Home | History | Annotate | Download | only in Analysis
      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/Support/CallSite.h"
     16 #include "llvm/CallingConv.h"
     17 #include "llvm/IntrinsicInst.h"
     18 #include "llvm/ADT/SmallPtrSet.h"
     19 
     20 using namespace llvm;
     21 
     22 /// callIsSmall - If a call is likely to lower to a single target instruction,
     23 /// or is otherwise deemed small return true.
     24 /// TODO: Perhaps calls like memcpy, strcpy, etc?
     25 bool llvm::callIsSmall(const Function *F) {
     26   if (!F) return false;
     27 
     28   if (F->hasLocalLinkage()) return false;
     29 
     30   if (!F->hasName()) return false;
     31 
     32   StringRef Name = F->getName();
     33 
     34   // These will all likely lower to a single selection DAG node.
     35   if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
     36       Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
     37       Name == "sin" || Name == "sinf" || Name == "sinl" ||
     38       Name == "cos" || Name == "cosf" || Name == "cosl" ||
     39       Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
     40     return true;
     41 
     42   // These are all likely to be optimized into something smaller.
     43   if (Name == "pow" || Name == "powf" || Name == "powl" ||
     44       Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
     45       Name == "floor" || Name == "floorf" || Name == "ceil" ||
     46       Name == "round" || Name == "ffs" || Name == "ffsl" ||
     47       Name == "abs" || Name == "labs" || Name == "llabs")
     48     return true;
     49 
     50   return false;
     51 }
     52 
     53 /// analyzeBasicBlock - Fill in the current structure with information gleaned
     54 /// from the specified block.
     55 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
     56   ++NumBlocks;
     57   unsigned NumInstsBeforeThisBB = NumInsts;
     58   for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
     59        II != E; ++II) {
     60     if (isa<PHINode>(II)) continue;           // PHI nodes don't count.
     61 
     62     // Special handling for calls.
     63     if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
     64       if (isa<DbgInfoIntrinsic>(II))
     65         continue;  // Debug intrinsics don't count as size.
     66 
     67       ImmutableCallSite CS(cast<Instruction>(II));
     68 
     69       if (const Function *F = CS.getCalledFunction()) {
     70         // If a function is both internal and has a single use, then it is
     71         // extremely likely to get inlined in the future (it was probably
     72         // exposed by an interleaved devirtualization pass).
     73         if (F->hasInternalLinkage() && F->hasOneUse())
     74           ++NumInlineCandidates;
     75 
     76         // If this call is to function itself, then the function is recursive.
     77         // Inlining it into other functions is a bad idea, because this is
     78         // basically just a form of loop peeling, and our metrics aren't useful
     79         // for that case.
     80         if (F == BB->getParent())
     81           isRecursive = true;
     82       }
     83 
     84       if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
     85         // Each argument to a call takes on average one instruction to set up.
     86         NumInsts += CS.arg_size();
     87 
     88         // We don't want inline asm to count as a call - that would prevent loop
     89         // unrolling. The argument setup cost is still real, though.
     90         if (!isa<InlineAsm>(CS.getCalledValue()))
     91           ++NumCalls;
     92       }
     93     }
     94 
     95     if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
     96       if (!AI->isStaticAlloca())
     97         this->usesDynamicAlloca = true;
     98     }
     99 
    100     if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy())
    101       ++NumVectorInsts;
    102 
    103     if (const CastInst *CI = dyn_cast<CastInst>(II)) {
    104       // Noop casts, including ptr <-> int,  don't count.
    105       if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
    106           isa<PtrToIntInst>(CI))
    107         continue;
    108       // Result of a cmp instruction is often extended (to be used by other
    109       // cmp instructions, logical or return instructions). These are usually
    110       // nop on most sane targets.
    111       if (isa<CmpInst>(CI->getOperand(0)))
    112         continue;
    113     } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
    114       // If a GEP has all constant indices, it will probably be folded with
    115       // a load/store.
    116       if (GEPI->hasAllConstantIndices())
    117         continue;
    118     }
    119 
    120     ++NumInsts;
    121   }
    122 
    123   if (isa<ReturnInst>(BB->getTerminator()))
    124     ++NumRets;
    125 
    126   // We never want to inline functions that contain an indirectbr.  This is
    127   // incorrect because all the blockaddress's (in static global initializers
    128   // for example) would be referring to the original function, and this indirect
    129   // jump would jump from the inlined copy of the function into the original
    130   // function which is extremely undefined behavior.
    131   if (isa<IndirectBrInst>(BB->getTerminator()))
    132     containsIndirectBr = true;
    133 
    134   // Remember NumInsts for this BB.
    135   NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB;
    136 }
    137 
    138 // CountCodeReductionForConstant - Figure out an approximation for how many
    139 // instructions will be constant folded if the specified value is constant.
    140 //
    141 unsigned CodeMetrics::CountCodeReductionForConstant(Value *V) {
    142   unsigned Reduction = 0;
    143   for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
    144     User *U = *UI;
    145     if (isa<BranchInst>(U) || isa<SwitchInst>(U)) {
    146       // We will be able to eliminate all but one of the successors.
    147       const TerminatorInst &TI = cast<TerminatorInst>(*U);
    148       const unsigned NumSucc = TI.getNumSuccessors();
    149       unsigned Instrs = 0;
    150       for (unsigned I = 0; I != NumSucc; ++I)
    151         Instrs += NumBBInsts[TI.getSuccessor(I)];
    152       // We don't know which blocks will be eliminated, so use the average size.
    153       Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
    154     } else {
    155       // Figure out if this instruction will be removed due to simple constant
    156       // propagation.
    157       Instruction &Inst = cast<Instruction>(*U);
    158 
    159       // We can't constant propagate instructions which have effects or
    160       // read memory.
    161       //
    162       // FIXME: It would be nice to capture the fact that a load from a
    163       // pointer-to-constant-global is actually a *really* good thing to zap.
    164       // Unfortunately, we don't know the pointer that may get propagated here,
    165       // so we can't make this decision.
    166       if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
    167           isa<AllocaInst>(Inst))
    168         continue;
    169 
    170       bool AllOperandsConstant = true;
    171       for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
    172         if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
    173           AllOperandsConstant = false;
    174           break;
    175         }
    176 
    177       if (AllOperandsConstant) {
    178         // We will get to remove this instruction...
    179         Reduction += InlineConstants::InstrCost;
    180 
    181         // And any other instructions that use it which become constants
    182         // themselves.
    183         Reduction += CountCodeReductionForConstant(&Inst);
    184       }
    185     }
    186   }
    187   return Reduction;
    188 }
    189 
    190 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
    191 // the function will be if it is inlined into a context where an argument
    192 // becomes an alloca.
    193 //
    194 unsigned CodeMetrics::CountCodeReductionForAlloca(Value *V) {
    195   if (!V->getType()->isPointerTy()) return 0;  // Not a pointer
    196   unsigned Reduction = 0;
    197   for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
    198     Instruction *I = cast<Instruction>(*UI);
    199     if (isa<LoadInst>(I) || isa<StoreInst>(I))
    200       Reduction += InlineConstants::InstrCost;
    201     else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
    202       // If the GEP has variable indices, we won't be able to do much with it.
    203       if (GEP->hasAllConstantIndices())
    204         Reduction += CountCodeReductionForAlloca(GEP);
    205     } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
    206       // Track pointer through bitcasts.
    207       Reduction += CountCodeReductionForAlloca(BCI);
    208     } else {
    209       // If there is some other strange instruction, we're not going to be able
    210       // to do much if we inline this.
    211       return 0;
    212     }
    213   }
    214 
    215   return Reduction;
    216 }
    217 
    218 /// analyzeFunction - Fill in the current structure with information gleaned
    219 /// from the specified function.
    220 void CodeMetrics::analyzeFunction(Function *F) {
    221   // If this function contains a call to setjmp or _setjmp, never inline
    222   // it.  This is a hack because we depend on the user marking their local
    223   // variables as volatile if they are live across a setjmp call, and they
    224   // probably won't do this in callers.
    225   if (F->callsFunctionThatReturnsTwice())
    226     callsSetJmp = true;
    227 
    228   // Look at the size of the callee.
    229   for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
    230     analyzeBasicBlock(&*BB);
    231 }
    232 
    233 /// analyzeFunction - Fill in the current structure with information gleaned
    234 /// from the specified function.
    235 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
    236   Metrics.analyzeFunction(F);
    237 
    238   // A function with exactly one return has it removed during the inlining
    239   // process (see InlineFunction), so don't count it.
    240   // FIXME: This knowledge should really be encoded outside of FunctionInfo.
    241   if (Metrics.NumRets==1)
    242     --Metrics.NumInsts;
    243 
    244   // Check out all of the arguments to the function, figuring out how much
    245   // code can be eliminated if one of the arguments is a constant.
    246   ArgumentWeights.reserve(F->arg_size());
    247   for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
    248     ArgumentWeights.push_back(ArgInfo(Metrics.CountCodeReductionForConstant(I),
    249                                       Metrics.CountCodeReductionForAlloca(I)));
    250 }
    251 
    252 /// NeverInline - returns true if the function should never be inlined into
    253 /// any caller
    254 bool InlineCostAnalyzer::FunctionInfo::NeverInline() {
    255   return (Metrics.callsSetJmp || Metrics.isRecursive ||
    256           Metrics.containsIndirectBr);
    257 }
    258 // getSpecializationBonus - The heuristic used to determine the per-call
    259 // performance boost for using a specialization of Callee with argument
    260 // specializedArgNo replaced by a constant.
    261 int InlineCostAnalyzer::getSpecializationBonus(Function *Callee,
    262          SmallVectorImpl<unsigned> &SpecializedArgNos)
    263 {
    264   if (Callee->mayBeOverridden())
    265     return 0;
    266 
    267   int Bonus = 0;
    268   // If this function uses the coldcc calling convention, prefer not to
    269   // specialize it.
    270   if (Callee->getCallingConv() == CallingConv::Cold)
    271     Bonus -= InlineConstants::ColdccPenalty;
    272 
    273   // Get information about the callee.
    274   FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
    275 
    276   // If we haven't calculated this information yet, do so now.
    277   if (CalleeFI->Metrics.NumBlocks == 0)
    278     CalleeFI->analyzeFunction(Callee);
    279 
    280   unsigned ArgNo = 0;
    281   unsigned i = 0;
    282   for (Function::arg_iterator I = Callee->arg_begin(), E = Callee->arg_end();
    283        I != E; ++I, ++ArgNo)
    284     if (ArgNo == SpecializedArgNos[i]) {
    285       ++i;
    286       Bonus += CountBonusForConstant(I);
    287     }
    288 
    289   // Calls usually take a long time, so they make the specialization gain
    290   // smaller.
    291   Bonus -= CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
    292 
    293   return Bonus;
    294 }
    295 
    296 // ConstantFunctionBonus - Figure out how much of a bonus we can get for
    297 // possibly devirtualizing a function. We'll subtract the size of the function
    298 // we may wish to inline from the indirect call bonus providing a limit on
    299 // growth. Leave an upper limit of 0 for the bonus - we don't want to penalize
    300 // inlining because we decide we don't want to give a bonus for
    301 // devirtualizing.
    302 int InlineCostAnalyzer::ConstantFunctionBonus(CallSite CS, Constant *C) {
    303 
    304   // This could just be NULL.
    305   if (!C) return 0;
    306 
    307   Function *F = dyn_cast<Function>(C);
    308   if (!F) return 0;
    309 
    310   int Bonus = InlineConstants::IndirectCallBonus + getInlineSize(CS, F);
    311   return (Bonus > 0) ? 0 : Bonus;
    312 }
    313 
    314 // CountBonusForConstant - Figure out an approximation for how much per-call
    315 // performance boost we can expect if the specified value is constant.
    316 int InlineCostAnalyzer::CountBonusForConstant(Value *V, Constant *C) {
    317   unsigned Bonus = 0;
    318   for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
    319     User *U = *UI;
    320     if (CallInst *CI = dyn_cast<CallInst>(U)) {
    321       // Turning an indirect call into a direct call is a BIG win
    322       if (CI->getCalledValue() == V)
    323         Bonus += ConstantFunctionBonus(CallSite(CI), C);
    324     } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
    325       // Turning an indirect call into a direct call is a BIG win
    326       if (II->getCalledValue() == V)
    327         Bonus += ConstantFunctionBonus(CallSite(II), C);
    328     }
    329     // FIXME: Eliminating conditional branches and switches should
    330     // also yield a per-call performance boost.
    331     else {
    332       // Figure out the bonuses that wll accrue due to simple constant
    333       // propagation.
    334       Instruction &Inst = cast<Instruction>(*U);
    335 
    336       // We can't constant propagate instructions which have effects or
    337       // read memory.
    338       //
    339       // FIXME: It would be nice to capture the fact that a load from a
    340       // pointer-to-constant-global is actually a *really* good thing to zap.
    341       // Unfortunately, we don't know the pointer that may get propagated here,
    342       // so we can't make this decision.
    343       if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
    344           isa<AllocaInst>(Inst))
    345         continue;
    346 
    347       bool AllOperandsConstant = true;
    348       for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
    349         if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
    350           AllOperandsConstant = false;
    351           break;
    352         }
    353 
    354       if (AllOperandsConstant)
    355         Bonus += CountBonusForConstant(&Inst);
    356     }
    357   }
    358 
    359   return Bonus;
    360 }
    361 
    362 int InlineCostAnalyzer::getInlineSize(CallSite CS, Function *Callee) {
    363   // Get information about the callee.
    364   FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
    365 
    366   // If we haven't calculated this information yet, do so now.
    367   if (CalleeFI->Metrics.NumBlocks == 0)
    368     CalleeFI->analyzeFunction(Callee);
    369 
    370   // InlineCost - This value measures how good of an inline candidate this call
    371   // site is to inline.  A lower inline cost make is more likely for the call to
    372   // be inlined.  This value may go negative.
    373   //
    374   int InlineCost = 0;
    375 
    376   // Compute any size reductions we can expect due to arguments being passed into
    377   // the function.
    378   //
    379   unsigned ArgNo = 0;
    380   CallSite::arg_iterator I = CS.arg_begin();
    381   for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
    382        FI != FE; ++I, ++FI, ++ArgNo) {
    383 
    384     // If an alloca is passed in, inlining this function is likely to allow
    385     // significant future optimization possibilities (like scalar promotion, and
    386     // scalarization), so encourage the inlining of the function.
    387     //
    388     if (isa<AllocaInst>(I))
    389       InlineCost -= CalleeFI->ArgumentWeights[ArgNo].AllocaWeight;
    390 
    391     // If this is a constant being passed into the function, use the argument
    392     // weights calculated for the callee to determine how much will be folded
    393     // away with this information.
    394     else if (isa<Constant>(I))
    395       InlineCost -= CalleeFI->ArgumentWeights[ArgNo].ConstantWeight;
    396   }
    397 
    398   // Each argument passed in has a cost at both the caller and the callee
    399   // sides.  Measurements show that each argument costs about the same as an
    400   // instruction.
    401   InlineCost -= (CS.arg_size() * InlineConstants::InstrCost);
    402 
    403   // Now that we have considered all of the factors that make the call site more
    404   // likely to be inlined, look at factors that make us not want to inline it.
    405 
    406   // Calls usually take a long time, so they make the inlining gain smaller.
    407   InlineCost += CalleeFI->Metrics.NumCalls * InlineConstants::CallPenalty;
    408 
    409   // Look at the size of the callee. Each instruction counts as 5.
    410   InlineCost += CalleeFI->Metrics.NumInsts*InlineConstants::InstrCost;
    411 
    412   return InlineCost;
    413 }
    414 
    415 int InlineCostAnalyzer::getInlineBonuses(CallSite CS, Function *Callee) {
    416   // Get information about the callee.
    417   FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
    418 
    419   // If we haven't calculated this information yet, do so now.
    420   if (CalleeFI->Metrics.NumBlocks == 0)
    421     CalleeFI->analyzeFunction(Callee);
    422 
    423   bool isDirectCall = CS.getCalledFunction() == Callee;
    424   Instruction *TheCall = CS.getInstruction();
    425   int Bonus = 0;
    426 
    427   // If there is only one call of the function, and it has internal linkage,
    428   // make it almost guaranteed to be inlined.
    429   //
    430   if (Callee->hasLocalLinkage() && Callee->hasOneUse() && isDirectCall)
    431     Bonus += InlineConstants::LastCallToStaticBonus;
    432 
    433   // If the instruction after the call, or if the normal destination of the
    434   // invoke is an unreachable instruction, the function is noreturn.  As such,
    435   // there is little point in inlining this.
    436   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
    437     if (isa<UnreachableInst>(II->getNormalDest()->begin()))
    438       Bonus += InlineConstants::NoreturnPenalty;
    439   } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
    440     Bonus += InlineConstants::NoreturnPenalty;
    441 
    442   // If this function uses the coldcc calling convention, prefer not to inline
    443   // it.
    444   if (Callee->getCallingConv() == CallingConv::Cold)
    445     Bonus += InlineConstants::ColdccPenalty;
    446 
    447   // Add to the inline quality for properties that make the call valuable to
    448   // inline.  This includes factors that indicate that the result of inlining
    449   // the function will be optimizable.  Currently this just looks at arguments
    450   // passed into the function.
    451   //
    452   CallSite::arg_iterator I = CS.arg_begin();
    453   for (Function::arg_iterator FI = Callee->arg_begin(), FE = Callee->arg_end();
    454        FI != FE; ++I, ++FI)
    455     // Compute any constant bonus due to inlining we want to give here.
    456     if (isa<Constant>(I))
    457       Bonus += CountBonusForConstant(FI, cast<Constant>(I));
    458 
    459   return Bonus;
    460 }
    461 
    462 // getInlineCost - The heuristic used to determine if we should inline the
    463 // function call or not.
    464 //
    465 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
    466                                SmallPtrSet<const Function*, 16> &NeverInline) {
    467   return getInlineCost(CS, CS.getCalledFunction(), NeverInline);
    468 }
    469 
    470 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
    471                                Function *Callee,
    472                                SmallPtrSet<const Function*, 16> &NeverInline) {
    473   Instruction *TheCall = CS.getInstruction();
    474   Function *Caller = TheCall->getParent()->getParent();
    475 
    476   // Don't inline functions which can be redefined at link-time to mean
    477   // something else.  Don't inline functions marked noinline or call sites
    478   // marked noinline.
    479   if (Callee->mayBeOverridden() ||
    480       Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
    481       CS.isNoInline())
    482     return llvm::InlineCost::getNever();
    483 
    484   // Get information about the callee.
    485   FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
    486 
    487   // If we haven't calculated this information yet, do so now.
    488   if (CalleeFI->Metrics.NumBlocks == 0)
    489     CalleeFI->analyzeFunction(Callee);
    490 
    491   // If we should never inline this, return a huge cost.
    492   if (CalleeFI->NeverInline())
    493     return InlineCost::getNever();
    494 
    495   // FIXME: It would be nice to kill off CalleeFI->NeverInline. Then we
    496   // could move this up and avoid computing the FunctionInfo for
    497   // things we are going to just return always inline for. This
    498   // requires handling setjmp somewhere else, however.
    499   if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
    500     return InlineCost::getAlways();
    501 
    502   if (CalleeFI->Metrics.usesDynamicAlloca) {
    503     // Get information about the caller.
    504     FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
    505 
    506     // If we haven't calculated this information yet, do so now.
    507     if (CallerFI.Metrics.NumBlocks == 0) {
    508       CallerFI.analyzeFunction(Caller);
    509 
    510       // Recompute the CalleeFI pointer, getting Caller could have invalidated
    511       // it.
    512       CalleeFI = &CachedFunctionInfo[Callee];
    513     }
    514 
    515     // Don't inline a callee with dynamic alloca into a caller without them.
    516     // Functions containing dynamic alloca's are inefficient in various ways;
    517     // don't create more inefficiency.
    518     if (!CallerFI.Metrics.usesDynamicAlloca)
    519       return InlineCost::getNever();
    520   }
    521 
    522   // InlineCost - This value measures how good of an inline candidate this call
    523   // site is to inline.  A lower inline cost make is more likely for the call to
    524   // be inlined.  This value may go negative due to the fact that bonuses
    525   // are negative numbers.
    526   //
    527   int InlineCost = getInlineSize(CS, Callee) + getInlineBonuses(CS, Callee);
    528   return llvm::InlineCost::get(InlineCost);
    529 }
    530 
    531 // getSpecializationCost - The heuristic used to determine the code-size
    532 // impact of creating a specialized version of Callee with argument
    533 // SpecializedArgNo replaced by a constant.
    534 InlineCost InlineCostAnalyzer::getSpecializationCost(Function *Callee,
    535                                SmallVectorImpl<unsigned> &SpecializedArgNos)
    536 {
    537   // Don't specialize functions which can be redefined at link-time to mean
    538   // something else.
    539   if (Callee->mayBeOverridden())
    540     return llvm::InlineCost::getNever();
    541 
    542   // Get information about the callee.
    543   FunctionInfo *CalleeFI = &CachedFunctionInfo[Callee];
    544 
    545   // If we haven't calculated this information yet, do so now.
    546   if (CalleeFI->Metrics.NumBlocks == 0)
    547     CalleeFI->analyzeFunction(Callee);
    548 
    549   int Cost = 0;
    550 
    551   // Look at the original size of the callee.  Each instruction counts as 5.
    552   Cost += CalleeFI->Metrics.NumInsts * InlineConstants::InstrCost;
    553 
    554   // Offset that with the amount of code that can be constant-folded
    555   // away with the given arguments replaced by constants.
    556   for (SmallVectorImpl<unsigned>::iterator an = SpecializedArgNos.begin(),
    557        ae = SpecializedArgNos.end(); an != ae; ++an)
    558     Cost -= CalleeFI->ArgumentWeights[*an].ConstantWeight;
    559 
    560   return llvm::InlineCost::get(Cost);
    561 }
    562 
    563 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
    564 // higher threshold to determine if the function call should be inlined.
    565 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
    566   Function *Callee = CS.getCalledFunction();
    567 
    568   // Get information about the callee.
    569   FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
    570 
    571   // If we haven't calculated this information yet, do so now.
    572   if (CalleeFI.Metrics.NumBlocks == 0)
    573     CalleeFI.analyzeFunction(Callee);
    574 
    575   float Factor = 1.0f;
    576   // Single BB functions are often written to be inlined.
    577   if (CalleeFI.Metrics.NumBlocks == 1)
    578     Factor += 0.5f;
    579 
    580   // Be more aggressive if the function contains a good chunk (if it mades up
    581   // at least 10% of the instructions) of vector instructions.
    582   if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
    583     Factor += 2.0f;
    584   else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
    585     Factor += 1.5f;
    586   return Factor;
    587 }
    588 
    589 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
    590 /// been inlined.
    591 void
    592 InlineCostAnalyzer::growCachedCostInfo(Function *Caller, Function *Callee) {
    593   CodeMetrics &CallerMetrics = CachedFunctionInfo[Caller].Metrics;
    594 
    595   // For small functions we prefer to recalculate the cost for better accuracy.
    596   if (CallerMetrics.NumBlocks < 10 && CallerMetrics.NumInsts < 1000) {
    597     resetCachedCostInfo(Caller);
    598     return;
    599   }
    600 
    601   // For large functions, we can save a lot of computation time by skipping
    602   // recalculations.
    603   if (CallerMetrics.NumCalls > 0)
    604     --CallerMetrics.NumCalls;
    605 
    606   if (Callee == 0) return;
    607 
    608   CodeMetrics &CalleeMetrics = CachedFunctionInfo[Callee].Metrics;
    609 
    610   // If we don't have metrics for the callee, don't recalculate them just to
    611   // update an approximation in the caller.  Instead, just recalculate the
    612   // caller info from scratch.
    613   if (CalleeMetrics.NumBlocks == 0) {
    614     resetCachedCostInfo(Caller);
    615     return;
    616   }
    617 
    618   // Since CalleeMetrics were already calculated, we know that the CallerMetrics
    619   // reference isn't invalidated: both were in the DenseMap.
    620   CallerMetrics.usesDynamicAlloca |= CalleeMetrics.usesDynamicAlloca;
    621 
    622   // FIXME: If any of these three are true for the callee, the callee was
    623   // not inlined into the caller, so I think they're redundant here.
    624   CallerMetrics.callsSetJmp |= CalleeMetrics.callsSetJmp;
    625   CallerMetrics.isRecursive |= CalleeMetrics.isRecursive;
    626   CallerMetrics.containsIndirectBr |= CalleeMetrics.containsIndirectBr;
    627 
    628   CallerMetrics.NumInsts += CalleeMetrics.NumInsts;
    629   CallerMetrics.NumBlocks += CalleeMetrics.NumBlocks;
    630   CallerMetrics.NumCalls += CalleeMetrics.NumCalls;
    631   CallerMetrics.NumVectorInsts += CalleeMetrics.NumVectorInsts;
    632   CallerMetrics.NumRets += CalleeMetrics.NumRets;
    633 
    634   // analyzeBasicBlock counts each function argument as an inst.
    635   if (CallerMetrics.NumInsts >= Callee->arg_size())
    636     CallerMetrics.NumInsts -= Callee->arg_size();
    637   else
    638     CallerMetrics.NumInsts = 0;
    639 
    640   // We are not updating the argument weights. We have already determined that
    641   // Caller is a fairly large function, so we accept the loss of precision.
    642 }
    643 
    644 /// clear - empty the cache of inline costs
    645 void InlineCostAnalyzer::clear() {
    646   CachedFunctionInfo.clear();
    647 }
    648