dyn_cast<FPMathOperator>(&I))
     963          12 :       SimpleV = SimplifyFPBinOp(I.getOpcode(), COps[0], COps[1],
     964        5161 :                                 FI->getFastMathFlags(), DL);
     965             :     else
     966       25795 :       SimpleV = SimplifyBinOp(I.getOpcode(), COps[0], COps[1], DL);
     967        5161 :     return dyn_cast_or_null<Constant>(SimpleV);
     968     2212238 :   };
     969             : 
     970     2212238 :   if (simplifyInstruction(I, Evaluate))
     971             :     return true;
     972             : 
     973             :   // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
     974     2207077 :   disableSROA(LHS);
     975     2207077 :   disableSROA(RHS);
     976             : 
     977     2207077 :   return false;
     978             : }
     979             : 
     980     2386679 : bool CallAnalyzer::visitLoad(LoadInst &I) {
     981             :   Value *SROAArg;
     982     2386679 :   DenseMap<Value *, int>::iterator CostIt;
     983     2386679 :   if (lookupSROAArgAndCost(I.getPointerOperand(), SROAArg, CostIt)) {
     984       94753 :     if (I.isSimple()) {
     985      189506 :       accumulateSROACost(CostIt, InlineConstants::InstrCost);
     986       94753 :       return true;
     987             :     }
     988             : 
     989         222 :     disableSROA(CostIt);
     990             :   }
     991             : 
     992             :   return false;
     993             : }
     994             : 
     995     2437283 : bool CallAnalyzer::visitStore(StoreInst &I) {
     996             :   Value *SROAArg;
     997     2437283 :   DenseMap<Value *, int>::iterator CostIt;
     998     2437283 :   if (lookupSROAArgAndCost(I.getPointerOperand(), SROAArg, CostIt)) {
     999       39177 :     if (I.isSimple()) {
    1000       78354 :       accumulateSROACost(CostIt, InlineConstants::InstrCost);
    1001       39177 :       return true;
    1002             :     }
    1003             : 
    1004          79 :     disableSROA(CostIt);
    1005             :   }
    1006             : 
    1007             :   return false;
    1008             : }
    1009             : 
    1010             : bool CallAnalyzer::visitExtractValue(ExtractValueInst &I) {
    1011             :   // Constant folding for extract value is trivial.
    1012      108103 :   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
    1013           6 :         return ConstantExpr::getExtractValue(COps[0], I.getIndices());
    1014           2 :       }))
    1015             :     return true;
    1016             : 
    1017             :   // SROA can look through these but give them a cost.
    1018             :   return false;
    1019             : }
    1020             : 
    1021             : bool CallAnalyzer::visitInsertValue(InsertValueInst &I) {
    1022             :   // Constant folding for insert value is trivial.
    1023        9850 :   if (simplifyInstruction(I, [&](SmallVectorImpl<Constant *> &COps) {
    1024           0 :         return ConstantExpr::getInsertValue(/*AggregateOperand*/ COps[0],
    1025           0 :                                             /*InsertedValueOperand*/ COps[1],
    1026           0 :                                             I.getIndices());
    1027           0 :       }))
    1028             :     return true;
    1029             : 
    1030             :   // SROA can look through these but give them a cost.
    1031             :   return false;
    1032             : }
    1033             : 
    1034             : /// \brief Try to simplify a call site.
    1035             : ///
    1036             : /// Takes a concrete function and callsite and tries to actually simplify it by
    1037             : /// analyzing the arguments and call itself with instsimplify. Returns true if
    1038             : /// it has simplified the callsite to some other entity (a constant), making it
    1039             : /// free.
    1040      986465 : bool CallAnalyzer::simplifyCallSite(Function *F, CallSite CS) {
    1041             :   // FIXME: Using the instsimplify logic directly for this is inefficient
    1042             :   // because we have to continually rebuild the argument list even when no
    1043             :   // simplifications can be performed. Until that is fixed with remapping
    1044             :   // inside of instsimplify, directly constant fold calls here.
    1045      986465 :   if (!canConstantFoldCallTo(CS, F))
    1046             :     return false;
    1047             : 
    1048             :   // Try to re-map the arguments to constants.
    1049         679 :   SmallVector<Constant *, 4> ConstantArgs;
    1050         679 :   ConstantArgs.reserve(CS.arg_size());
    1051         703 :   for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E;
    1052             :        ++I) {
    1053         697 :     Constant *C = dyn_cast<Constant>(*I);
    1054         697 :     if (!C)
    1055        2043 :       C = dyn_cast_or_null<Constant>(SimplifiedValues.lookup(*I));
    1056         697 :     if (!C)
    1057         673 :       return false; // This argument doesn't map to a constant.
    1058             : 
    1059          24 :     ConstantArgs.push_back(C);
    1060             :   }
    1061          12 :   if (Constant *C = ConstantFoldCall(CS, F, ConstantArgs)) {
    1062          12 :     SimplifiedValues[CS.getInstruction()] = C;
    1063           6 :     return true;
    1064             :   }
    1065             : 
    1066             :   return false;
    1067             : }
    1068             : 
    1069     1002239 : bool CallAnalyzer::visitCallSite(CallSite CS) {
    1070     1002247 :   if (CS.hasFnAttr(Attribute::ReturnsTwice) &&
    1071          16 :       !F.hasFnAttribute(Attribute::ReturnsTwice)) {
    1072             :     // This aborts the entire analysis.
    1073           4 :     ExposesReturnsTwice = true;
    1074           4 :     return false;
    1075             :   }
    1076     3640720 :   if (CS.isCall() && cast<CallInst>(CS.getInstruction())->cannotDuplicate())
    1077          11 :     ContainsNoDuplicateCall = true;
    1078             : 
    1079      986465 :   if (Function *F = CS.getCalledFunction()) {
    1080             :     // When we have a concrete function, first try to simplify it directly.
    1081      986465 :     if (simplifyCallSite(F, CS))
    1082             :       return true;
    1083             : 
    1084             :     // Next check if it is an intrinsic we know about.
    1085             :     // FIXME: Lift this into part of the InstVisitor.
    1086     1202971 :     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) {
    1087      216512 :       switch (II->getIntrinsicID()) {
    1088      203684 :       default:
    1089      203684 :         return Base::visitCallSite(CS);
    1090             : 
    1091           0 :       case Intrinsic::load_relative:
    1092             :         // This is normally lowered to 4 LLVM instructions.
    1093           0 :         Cost += 3 * InlineConstants::InstrCost;
    1094           0 :         return false;
    1095             : 
    1096             :       case Intrinsic::memset:
    1097             :       case Intrinsic::memcpy:
    1098             :       case Intrinsic::memmove:
    1099             :         // SROA can usually chew through these intrinsics, but they aren't free.
    1100             :         return false;
    1101           4 :       case Intrinsic::localescape:
    1102           4 :         HasFrameEscape = true;
    1103           4 :         return false;
    1104             :       }
    1105             :     }
    1106             : 
    1107      769947 :     if (F == CS.getInstruction()->getParent()->getParent()) {
    1108             :       // This flag will fully abort the analysis, so don't bother with anything
    1109             :       // else.
    1110        1393 :       IsRecursiveCall = true;
    1111        1393 :       return false;
    1112             :     }
    1113             : 
    1114      768554 :     if (TTI.isLoweredToCall(F)) {
    1115             :       // We account for the average 1 instruction per call argument setup
    1116             :       // here.
    1117      768552 :       Cost += CS.arg_size() * InlineConstants::InstrCost;
    1118             : 
    1119             :       // Everything other than inline ASM will also have a significant cost
    1120             :       // merely from making the call.
    1121     1537104 :       if (!isa<InlineAsm>(CS.getCalledValue()))
    1122      768552 :         Cost += InlineConstants::CallPenalty;
    1123             :     }
    1124             : 
    1125      768554 :     return Base::visitCallSite(CS);
    1126             :   }
    1127             : 
    1128             :   // Otherwise we're in a very special case -- an indirect function call. See
    1129             :   // if we can be particularly clever about this.
    1130       15770 :   Value *Callee = CS.getCalledValue();
    1131             : 
    1132             :   // First, pay the price of the argument setup. We account for the average
    1133             :   // 1 instruction per call argument setup here.
    1134       15770 :   Cost += CS.arg_size() * InlineConstants::InstrCost;
    1135             : 
    1136             :   // Next, check if this happens to be an indirect function call to a known
    1137             :   // function in this inline context. If not, we've done all we can.
    1138       31911 :   Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee));
    1139             :   if (!F)
    1140       15399 :     return Base::visitCallSite(CS);
    1141             : 
    1142             :   // If we have a constant that we are calling as a function, we can peer
    1143             :   // through it and see the function target. This happens not infrequently
    1144             :   // during devirtualization and so we want to give it a hefty bonus for
    1145             :   // inlining, but cap that bonus in the event that inlining wouldn't pan
    1146             :   // out. Pretend to inline the function, with a custom threshold.
    1147         371 :   auto IndirectCallParams = Params;
    1148         371 :   IndirectCallParams.DefaultThreshold = InlineConstants::IndirectCallThreshold;
    1149             :   CallAnalyzer CA(TTI, GetAssumptionCache, GetBFI, PSI, ORE, *F, CS,
    1150         742 :                   IndirectCallParams);
    1151         371 :   if (CA.analyzeCall(CS)) {
    1152             :     // We were able to inline the indirect call! Subtract the cost from the
    1153             :     // threshold to get the bonus we want to apply, but don't go below zero.
    1154         456 :     Cost -= std::max(0, CA.getThreshold() - CA.getCost());
    1155             :   }
    1156             : 
    1157         371 :   return Base::visitCallSite(CS);
    1158             : }
    1159             : 
    1160             : bool CallAnalyzer::visitReturnInst(ReturnInst &RI) {
    1161             :   // At least one return instruction will be free after inlining.
    1162      190763 :   bool Free = !HasReturn;
    1163      190763 :   HasReturn = true;
    1164             :   return Free;
    1165             : }
    1166             : 
    1167      424901 : bool CallAnalyzer::visitBranchInst(BranchInst &BI) {
    1168             :   // We model unconditional branches as essentially free -- they really
    1169             :   // shouldn't exist at all, but handling them makes the behavior of the
    1170             :   // inliner more regular and predictable. Interestingly, conditional branches
    1171             :   // which will fold away are also free.
    1172     1071812 :   return BI.isUnconditional() || isa<ConstantInt>(BI.getCondition()) ||
    1173      645943 :          dyn_cast_or_null<ConstantInt>(
    1174      640216 :              SimplifiedValues.lookup(BI.getCondition()));
    1175             : }
    1176             : 
    1177        1303 : bool CallAnalyzer::visitSwitchInst(SwitchInst &SI) {
    1178             :   // We model unconditional switches as free, see the comments on handling
    1179             :   // branches.
    1180        2606 :   if (isa<ConstantInt>(SI.getCondition()))
    1181             :     return true;
    1182        3909 :   if (Value *V = SimplifiedValues.lookup(SI.getCondition()))
    1183         128 :     if (isa<ConstantInt>(V))
    1184             :       return true;
    1185             : 
    1186             :   // Assume the most general case where the switch is lowered into
    1187             :   // either a jump table, bit test, or a balanced binary tree consisting of
    1188             :   // case clusters without merging adjacent clusters with the same
    1189             :   // destination. We do not consider the switches that are lowered with a mix
    1190             :   // of jump table/bit test/binary search tree. The cost of the switch is
    1191             :   // proportional to the size of the tree or the size of jump table range.
    1192             :   //
    1193             :   // NB: We convert large switches which are just used to initialize large phi
    1194             :   // nodes to lookup tables instead in simplify-cfg, so this shouldn't prevent
    1195             :   // inlining those. It will prevent inlining in cases where the optimization
    1196             :   // does not (yet) fire.
    1197             : 
    1198             :   // Maximum valid cost increased in this function.
    1199        1239 :   int CostUpperBound = INT_MAX - InlineConstants::InstrCost - 1;
    1200             : 
    1201             :   // Exit early for a large switch, assuming one case needs at least one
    1202             :   // instruction.
    1203             :   // FIXME: This is not true for a bit test, but ignore such case for now to
    1204             :   // save compile-time.
    1205             :   int64_t CostLowerBound =
    1206        2478 :       std::min((int64_t)CostUpperBound,
    1207        3717 :                (int64_t)SI.getNumCases() * InlineConstants::InstrCost + Cost);
    1208             : 
    1209        1242 :   if (CostLowerBound > Threshold && !ComputeFullInlineCost) {
    1210           3 :     Cost = CostLowerBound;
    1211           3 :     return false;
    1212             :   }
    1213             : 
    1214        1236 :   unsigned JumpTableSize = 0;
    1215             :   unsigned NumCaseCluster =
    1216        1236 :       TTI.getEstimatedNumberOfCaseClusters(SI, JumpTableSize);
    1217             : 
    1218             :   // If suitable for a jump table, consider the cost for the table size and
    1219             :   // branch to destination.
    1220        1236 :   if (JumpTableSize) {
    1221         202 :     int64_t JTCost = (int64_t)JumpTableSize * InlineConstants::InstrCost +
    1222             :                      4 * InlineConstants::InstrCost;
    1223             : 
    1224         404 :     Cost = std::min((int64_t)CostUpperBound, JTCost + Cost);
    1225         202 :     return false;
    1226             :   }
    1227             : 
    1228             :   // Considering forming a binary search, we should find the number of nodes
    1229             :   // which is same as the number of comparisons when lowered. For a given
    1230             :   // number of clusters, n, we can define a recursive function, f(n), to find
    1231             :   // the number of nodes in the tree. The recursion is :
    1232             :   // f(n) = 1 + f(n/2) + f (n - n/2), when n > 3,
    1233             :   // and f(n) = n, when n <= 3.
    1234             :   // This will lead a binary tree where the leaf should be either f(2) or f(3)
    1235             :   // when n > 3.  So, the number of comparisons from leaves should be n, while
    1236             :   // the number of non-leaf should be :
    1237             :   //   2^(log2(n) - 1) - 1
    1238             :   //   = 2^log2(n) * 2^-1 - 1
    1239             :   //   = n / 2 - 1.
    1240             :   // Considering comparisons from leaf and non-leaf nodes, we can estimate the
    1241             :   // number of comparisons in a simple closed form :
    1242             :   //   n + n / 2 - 1 = n * 3 / 2 - 1
    1243        1034 :   if (NumCaseCluster <= 3) {
    1244             :     // Suppose a comparison includes one compare and one conditional branch.
    1245         980 :     Cost += NumCaseCluster * 2 * InlineConstants::InstrCost;
    1246         980 :     return false;
    1247             :   }
    1248             : 
    1249          54 :   int64_t ExpectedNumberOfCompare = 3 * (int64_t)NumCaseCluster / 2 - 1;
    1250          54 :   int64_t SwitchCost =
    1251             :       ExpectedNumberOfCompare * 2 * InlineConstants::InstrCost;
    1252             : 
    1253         108 :   Cost = std::min((int64_t)CostUpperBound, SwitchCost + Cost);
    1254          54 :   return false;
    1255             : }
    1256             : 
    1257             : bool CallAnalyzer::visitIndirectBrInst(IndirectBrInst &IBI) {
    1258             :   // We never want to inline functions that contain an indirectbr.  This is
    1259             :   // incorrect because all the blockaddress's (in static global initializers
    1260             :   // for example) would be referring to the original function, and this
    1261             :   // indirect jump would jump from the inlined copy of the function into the
    1262             :   // original function which is extremely undefined behavior.
    1263             :   // FIXME: This logic isn't really right; we can safely inline functions with
    1264             :   // indirectbr's as long as no other function or global references the
    1265             :   // blockaddress of a block within the current function.
    1266           0 :   HasIndirectBr = true;
    1267             :   return false;
    1268             : }
    1269             : 
    1270             : bool CallAnalyzer::visitResumeInst(ResumeInst &RI) {
    1271             :   // FIXME: It's not clear that a single instruction is an accurate model for
    1272             :   // the inline cost of a resume instruction.
    1273             :   return false;
    1274             : }
    1275             : 
    1276             : bool CallAnalyzer::visitCleanupReturnInst(CleanupReturnInst &CRI) {
    1277             :   // FIXME: It's not clear that a single instruction is an accurate model for
    1278             :   // the inline cost of a cleanupret instruction.
    1279             :   return false;
    1280             : }
    1281             : 
    1282             : bool CallAnalyzer::visitCatchReturnInst(CatchReturnInst &CRI) {
    1283             :   // FIXME: It's not clear that a single instruction is an accurate model for
    1284             :   // the inline cost of a catchret instruction.
    1285             :   return false;
    1286             : }
    1287             : 
    1288             : bool CallAnalyzer::visitUnreachableInst(UnreachableInst &I) {
    1289             :   // FIXME: It might be reasonably to discount the cost of instructions leading
    1290             :   // to unreachable as they have the lowest possible impact on both runtime and
    1291             :   // code size.
    1292             :   return true; // No actual code is needed for unreachable.
    1293             : }
    1294             : 
    1295     1267924 : bool CallAnalyzer::visitInstruction(Instruction &I) {
    1296             :   // Some instructions are free. All of the free intrinsics can also be
    1297             :   // handled by SROA, etc.
    1298     1267924 :   if (TargetTransformInfo::TCC_Free == TTI.getUserCost(&I))
    1299             :     return true;
    1300             : 
    1301             :   // We found something we don't understand or can't handle. Mark any SROA-able
    1302             :   // values in the operand list as no longer viable.
    1303     6223072 :   for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI)
    1304     3025039 :     disableSROA(*OI);
    1305             : 
    1306             :   return false;
    1307             : }
    1308             : 
    1309             : /// \brief Analyze a basic block for its contribution to the inline cost.
    1310             : ///
    1311             : /// This method walks the analyzer over every instruction in the given basic
    1312             : /// block and accounts for their cost during inlining at this callsite. It
    1313             : /// aborts early if the threshold has been exceeded or an impossible to inline
    1314             : /// construct has been detected. It returns false if inlining is no longer
    1315             : /// viable, and true if inlining remains viable.
    1316      868156 : bool CallAnalyzer::analyzeBlock(BasicBlock *BB,
    1317             :                                 SmallPtrSetImpl<const Value *> &EphValues) {
    1318    14666571 :   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
    1319             :     // FIXME: Currently, the number of instructions in a function regardless of
    1320             :     // our ability to simplify them during inline to constants or dead code,
    1321             :     // are actually used by the vector bonus heuristic. As long as that's true,
    1322             :     // we have to special case debug intrinsics here to prevent differences in
    1323             :     // inlining due to debug symbols. Eventually, the number of unsimplified
    1324             :     // instructions shouldn't factor into the cost computation, but until then,
    1325             :     // hack around it here.
    1326    12112805 :     if (isa<DbgInfoIntrinsic>(I))
    1327      328621 :       continue;
    1328             : 
    1329             :     // Skip ephemeral values.
    1330    11784184 :     if (EphValues.count(&*I))
    1331           3 :       continue;
    1332             : 
    1333    11784181 :     ++NumInstructions;
    1334    35352527 :     if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy())
    1335          20 :       ++NumVectorInstructions;
    1336             : 
    1337             :     // If the instruction is floating point, and the target says this operation
    1338             :     // is expensive or the function has the "use-soft-float" attribute, this may
    1339             :     // eventually become a library call. Treat the cost as such.
    1340    11784181 :     if (I->getType()->isFloatingPointTy()) {
    1341             :       // If the function has the "use-soft-float" attribute, mark it as
    1342             :       // expensive.
    1343        2052 :       if (TTI.getFPOpCost(I->getType()) == TargetTransformInfo::TCC_Expensive ||
    1344        3042 :           (F.getFnAttribute("use-soft-float").getValueAsString() == "true"))
    1345          36 :         Cost += InlineConstants::CallPenalty;
    1346             :     }
    1347             : 
    1348             :     // If the instruction simplified to a constant, there is no cost to this
    1349             :     // instruction. Visit the instructions using our InstVisitor to account for
    1350             :     // all of the per-instruction logic. The visit tree returns true if we
    1351             :     // consumed the instruction in any way, and false if the instruction's base
    1352             :     // cost should count against inlining.
    1353    23568362 :     if (Base::visit(&*I))
    1354     1840106 :       ++NumInstructionsSimplified;
    1355             :     else
    1356     9944075 :       Cost += InlineConstants::InstrCost;
    1357             : 
    1358             :     using namespace ore;
    1359             :     // If the visit this instruction detected an uninlinable pattern, abort.
    1360    23566958 :     if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca ||
    1361    23565554 :         HasIndirectBr || HasFrameEscape) {
    1362        1408 :       if (ORE)
    1363         108 :         ORE->emit(OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline",
    1364          72 :                                            CandidateCS.getInstruction())
    1365         108 :                   << NV("Callee", &F)
    1366             :                   << " has uninlinable pattern and cost is not fully computed");
    1367       50702 :       return false;
    1368             :     }
    1369             : 
    1370             :     // If the caller is a recursive function then we don't want to inline
    1371             :     // functions which allocate a lot of stack space because it would increase
    1372             :     // the caller stack usage dramatically.
    1373    11872760 :     if (IsCallerRecursive &&
    1374       89987 :         AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller) {
    1375           8 :       if (ORE)
    1376           3 :         ORE->emit(
    1377           6 :             OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline",
    1378           6 :                                      CandidateCS.getInstruction())
    1379           9 :             << NV("Callee", &F)
    1380             :             << " is recursive and allocates too much stack space. Cost is "
    1381             :                "not fully computed");
    1382             :       return false;
    1383             :     }
    1384             : 
    1385             :     // Check if we've past the maximum possible threshold so we don't spin in
    1386             :     // huge basic blocks that will never inline.
    1387    11833861 :     if (Cost >= Threshold && !ComputeFullInlineCost)
    1388             :       return false;
    1389             :   }
    1390             : 
    1391      817454 :   return true;
    1392             : }
    1393             : 
    1394             : /// \brief Compute the base pointer and cumulative constant offsets for V.
    1395             : ///
    1396             : /// This strips all constant offsets off of V, leaving it the base pointer, and
    1397             : /// accumulates the total constant offset applied in the returned constant. It
    1398             : /// returns 0 if V is not a pointer, and returns the constant '0' if there are
    1399             : /// no constant offsets applied.
    1400      395210 : ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) {
    1401      790420 :   if (!V->getType()->isPointerTy())
    1402             :     return nullptr;
    1403             : 
    1404      705470 :   unsigned IntPtrWidth = DL.getPointerSizeInBits();
    1405      352735 :   APInt Offset = APInt::getNullValue(IntPtrWidth);
    1406             : 
    1407             :   // Even though we don't look through PHI nodes, we could be called on an
    1408             :   // instruction in an unreachable block, which may be on a cycle.
    1409      705470 :   SmallPtrSet<Value *, 4> Visited;
    1410      352735 :   Visited.insert(V);
    1411             :   do {
    1412      499650 :     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
    1413       70044 :       if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset))
    1414             :         return nullptr;
    1415       66463 :       V = GEP->getPointerOperand();
    1416      587898 :     } else if (Operator::getOpcode(V) == Instruction::BitCast) {
    1417       31224 :       V = cast<Operator>(V)->getOperand(0);
    1418      349154 :     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
    1419           0 :       if (GA->isInterposable())
    1420             :         break;
    1421           0 :       V = GA->getAliasee();
    1422             :     } else {
    1423             :       break;
    1424             :     }
    1425             :     assert(V->getType()->isPointerTy() && "Unexpected operand type!");
    1426       76871 :   } while (Visited.insert(V).second);
    1427             : 
    1428      349154 :   Type *IntPtrTy = DL.getIntPtrType(V->getContext());
    1429      698308 :   return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset));
    1430             : }
    1431             : 
    1432             : /// \brief Analyze a call site for potential inlining.
    1433             : ///
    1434             : /// Returns true if inlining this call is viable, and false if it is not
    1435             : /// viable. It computes the cost and adjusts the threshold based on numerous
    1436             : /// factors and heuristics. If this method returns false but the computed cost
    1437             : /// is below the computed threshold, then inlining was forcibly disabled by
    1438             : /// some artifact of the routine.
    1439      228149 : bool CallAnalyzer::analyzeCall(CallSite CS) {
    1440      228149 :   ++NumCallsAnalyzed;
    1441             : 
    1442             :   // Perform some tweaks to the cost and threshold based on the direct
    1443             :   // callsite information.
    1444             : 
    1445             :   // We want to more aggressively inline vector-dense kernels, so up the
    1446             :   // threshold, and we'll lower it if the % of vector instructions gets too
    1447             :   // low. Note that these bonuses are some what arbitrary and evolved over time
    1448             :   // by accident as much as because they are principled bonuses.
    1449             :   //
    1450             :   // FIXME: It would be nice to remove all such bonuses. At least it would be
    1451             :   // nice to base the bonus values on something more scientific.
    1452             :   assert(NumInstructions == 0);
    1453             :   assert(NumVectorInstructions == 0);
    1454             : 
    1455             :   // Update the threshold based on callsite properties
    1456      228149 :   updateThreshold(CS, F);
    1457             : 
    1458             :   // Speculatively apply all possible bonuses to Threshold. If cost exceeds
    1459             :   // this Threshold any time, and cost cannot decrease, we can stop processing
    1460             :   // the rest of the function body.
    1461      228149 :   Threshold += (SingleBBBonus + VectorBonus);
    1462             : 
    1463             :   // Give out bonuses for the callsite, as the instructions setting them up
    1464             :   // will be gone after inlining.
    1465      228149 :   Cost -= getCallsiteCost(CS, DL);
    1466             : 
    1467             :   // If this function uses the coldcc calling convention, prefer not to inline
    1468             :   // it.
    1469      456298 :   if (F.getCallingConv() == CallingConv::Cold)
    1470           0 :     Cost += InlineConstants::ColdccPenalty;
    1471             : 
    1472             :   // Check if we're done. This can happen due to bonuses and penalties.
    1473      228149 :   if (Cost >= Threshold && !ComputeFullInlineCost)
    1474             :     return false;
    1475             : 
    1476      456298 :   if (F.empty())
    1477             :     return true;
    1478             : 
    1479      228137 :   Function *Caller = CS.getInstruction()->getParent()->getParent();
    1480             :   // Check if the caller function is recursive itself.
    1481     1601139 :   for (User *U : Caller->users()) {
    1482      459114 :     CallSite Site(U);
    1483      459114 :     if (!Site)
    1484       53146 :       continue;
    1485      405968 :     Instruction *I = Site.getInstruction();
    1486      405968 :     if (I->getParent()->getParent() == Caller) {
    1487        1500 :       IsCallerRecursive = true;
    1488        1500 :       break;
    1489             :     }
    1490             :   }
    1491             : 
    1492             :   // Populate our simplified values by mapping from function arguments to call
    1493             :   // arguments with known important simplifications.
    1494      228137 :   CallSite::arg_iterator CAI = CS.arg_begin();
    1495     1079621 :   for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end();
    1496      623347 :        FAI != FAE; ++FAI, ++CAI) {
    1497             :     assert(CAI != CS.arg_end());
    1498       55850 :     if (Constant *C = dyn_cast<Constant>(CAI))
    1499      111700 :       SimplifiedValues[&*FAI] = C;
    1500             : 
    1501      395210 :     Value *PtrArg = *CAI;
    1502      395210 :     if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) {
    1503     1396616 :       ConstantOffsetPtrs[&*FAI] = std::make_pair(PtrArg, C->getValue());
    1504             : 
    1505             :       // We can SROA any pointer arguments derived from alloca instructions.
    1506      349154 :       if (isa<AllocaInst>(PtrArg)) {
    1507      337068 :         SROAArgValues[&*FAI] = PtrArg;
    1508      337068 :         SROAArgCosts[PtrArg] = 0;
    1509             :       }
    1510             :     }
    1511             :   }
    1512      456274 :   NumConstantArgs = SimplifiedValues.size();
    1513      456274 :   NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size();
    1514      456274 :   NumAllocaArgs = SROAArgValues.size();
    1515             : 
    1516             :   // FIXME: If a caller has multiple calls to a callee, we end up recomputing
    1517             :   // the ephemeral values multiple times (and they're completely determined by
    1518             :   // the callee, so this is purely duplicate work).
    1519      228137 :   SmallPtrSet<const Value *, 32> EphValues;
    1520      456274 :   CodeMetrics::collectEphemeralValues(&F, &GetAssumptionCache(F), EphValues);
    1521             : 
    1522             :   // The worklist of live basic blocks in the callee *after* inlining. We avoid
    1523             :   // adding basic blocks of the callee which can be proven to be dead for this
    1524             :   // particular call site in order to get more accurate cost estimates. This
    1525             :   // requires a somewhat heavyweight iteration pattern: we need to walk the
    1526             :   // basic blocks in a breadth-first order as we insert live successors. To
    1527             :   // accomplish this, prioritizing for small iterations because we exit after
    1528             :   // crossing our threshold, we use a small-size optimized SetVector.
    1529             :   typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>,
    1530             :                     SmallPtrSet<BasicBlock *, 16>>
    1531             :       BBSetVector;
    1532      456274 :   BBSetVector BBWorklist;
    1533      456274 :   BBWorklist.insert(&F.getEntryBlock());
    1534      228137 :   bool SingleBB = true;
    1535             :   // Note that we *must not* cache the size, this loop grows the worklist.
    1536     2091182 :   for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) {
    1537             :     // Bail out the moment we cross the threshold. This means we'll under-count
    1538             :     // the cost, but only when undercounting doesn't matter.
    1539      873513 :     if (Cost >= Threshold && !ComputeFullInlineCost)
    1540             :       break;
    1541             : 
    1542     1736316 :     BasicBlock *BB = BBWorklist[Idx];
    1543      868158 :     if (BB->empty())
    1544           0 :       continue;
    1545             : 
    1546             :     // Disallow inlining a blockaddress. A blockaddress only has defined
    1547             :     // behavior for an indirect branch in the same function, and we do not
    1548             :     // currently support inlining indirect branches. But, the inliner may not
    1549             :     // see an indirect branch that ends up being dead code at a particular call
    1550             :     // site. If the blockaddress escapes the function, e.g., via a global
    1551             :     // variable, inlining may lead to an invalid cross-function reference.
    1552      868158 :     if (BB->hasAddressTaken())
    1553             :       return false;
    1554             : 
    1555             :     // Analyze the cost of this block. If we blow through the threshold, this
    1556             :     // returns false, and we can bail on out.
    1557      868156 :     if (!analyzeBlock(BB, EphValues))
    1558             :       return false;
    1559             : 
    1560      817454 :     TerminatorInst *TI = BB->getTerminator();
    1561             : 
    1562             :     // Add in the live successors by first checking whether we have terminator
    1563             :     // that may be simplified based on the values simplified by this call.
    1564      424264 :     if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
    1565      424264 :       if (BI->isConditional()) {
    1566      215162 :         Value *Cond = BI->getCondition();
    1567        6930 :         if (ConstantInt *SimpleCond =
    1568      437254 :                 dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
    1569       13860 :           BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0));
    1570        6930 :           continue;
    1571             :         }
    1572             :       }
    1573        1269 :     } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
    1574        1269 :       Value *Cond = SI->getCondition();
    1575          64 :       if (ConstantInt *SimpleCond =
    1576        2602 :               dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) {
    1577         128 :         BBWorklist.insert(SI->findCaseValue(SimpleCond)->getCaseSuccessor());
    1578          64 :         continue;
    1579             :       }
    1580             :     }
    1581             : 
    1582             :     // If we're unable to select a particular successor, just count all of
    1583             :     // them.
    1584     1681915 :     for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize;
    1585             :          ++TIdx)
    1586      871455 :       BBWorklist.insert(TI->getSuccessor(TIdx));
    1587             : 
    1588             :     // If we had any successors at this point, than post-inlining is likely to
    1589             :     // have them as well. Note that we assume any basic blocks which existed
    1590             :     // due to branches or switches which folded above will also fold after
    1591             :     // inlining.
    1592      810460 :     if (SingleBB && TI->getNumSuccessors() > 1) {
    1593             :       // Take off the bonus we applied to the threshold.
    1594      137332 :       Threshold -= SingleBBBonus;
    1595      137332 :       SingleBB = false;
    1596             :     }
    1597             :   }
    1598             : 
    1599             :   bool OnlyOneCallAndLocalLinkage =
    1600      194702 :       F.hasLocalLinkage() && F.hasOneUse() && &F == CS.getCalledFunction();
    1601             :   // If this is a noduplicate call, we can still inline as long as
    1602             :   // inlining this would cause the removal of the caller (so the instruction
    1603             :   // is not actually duplicated, just moved).
    1604      173593 :   if (!OnlyOneCallAndLocalLinkage && ContainsNoDuplicateCall)
    1605             :     return false;
    1606             : 
    1607             :   // We applied the maximum possible vector bonus at the beginning. Now,
    1608             :   // subtract the excess bonus, if any, from the Threshold before
    1609             :   // comparing against Cost.
    1610      177424 :   if (NumVectorInstructions <= NumInstructions / 10)
    1611      177416 :     Threshold -= VectorBonus;
    1612           8 :   else if (NumVectorInstructions <= NumInstructions / 2)
    1613           4 :     Threshold -= VectorBonus/2;
    1614             : 
    1615      354848 :   return Cost < std::max(1, Threshold);
    1616             : }
    1617             : 
    1618             : #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    1619             : /// \brief Dump stats about this call's analysis.
    1620             : LLVM_DUMP_METHOD void CallAnalyzer::dump() {
    1621             : #define DEBUG_PRINT_STAT(x) dbgs() << "      " #x ": " << x << "\n"
    1622             :   DEBUG_PRINT_STAT(NumConstantArgs);
    1623             :   DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs);
    1624             :   DEBUG_PRINT_STAT(NumAllocaArgs);
    1625             :   DEBUG_PRINT_STAT(NumConstantPtrCmps);
    1626             :   DEBUG_PRINT_STAT(NumConstantPtrDiffs);
    1627             :   DEBUG_PRINT_STAT(NumInstructionsSimplified);
    1628             :   DEBUG_PRINT_STAT(NumInstructions);
    1629             :   DEBUG_PRINT_STAT(SROACostSavings);
    1630             :   DEBUG_PRINT_STAT(SROACostSavingsLost);
    1631             :   DEBUG_PRINT_STAT(ContainsNoDuplicateCall);
    1632             :   DEBUG_PRINT_STAT(Cost);
    1633             :   DEBUG_PRINT_STAT(Threshold);
    1634             : #undef DEBUG_PRINT_STAT
    1635             : }
    1636             : #endif
    1637             : 
    1638             : /// \brief Test that there are no attribute conflicts between Caller and Callee
    1639             : ///        that prevent inlining.
    1640      353239 : static bool functionsHaveCompatibleAttributes(Function *Caller,
    1641             :                                               Function *Callee,
    1642             :                                               TargetTransformInfo &TTI) {
    1643      706461 :   return TTI.areInlineCompatible(Caller, Callee) &&
    1644      706461 :          AttributeFuncs::areInlineCompatible(*Caller, *Callee);
    1645             : }
    1646             : 
    1647      228962 : int llvm::getCallsiteCost(CallSite CS, const DataLayout &DL) {
    1648      228962 :   int Cost = 0;
    1649      625255 :   for (unsigned I = 0, E = CS.arg_size(); I != E; ++I) {
    1650      396293 :     if (CS.isByValArgument(I)) {
    1651             :       // We approximate the number of loads and stores needed by dividing the
    1652             :       // size of the byval type by the target's pointer size.
    1653       10058 :       PointerType *PTy = cast<PointerType>(CS.getArgument(I)->getType());
    1654        5029 :       unsigned TypeSize = DL.getTypeSizeInBits(PTy->getElementType());
    1655        5029 :       unsigned PointerSize = DL.getPointerSizeInBits();
    1656             :       // Ceiling division.
    1657        5029 :       unsigned NumStores = (TypeSize + PointerSize - 1) / PointerSize;
    1658             : 
    1659             :       // If it generates more than 8 stores it is likely to be expanded as an
    1660             :       // inline memcpy so we take that as an upper bound. Otherwise we assume
    1661             :       // one load and one store per word copied.
    1662             :       // FIXME: The maxStoresPerMemcpy setting from the target should be used
    1663             :       // here instead of a magic number of 8, but it's not available via
    1664             :       // DataLayout.
    1665       10058 :       NumStores = std::min(NumStores, 8U);
    1666             : 
    1667        5029 :       Cost += 2 * NumStores * InlineConstants::InstrCost;
    1668             :     } else {
    1669             :       // For non-byval arguments subtract off one instruction per call
    1670             :       // argument.
    1671      391264 :       Cost += InlineConstants::InstrCost;
    1672             :     }
    1673             :   }
    1674             :   // The call instruction also disappears after inlining.
    1675      228962 :   Cost += InlineConstants::InstrCost + InlineConstants::CallPenalty;
    1676      228962 :   return Cost;
    1677             : }
    1678             : 
    1679      356957 : InlineCost llvm::getInlineCost(
    1680             :     CallSite CS, const InlineParams &Params, TargetTransformInfo &CalleeTTI,
    1681             :     std::function<AssumptionCache &(Function &)> &GetAssumptionCache,
    1682             :     Optional<function_ref<BlockFrequencyInfo &(Function &)>> GetBFI,
    1683             :     ProfileSummaryInfo *PSI, OptimizationRemarkEmitter *ORE) {
    1684             :   return getInlineCost(CS, CS.getCalledFunction(), Params, CalleeTTI,
    1685     1070871 :                        GetAssumptionCache, GetBFI, PSI, ORE);
    1686             : }
    1687             : 
    1688      356957 : InlineCost llvm::getInlineCost(
    1689             :     CallSite CS, Function *Callee, const InlineParams &Params,
    1690             :     TargetTransformInfo &CalleeTTI,
    1691             :     std::function<AssumptionCache &(Function &)> &GetAssumptionCache,
    1692             :     Optional<function_ref<BlockFrequencyInfo &(Function &)>> GetBFI,
    1693             :     ProfileSummaryInfo *PSI, OptimizationRemarkEmitter *ORE) {
    1694             : 
    1695             :   // Cannot inline indirect calls.
    1696      356957 :   if (!Callee)
    1697             :     return llvm::InlineCost::getNever();
    1698             : 
    1699             :   // Calls to functions with always-inline attributes should be inlined
    1700             :   // whenever possible.
    1701      356957 :   if (CS.hasFnAttr(Attribute::AlwaysInline)) {
    1702        3718 :     if (isInlineViable(*Callee))
    1703             :       return llvm::InlineCost::getAlways();
    1704             :     return llvm::InlineCost::getNever();
    1705             :   }
    1706             : 
    1707             :   // Never inline functions with conflicting attributes (unless callee has
    1708             :   // always-inline attribute).
    1709      353239 :   Function *Caller = CS.getCaller();
    1710      353239 :   if (!functionsHaveCompatibleAttributes(Caller, Callee, CalleeTTI))
    1711             :     return llvm::InlineCost::getNever();
    1712             : 
    1713             :   // Don't inline this call if the caller has the optnone attribute.
    1714      353201 :   if (Caller->hasFnAttribute(Attribute::OptimizeNone))
    1715             :     return llvm::InlineCost::getNever();
    1716             : 
    1717             :   // Don't inline functions which can be interposed at link-time.  Don't inline
    1718             :   // functions marked noinline or call sites marked noinline.
    1719             :   // Note: inlining non-exact non-interposable functions is fine, since we know
    1720             :   // we have *a* correct implementation of the source level function.
    1721     1287350 :   if (Callee->isInterposable() || Callee->hasFnAttribute(Attribute::NoInline) ||
    1722      227800 :       CS.isNoInline())
    1723             :     return llvm::InlineCost::getNever();
    1724             : 
    1725      227778 :   if (ORE)
    1726         914 :     ComputeFullInlineCost = true;
    1727             : 
    1728             :   DEBUG(llvm::dbgs() << "      Analyzing call of " << Callee->getName()
    1729             :                      << "... (caller:" << Caller->getName() << ")\n");
    1730             : 
    1731             :   CallAnalyzer CA(CalleeTTI, GetAssumptionCache, GetBFI, PSI, ORE, *Callee, CS,
    1732      455556 :                   Params);
    1733      227778 :   bool ShouldInline = CA.analyzeCall(CS);
    1734             : 
    1735             :   DEBUG(CA.dump());
    1736             : 
    1737             :   // Check if there was a reason to force inlining or no inlining.
    1738      227778 :   if (!ShouldInline && CA.getCost() < CA.getThreshold())
    1739             :     return InlineCost::getNever();
    1740      336420 :   if (ShouldInline && CA.getCost() >= CA.getThreshold())
    1741             :     return InlineCost::getAlways();
    1742             : 
    1743      226370 :   return llvm::InlineCost::get(CA.getCost(), CA.getThreshold());
    1744             : }
    1745             : 
    1746       26956 : bool llvm::isInlineViable(Function &F) {
    1747       26956 :   bool ReturnsTwice = F.hasFnAttribute(Attribute::ReturnsTwice);
    1748       53912 :   for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
    1749             :     // Disallow inlining of functions which contain indirect branches or
    1750             :     // blockaddresses.
    1751      140612 :     if (isa<IndirectBrInst>(BI->getTerminator()) || BI->hasAddressTaken())
    1752             :       return false;
    1753             : 
    1754      470676 :     for (auto &II : *BI) {
    1755      386316 :       CallSite CS(&II);
    1756      386316 :       if (!CS)
    1757      361075 :         continue;
    1758             : 
    1759             :       // Disallow recursive calls.
    1760       25241 :       if (&F == CS.getCalledFunction())
    1761             :         return false;
    1762             : 
    1763             :       // Disallow calls which expose returns-twice to a function not previously
    1764             :       // attributed as such.
    1765       75575 :       if (!ReturnsTwice && CS.isCall() &&
    1766       75375 :           cast<CallInst>(CS.getInstruction())->canReturnTwice())
    1767             :         return false;
    1768             : 
    1769             :       // Disallow inlining functions that call @llvm.localescape. Doing this
    1770             :       // correctly would require major changes to the inliner.
    1771       50240 :       if (CS.getCalledFunction() &&
    1772       25120 :           CS.getCalledFunction()->getIntrinsicID() ==
    1773             :               llvm::Intrinsic::localescape)
    1774             :         return false;
    1775             :     }
    1776             :   }
    1777             : 
    1778             :   return true;
    1779             : }
    1780             : 
    1781             : // APIs to create InlineParams based on command line flags and/or other
    1782             : // parameters.
    1783             : 
    1784        1350 : InlineParams llvm::getInlineParams(int Threshold) {
    1785        1350 :   InlineParams Params;
    1786             : 
    1787             :   // This field is the threshold to use for a callee by default. This is
    1788             :   // derived from one or more of:
    1789             :   //  * optimization or size-optimization levels,
    1790             :   //  * a value passed to createFunctionInliningPass function, or
    1791             :   //  * the -inline-threshold flag.
    1792             :   //  If the -inline-threshold flag is explicitly specified, that is used
    1793             :   //  irrespective of anything else.
    1794        1350 :   if (InlineThreshold.getNumOccurrences() > 0)
    1795          49 :     Params.DefaultThreshold = InlineThreshold;
    1796             :   else
    1797        1301 :     Params.DefaultThreshold = Threshold;
    1798             : 
    1799             :   // Set the HintThreshold knob from the -inlinehint-threshold.
    1800        2700 :   Params.HintThreshold = HintThreshold;
    1801             : 
    1802             :   // Set the HotCallSiteThreshold knob from the -hot-callsite-threshold.
    1803        2700 :   Params.HotCallSiteThreshold = HotCallSiteThreshold;
    1804             : 
    1805             :   // If the -locally-hot-callsite-threshold is explicitly specified, use it to
    1806             :   // populate LocallyHotCallSiteThreshold. Later, we populate
    1807             :   // Params.LocallyHotCallSiteThreshold from -locally-hot-callsite-threshold if
    1808             :   // we know that optimization level is O3 (in the getInlineParams variant that
    1809             :   // takes the opt and size levels).
    1810             :   // FIXME: Remove this check (and make the assignment unconditional) after
    1811             :   // addressing size regression issues at O2.
    1812        1350 :   if (LocallyHotCallSiteThreshold.getNumOccurrences() > 0)
    1813           0 :     Params.LocallyHotCallSiteThreshold = LocallyHotCallSiteThreshold;
    1814             : 
    1815             :   // Set the ColdCallSiteThreshold knob from the -inline-cold-callsite-threshold.
    1816        2700 :   Params.ColdCallSiteThreshold = ColdCallSiteThreshold;
    1817             : 
    1818             :   // Set the OptMinSizeThreshold and OptSizeThreshold params only if the
    1819             :   // -inlinehint-threshold commandline option is not explicitly given. If that
    1820             :   // option is present, then its value applies even for callees with size and
    1821             :   // minsize attributes.
    1822             :   // If the -inline-threshold is not specified, set the ColdThreshold from the
    1823             :   // -inlinecold-threshold even if it is not explicitly passed. If
    1824             :   // -inline-threshold is specified, then -inlinecold-threshold needs to be
    1825             :   // explicitly specified to set the ColdThreshold knob
    1826        1350 :   if (InlineThreshold.getNumOccurrences() == 0) {
    1827        2602 :     Params.OptMinSizeThreshold = InlineConstants::OptMinSizeThreshold;
    1828        2602 :     Params.OptSizeThreshold = InlineConstants::OptSizeThreshold;
    1829        1301 :     Params.ColdThreshold = ColdThreshold;
    1830          49 :   } else if (ColdThreshold.getNumOccurrences() > 0) {
    1831           0 :     Params.ColdThreshold = ColdThreshold;
    1832             :   }
    1833        1350 :   return Params;
    1834             : }
    1835             : 
    1836         737 : InlineParams llvm::getInlineParams() {
    1837         737 :   return getInlineParams(InlineThreshold);
    1838             : }
    1839             : 
    1840             : // Compute the default threshold for inlining based on the opt level and the
    1841             : // size opt level.
    1842             : static int computeThresholdFromOptLevels(unsigned OptLevel,
    1843             :                                          unsigned SizeOptLevel) {
    1844         613 :   if (OptLevel > 2)
    1845             :     return InlineConstants::OptAggressiveThreshold;
    1846         430 :   if (SizeOptLevel == 1) // -Os
    1847             :     return InlineConstants::OptSizeThreshold;
    1848         382 :   if (SizeOptLevel == 2) // -Oz
    1849             :     return InlineConstants::OptMinSizeThreshold;
    1850         359 :   return InlineThreshold;
    1851             : }
    1852             : 
    1853         613 : InlineParams llvm::getInlineParams(unsigned OptLevel, unsigned SizeOptLevel) {
    1854             :   auto Params =
    1855         613 :       getInlineParams(computeThresholdFromOptLevels(OptLevel, SizeOptLevel));
    1856             :   // At O3, use the value of -locally-hot-callsite-threshold option to populate
    1857             :   // Params.LocallyHotCallSiteThreshold. Below O3, this flag has effect only
    1858             :   // when it is specified explicitly.
    1859         613 :   if (OptLevel > 2)
    1860         183 :     Params.LocallyHotCallSiteThreshold = LocallyHotCallSiteThreshold;
    1861         613 :   return Params;
    1862      216918 : }

Generated by: LCOV version 1.13