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/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallPtrSet.h" 17 #include "llvm/IR/Constants.h" 18 #include "llvm/IR/DerivedTypes.h" 19 #include "llvm/IR/Instructions.h" 20 #include "llvm/IR/Module.h" 21 #include "llvm/IR/ValueSymbolTable.h" 22 #include "llvm/Support/Debug.h" 23 #include "llvm/Support/raw_ostream.h" 24 #include <algorithm> 25 using namespace llvm; 26 27 static bool isIntOrIntVectorValue(const std::pair<const Value*, unsigned> &V) { 28 return V.first->getType()->isIntOrIntVectorTy(); 29 } 30 31 /// ValueEnumerator - Enumerate module-level information. 32 ValueEnumerator::ValueEnumerator(const Module *M) { 33 // Enumerate the global variables. 34 for (Module::const_global_iterator I = M->global_begin(), 35 E = M->global_end(); I != E; ++I) 36 EnumerateValue(I); 37 38 // Enumerate the functions. 39 for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I) { 40 EnumerateValue(I); 41 EnumerateAttributes(cast<Function>(I)->getAttributes()); 42 } 43 44 // Enumerate the aliases. 45 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end(); 46 I != E; ++I) 47 EnumerateValue(I); 48 49 // Remember what is the cutoff between globalvalue's and other constants. 50 unsigned FirstConstant = Values.size(); 51 52 // Enumerate the global variable initializers. 53 for (Module::const_global_iterator I = M->global_begin(), 54 E = M->global_end(); I != E; ++I) 55 if (I->hasInitializer()) 56 EnumerateValue(I->getInitializer()); 57 58 // Enumerate the aliasees. 59 for (Module::const_alias_iterator I = M->alias_begin(), E = M->alias_end(); 60 I != E; ++I) 61 EnumerateValue(I->getAliasee()); 62 63 // Insert constants and metadata that are named at module level into the slot 64 // pool so that the module symbol table can refer to them... 65 EnumerateValueSymbolTable(M->getValueSymbolTable()); 66 EnumerateNamedMetadata(M); 67 68 SmallVector<std::pair<unsigned, MDNode*>, 8> MDs; 69 70 // Enumerate types used by function bodies and argument lists. 71 for (Module::const_iterator F = M->begin(), E = M->end(); F != E; ++F) { 72 73 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); 74 I != E; ++I) 75 EnumerateType(I->getType()); 76 77 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 78 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E;++I){ 79 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 80 OI != E; ++OI) { 81 if (MDNode *MD = dyn_cast<MDNode>(*OI)) 82 if (MD->isFunctionLocal() && MD->getFunction()) 83 // These will get enumerated during function-incorporation. 84 continue; 85 EnumerateOperandType(*OI); 86 } 87 EnumerateType(I->getType()); 88 if (const CallInst *CI = dyn_cast<CallInst>(I)) 89 EnumerateAttributes(CI->getAttributes()); 90 else if (const InvokeInst *II = dyn_cast<InvokeInst>(I)) 91 EnumerateAttributes(II->getAttributes()); 92 93 // Enumerate metadata attached with this instruction. 94 MDs.clear(); 95 I->getAllMetadataOtherThanDebugLoc(MDs); 96 for (unsigned i = 0, e = MDs.size(); i != e; ++i) 97 EnumerateMetadata(MDs[i].second); 98 99 if (!I->getDebugLoc().isUnknown()) { 100 MDNode *Scope, *IA; 101 I->getDebugLoc().getScopeAndInlinedAt(Scope, IA, I->getContext()); 102 if (Scope) EnumerateMetadata(Scope); 103 if (IA) EnumerateMetadata(IA); 104 } 105 } 106 } 107 108 // Optimize constant ordering. 109 OptimizeConstants(FirstConstant, Values.size()); 110 } 111 112 unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const { 113 InstructionMapType::const_iterator I = InstructionMap.find(Inst); 114 assert(I != InstructionMap.end() && "Instruction is not mapped!"); 115 return I->second; 116 } 117 118 void ValueEnumerator::setInstructionID(const Instruction *I) { 119 InstructionMap[I] = InstructionCount++; 120 } 121 122 unsigned ValueEnumerator::getValueID(const Value *V) const { 123 if (isa<MDNode>(V) || isa<MDString>(V)) { 124 ValueMapType::const_iterator I = MDValueMap.find(V); 125 assert(I != MDValueMap.end() && "Value not in slotcalculator!"); 126 return I->second-1; 127 } 128 129 ValueMapType::const_iterator I = ValueMap.find(V); 130 assert(I != ValueMap.end() && "Value not in slotcalculator!"); 131 return I->second-1; 132 } 133 134 void ValueEnumerator::dump() const { 135 print(dbgs(), ValueMap, "Default"); 136 dbgs() << '\n'; 137 print(dbgs(), MDValueMap, "MetaData"); 138 dbgs() << '\n'; 139 } 140 141 void ValueEnumerator::print(raw_ostream &OS, const ValueMapType &Map, 142 const char *Name) const { 143 144 OS << "Map Name: " << Name << "\n"; 145 OS << "Size: " << Map.size() << "\n"; 146 for (ValueMapType::const_iterator I = Map.begin(), 147 E = Map.end(); I != E; ++I) { 148 149 const Value *V = I->first; 150 if (V->hasName()) 151 OS << "Value: " << V->getName(); 152 else 153 OS << "Value: [null]\n"; 154 V->dump(); 155 156 OS << " Uses(" << std::distance(V->use_begin(),V->use_end()) << "):"; 157 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end(); 158 UI != UE; ++UI) { 159 if (UI != V->use_begin()) 160 OS << ","; 161 if((*UI)->hasName()) 162 OS << " " << (*UI)->getName(); 163 else 164 OS << " [null]"; 165 166 } 167 OS << "\n\n"; 168 } 169 } 170 171 // Optimize constant ordering. 172 namespace { 173 struct CstSortPredicate { 174 ValueEnumerator &VE; 175 explicit CstSortPredicate(ValueEnumerator &ve) : VE(ve) {} 176 bool operator()(const std::pair<const Value*, unsigned> &LHS, 177 const std::pair<const Value*, unsigned> &RHS) { 178 // Sort by plane. 179 if (LHS.first->getType() != RHS.first->getType()) 180 return VE.getTypeID(LHS.first->getType()) < 181 VE.getTypeID(RHS.first->getType()); 182 // Then by frequency. 183 return LHS.second > RHS.second; 184 } 185 }; 186 } 187 188 /// OptimizeConstants - Reorder constant pool for denser encoding. 189 void ValueEnumerator::OptimizeConstants(unsigned CstStart, unsigned CstEnd) { 190 if (CstStart == CstEnd || CstStart+1 == CstEnd) return; 191 192 CstSortPredicate P(*this); 193 std::stable_sort(Values.begin()+CstStart, Values.begin()+CstEnd, P); 194 195 // Ensure that integer and vector of integer constants are at the start of the 196 // constant pool. This is important so that GEP structure indices come before 197 // gep constant exprs. 198 std::partition(Values.begin()+CstStart, Values.begin()+CstEnd, 199 isIntOrIntVectorValue); 200 201 // Rebuild the modified portion of ValueMap. 202 for (; CstStart != CstEnd; ++CstStart) 203 ValueMap[Values[CstStart].first] = CstStart+1; 204 } 205 206 207 /// EnumerateValueSymbolTable - Insert all of the values in the specified symbol 208 /// table into the values table. 209 void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) { 210 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end(); 211 VI != VE; ++VI) 212 EnumerateValue(VI->getValue()); 213 } 214 215 /// EnumerateNamedMetadata - Insert all of the values referenced by 216 /// named metadata in the specified module. 217 void ValueEnumerator::EnumerateNamedMetadata(const Module *M) { 218 for (Module::const_named_metadata_iterator I = M->named_metadata_begin(), 219 E = M->named_metadata_end(); I != E; ++I) 220 EnumerateNamedMDNode(I); 221 } 222 223 void ValueEnumerator::EnumerateNamedMDNode(const NamedMDNode *MD) { 224 for (unsigned i = 0, e = MD->getNumOperands(); i != e; ++i) 225 EnumerateMetadata(MD->getOperand(i)); 226 } 227 228 /// EnumerateMDNodeOperands - Enumerate all non-function-local values 229 /// and types referenced by the given MDNode. 230 void ValueEnumerator::EnumerateMDNodeOperands(const MDNode *N) { 231 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 232 if (Value *V = N->getOperand(i)) { 233 if (isa<MDNode>(V) || isa<MDString>(V)) 234 EnumerateMetadata(V); 235 else if (!isa<Instruction>(V) && !isa<Argument>(V)) 236 EnumerateValue(V); 237 } else 238 EnumerateType(Type::getVoidTy(N->getContext())); 239 } 240 } 241 242 void ValueEnumerator::EnumerateMetadata(const Value *MD) { 243 assert((isa<MDNode>(MD) || isa<MDString>(MD)) && "Invalid metadata kind"); 244 245 // Enumerate the type of this value. 246 EnumerateType(MD->getType()); 247 248 const MDNode *N = dyn_cast<MDNode>(MD); 249 250 // In the module-level pass, skip function-local nodes themselves, but 251 // do walk their operands. 252 if (N && N->isFunctionLocal() && N->getFunction()) { 253 EnumerateMDNodeOperands(N); 254 return; 255 } 256 257 // Check to see if it's already in! 258 unsigned &MDValueID = MDValueMap[MD]; 259 if (MDValueID) { 260 // Increment use count. 261 MDValues[MDValueID-1].second++; 262 return; 263 } 264 MDValues.push_back(std::make_pair(MD, 1U)); 265 MDValueID = MDValues.size(); 266 267 // Enumerate all non-function-local operands. 268 if (N) 269 EnumerateMDNodeOperands(N); 270 } 271 272 /// EnumerateFunctionLocalMetadataa - Incorporate function-local metadata 273 /// information reachable from the given MDNode. 274 void ValueEnumerator::EnumerateFunctionLocalMetadata(const MDNode *N) { 275 assert(N->isFunctionLocal() && N->getFunction() && 276 "EnumerateFunctionLocalMetadata called on non-function-local mdnode!"); 277 278 // Enumerate the type of this value. 279 EnumerateType(N->getType()); 280 281 // Check to see if it's already in! 282 unsigned &MDValueID = MDValueMap[N]; 283 if (MDValueID) { 284 // Increment use count. 285 MDValues[MDValueID-1].second++; 286 return; 287 } 288 MDValues.push_back(std::make_pair(N, 1U)); 289 MDValueID = MDValues.size(); 290 291 // To incoroporate function-local information visit all function-local 292 // MDNodes and all function-local values they reference. 293 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 294 if (Value *V = N->getOperand(i)) { 295 if (MDNode *O = dyn_cast<MDNode>(V)) { 296 if (O->isFunctionLocal() && O->getFunction()) 297 EnumerateFunctionLocalMetadata(O); 298 } else if (isa<Instruction>(V) || isa<Argument>(V)) 299 EnumerateValue(V); 300 } 301 302 // Also, collect all function-local MDNodes for easy access. 303 FunctionLocalMDs.push_back(N); 304 } 305 306 void ValueEnumerator::EnumerateValue(const Value *V) { 307 assert(!V->getType()->isVoidTy() && "Can't insert void values!"); 308 assert(!isa<MDNode>(V) && !isa<MDString>(V) && 309 "EnumerateValue doesn't handle Metadata!"); 310 311 // Check to see if it's already in! 312 unsigned &ValueID = ValueMap[V]; 313 if (ValueID) { 314 // Increment use count. 315 Values[ValueID-1].second++; 316 return; 317 } 318 319 // Enumerate the type of this value. 320 EnumerateType(V->getType()); 321 322 if (const Constant *C = dyn_cast<Constant>(V)) { 323 if (isa<GlobalValue>(C)) { 324 // Initializers for globals are handled explicitly elsewhere. 325 } else if (C->getNumOperands()) { 326 // If a constant has operands, enumerate them. This makes sure that if a 327 // constant has uses (for example an array of const ints), that they are 328 // inserted also. 329 330 // We prefer to enumerate them with values before we enumerate the user 331 // itself. This makes it more likely that we can avoid forward references 332 // in the reader. We know that there can be no cycles in the constants 333 // graph that don't go through a global variable. 334 for (User::const_op_iterator I = C->op_begin(), E = C->op_end(); 335 I != E; ++I) 336 if (!isa<BasicBlock>(*I)) // Don't enumerate BB operand to BlockAddress. 337 EnumerateValue(*I); 338 339 // Finally, add the value. Doing this could make the ValueID reference be 340 // dangling, don't reuse it. 341 Values.push_back(std::make_pair(V, 1U)); 342 ValueMap[V] = Values.size(); 343 return; 344 } 345 } 346 347 // Add the value. 348 Values.push_back(std::make_pair(V, 1U)); 349 ValueID = Values.size(); 350 } 351 352 353 void ValueEnumerator::EnumerateType(Type *Ty) { 354 unsigned *TypeID = &TypeMap[Ty]; 355 356 // We've already seen this type. 357 if (*TypeID) 358 return; 359 360 // If it is a non-anonymous struct, mark the type as being visited so that we 361 // don't recursively visit it. This is safe because we allow forward 362 // references of these in the bitcode reader. 363 if (StructType *STy = dyn_cast<StructType>(Ty)) 364 if (!STy->isLiteral()) 365 *TypeID = ~0U; 366 367 // Enumerate all of the subtypes before we enumerate this type. This ensures 368 // that the type will be enumerated in an order that can be directly built. 369 for (Type::subtype_iterator I = Ty->subtype_begin(), E = Ty->subtype_end(); 370 I != E; ++I) 371 EnumerateType(*I); 372 373 // Refresh the TypeID pointer in case the table rehashed. 374 TypeID = &TypeMap[Ty]; 375 376 // Check to see if we got the pointer another way. This can happen when 377 // enumerating recursive types that hit the base case deeper than they start. 378 // 379 // If this is actually a struct that we are treating as forward ref'able, 380 // then emit the definition now that all of its contents are available. 381 if (*TypeID && *TypeID != ~0U) 382 return; 383 384 // Add this type now that its contents are all happily enumerated. 385 Types.push_back(Ty); 386 387 *TypeID = Types.size(); 388 } 389 390 // Enumerate the types for the specified value. If the value is a constant, 391 // walk through it, enumerating the types of the constant. 392 void ValueEnumerator::EnumerateOperandType(const Value *V) { 393 EnumerateType(V->getType()); 394 395 if (const Constant *C = dyn_cast<Constant>(V)) { 396 // If this constant is already enumerated, ignore it, we know its type must 397 // be enumerated. 398 if (ValueMap.count(V)) return; 399 400 // This constant may have operands, make sure to enumerate the types in 401 // them. 402 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) { 403 const Value *Op = C->getOperand(i); 404 405 // Don't enumerate basic blocks here, this happens as operands to 406 // blockaddress. 407 if (isa<BasicBlock>(Op)) continue; 408 409 EnumerateOperandType(Op); 410 } 411 412 if (const MDNode *N = dyn_cast<MDNode>(V)) { 413 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 414 if (Value *Elem = N->getOperand(i)) 415 EnumerateOperandType(Elem); 416 } 417 } else if (isa<MDString>(V) || isa<MDNode>(V)) 418 EnumerateMetadata(V); 419 } 420 421 void ValueEnumerator::EnumerateAttributes(AttributeSet PAL) { 422 if (PAL.isEmpty()) return; // null is always 0. 423 424 // Do a lookup. 425 unsigned &Entry = AttributeMap[PAL]; 426 if (Entry == 0) { 427 // Never saw this before, add it. 428 Attribute.push_back(PAL); 429 Entry = Attribute.size(); 430 } 431 432 // Do lookups for all attribute groups. 433 for (unsigned i = 0, e = PAL.getNumSlots(); i != e; ++i) { 434 AttributeSet AS = PAL.getSlotAttributes(i); 435 unsigned &Entry = AttributeGroupMap[AS]; 436 if (Entry == 0) { 437 AttributeGroups.push_back(AS); 438 Entry = AttributeGroups.size(); 439 } 440 } 441 } 442 443 void ValueEnumerator::incorporateFunction(const Function &F) { 444 InstructionCount = 0; 445 NumModuleValues = Values.size(); 446 NumModuleMDValues = MDValues.size(); 447 448 // Adding function arguments to the value table. 449 for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end(); 450 I != E; ++I) 451 EnumerateValue(I); 452 453 FirstFuncConstantID = Values.size(); 454 455 // Add all function-level constants to the value table. 456 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 457 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) 458 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 459 OI != E; ++OI) { 460 if ((isa<Constant>(*OI) && !isa<GlobalValue>(*OI)) || 461 isa<InlineAsm>(*OI)) 462 EnumerateValue(*OI); 463 } 464 BasicBlocks.push_back(BB); 465 ValueMap[BB] = BasicBlocks.size(); 466 } 467 468 // Optimize the constant layout. 469 OptimizeConstants(FirstFuncConstantID, Values.size()); 470 471 // Add the function's parameter attributes so they are available for use in 472 // the function's instruction. 473 EnumerateAttributes(F.getAttributes()); 474 475 FirstInstID = Values.size(); 476 477 SmallVector<MDNode *, 8> FnLocalMDVector; 478 // Add all of the instructions. 479 for (Function::const_iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 480 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E; ++I) { 481 for (User::const_op_iterator OI = I->op_begin(), E = I->op_end(); 482 OI != E; ++OI) { 483 if (MDNode *MD = dyn_cast<MDNode>(*OI)) 484 if (MD->isFunctionLocal() && MD->getFunction()) 485 // Enumerate metadata after the instructions they might refer to. 486 FnLocalMDVector.push_back(MD); 487 } 488 489 SmallVector<std::pair<unsigned, MDNode*>, 8> MDs; 490 I->getAllMetadataOtherThanDebugLoc(MDs); 491 for (unsigned i = 0, e = MDs.size(); i != e; ++i) { 492 MDNode *N = MDs[i].second; 493 if (N->isFunctionLocal() && N->getFunction()) 494 FnLocalMDVector.push_back(N); 495 } 496 497 if (!I->getType()->isVoidTy()) 498 EnumerateValue(I); 499 } 500 } 501 502 // Add all of the function-local metadata. 503 for (unsigned i = 0, e = FnLocalMDVector.size(); i != e; ++i) 504 EnumerateFunctionLocalMetadata(FnLocalMDVector[i]); 505 } 506 507 void ValueEnumerator::purgeFunction() { 508 /// Remove purged values from the ValueMap. 509 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i) 510 ValueMap.erase(Values[i].first); 511 for (unsigned i = NumModuleMDValues, e = MDValues.size(); i != e; ++i) 512 MDValueMap.erase(MDValues[i].first); 513 for (unsigned i = 0, e = BasicBlocks.size(); i != e; ++i) 514 ValueMap.erase(BasicBlocks[i]); 515 516 Values.resize(NumModuleValues); 517 MDValues.resize(NumModuleMDValues); 518 BasicBlocks.clear(); 519 FunctionLocalMDs.clear(); 520 } 521 522 static void IncorporateFunctionInfoGlobalBBIDs(const Function *F, 523 DenseMap<const BasicBlock*, unsigned> &IDMap) { 524 unsigned Counter = 0; 525 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 526 IDMap[BB] = ++Counter; 527 } 528 529 /// getGlobalBasicBlockID - This returns the function-specific ID for the 530 /// specified basic block. This is relatively expensive information, so it 531 /// should only be used by rare constructs such as address-of-label. 532 unsigned ValueEnumerator::getGlobalBasicBlockID(const BasicBlock *BB) const { 533 unsigned &Idx = GlobalBasicBlockIDs[BB]; 534 if (Idx != 0) 535 return Idx-1; 536 537 IncorporateFunctionInfoGlobalBBIDs(BB->getParent(), GlobalBasicBlockIDs); 538 return getGlobalBasicBlockID(BB); 539 } 540 541