ValueEnumerator.cpp revision cc7b011728b9e8c3574247b81f79689840b3d33a
1//===-- ValueEnumerator.cpp - Number values and types for bitcode writer --===// 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 the ValueEnumerator class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "ValueEnumerator.h" 15#include "llvm/Constants.h" 16#include "llvm/DerivedTypes.h" 17#include "llvm/Module.h" 18#include "llvm/TypeSymbolTable.h" 19#include "llvm/ValueSymbolTable.h" 20#include "llvm/Instructions.h" 21#include <algorithm> 22using namespace llvm; 23 24static bool isSingleValueType(const std::pair<const llvm::Type*, 25 unsigned int> &P) { 26 return P.first->isSingleValueType(); 27} 28 29static bool isIntegerValue(const std::pair<const Value*, unsigned> &V) { 30 return isa<IntegerType>(V.first->getType()); 31} 32 33static bool CompareByFrequency(const std::pair<const llvm::Type*, 34 unsigned int> &P1, 35 const std::pair<const llvm::Type*, 36 unsigned int> &P2) { 37 return P1.second > P2.second; 38} 39 40/// ValueEnumerator - Enumerate module-level information. 41ValueEnumerator::ValueEnumerator(const Module *M) { 42 InstructionCount = 0; 43 44 // Enumerate the global variables. 45 for (Module::const_global_iterator I = M->global_begin(), 46 E = M->global_end(); I != E; ++I) 47 EnumerateValue(I); 48 49 // Enumerate the functions. 50 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I) { 51 EnumerateValue(I); 52 EnumerateAttributes(cast<Function>(I)->getAttributes()); 53 } 54 55 // Enumerate the aliases. 56 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end(); 57 I != E; ++I) 58 EnumerateValue(I); 59 60 // Remember what is the cutoff between globalvalue's and other constants. 61 unsigned FirstConstant = Values.size(); 62 63 // Enumerate the global variable initializers. 64 for (Module::const_global_iterator I = M->global_begin(), 65 E = M->global_end(); I != E; ++I) 66 if (I->hasInitializer()) 67 EnumerateValue(I->getInitializer()); 68 69 // Enumerate the aliasees. 70 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end(); 71 I != E; ++I) 72 EnumerateValue(I->getAliasee()); 73 74 // Enumerate types used by the type symbol table. 75 EnumerateTypeSymbolTable(M->getTypeSymbolTable()); 76 77 // Insert constants that are named at module level into the slot pool so that 78 // the module symbol table can refer to them... 79 EnumerateValueSymbolTable(M->getValueSymbolTable()); 80 81 SmallVector<std::pair<unsigned, MDNode*>, 8> MDs; 82 83 // Enumerate types used by function bodies and argument lists. 84 for (Module::const_iterator F = M->begin(), E = M->end(); F != E; ++F) { 85 86 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); 87 I != E; ++I) 88 EnumerateType(I->getType()); 89 90 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 91 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E;++I){ 92 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 93 OI != E; ++OI) 94 EnumerateOperandType(*OI); 95 EnumerateType(I->getType()); 96 if (const CallInst *CI = dyn_cast<CallInst>(I)) 97 EnumerateAttributes(CI->getAttributes()); 98 else if (const InvokeInst *II = dyn_cast<InvokeInst>(I)) 99 EnumerateAttributes(II->getAttributes()); 100 101 // Enumerate metadata attached with this instruction. 102 MDs.clear(); 103 I->getAllMetadata(MDs); 104 for (unsigned i = 0, e = MDs.size(); i != e; ++i) 105 EnumerateMetadata(MDs[i].second); 106 } 107 } 108 109 // Optimize constant ordering. 110 OptimizeConstants(FirstConstant, Values.size()); 111 112 // Sort the type table by frequency so that most commonly used types are early 113 // in the table (have low bit-width). 114 std::stable_sort(Types.begin(), Types.end(), CompareByFrequency); 115 116 // Partition the Type ID's so that the single-value types occur before the 117 // aggregate types. This allows the aggregate types to be dropped from the 118 // type table after parsing the global variable initializers. 119 std::partition(Types.begin(), Types.end(), isSingleValueType); 120 121 // Now that we rearranged the type table, rebuild TypeMap. 122 for (unsigned i = 0, e = Types.size(); i != e; ++i) 123 TypeMap[Types[i].first] = i+1; 124} 125 126unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const { 127 InstructionMapType::const_iterator I = InstructionMap.find(Inst); 128 assert (I != InstructionMap.end() && "Instruction is not mapped!"); 129 return I->second; 130} 131 132void ValueEnumerator::setInstructionID(const Instruction *I) { 133 InstructionMap[I] = InstructionCount++; 134} 135 136unsigned ValueEnumerator::getValueID(const Value *V) const { 137 if (isa<MetadataBase>(V)) { 138 ValueMapType::const_iterator I = MDValueMap.find(V); 139 assert(I != MDValueMap.end() && "Value not in slotcalculator!"); 140 return I->second-1; 141 } 142 143 ValueMapType::const_iterator I = ValueMap.find(V); 144 assert(I != ValueMap.end() && "Value not in slotcalculator!"); 145 return I->second-1; 146} 147 148// Optimize constant ordering. 149namespace { 150 struct CstSortPredicate { 151 ValueEnumerator &VE; 152 explicit CstSortPredicate(ValueEnumerator &ve) : VE(ve) {} 153 bool operator()(const std::pair<const Value*, unsigned> &LHS, 154 const std::pair<const Value*, unsigned> &RHS) { 155 // Sort by plane. 156 if (LHS.first->getType() != RHS.first->getType()) 157 return VE.getTypeID(LHS.first->getType()) < 158 VE.getTypeID(RHS.first->getType()); 159 // Then by frequency. 160 return LHS.second > RHS.second; 161 } 162 }; 163} 164 165/// OptimizeConstants - Reorder constant pool for denser encoding. 166void ValueEnumerator::OptimizeConstants(unsigned CstStart, unsigned CstEnd) { 167 if (CstStart == CstEnd || CstStart+1 == CstEnd) return; 168 169 CstSortPredicate P(*this); 170 std::stable_sort(Values.begin()+CstStart, Values.begin()+CstEnd, P); 171 172 // Ensure that integer constants are at the start of the constant pool. This 173 // is important so that GEP structure indices come before gep constant exprs. 174 std::partition(Values.begin()+CstStart, Values.begin()+CstEnd, 175 isIntegerValue); 176 177 // Rebuild the modified portion of ValueMap. 178 for (; CstStart != CstEnd; ++CstStart) 179 ValueMap[Values[CstStart].first] = CstStart+1; 180} 181 182 183/// EnumerateTypeSymbolTable - Insert all of the types in the specified symbol 184/// table. 185void ValueEnumerator::EnumerateTypeSymbolTable(const TypeSymbolTable &TST) { 186 for (TypeSymbolTable::const_iterator TI = TST.begin(), TE = TST.end(); 187 TI != TE; ++TI) 188 EnumerateType(TI->second); 189} 190 191/// EnumerateValueSymbolTable - Insert all of the values in the specified symbol 192/// table into the values table. 193void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) { 194 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end(); 195 VI != VE; ++VI) 196 EnumerateValue(VI->getValue()); 197} 198 199void ValueEnumerator::EnumerateMetadata(const MetadataBase *MD) { 200 // Check to see if it's already in! 201 unsigned &MDValueID = MDValueMap[MD]; 202 if (MDValueID) { 203 // Increment use count. 204 MDValues[MDValueID-1].second++; 205 return; 206 } 207 208 // Enumerate the type of this value. 209 EnumerateType(MD->getType()); 210 211 if (const MDNode *N = dyn_cast<MDNode>(MD)) { 212 MDValues.push_back(std::make_pair(MD, 1U)); 213 MDValueMap[MD] = MDValues.size(); 214 MDValueID = MDValues.size(); 215 for (unsigned i = 0, e = N->getNumElements(); i != e; ++i) { 216 if (Value *V = N->getElement(i)) 217 EnumerateValue(V); 218 else 219 EnumerateType(Type::getVoidTy(MD->getContext())); 220 } 221 return; 222 } 223 224 if (const NamedMDNode *N = dyn_cast<NamedMDNode>(MD)) { 225 for (unsigned i = 0, e = N->getNumElements(); i != e; ++i) 226 EnumerateValue(N->getElement(i)); 227 MDValues.push_back(std::make_pair(MD, 1U)); 228 MDValueMap[MD] = Values.size(); 229 return; 230 } 231 232 // Add the value. 233 assert(isa<MDString>(MD) && "Unknown metadata kind"); 234 MDValues.push_back(std::make_pair(MD, 1U)); 235 MDValueID = MDValues.size(); 236} 237 238void ValueEnumerator::EnumerateValue(const Value *V) { 239 assert(!V->getType()->isVoidTy() && "Can't insert void values!"); 240 if (const MetadataBase *MB = dyn_cast<MetadataBase>(V)) 241 return EnumerateMetadata(MB); 242 243 // Check to see if it's already in! 244 unsigned &ValueID = ValueMap[V]; 245 if (ValueID) { 246 // Increment use count. 247 Values[ValueID-1].second++; 248 return; 249 } 250 251 // Enumerate the type of this value. 252 EnumerateType(V->getType()); 253 254 if (const Constant *C = dyn_cast<Constant>(V)) { 255 if (isa<GlobalValue>(C)) { 256 // Initializers for globals are handled explicitly elsewhere. 257 } else if (isa<ConstantArray>(C) && cast<ConstantArray>(C)->isString()) { 258 // Do not enumerate the initializers for an array of simple characters. 259 // The initializers just polute the value table, and we emit the strings 260 // specially. 261 } else if (C->getNumOperands()) { 262 // If a constant has operands, enumerate them. This makes sure that if a 263 // constant has uses (for example an array of const ints), that they are 264 // inserted also. 265 266 // We prefer to enumerate them with values before we enumerate the user 267 // itself. This makes it more likely that we can avoid forward references 268 // in the reader. We know that there can be no cycles in the constants 269 // graph that don't go through a global variable. 270 for (User::const_op_iterator I = C->op_begin(), E = C->op_end(); 271 I != E; ++I) 272 if (!isa<BasicBlock>(*I)) // Don't enumerate BB operand to BlockAddress. 273 EnumerateValue(*I); 274 275 // Finally, add the value. Doing this could make the ValueID reference be 276 // dangling, don't reuse it. 277 Values.push_back(std::make_pair(V, 1U)); 278 ValueMap[V] = Values.size(); 279 return; 280 } 281 } 282 283 // Add the value. 284 Values.push_back(std::make_pair(V, 1U)); 285 ValueID = Values.size(); 286} 287 288 289void ValueEnumerator::EnumerateType(const Type *Ty) { 290 unsigned &TypeID = TypeMap[Ty]; 291 292 if (TypeID) { 293 // If we've already seen this type, just increase its occurrence count. 294 Types[TypeID-1].second++; 295 return; 296 } 297 298 // First time we saw this type, add it. 299 Types.push_back(std::make_pair(Ty, 1U)); 300 TypeID = Types.size(); 301 302 // Enumerate subtypes. 303 for (Type::subtype_iterator I = Ty->subtype_begin(), E = Ty->subtype_end(); 304 I != E; ++I) 305 EnumerateType(*I); 306} 307 308// Enumerate the types for the specified value. If the value is a constant, 309// walk through it, enumerating the types of the constant. 310void ValueEnumerator::EnumerateOperandType(const Value *V) { 311 EnumerateType(V->getType()); 312 if (const Constant *C = dyn_cast<Constant>(V)) { 313 // If this constant is already enumerated, ignore it, we know its type must 314 // be enumerated. 315 if (ValueMap.count(V)) return; 316 317 // This constant may have operands, make sure to enumerate the types in 318 // them. 319 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) { 320 const User *Op = C->getOperand(i); 321 322 // Don't enumerate basic blocks here, this happens as operands to 323 // blockaddress. 324 if (isa<BasicBlock>(Op)) continue; 325 326 EnumerateOperandType(cast<Constant>(Op)); 327 } 328 329 if (const MDNode *N = dyn_cast<MDNode>(V)) { 330 for (unsigned i = 0, e = N->getNumElements(); i != e; ++i) 331 if (Value *Elem = N->getElement(i)) 332 EnumerateOperandType(Elem); 333 } 334 } else if (isa<MDString>(V) || isa<MDNode>(V)) 335 EnumerateValue(V); 336} 337 338void ValueEnumerator::EnumerateAttributes(const AttrListPtr &PAL) { 339 if (PAL.isEmpty()) return; // null is always 0. 340 // Do a lookup. 341 unsigned &Entry = AttributeMap[PAL.getRawPointer()]; 342 if (Entry == 0) { 343 // Never saw this before, add it. 344 Attributes.push_back(PAL); 345 Entry = Attributes.size(); 346 } 347} 348 349 350void ValueEnumerator::incorporateFunction(const Function &F) { 351 NumModuleValues = Values.size(); 352 353 // Adding function arguments to the value table. 354 for(Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); 355 I != E; ++I) 356 EnumerateValue(I); 357 358 FirstFuncConstantID = Values.size(); 359 360 // Add all function-level constants to the value table. 361 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 362 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) 363 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 364 OI != E; ++OI) { 365 if ((isa<Constant>(*OI) && !isa<GlobalValue>(*OI)) || 366 isa<InlineAsm>(*OI)) 367 EnumerateValue(*OI); 368 } 369 BasicBlocks.push_back(BB); 370 ValueMap[BB] = BasicBlocks.size(); 371 } 372 373 // Optimize the constant layout. 374 OptimizeConstants(FirstFuncConstantID, Values.size()); 375 376 // Add the function's parameter attributes so they are available for use in 377 // the function's instruction. 378 EnumerateAttributes(F.getAttributes()); 379 380 FirstInstID = Values.size(); 381 382 // Add all of the instructions. 383 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 384 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) { 385 if (I->getType() != Type::getVoidTy(F.getContext())) 386 EnumerateValue(I); 387 } 388 } 389} 390 391void ValueEnumerator::purgeFunction() { 392 /// Remove purged values from the ValueMap. 393 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i) 394 ValueMap.erase(Values[i].first); 395 for (unsigned i = 0, e = BasicBlocks.size(); i != e; ++i) 396 ValueMap.erase(BasicBlocks[i]); 397 398 Values.resize(NumModuleValues); 399 BasicBlocks.clear(); 400} 401 402static void IncorporateFunctionInfoGlobalBBIDs(const Function *F, 403 DenseMap<const BasicBlock*, unsigned> &IDMap) { 404 unsigned Counter = 0; 405 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 406 IDMap[BB] = ++Counter; 407} 408 409/// getGlobalBasicBlockID - This returns the function-specific ID for the 410/// specified basic block. This is relatively expensive information, so it 411/// should only be used by rare constructs such as address-of-label. 412unsigned ValueEnumerator::getGlobalBasicBlockID(const BasicBlock *BB) const { 413 unsigned &Idx = GlobalBasicBlockIDs[BB]; 414 if (Idx != 0) 415 return Idx-1; 416 417 IncorporateFunctionInfoGlobalBBIDs(BB->getParent(), GlobalBasicBlockIDs); 418 return getGlobalBasicBlockID(BB); 419} 420 421