LLVM API Documentation
00001 //===- InlineCost.cpp - Cost analysis for inliner -------------------------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This file implements inline cost analysis. 00011 // 00012 //===----------------------------------------------------------------------===// 00013 00014 00015 #include "llvm/Transforms/Utils/InlineCost.h" 00016 #include "llvm/Support/CallSite.h" 00017 #include "llvm/CallingConv.h" 00018 #include "llvm/IntrinsicInst.h" 00019 00020 using namespace llvm; 00021 00022 // CountCodeReductionForConstant - Figure out an approximation for how many 00023 // instructions will be constant folded if the specified value is constant. 00024 // 00025 unsigned InlineCostAnalyzer::FunctionInfo:: 00026 CountCodeReductionForConstant(Value *V) { 00027 unsigned Reduction = 0; 00028 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 00029 if (isa<BranchInst>(*UI)) 00030 Reduction += 40; // Eliminating a conditional branch is a big win 00031 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI)) 00032 // Eliminating a switch is a big win, proportional to the number of edges 00033 // deleted. 00034 Reduction += (SI->getNumSuccessors()-1) * 40; 00035 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 00036 // Turning an indirect call into a direct call is a BIG win 00037 Reduction += CI->getCalledValue() == V ? 500 : 0; 00038 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 00039 // Turning an indirect call into a direct call is a BIG win 00040 Reduction += II->getCalledValue() == V ? 500 : 0; 00041 } else { 00042 // Figure out if this instruction will be removed due to simple constant 00043 // propagation. 00044 Instruction &Inst = cast<Instruction>(**UI); 00045 00046 // We can't constant propagate instructions which have effects or 00047 // read memory. 00048 // 00049 // FIXME: It would be nice to capture the fact that a load from a 00050 // pointer-to-constant-global is actually a *really* good thing to zap. 00051 // Unfortunately, we don't know the pointer that may get propagated here, 00052 // so we can't make this decision. 00053 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() || 00054 isa<AllocationInst>(Inst)) 00055 continue; 00056 00057 bool AllOperandsConstant = true; 00058 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i) 00059 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) { 00060 AllOperandsConstant = false; 00061 break; 00062 } 00063 00064 if (AllOperandsConstant) { 00065 // We will get to remove this instruction... 00066 Reduction += 7; 00067 00068 // And any other instructions that use it which become constants 00069 // themselves. 00070 Reduction += CountCodeReductionForConstant(&Inst); 00071 } 00072 } 00073 00074 return Reduction; 00075 } 00076 00077 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller 00078 // the function will be if it is inlined into a context where an argument 00079 // becomes an alloca. 00080 // 00081 unsigned InlineCostAnalyzer::FunctionInfo:: 00082 CountCodeReductionForAlloca(Value *V) { 00083 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer 00084 unsigned Reduction = 0; 00085 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 00086 Instruction *I = cast<Instruction>(*UI); 00087 if (isa<LoadInst>(I) || isa<StoreInst>(I)) 00088 Reduction += 10; 00089 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) { 00090 // If the GEP has variable indices, we won't be able to do much with it. 00091 if (!GEP->hasAllConstantIndices()) 00092 Reduction += CountCodeReductionForAlloca(GEP)+15; 00093 } else { 00094 // If there is some other strange instruction, we're not going to be able 00095 // to do much if we inline this. 00096 return 0; 00097 } 00098 } 00099 00100 return Reduction; 00101 } 00102 00103 /// analyzeFunction - Fill in the current structure with information gleaned 00104 /// from the specified function. 00105 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) { 00106 unsigned NumInsts = 0, NumBlocks = 0, NumVectorInsts = 0; 00107 00108 // Look at the size of the callee. Each basic block counts as 20 units, and 00109 // each instruction counts as 5. 00110 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) { 00111 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end(); 00112 II != E; ++II) { 00113 if (isa<PHINode>(II)) continue; // PHI nodes don't count. 00114 00115 // Special handling for calls. 00116 if (isa<CallInst>(II) || isa<InvokeInst>(II)) { 00117 if (isa<DbgInfoIntrinsic>(II)) 00118 continue; // Debug intrinsics don't count as size. 00119 00120 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II)); 00121 00122 // If this function contains a call to setjmp or _setjmp, never inline 00123 // it. This is a hack because we depend on the user marking their local 00124 // variables as volatile if they are live across a setjmp call, and they 00125 // probably won't do this in callers. 00126 if (Function *F = CS.getCalledFunction()) 00127 if (F->isDeclaration() && 00128 (F->getName() == "setjmp" || F->getName() == "_setjmp")) { 00129 NeverInline = true; 00130 return; 00131 } 00132 00133 // Calls often compile into many machine instructions. Bump up their 00134 // cost to reflect this. 00135 if (!isa<IntrinsicInst>(II)) 00136 NumInsts += 5; 00137 } 00138 00139 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) { 00140 if (!AI->isStaticAlloca()) 00141 this->usesDynamicAlloca = true; 00142 } 00143 00144 if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType())) 00145 ++NumVectorInsts; 00146 00147 // Noop casts, including ptr <-> int, don't count. 00148 if (const CastInst *CI = dyn_cast<CastInst>(II)) { 00149 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) || 00150 isa<PtrToIntInst>(CI)) 00151 continue; 00152 } else if (const GetElementPtrInst *GEPI = 00153 dyn_cast<GetElementPtrInst>(II)) { 00154 // If a GEP has all constant indices, it will probably be folded with 00155 // a load/store. 00156 if (GEPI->hasAllConstantIndices()) 00157 continue; 00158 } 00159 00160 ++NumInsts; 00161 } 00162 00163 ++NumBlocks; 00164 } 00165 00166 this->NumBlocks = NumBlocks; 00167 this->NumInsts = NumInsts; 00168 this->NumVectorInsts = NumVectorInsts; 00169 00170 // Check out all of the arguments to the function, figuring out how much 00171 // code can be eliminated if one of the arguments is a constant. 00172 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I) 00173 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I), 00174 CountCodeReductionForAlloca(I))); 00175 } 00176 00177 00178 00179 // getInlineCost - The heuristic used to determine if we should inline the 00180 // function call or not. 00181 // 00182 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS, 00183 SmallPtrSet<const Function *, 16> &NeverInline) { 00184 Instruction *TheCall = CS.getInstruction(); 00185 Function *Callee = CS.getCalledFunction(); 00186 Function *Caller = TheCall->getParent()->getParent(); 00187 00188 // Don't inline functions which can be redefined at link-time to mean 00189 // something else. 00190 if (Callee->mayBeOverridden() || 00191 // Don't inline functions marked noinline. 00192 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee)) 00193 return llvm::InlineCost::getNever(); 00194 00195 // InlineCost - This value measures how good of an inline candidate this call 00196 // site is to inline. A lower inline cost make is more likely for the call to 00197 // be inlined. This value may go negative. 00198 // 00199 int InlineCost = 0; 00200 00201 // If there is only one call of the function, and it has internal linkage, 00202 // make it almost guaranteed to be inlined. 00203 // 00204 if (Callee->hasLocalLinkage() && Callee->hasOneUse()) 00205 InlineCost -= 15000; 00206 00207 // If this function uses the coldcc calling convention, prefer not to inline 00208 // it. 00209 if (Callee->getCallingConv() == CallingConv::Cold) 00210 InlineCost += 2000; 00211 00212 // If the instruction after the call, or if the normal destination of the 00213 // invoke is an unreachable instruction, the function is noreturn. As such, 00214 // there is little point in inlining this. 00215 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 00216 if (isa<UnreachableInst>(II->getNormalDest()->begin())) 00217 InlineCost += 10000; 00218 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall))) 00219 InlineCost += 10000; 00220 00221 // Get information about the callee... 00222 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee]; 00223 00224 // If we haven't calculated this information yet, do so now. 00225 if (CalleeFI.NumBlocks == 0) 00226 CalleeFI.analyzeFunction(Callee); 00227 00228 // If we should never inline this, return a huge cost. 00229 if (CalleeFI.NeverInline) 00230 return InlineCost::getNever(); 00231 00232 // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we 00233 // could move this up and avoid computing the FunctionInfo for 00234 // things we are going to just return always inline for. This 00235 // requires handling setjmp somewhere else, however. 00236 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline)) 00237 return InlineCost::getAlways(); 00238 00239 if (CalleeFI.usesDynamicAlloca) { 00240 // Get infomation about the caller... 00241 FunctionInfo &CallerFI = CachedFunctionInfo[Caller]; 00242 00243 // If we haven't calculated this information yet, do so now. 00244 if (CallerFI.NumBlocks == 0) 00245 CallerFI.analyzeFunction(Caller); 00246 00247 // Don't inline a callee with dynamic alloca into a caller without them. 00248 // Functions containing dynamic alloca's are inefficient in various ways; 00249 // don't create more inefficiency. 00250 if (!CallerFI.usesDynamicAlloca) 00251 return InlineCost::getNever(); 00252 } 00253 00254 // Add to the inline quality for properties that make the call valuable to 00255 // inline. This includes factors that indicate that the result of inlining 00256 // the function will be optimizable. Currently this just looks at arguments 00257 // passed into the function. 00258 // 00259 unsigned ArgNo = 0; 00260 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 00261 I != E; ++I, ++ArgNo) { 00262 // Each argument passed in has a cost at both the caller and the callee 00263 // sides. This favors functions that take many arguments over functions 00264 // that take few arguments. 00265 InlineCost -= 20; 00266 00267 // If this is a function being passed in, it is very likely that we will be 00268 // able to turn an indirect function call into a direct function call. 00269 if (isa<Function>(I)) 00270 InlineCost -= 100; 00271 00272 // If an alloca is passed in, inlining this function is likely to allow 00273 // significant future optimization possibilities (like scalar promotion, and 00274 // scalarization), so encourage the inlining of the function. 00275 // 00276 else if (isa<AllocaInst>(I)) { 00277 if (ArgNo < CalleeFI.ArgumentWeights.size()) 00278 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight; 00279 00280 // If this is a constant being passed into the function, use the argument 00281 // weights calculated for the callee to determine how much will be folded 00282 // away with this information. 00283 } else if (isa<Constant>(I)) { 00284 if (ArgNo < CalleeFI.ArgumentWeights.size()) 00285 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight; 00286 } 00287 } 00288 00289 // Now that we have considered all of the factors that make the call site more 00290 // likely to be inlined, look at factors that make us not want to inline it. 00291 00292 // Don't inline into something too big, which would make it bigger. 00293 // 00294 InlineCost += Caller->size()/15; 00295 00296 // Look at the size of the callee. Each instruction counts as 5. 00297 InlineCost += CalleeFI.NumInsts*5; 00298 00299 return llvm::InlineCost::get(InlineCost); 00300 } 00301 00302 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a 00303 // higher threshold to determine if the function call should be inlined. 00304 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) { 00305 Function *Callee = CS.getCalledFunction(); 00306 00307 // Get information about the callee... 00308 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee]; 00309 00310 // If we haven't calculated this information yet, do so now. 00311 if (CalleeFI.NumBlocks == 0) 00312 CalleeFI.analyzeFunction(Callee); 00313 00314 float Factor = 1.0f; 00315 // Single BB functions are often written to be inlined. 00316 if (CalleeFI.NumBlocks == 1) 00317 Factor += 0.5f; 00318 00319 // Be more aggressive if the function contains a good chunk (if it mades up 00320 // at least 10% of the instructions) of vector instructions. 00321 if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/2) 00322 Factor += 2.0f; 00323 else if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10) 00324 Factor += 1.5f; 00325 return Factor; 00326 }