InlineSimple.cpp revision 3e15bf33e024b9df9e89351a165acfdb1dde51ed
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/Support/Compiler.h" 22#include "llvm/Transforms/IPO.h" 23using namespace llvm; 24 25namespace { 26 struct VISIBILITY_HIDDEN ArgInfo { 27 unsigned ConstantWeight; 28 unsigned AllocaWeight; 29 30 ArgInfo(unsigned CWeight, unsigned AWeight) 31 : ConstantWeight(CWeight), AllocaWeight(AWeight) {} 32 }; 33 34 // FunctionInfo - For each function, calculate the size of it in blocks and 35 // instructions. 36 struct VISIBILITY_HIDDEN FunctionInfo { 37 // NumInsts, NumBlocks - Keep track of how large each function is, which is 38 // used to estimate the code size cost of inlining it. 39 unsigned NumInsts, NumBlocks; 40 41 // ArgumentWeights - Each formal argument of the function is inspected to 42 // see if it is used in any contexts where making it a constant or alloca 43 // would reduce the code size. If so, we add some value to the argument 44 // entry here. 45 std::vector<ArgInfo> ArgumentWeights; 46 47 FunctionInfo() : NumInsts(0), NumBlocks(0) {} 48 49 /// analyzeFunction - Fill in the current structure with information gleaned 50 /// from the specified function. 51 void analyzeFunction(Function *F); 52 }; 53 54 class VISIBILITY_HIDDEN SimpleInliner : public Inliner { 55 std::map<const Function*, FunctionInfo> CachedFunctionInfo; 56 public: 57 static const char ID; // Pass identifcation, replacement for typeid 58 int getInlineCost(CallSite CS); 59 }; 60 const char SimpleInliner::ID = 0; 61 RegisterPass<SimpleInliner> X("inline", "Function Integration/Inlining"); 62} 63 64Pass *llvm::createFunctionInliningPass() { return new SimpleInliner(); } 65 66// CountCodeReductionForConstant - Figure out an approximation for how many 67// instructions will be constant folded if the specified value is constant. 68// 69static unsigned CountCodeReductionForConstant(Value *V) { 70 unsigned Reduction = 0; 71 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI) 72 if (isa<BranchInst>(*UI)) 73 Reduction += 40; // Eliminating a conditional branch is a big win 74 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI)) 75 // Eliminating a switch is a big win, proportional to the number of edges 76 // deleted. 77 Reduction += (SI->getNumSuccessors()-1) * 40; 78 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 79 // Turning an indirect call into a direct call is a BIG win 80 Reduction += CI->getCalledValue() == V ? 500 : 0; 81 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) { 82 // Turning an indirect call into a direct call is a BIG win 83 Reduction += II->getCalledValue() == V ? 500 : 0; 84 } else { 85 // Figure out if this instruction will be removed due to simple constant 86 // propagation. 87 Instruction &Inst = cast<Instruction>(**UI); 88 bool AllOperandsConstant = true; 89 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i) 90 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) { 91 AllOperandsConstant = false; 92 break; 93 } 94 95 if (AllOperandsConstant) { 96 // We will get to remove this instruction... 97 Reduction += 7; 98 99 // And any other instructions that use it which become constants 100 // themselves. 101 Reduction += CountCodeReductionForConstant(&Inst); 102 } 103 } 104 105 return Reduction; 106} 107 108// CountCodeReductionForAlloca - Figure out an approximation of how much smaller 109// the function will be if it is inlined into a context where an argument 110// becomes an alloca. 111// 112static unsigned CountCodeReductionForAlloca(Value *V) { 113 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer 114 unsigned Reduction = 0; 115 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 116 Instruction *I = cast<Instruction>(*UI); 117 if (isa<LoadInst>(I) || isa<StoreInst>(I)) 118 Reduction += 10; 119 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) { 120 // If the GEP has variable indices, we won't be able to do much with it. 121 for (Instruction::op_iterator I = GEP->op_begin()+1, E = GEP->op_end(); 122 I != E; ++I) 123 if (!isa<Constant>(*I)) return 0; 124 Reduction += CountCodeReductionForAlloca(GEP)+15; 125 } else { 126 // If there is some other strange instruction, we're not going to be able 127 // to do much if we inline this. 128 return 0; 129 } 130 } 131 132 return Reduction; 133} 134 135/// analyzeFunction - Fill in the current structure with information gleaned 136/// from the specified function. 137void FunctionInfo::analyzeFunction(Function *F) { 138 unsigned NumInsts = 0, NumBlocks = 0; 139 140 // Look at the size of the callee. Each basic block counts as 20 units, and 141 // each instruction counts as 10. 142 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) { 143 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end(); 144 II != E; ++II) { 145 if (isa<DbgInfoIntrinsic>(II)) continue; // Debug intrinsics don't count. 146 147 // Noop casts, including ptr <-> int, don't count. 148 if (const CastInst *CI = dyn_cast<CastInst>(II)) { 149 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) || 150 isa<PtrToIntInst>(CI)) 151 continue; 152 } else if (const GetElementPtrInst *GEPI = 153 dyn_cast<GetElementPtrInst>(II)) { 154 // If a GEP has all constant indices, it will probably be folded with 155 // a load/store. 156 bool AllConstant = true; 157 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i) 158 if (!isa<ConstantInt>(GEPI->getOperand(i))) { 159 AllConstant = false; 160 break; 161 } 162 if (AllConstant) continue; 163 } 164 165 ++NumInsts; 166 } 167 168 ++NumBlocks; 169 } 170 171 this->NumBlocks = NumBlocks; 172 this->NumInsts = NumInsts; 173 174 // Check out all of the arguments to the function, figuring out how much 175 // code can be eliminated if one of the arguments is a constant. 176 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I) 177 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I), 178 CountCodeReductionForAlloca(I))); 179} 180 181 182// getInlineCost - The heuristic used to determine if we should inline the 183// function call or not. 184// 185int SimpleInliner::getInlineCost(CallSite CS) { 186 Instruction *TheCall = CS.getInstruction(); 187 Function *Callee = CS.getCalledFunction(); 188 const Function *Caller = TheCall->getParent()->getParent(); 189 190 // Don't inline a directly recursive call. 191 if (Caller == Callee) return 2000000000; 192 193 // InlineCost - This value measures how good of an inline candidate this call 194 // site is to inline. A lower inline cost make is more likely for the call to 195 // be inlined. This value may go negative. 196 // 197 int InlineCost = 0; 198 199 // If there is only one call of the function, and it has internal linkage, 200 // make it almost guaranteed to be inlined. 201 // 202 if (Callee->hasInternalLinkage() && Callee->hasOneUse()) 203 InlineCost -= 30000; 204 205 // If this function uses the coldcc calling convention, prefer not to inline 206 // it. 207 if (Callee->getCallingConv() == CallingConv::Cold) 208 InlineCost += 2000; 209 210 // If the instruction after the call, or if the normal destination of the 211 // invoke is an unreachable instruction, the function is noreturn. As such, 212 // there is little point in inlining this. 213 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 214 if (isa<UnreachableInst>(II->getNormalDest()->begin())) 215 InlineCost += 10000; 216 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall))) 217 InlineCost += 10000; 218 219 // Get information about the callee... 220 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee]; 221 222 // If we haven't calculated this information yet, do so now. 223 if (CalleeFI.NumBlocks == 0) 224 CalleeFI.analyzeFunction(Callee); 225 226 // Add to the inline quality for properties that make the call valuable to 227 // inline. This includes factors that indicate that the result of inlining 228 // the function will be optimizable. Currently this just looks at arguments 229 // passed into the function. 230 // 231 unsigned ArgNo = 0; 232 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 233 I != E; ++I, ++ArgNo) { 234 // Each argument passed in has a cost at both the caller and the callee 235 // sides. This favors functions that take many arguments over functions 236 // that take few arguments. 237 InlineCost -= 20; 238 239 // If this is a function being passed in, it is very likely that we will be 240 // able to turn an indirect function call into a direct function call. 241 if (isa<Function>(I)) 242 InlineCost -= 100; 243 244 // If an alloca is passed in, inlining this function is likely to allow 245 // significant future optimization possibilities (like scalar promotion, and 246 // scalarization), so encourage the inlining of the function. 247 // 248 else if (isa<AllocaInst>(I)) { 249 if (ArgNo < CalleeFI.ArgumentWeights.size()) 250 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight; 251 252 // If this is a constant being passed into the function, use the argument 253 // weights calculated for the callee to determine how much will be folded 254 // away with this information. 255 } else if (isa<Constant>(I)) { 256 if (ArgNo < CalleeFI.ArgumentWeights.size()) 257 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight; 258 } 259 } 260 261 // Now that we have considered all of the factors that make the call site more 262 // likely to be inlined, look at factors that make us not want to inline it. 263 264 // Don't inline into something too big, which would make it bigger. Here, we 265 // count each basic block as a single unit. 266 // 267 InlineCost += Caller->size()/20; 268 269 270 // Look at the size of the callee. Each basic block counts as 20 units, and 271 // each instruction counts as 5. 272 InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20; 273 return InlineCost; 274} 275 276