InlineSimple.cpp revision ccca3ca85f046bf7c99aa954ac121fdf59722499
1//===- InlineSimple.cpp - Code to perform simple function inlining --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements bottom-up inlining of functions into callees. 11// 12//===----------------------------------------------------------------------===// 13 14#include "Inliner.h" 15#include "llvm/CallingConv.h" 16#include "llvm/Instructions.h" 17#include "llvm/IntrinsicInst.h" 18#include "llvm/Function.h" 19#include "llvm/Type.h" 20#include "llvm/Support/CallSite.h" 21#include "llvm/Transforms/IPO.h" 22using namespace llvm; 23 24namespace { 25 struct ArgInfo { 26 unsigned ConstantWeight; 27 unsigned AllocaWeight; 28 29 ArgInfo(unsigned CWeight, unsigned AWeight) 30 : ConstantWeight(CWeight), AllocaWeight(AWeight) {} 31 }; 32 33 // FunctionInfo - For each function, calculate the size of it in blocks and 34 // instructions. 35 struct FunctionInfo { 36 // NumInsts, NumBlocks - Keep track of how large each function is, which is 37 // used to estimate the code size cost of inlining it. 38 unsigned NumInsts, NumBlocks; 39 40 // ArgumentWeights - Each formal argument of the function is inspected to 41 // see if it is used in any contexts where making it a constant or alloca 42 // would reduce the code size. If so, we add some value to the argument 43 // entry here. 44 std::vector<ArgInfo> ArgumentWeights; 45 46 FunctionInfo() : NumInsts(0), NumBlocks(0) {} 47 48 /// analyzeFunction - Fill in the current structure with information gleaned 49 /// from the specified function. 50 void analyzeFunction(Function *F); 51 }; 52 53 class SimpleInliner : public Inliner { 54 std::map<const Function*, FunctionInfo> CachedFunctionInfo; 55 public: 56 int getInlineCost(CallSite CS); 57 }; 58 RegisterOpt<SimpleInliner> X("inline", "Function Integration/Inlining"); 59} 60 61ModulePass *llvm::createFunctionInliningPass() { return new SimpleInliner(); } 62 63// CountCodeReductionForConstant - Figure out an approximation for how many 64// instructions will be constant folded if the specified value is constant. 65// 66static unsigned CountCodeReductionForConstant(Value *V) { 67 unsigned Reduction = 0; 68 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 69 if (isa<BranchInst>(*UI)) 70 Reduction += 40; // Eliminating a conditional branch is a big win 71 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI)) 72 // Eliminating a switch is a big win, proportional to the number of edges 73 // deleted. 74 Reduction += (SI->getNumSuccessors()-1) * 40; 75 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 76 // Turning an indirect call into a direct call is a BIG win 77 Reduction += CI->getCalledValue() == V ? 500 : 0; 78 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 79 // Turning an indirect call into a direct call is a BIG win 80 Reduction += II->getCalledValue() == V ? 500 : 0; 81 } else { 82 // Figure out if this instruction will be removed due to simple constant 83 // propagation. 84 Instruction &Inst = cast<Instruction>(**UI); 85 bool AllOperandsConstant = true; 86 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i) 87 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) { 88 AllOperandsConstant = false; 89 break; 90 } 91 92 if (AllOperandsConstant) { 93 // We will get to remove this instruction... 94 Reduction += 7; 95 96 // And any other instructions that use it which become constants 97 // themselves. 98 Reduction += CountCodeReductionForConstant(&Inst); 99 } 100 } 101 102 return Reduction; 103} 104 105// CountCodeReductionForAlloca - Figure out an approximation of how much smaller 106// the function will be if it is inlined into a context where an argument 107// becomes an alloca. 108// 109static unsigned CountCodeReductionForAlloca(Value *V) { 110 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer 111 unsigned Reduction = 0; 112 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 113 Instruction *I = cast<Instruction>(*UI); 114 if (isa<LoadInst>(I) || isa<StoreInst>(I)) 115 Reduction += 10; 116 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) { 117 // If the GEP has variable indices, we won't be able to do much with it. 118 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end(); 119 I != E; ++I) 120 if (!isa<Constant>(*I)) return 0; 121 Reduction += CountCodeReductionForAlloca(GEP)+15; 122 } else { 123 // If there is some other strange instruction, we're not going to be able 124 // to do much if we inline this. 125 return 0; 126 } 127 } 128 129 return Reduction; 130} 131 132/// analyzeFunction - Fill in the current structure with information gleaned 133/// from the specified function. 134void FunctionInfo::analyzeFunction(Function *F) { 135 unsigned NumInsts = 0, NumBlocks = 0; 136 137 // Look at the size of the callee. Each basic block counts as 20 units, and 138 // each instruction counts as 10. 139 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) { 140 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end(); 141 II != E; ++II) 142 if (!isa<DbgInfoIntrinsic>(II)) 143 ++NumInsts; 144 145 ++NumBlocks; 146 } 147 148 this->NumBlocks = NumBlocks; 149 this->NumInsts = NumInsts; 150 151 // Check out all of the arguments to the function, figuring out how much 152 // code can be eliminated if one of the arguments is a constant. 153 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I) 154 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I), 155 CountCodeReductionForAlloca(I))); 156} 157 158 159// getInlineCost - The heuristic used to determine if we should inline the 160// function call or not. 161// 162int SimpleInliner::getInlineCost(CallSite CS) { 163 Instruction *TheCall = CS.getInstruction(); 164 Function *Callee = CS.getCalledFunction(); 165 const Function *Caller = TheCall->getParent()->getParent(); 166 167 // Don't inline a directly recursive call. 168 if (Caller == Callee) return 2000000000; 169 170 // InlineCost - This value measures how good of an inline candidate this call 171 // site is to inline. A lower inline cost make is more likely for the call to 172 // be inlined. This value may go negative. 173 // 174 int InlineCost = 0; 175 176 // If there is only one call of the function, and it has internal linkage, 177 // make it almost guaranteed to be inlined. 178 // 179 if (Callee->hasInternalLinkage() && Callee->hasOneUse()) 180 InlineCost -= 30000; 181 182 // If this function uses the coldcc calling convention, prefer not to inline 183 // it. 184 if (Callee->getCallingConv() == CallingConv::Cold) 185 InlineCost += 2000; 186 187 // If the instruction after the call, or if the normal destination of the 188 // invoke is an unreachable instruction, the function is noreturn. As such, 189 // there is little point in inlining this. 190 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 191 if (isa<UnreachableInst>(II->getNormalDest()->begin())) 192 InlineCost += 10000; 193 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall))) 194 InlineCost += 10000; 195 196 // Get information about the callee... 197 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee]; 198 199 // If we haven't calculated this information yet, do so now. 200 if (CalleeFI.NumBlocks == 0) 201 CalleeFI.analyzeFunction(Callee); 202 203 // Add to the inline quality for properties that make the call valuable to 204 // inline. This includes factors that indicate that the result of inlining 205 // the function will be optimizable. Currently this just looks at arguments 206 // passed into the function. 207 // 208 unsigned ArgNo = 0; 209 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 210 I != E; ++I, ++ArgNo) { 211 // Each argument passed in has a cost at both the caller and the callee 212 // sides. This favors functions that take many arguments over functions 213 // that take few arguments. 214 InlineCost -= 20; 215 216 // If this is a function being passed in, it is very likely that we will be 217 // able to turn an indirect function call into a direct function call. 218 if (isa<Function>(I)) 219 InlineCost -= 100; 220 221 // If an alloca is passed in, inlining this function is likely to allow 222 // significant future optimization possibilities (like scalar promotion, and 223 // scalarization), so encourage the inlining of the function. 224 // 225 else if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) { 226 if (ArgNo < CalleeFI.ArgumentWeights.size()) 227 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight; 228 229 // If this is a constant being passed into the function, use the argument 230 // weights calculated for the callee to determine how much will be folded 231 // away with this information. 232 } else if (isa<Constant>(I)) { 233 if (ArgNo < CalleeFI.ArgumentWeights.size()) 234 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight; 235 } 236 } 237 238 // Now that we have considered all of the factors that make the call site more 239 // likely to be inlined, look at factors that make us not want to inline it. 240 241 // Don't inline into something too big, which would make it bigger. Here, we 242 // count each basic block as a single unit. 243 // 244 InlineCost += Caller->size()/20; 245 246 247 // Look at the size of the callee. Each basic block counts as 20 units, and 248 // each instruction counts as 5. 249 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20; 250 return InlineCost; 251} 252 253