ValueEnumerator.cpp revision 206124
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 V.first->getType()->isIntegerTy(); 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 // Enumerate the global variables. 43 for (Module::const_global_iterator I = M->global_begin(), 44 E = M->global_end(); I != E; ++I) 45 EnumerateValue(I); 46 47 // Enumerate the functions. 48 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I) { 49 EnumerateValue(I); 50 EnumerateAttributes(cast<Function>(I)->getAttributes()); 51 } 52 53 // Enumerate the aliases. 54 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end(); 55 I != E; ++I) 56 EnumerateValue(I); 57 58 // Remember what is the cutoff between globalvalue's and other constants. 59 unsigned FirstConstant = Values.size(); 60 61 // Enumerate the global variable initializers. 62 for (Module::const_global_iterator I = M->global_begin(), 63 E = M->global_end(); I != E; ++I) 64 if (I->hasInitializer()) 65 EnumerateValue(I->getInitializer()); 66 67 // Enumerate the aliasees. 68 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end(); 69 I != E; ++I) 70 EnumerateValue(I->getAliasee()); 71 72 // Enumerate types used by the type symbol table. 73 EnumerateTypeSymbolTable(M->getTypeSymbolTable()); 74 75 // Insert constants and metadata that are named at module level into the slot 76 // pool so that the module symbol table can refer to them... 77 EnumerateValueSymbolTable(M->getValueSymbolTable()); 78 EnumerateMDSymbolTable(M->getMDSymbolTable()); 79 80 SmallVector<std::pair<unsigned, MDNode*>, 8> MDs; 81 82 // Enumerate types used by function bodies and argument lists. 83 for (Module::const_iterator F = M->begin(), E = M->end(); F != E; ++F) { 84 85 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); 86 I != E; ++I) 87 EnumerateType(I->getType()); 88 89 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 90 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E;++I){ 91 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 92 OI != E; ++OI) { 93 if (MDNode *MD = dyn_cast<MDNode>(*OI)) 94 if (MD->isFunctionLocal() && MD->getFunction()) 95 // These will get enumerated during function-incorporation. 96 continue; 97 EnumerateOperandType(*OI); 98 } 99 EnumerateType(I->getType()); 100 if (const CallInst *CI = dyn_cast<CallInst>(I)) 101 EnumerateAttributes(CI->getAttributes()); 102 else if (const InvokeInst *II = dyn_cast<InvokeInst>(I)) 103 EnumerateAttributes(II->getAttributes()); 104 105 // Enumerate metadata attached with this instruction. 106 MDs.clear(); 107 I->getAllMetadataOtherThanDebugLoc(MDs); 108 for (unsigned i = 0, e = MDs.size(); i != e; ++i) 109 EnumerateMetadata(MDs[i].second); 110 111 if (!I->getDebugLoc().isUnknown()) { 112 MDNode *Scope, *IA; 113 I->getDebugLoc().getScopeAndInlinedAt(Scope, IA, I->getContext()); 114 if (Scope) EnumerateMetadata(Scope); 115 if (IA) EnumerateMetadata(IA); 116 } 117 } 118 } 119 120 // Optimize constant ordering. 121 OptimizeConstants(FirstConstant, Values.size()); 122 123 // Sort the type table by frequency so that most commonly used types are early 124 // in the table (have low bit-width). 125 std::stable_sort(Types.begin(), Types.end(), CompareByFrequency); 126 127 // Partition the Type ID's so that the single-value types occur before the 128 // aggregate types. This allows the aggregate types to be dropped from the 129 // type table after parsing the global variable initializers. 130 std::partition(Types.begin(), Types.end(), isSingleValueType); 131 132 // Now that we rearranged the type table, rebuild TypeMap. 133 for (unsigned i = 0, e = Types.size(); i != e; ++i) 134 TypeMap[Types[i].first] = i+1; 135} 136 137unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const { 138 InstructionMapType::const_iterator I = InstructionMap.find(Inst); 139 assert (I != InstructionMap.end() && "Instruction is not mapped!"); 140 return I->second; 141} 142 143void ValueEnumerator::setInstructionID(const Instruction *I) { 144 InstructionMap[I] = InstructionCount++; 145} 146 147unsigned ValueEnumerator::getValueID(const Value *V) const { 148 if (isa<MDNode>(V) || isa<MDString>(V)) { 149 ValueMapType::const_iterator I = MDValueMap.find(V); 150 assert(I != MDValueMap.end() && "Value not in slotcalculator!"); 151 return I->second-1; 152 } 153 154 ValueMapType::const_iterator I = ValueMap.find(V); 155 assert(I != ValueMap.end() && "Value not in slotcalculator!"); 156 return I->second-1; 157} 158 159// Optimize constant ordering. 160namespace { 161 struct CstSortPredicate { 162 ValueEnumerator &VE; 163 explicit CstSortPredicate(ValueEnumerator &ve) : VE(ve) {} 164 bool operator()(const std::pair<const Value*, unsigned> &LHS, 165 const std::pair<const Value*, unsigned> &RHS) { 166 // Sort by plane. 167 if (LHS.first->getType() != RHS.first->getType()) 168 return VE.getTypeID(LHS.first->getType()) < 169 VE.getTypeID(RHS.first->getType()); 170 // Then by frequency. 171 return LHS.second > RHS.second; 172 } 173 }; 174} 175 176/// OptimizeConstants - Reorder constant pool for denser encoding. 177void ValueEnumerator::OptimizeConstants(unsigned CstStart, unsigned CstEnd) { 178 if (CstStart == CstEnd || CstStart+1 == CstEnd) return; 179 180 CstSortPredicate P(*this); 181 std::stable_sort(Values.begin()+CstStart, Values.begin()+CstEnd, P); 182 183 // Ensure that integer constants are at the start of the constant pool. This 184 // is important so that GEP structure indices come before gep constant exprs. 185 std::partition(Values.begin()+CstStart, Values.begin()+CstEnd, 186 isIntegerValue); 187 188 // Rebuild the modified portion of ValueMap. 189 for (; CstStart != CstEnd; ++CstStart) 190 ValueMap[Values[CstStart].first] = CstStart+1; 191} 192 193 194/// EnumerateTypeSymbolTable - Insert all of the types in the specified symbol 195/// table. 196void ValueEnumerator::EnumerateTypeSymbolTable(const TypeSymbolTable &TST) { 197 for (TypeSymbolTable::const_iterator TI = TST.begin(), TE = TST.end(); 198 TI != TE; ++TI) 199 EnumerateType(TI->second); 200} 201 202/// EnumerateValueSymbolTable - Insert all of the values in the specified symbol 203/// table into the values table. 204void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) { 205 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end(); 206 VI != VE; ++VI) 207 EnumerateValue(VI->getValue()); 208} 209 210/// EnumerateMDSymbolTable - Insert all of the values in the specified metadata 211/// table. 212void ValueEnumerator::EnumerateMDSymbolTable(const MDSymbolTable &MST) { 213 for (MDSymbolTable::const_iterator MI = MST.begin(), ME = MST.end(); 214 MI != ME; ++MI) 215 EnumerateValue(MI->getValue()); 216} 217 218void ValueEnumerator::EnumerateNamedMDNode(const NamedMDNode *MD) { 219 // Check to see if it's already in! 220 unsigned &MDValueID = MDValueMap[MD]; 221 if (MDValueID) { 222 // Increment use count. 223 MDValues[MDValueID-1].second++; 224 return; 225 } 226 227 // Enumerate the type of this value. 228 EnumerateType(MD->getType()); 229 230 for (unsigned i = 0, e = MD->getNumOperands(); i != e; ++i) 231 if (MDNode *E = MD->getOperand(i)) 232 EnumerateValue(E); 233 MDValues.push_back(std::make_pair(MD, 1U)); 234 MDValueMap[MD] = Values.size(); 235} 236 237void ValueEnumerator::EnumerateMetadata(const Value *MD) { 238 assert((isa<MDNode>(MD) || isa<MDString>(MD)) && "Invalid metadata kind"); 239 // Check to see if it's already in! 240 unsigned &MDValueID = MDValueMap[MD]; 241 if (MDValueID) { 242 // Increment use count. 243 MDValues[MDValueID-1].second++; 244 return; 245 } 246 247 // Enumerate the type of this value. 248 EnumerateType(MD->getType()); 249 250 if (const MDNode *N = dyn_cast<MDNode>(MD)) { 251 MDValues.push_back(std::make_pair(MD, 1U)); 252 MDValueMap[MD] = MDValues.size(); 253 MDValueID = MDValues.size(); 254 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 255 if (Value *V = N->getOperand(i)) 256 EnumerateValue(V); 257 else 258 EnumerateType(Type::getVoidTy(MD->getContext())); 259 } 260 return; 261 } 262 263 // Add the value. 264 assert(isa<MDString>(MD) && "Unknown metadata kind"); 265 MDValues.push_back(std::make_pair(MD, 1U)); 266 MDValueID = MDValues.size(); 267} 268 269void ValueEnumerator::EnumerateValue(const Value *V) { 270 assert(!V->getType()->isVoidTy() && "Can't insert void values!"); 271 if (isa<MDNode>(V) || isa<MDString>(V)) 272 return EnumerateMetadata(V); 273 else if (const NamedMDNode *NMD = dyn_cast<NamedMDNode>(V)) 274 return EnumerateNamedMDNode(NMD); 275 276 // Check to see if it's already in! 277 unsigned &ValueID = ValueMap[V]; 278 if (ValueID) { 279 // Increment use count. 280 Values[ValueID-1].second++; 281 return; 282 } 283 284 // Enumerate the type of this value. 285 EnumerateType(V->getType()); 286 287 if (const Constant *C = dyn_cast<Constant>(V)) { 288 if (isa<GlobalValue>(C)) { 289 // Initializers for globals are handled explicitly elsewhere. 290 } else if (isa<ConstantArray>(C) && cast<ConstantArray>(C)->isString()) { 291 // Do not enumerate the initializers for an array of simple characters. 292 // The initializers just polute the value table, and we emit the strings 293 // specially. 294 } else if (C->getNumOperands()) { 295 // If a constant has operands, enumerate them. This makes sure that if a 296 // constant has uses (for example an array of const ints), that they are 297 // inserted also. 298 299 // We prefer to enumerate them with values before we enumerate the user 300 // itself. This makes it more likely that we can avoid forward references 301 // in the reader. We know that there can be no cycles in the constants 302 // graph that don't go through a global variable. 303 for (User::const_op_iterator I = C->op_begin(), E = C->op_end(); 304 I != E; ++I) 305 if (!isa<BasicBlock>(*I)) // Don't enumerate BB operand to BlockAddress. 306 EnumerateValue(*I); 307 308 // Finally, add the value. Doing this could make the ValueID reference be 309 // dangling, don't reuse it. 310 Values.push_back(std::make_pair(V, 1U)); 311 ValueMap[V] = Values.size(); 312 return; 313 } 314 } 315 316 // Add the value. 317 Values.push_back(std::make_pair(V, 1U)); 318 ValueID = Values.size(); 319} 320 321 322void ValueEnumerator::EnumerateType(const Type *Ty) { 323 unsigned &TypeID = TypeMap[Ty]; 324 325 if (TypeID) { 326 // If we've already seen this type, just increase its occurrence count. 327 Types[TypeID-1].second++; 328 return; 329 } 330 331 // First time we saw this type, add it. 332 Types.push_back(std::make_pair(Ty, 1U)); 333 TypeID = Types.size(); 334 335 // Enumerate subtypes. 336 for (Type::subtype_iterator I = Ty->subtype_begin(), E = Ty->subtype_end(); 337 I != E; ++I) 338 EnumerateType(*I); 339} 340 341// Enumerate the types for the specified value. If the value is a constant, 342// walk through it, enumerating the types of the constant. 343void ValueEnumerator::EnumerateOperandType(const Value *V) { 344 EnumerateType(V->getType()); 345 346 if (const Constant *C = dyn_cast<Constant>(V)) { 347 // If this constant is already enumerated, ignore it, we know its type must 348 // be enumerated. 349 if (ValueMap.count(V)) return; 350 351 // This constant may have operands, make sure to enumerate the types in 352 // them. 353 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) { 354 const User *Op = C->getOperand(i); 355 356 // Don't enumerate basic blocks here, this happens as operands to 357 // blockaddress. 358 if (isa<BasicBlock>(Op)) continue; 359 360 EnumerateOperandType(cast<Constant>(Op)); 361 } 362 363 if (const MDNode *N = dyn_cast<MDNode>(V)) { 364 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 365 if (Value *Elem = N->getOperand(i)) 366 EnumerateOperandType(Elem); 367 } 368 } else if (isa<MDString>(V) || isa<MDNode>(V)) 369 EnumerateValue(V); 370} 371 372void ValueEnumerator::EnumerateAttributes(const AttrListPtr &PAL) { 373 if (PAL.isEmpty()) return; // null is always 0. 374 // Do a lookup. 375 unsigned &Entry = AttributeMap[PAL.getRawPointer()]; 376 if (Entry == 0) { 377 // Never saw this before, add it. 378 Attributes.push_back(PAL); 379 Entry = Attributes.size(); 380 } 381} 382 383 384void ValueEnumerator::incorporateFunction(const Function &F) { 385 InstructionCount = 0; 386 NumModuleValues = Values.size(); 387 388 // Adding function arguments to the value table. 389 for(Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); 390 I != E; ++I) 391 EnumerateValue(I); 392 393 FirstFuncConstantID = Values.size(); 394 395 // Add all function-level constants to the value table. 396 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 397 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) 398 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 399 OI != E; ++OI) { 400 if ((isa<Constant>(*OI) && !isa<GlobalValue>(*OI)) || 401 isa<InlineAsm>(*OI)) 402 EnumerateValue(*OI); 403 } 404 BasicBlocks.push_back(BB); 405 ValueMap[BB] = BasicBlocks.size(); 406 } 407 408 // Optimize the constant layout. 409 OptimizeConstants(FirstFuncConstantID, Values.size()); 410 411 // Add the function's parameter attributes so they are available for use in 412 // the function's instruction. 413 EnumerateAttributes(F.getAttributes()); 414 415 FirstInstID = Values.size(); 416 417 SmallVector<MDNode *, 8> FunctionLocalMDs; 418 // Add all of the instructions. 419 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 420 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) { 421 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 422 OI != E; ++OI) { 423 if (MDNode *MD = dyn_cast<MDNode>(*OI)) 424 if (MD->isFunctionLocal() && MD->getFunction()) 425 // Enumerate metadata after the instructions they might refer to. 426 FunctionLocalMDs.push_back(MD); 427 } 428 if (!I->getType()->isVoidTy()) 429 EnumerateValue(I); 430 } 431 } 432 433 // Add all of the function-local metadata. 434 for (unsigned i = 0, e = FunctionLocalMDs.size(); i != e; ++i) 435 EnumerateOperandType(FunctionLocalMDs[i]); 436} 437 438void ValueEnumerator::purgeFunction() { 439 /// Remove purged values from the ValueMap. 440 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i) 441 ValueMap.erase(Values[i].first); 442 for (unsigned i = 0, e = BasicBlocks.size(); i != e; ++i) 443 ValueMap.erase(BasicBlocks[i]); 444 445 Values.resize(NumModuleValues); 446 BasicBlocks.clear(); 447} 448 449static void IncorporateFunctionInfoGlobalBBIDs(const Function *F, 450 DenseMap<const BasicBlock*, unsigned> &IDMap) { 451 unsigned Counter = 0; 452 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 453 IDMap[BB] = ++Counter; 454} 455 456/// getGlobalBasicBlockID - This returns the function-specific ID for the 457/// specified basic block. This is relatively expensive information, so it 458/// should only be used by rare constructs such as address-of-label. 459unsigned ValueEnumerator::getGlobalBasicBlockID(const BasicBlock *BB) const { 460 unsigned &Idx = GlobalBasicBlockIDs[BB]; 461 if (Idx != 0) 462 return Idx-1; 463 464 IncorporateFunctionInfoGlobalBBIDs(BB->getParent(), GlobalBasicBlockIDs); 465 return getGlobalBasicBlockID(BB); 466} 467 468