ValueEnumerator.cpp revision 203954
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 and metadata that are named at module level into the slot 78 // pool so that the module symbol table can refer to them... 79 EnumerateValueSymbolTable(M->getValueSymbolTable()); 80 EnumerateMDSymbolTable(M->getMDSymbolTable()); 81 82 SmallVector<std::pair<unsigned, MDNode*>, 8> MDs; 83 84 // Enumerate types used by function bodies and argument lists. 85 for (Module::const_iterator F = M->begin(), E = M->end(); F != E; ++F) { 86 87 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); 88 I != E; ++I) 89 EnumerateType(I->getType()); 90 91 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 92 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E;++I){ 93 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 94 OI != E; ++OI) { 95 if (MDNode *MD = dyn_cast<MDNode>(*OI)) 96 if (MD->isFunctionLocal() && MD->getFunction()) 97 // These will get enumerated during function-incorporation. 98 continue; 99 EnumerateOperandType(*OI); 100 } 101 EnumerateType(I->getType()); 102 if (const CallInst *CI = dyn_cast<CallInst>(I)) 103 EnumerateAttributes(CI->getAttributes()); 104 else if (const InvokeInst *II = dyn_cast<InvokeInst>(I)) 105 EnumerateAttributes(II->getAttributes()); 106 107 // Enumerate metadata attached with this instruction. 108 MDs.clear(); 109 I->getAllMetadata(MDs); 110 for (unsigned i = 0, e = MDs.size(); i != e; ++i) 111 EnumerateMetadata(MDs[i].second); 112 } 113 } 114 115 // Optimize constant ordering. 116 OptimizeConstants(FirstConstant, Values.size()); 117 118 // Sort the type table by frequency so that most commonly used types are early 119 // in the table (have low bit-width). 120 std::stable_sort(Types.begin(), Types.end(), CompareByFrequency); 121 122 // Partition the Type ID's so that the single-value types occur before the 123 // aggregate types. This allows the aggregate types to be dropped from the 124 // type table after parsing the global variable initializers. 125 std::partition(Types.begin(), Types.end(), isSingleValueType); 126 127 // Now that we rearranged the type table, rebuild TypeMap. 128 for (unsigned i = 0, e = Types.size(); i != e; ++i) 129 TypeMap[Types[i].first] = i+1; 130} 131 132unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const { 133 InstructionMapType::const_iterator I = InstructionMap.find(Inst); 134 assert (I != InstructionMap.end() && "Instruction is not mapped!"); 135 return I->second; 136} 137 138void ValueEnumerator::setInstructionID(const Instruction *I) { 139 InstructionMap[I] = InstructionCount++; 140} 141 142unsigned ValueEnumerator::getValueID(const Value *V) const { 143 if (isa<MDNode>(V) || isa<MDString>(V)) { 144 ValueMapType::const_iterator I = MDValueMap.find(V); 145 assert(I != MDValueMap.end() && "Value not in slotcalculator!"); 146 return I->second-1; 147 } 148 149 ValueMapType::const_iterator I = ValueMap.find(V); 150 assert(I != ValueMap.end() && "Value not in slotcalculator!"); 151 return I->second-1; 152} 153 154// Optimize constant ordering. 155namespace { 156 struct CstSortPredicate { 157 ValueEnumerator &VE; 158 explicit CstSortPredicate(ValueEnumerator &ve) : VE(ve) {} 159 bool operator()(const std::pair<const Value*, unsigned> &LHS, 160 const std::pair<const Value*, unsigned> &RHS) { 161 // Sort by plane. 162 if (LHS.first->getType() != RHS.first->getType()) 163 return VE.getTypeID(LHS.first->getType()) < 164 VE.getTypeID(RHS.first->getType()); 165 // Then by frequency. 166 return LHS.second > RHS.second; 167 } 168 }; 169} 170 171/// OptimizeConstants - Reorder constant pool for denser encoding. 172void ValueEnumerator::OptimizeConstants(unsigned CstStart, unsigned CstEnd) { 173 if (CstStart == CstEnd || CstStart+1 == CstEnd) return; 174 175 CstSortPredicate P(*this); 176 std::stable_sort(Values.begin()+CstStart, Values.begin()+CstEnd, P); 177 178 // Ensure that integer constants are at the start of the constant pool. This 179 // is important so that GEP structure indices come before gep constant exprs. 180 std::partition(Values.begin()+CstStart, Values.begin()+CstEnd, 181 isIntegerValue); 182 183 // Rebuild the modified portion of ValueMap. 184 for (; CstStart != CstEnd; ++CstStart) 185 ValueMap[Values[CstStart].first] = CstStart+1; 186} 187 188 189/// EnumerateTypeSymbolTable - Insert all of the types in the specified symbol 190/// table. 191void ValueEnumerator::EnumerateTypeSymbolTable(const TypeSymbolTable &TST) { 192 for (TypeSymbolTable::const_iterator TI = TST.begin(), TE = TST.end(); 193 TI != TE; ++TI) 194 EnumerateType(TI->second); 195} 196 197/// EnumerateValueSymbolTable - Insert all of the values in the specified symbol 198/// table into the values table. 199void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) { 200 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end(); 201 VI != VE; ++VI) 202 EnumerateValue(VI->getValue()); 203} 204 205/// EnumerateMDSymbolTable - Insert all of the values in the specified metadata 206/// table. 207void ValueEnumerator::EnumerateMDSymbolTable(const MDSymbolTable &MST) { 208 for (MDSymbolTable::const_iterator MI = MST.begin(), ME = MST.end(); 209 MI != ME; ++MI) 210 EnumerateValue(MI->getValue()); 211} 212 213void ValueEnumerator::EnumerateNamedMDNode(const NamedMDNode *MD) { 214 // Check to see if it's already in! 215 unsigned &MDValueID = MDValueMap[MD]; 216 if (MDValueID) { 217 // Increment use count. 218 MDValues[MDValueID-1].second++; 219 return; 220 } 221 222 // Enumerate the type of this value. 223 EnumerateType(MD->getType()); 224 225 for (unsigned i = 0, e = MD->getNumOperands(); i != e; ++i) 226 if (MDNode *E = MD->getOperand(i)) 227 EnumerateValue(E); 228 MDValues.push_back(std::make_pair(MD, 1U)); 229 MDValueMap[MD] = Values.size(); 230} 231 232void ValueEnumerator::EnumerateMetadata(const Value *MD) { 233 assert((isa<MDNode>(MD) || isa<MDString>(MD)) && "Invalid metadata kind"); 234 // Check to see if it's already in! 235 unsigned &MDValueID = MDValueMap[MD]; 236 if (MDValueID) { 237 // Increment use count. 238 MDValues[MDValueID-1].second++; 239 return; 240 } 241 242 // Enumerate the type of this value. 243 EnumerateType(MD->getType()); 244 245 if (const MDNode *N = dyn_cast<MDNode>(MD)) { 246 MDValues.push_back(std::make_pair(MD, 1U)); 247 MDValueMap[MD] = MDValues.size(); 248 MDValueID = MDValues.size(); 249 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 250 if (Value *V = N->getOperand(i)) 251 EnumerateValue(V); 252 else 253 EnumerateType(Type::getVoidTy(MD->getContext())); 254 } 255 return; 256 } 257 258 // Add the value. 259 assert(isa<MDString>(MD) && "Unknown metadata kind"); 260 MDValues.push_back(std::make_pair(MD, 1U)); 261 MDValueID = MDValues.size(); 262} 263 264void ValueEnumerator::EnumerateValue(const Value *V) { 265 assert(!V->getType()->isVoidTy() && "Can't insert void values!"); 266 if (isa<MDNode>(V) || isa<MDString>(V)) 267 return EnumerateMetadata(V); 268 else if (const NamedMDNode *NMD = dyn_cast<NamedMDNode>(V)) 269 return EnumerateNamedMDNode(NMD); 270 271 // Check to see if it's already in! 272 unsigned &ValueID = ValueMap[V]; 273 if (ValueID) { 274 // Increment use count. 275 Values[ValueID-1].second++; 276 return; 277 } 278 279 // Enumerate the type of this value. 280 EnumerateType(V->getType()); 281 282 if (const Constant *C = dyn_cast<Constant>(V)) { 283 if (isa<GlobalValue>(C)) { 284 // Initializers for globals are handled explicitly elsewhere. 285 } else if (isa<ConstantArray>(C) && cast<ConstantArray>(C)->isString()) { 286 // Do not enumerate the initializers for an array of simple characters. 287 // The initializers just polute the value table, and we emit the strings 288 // specially. 289 } else if (C->getNumOperands()) { 290 // If a constant has operands, enumerate them. This makes sure that if a 291 // constant has uses (for example an array of const ints), that they are 292 // inserted also. 293 294 // We prefer to enumerate them with values before we enumerate the user 295 // itself. This makes it more likely that we can avoid forward references 296 // in the reader. We know that there can be no cycles in the constants 297 // graph that don't go through a global variable. 298 for (User::const_op_iterator I = C->op_begin(), E = C->op_end(); 299 I != E; ++I) 300 if (!isa<BasicBlock>(*I)) // Don't enumerate BB operand to BlockAddress. 301 EnumerateValue(*I); 302 303 // Finally, add the value. Doing this could make the ValueID reference be 304 // dangling, don't reuse it. 305 Values.push_back(std::make_pair(V, 1U)); 306 ValueMap[V] = Values.size(); 307 return; 308 } 309 } 310 311 // Add the value. 312 Values.push_back(std::make_pair(V, 1U)); 313 ValueID = Values.size(); 314} 315 316 317void ValueEnumerator::EnumerateType(const Type *Ty) { 318 unsigned &TypeID = TypeMap[Ty]; 319 320 if (TypeID) { 321 // If we've already seen this type, just increase its occurrence count. 322 Types[TypeID-1].second++; 323 return; 324 } 325 326 // First time we saw this type, add it. 327 Types.push_back(std::make_pair(Ty, 1U)); 328 TypeID = Types.size(); 329 330 // Enumerate subtypes. 331 for (Type::subtype_iterator I = Ty->subtype_begin(), E = Ty->subtype_end(); 332 I != E; ++I) 333 EnumerateType(*I); 334} 335 336// Enumerate the types for the specified value. If the value is a constant, 337// walk through it, enumerating the types of the constant. 338void ValueEnumerator::EnumerateOperandType(const Value *V) { 339 EnumerateType(V->getType()); 340 341 if (const Constant *C = dyn_cast<Constant>(V)) { 342 // If this constant is already enumerated, ignore it, we know its type must 343 // be enumerated. 344 if (ValueMap.count(V)) return; 345 346 // This constant may have operands, make sure to enumerate the types in 347 // them. 348 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) { 349 const User *Op = C->getOperand(i); 350 351 // Don't enumerate basic blocks here, this happens as operands to 352 // blockaddress. 353 if (isa<BasicBlock>(Op)) continue; 354 355 EnumerateOperandType(cast<Constant>(Op)); 356 } 357 358 if (const MDNode *N = dyn_cast<MDNode>(V)) { 359 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 360 if (Value *Elem = N->getOperand(i)) 361 EnumerateOperandType(Elem); 362 } 363 } else if (isa<MDString>(V) || isa<MDNode>(V)) 364 EnumerateValue(V); 365} 366 367void ValueEnumerator::EnumerateAttributes(const AttrListPtr &PAL) { 368 if (PAL.isEmpty()) return; // null is always 0. 369 // Do a lookup. 370 unsigned &Entry = AttributeMap[PAL.getRawPointer()]; 371 if (Entry == 0) { 372 // Never saw this before, add it. 373 Attributes.push_back(PAL); 374 Entry = Attributes.size(); 375 } 376} 377 378 379void ValueEnumerator::incorporateFunction(const Function &F) { 380 NumModuleValues = Values.size(); 381 382 // Adding function arguments to the value table. 383 for(Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); 384 I != E; ++I) 385 EnumerateValue(I); 386 387 FirstFuncConstantID = Values.size(); 388 389 // Add all function-level constants to the value table. 390 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 391 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) 392 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 393 OI != E; ++OI) { 394 if ((isa<Constant>(*OI) && !isa<GlobalValue>(*OI)) || 395 isa<InlineAsm>(*OI)) 396 EnumerateValue(*OI); 397 } 398 BasicBlocks.push_back(BB); 399 ValueMap[BB] = BasicBlocks.size(); 400 } 401 402 // Optimize the constant layout. 403 OptimizeConstants(FirstFuncConstantID, Values.size()); 404 405 // Add the function's parameter attributes so they are available for use in 406 // the function's instruction. 407 EnumerateAttributes(F.getAttributes()); 408 409 FirstInstID = Values.size(); 410 411 SmallVector<MDNode *, 8> FunctionLocalMDs; 412 // Add all of the instructions. 413 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 414 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) { 415 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 416 OI != E; ++OI) { 417 if (MDNode *MD = dyn_cast<MDNode>(*OI)) 418 if (MD->isFunctionLocal() && MD->getFunction()) 419 // Enumerate metadata after the instructions they might refer to. 420 FunctionLocalMDs.push_back(MD); 421 } 422 if (!I->getType()->isVoidTy()) 423 EnumerateValue(I); 424 } 425 } 426 427 // Add all of the function-local metadata. 428 for (unsigned i = 0, e = FunctionLocalMDs.size(); i != e; ++i) 429 EnumerateOperandType(FunctionLocalMDs[i]); 430} 431 432void ValueEnumerator::purgeFunction() { 433 /// Remove purged values from the ValueMap. 434 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i) 435 ValueMap.erase(Values[i].first); 436 for (unsigned i = 0, e = BasicBlocks.size(); i != e; ++i) 437 ValueMap.erase(BasicBlocks[i]); 438 439 Values.resize(NumModuleValues); 440 BasicBlocks.clear(); 441} 442 443static void IncorporateFunctionInfoGlobalBBIDs(const Function *F, 444 DenseMap<const BasicBlock*, unsigned> &IDMap) { 445 unsigned Counter = 0; 446 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 447 IDMap[BB] = ++Counter; 448} 449 450/// getGlobalBasicBlockID - This returns the function-specific ID for the 451/// specified basic block. This is relatively expensive information, so it 452/// should only be used by rare constructs such as address-of-label. 453unsigned ValueEnumerator::getGlobalBasicBlockID(const BasicBlock *BB) const { 454 unsigned &Idx = GlobalBasicBlockIDs[BB]; 455 if (Idx != 0) 456 return Idx-1; 457 458 IncorporateFunctionInfoGlobalBBIDs(BB->getParent(), GlobalBasicBlockIDs); 459 return getGlobalBasicBlockID(BB); 460} 461 462