1 //===- ValueEnumerator.cpp - Number values and types for bitcode writer ---===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements the ValueEnumerator class. 10 // Forked from lib/Bitcode/Writer 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "DXILValueEnumerator.h" 15 #include "DXILPointerType.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/Config/llvm-config.h" 18 #include "llvm/IR/Argument.h" 19 #include "llvm/IR/BasicBlock.h" 20 #include "llvm/IR/Constant.h" 21 #include "llvm/IR/DebugInfoMetadata.h" 22 #include "llvm/IR/DerivedTypes.h" 23 #include "llvm/IR/Function.h" 24 #include "llvm/IR/GlobalAlias.h" 25 #include "llvm/IR/GlobalIFunc.h" 26 #include "llvm/IR/GlobalObject.h" 27 #include "llvm/IR/GlobalValue.h" 28 #include "llvm/IR/GlobalVariable.h" 29 #include "llvm/IR/Instruction.h" 30 #include "llvm/IR/Instructions.h" 31 #include "llvm/IR/Metadata.h" 32 #include "llvm/IR/Module.h" 33 #include "llvm/IR/Operator.h" 34 #include "llvm/IR/Type.h" 35 #include "llvm/IR/Use.h" 36 #include "llvm/IR/User.h" 37 #include "llvm/IR/Value.h" 38 #include "llvm/IR/ValueSymbolTable.h" 39 #include "llvm/Support/Casting.h" 40 #include "llvm/Support/Compiler.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/MathExtras.h" 43 #include "llvm/Support/raw_ostream.h" 44 #include <algorithm> 45 #include <cstddef> 46 #include <iterator> 47 #include <tuple> 48 49 using namespace llvm; 50 using namespace llvm::dxil; 51 52 namespace { 53 54 struct OrderMap { 55 DenseMap<const Value *, std::pair<unsigned, bool>> IDs; 56 unsigned LastGlobalConstantID = 0; 57 unsigned LastGlobalValueID = 0; 58 59 OrderMap() = default; 60 61 bool isGlobalConstant(unsigned ID) const { 62 return ID <= LastGlobalConstantID; 63 } 64 65 bool isGlobalValue(unsigned ID) const { 66 return ID <= LastGlobalValueID && !isGlobalConstant(ID); 67 } 68 69 unsigned size() const { return IDs.size(); } 70 std::pair<unsigned, bool> &operator[](const Value *V) { return IDs[V]; } 71 72 std::pair<unsigned, bool> lookup(const Value *V) const { 73 return IDs.lookup(V); 74 } 75 76 void index(const Value *V) { 77 // Explicitly sequence get-size and insert-value operations to avoid UB. 78 unsigned ID = IDs.size() + 1; 79 IDs[V].first = ID; 80 } 81 }; 82 83 } // end anonymous namespace 84 85 static void orderValue(const Value *V, OrderMap &OM) { 86 if (OM.lookup(V).first) 87 return; 88 89 if (const Constant *C = dyn_cast<Constant>(V)) { 90 if (C->getNumOperands() && !isa<GlobalValue>(C)) { 91 for (const Value *Op : C->operands()) 92 if (!isa<BasicBlock>(Op) && !isa<GlobalValue>(Op)) 93 orderValue(Op, OM); 94 if (auto *CE = dyn_cast<ConstantExpr>(C)) 95 if (CE->getOpcode() == Instruction::ShuffleVector) 96 orderValue(CE->getShuffleMaskForBitcode(), OM); 97 } 98 } 99 100 // Note: we cannot cache this lookup above, since inserting into the map 101 // changes the map's size, and thus affects the other IDs. 102 OM.index(V); 103 } 104 105 static OrderMap orderModule(const Module &M) { 106 // This needs to match the order used by ValueEnumerator::ValueEnumerator() 107 // and ValueEnumerator::incorporateFunction(). 108 OrderMap OM; 109 110 // In the reader, initializers of GlobalValues are set *after* all the 111 // globals have been read. Rather than awkwardly modeling this behaviour 112 // directly in predictValueUseListOrderImpl(), just assign IDs to 113 // initializers of GlobalValues before GlobalValues themselves to model this 114 // implicitly. 115 for (const GlobalVariable &G : M.globals()) 116 if (G.hasInitializer()) 117 if (!isa<GlobalValue>(G.getInitializer())) 118 orderValue(G.getInitializer(), OM); 119 for (const GlobalAlias &A : M.aliases()) 120 if (!isa<GlobalValue>(A.getAliasee())) 121 orderValue(A.getAliasee(), OM); 122 for (const GlobalIFunc &I : M.ifuncs()) 123 if (!isa<GlobalValue>(I.getResolver())) 124 orderValue(I.getResolver(), OM); 125 for (const Function &F : M) { 126 for (const Use &U : F.operands()) 127 if (!isa<GlobalValue>(U.get())) 128 orderValue(U.get(), OM); 129 } 130 131 // As constants used in metadata operands are emitted as module-level 132 // constants, we must order them before other operands. Also, we must order 133 // these before global values, as these will be read before setting the 134 // global values' initializers. The latter matters for constants which have 135 // uses towards other constants that are used as initializers. 136 auto orderConstantValue = [&OM](const Value *V) { 137 if ((isa<Constant>(V) && !isa<GlobalValue>(V)) || isa<InlineAsm>(V)) 138 orderValue(V, OM); 139 }; 140 for (const Function &F : M) { 141 if (F.isDeclaration()) 142 continue; 143 for (const BasicBlock &BB : F) 144 for (const Instruction &I : BB) 145 for (const Value *V : I.operands()) { 146 if (const auto *MAV = dyn_cast<MetadataAsValue>(V)) { 147 if (const auto *VAM = 148 dyn_cast<ValueAsMetadata>(MAV->getMetadata())) { 149 orderConstantValue(VAM->getValue()); 150 } else if (const auto *AL = 151 dyn_cast<DIArgList>(MAV->getMetadata())) { 152 for (const auto *VAM : AL->getArgs()) 153 orderConstantValue(VAM->getValue()); 154 } 155 } 156 } 157 } 158 OM.LastGlobalConstantID = OM.size(); 159 160 // Initializers of GlobalValues are processed in 161 // BitcodeReader::ResolveGlobalAndAliasInits(). Match the order there rather 162 // than ValueEnumerator, and match the code in predictValueUseListOrderImpl() 163 // by giving IDs in reverse order. 164 // 165 // Since GlobalValues never reference each other directly (just through 166 // initializers), their relative IDs only matter for determining order of 167 // uses in their initializers. 168 for (const Function &F : M) 169 orderValue(&F, OM); 170 for (const GlobalAlias &A : M.aliases()) 171 orderValue(&A, OM); 172 for (const GlobalIFunc &I : M.ifuncs()) 173 orderValue(&I, OM); 174 for (const GlobalVariable &G : M.globals()) 175 orderValue(&G, OM); 176 OM.LastGlobalValueID = OM.size(); 177 178 for (const Function &F : M) { 179 if (F.isDeclaration()) 180 continue; 181 // Here we need to match the union of ValueEnumerator::incorporateFunction() 182 // and WriteFunction(). Basic blocks are implicitly declared before 183 // anything else (by declaring their size). 184 for (const BasicBlock &BB : F) 185 orderValue(&BB, OM); 186 for (const Argument &A : F.args()) 187 orderValue(&A, OM); 188 for (const BasicBlock &BB : F) 189 for (const Instruction &I : BB) { 190 for (const Value *Op : I.operands()) 191 if ((isa<Constant>(*Op) && !isa<GlobalValue>(*Op)) || 192 isa<InlineAsm>(*Op)) 193 orderValue(Op, OM); 194 if (auto *SVI = dyn_cast<ShuffleVectorInst>(&I)) 195 orderValue(SVI->getShuffleMaskForBitcode(), OM); 196 } 197 for (const BasicBlock &BB : F) 198 for (const Instruction &I : BB) 199 orderValue(&I, OM); 200 } 201 return OM; 202 } 203 204 static void predictValueUseListOrderImpl(const Value *V, const Function *F, 205 unsigned ID, const OrderMap &OM, 206 UseListOrderStack &Stack) { 207 // Predict use-list order for this one. 208 using Entry = std::pair<const Use *, unsigned>; 209 SmallVector<Entry, 64> List; 210 for (const Use &U : V->uses()) 211 // Check if this user will be serialized. 212 if (OM.lookup(U.getUser()).first) 213 List.push_back(std::make_pair(&U, List.size())); 214 215 if (List.size() < 2) 216 // We may have lost some users. 217 return; 218 219 bool IsGlobalValue = OM.isGlobalValue(ID); 220 llvm::sort(List, [&](const Entry &L, const Entry &R) { 221 const Use *LU = L.first; 222 const Use *RU = R.first; 223 if (LU == RU) 224 return false; 225 226 auto LID = OM.lookup(LU->getUser()).first; 227 auto RID = OM.lookup(RU->getUser()).first; 228 229 // Global values are processed in reverse order. 230 // 231 // Moreover, initializers of GlobalValues are set *after* all the globals 232 // have been read (despite having earlier IDs). Rather than awkwardly 233 // modeling this behaviour here, orderModule() has assigned IDs to 234 // initializers of GlobalValues before GlobalValues themselves. 235 if (OM.isGlobalValue(LID) && OM.isGlobalValue(RID)) { 236 if (LID == RID) 237 return LU->getOperandNo() > RU->getOperandNo(); 238 return LID < RID; 239 } 240 241 // If ID is 4, then expect: 7 6 5 1 2 3. 242 if (LID < RID) { 243 if (RID <= ID) 244 if (!IsGlobalValue) // GlobalValue uses don't get reversed. 245 return true; 246 return false; 247 } 248 if (RID < LID) { 249 if (LID <= ID) 250 if (!IsGlobalValue) // GlobalValue uses don't get reversed. 251 return false; 252 return true; 253 } 254 255 // LID and RID are equal, so we have different operands of the same user. 256 // Assume operands are added in order for all instructions. 257 if (LID <= ID) 258 if (!IsGlobalValue) // GlobalValue uses don't get reversed. 259 return LU->getOperandNo() < RU->getOperandNo(); 260 return LU->getOperandNo() > RU->getOperandNo(); 261 }); 262 263 if (llvm::is_sorted(List, [](const Entry &L, const Entry &R) { 264 return L.second < R.second; 265 })) 266 // Order is already correct. 267 return; 268 269 // Store the shuffle. 270 Stack.emplace_back(V, F, List.size()); 271 assert(List.size() == Stack.back().Shuffle.size() && "Wrong size"); 272 for (size_t I = 0, E = List.size(); I != E; ++I) 273 Stack.back().Shuffle[I] = List[I].second; 274 } 275 276 static void predictValueUseListOrder(const Value *V, const Function *F, 277 OrderMap &OM, UseListOrderStack &Stack) { 278 auto &IDPair = OM[V]; 279 assert(IDPair.first && "Unmapped value"); 280 if (IDPair.second) 281 // Already predicted. 282 return; 283 284 // Do the actual prediction. 285 IDPair.second = true; 286 if (!V->use_empty() && std::next(V->use_begin()) != V->use_end()) 287 predictValueUseListOrderImpl(V, F, IDPair.first, OM, Stack); 288 289 // Recursive descent into constants. 290 if (const Constant *C = dyn_cast<Constant>(V)) { 291 if (C->getNumOperands()) { // Visit GlobalValues. 292 for (const Value *Op : C->operands()) 293 if (isa<Constant>(Op)) // Visit GlobalValues. 294 predictValueUseListOrder(Op, F, OM, Stack); 295 if (auto *CE = dyn_cast<ConstantExpr>(C)) 296 if (CE->getOpcode() == Instruction::ShuffleVector) 297 predictValueUseListOrder(CE->getShuffleMaskForBitcode(), F, OM, 298 Stack); 299 } 300 } 301 } 302 303 static UseListOrderStack predictUseListOrder(const Module &M) { 304 OrderMap OM = orderModule(M); 305 306 // Use-list orders need to be serialized after all the users have been added 307 // to a value, or else the shuffles will be incomplete. Store them per 308 // function in a stack. 309 // 310 // Aside from function order, the order of values doesn't matter much here. 311 UseListOrderStack Stack; 312 313 // We want to visit the functions backward now so we can list function-local 314 // constants in the last Function they're used in. Module-level constants 315 // have already been visited above. 316 for (const Function &F : llvm::reverse(M)) { 317 if (F.isDeclaration()) 318 continue; 319 for (const BasicBlock &BB : F) 320 predictValueUseListOrder(&BB, &F, OM, Stack); 321 for (const Argument &A : F.args()) 322 predictValueUseListOrder(&A, &F, OM, Stack); 323 for (const BasicBlock &BB : F) 324 for (const Instruction &I : BB) { 325 for (const Value *Op : I.operands()) 326 if (isa<Constant>(*Op) || isa<InlineAsm>(*Op)) // Visit GlobalValues. 327 predictValueUseListOrder(Op, &F, OM, Stack); 328 if (auto *SVI = dyn_cast<ShuffleVectorInst>(&I)) 329 predictValueUseListOrder(SVI->getShuffleMaskForBitcode(), &F, OM, 330 Stack); 331 } 332 for (const BasicBlock &BB : F) 333 for (const Instruction &I : BB) 334 predictValueUseListOrder(&I, &F, OM, Stack); 335 } 336 337 // Visit globals last, since the module-level use-list block will be seen 338 // before the function bodies are processed. 339 for (const GlobalVariable &G : M.globals()) 340 predictValueUseListOrder(&G, nullptr, OM, Stack); 341 for (const Function &F : M) 342 predictValueUseListOrder(&F, nullptr, OM, Stack); 343 for (const GlobalAlias &A : M.aliases()) 344 predictValueUseListOrder(&A, nullptr, OM, Stack); 345 for (const GlobalIFunc &I : M.ifuncs()) 346 predictValueUseListOrder(&I, nullptr, OM, Stack); 347 for (const GlobalVariable &G : M.globals()) 348 if (G.hasInitializer()) 349 predictValueUseListOrder(G.getInitializer(), nullptr, OM, Stack); 350 for (const GlobalAlias &A : M.aliases()) 351 predictValueUseListOrder(A.getAliasee(), nullptr, OM, Stack); 352 for (const GlobalIFunc &I : M.ifuncs()) 353 predictValueUseListOrder(I.getResolver(), nullptr, OM, Stack); 354 for (const Function &F : M) { 355 for (const Use &U : F.operands()) 356 predictValueUseListOrder(U.get(), nullptr, OM, Stack); 357 } 358 359 return Stack; 360 } 361 362 ValueEnumerator::ValueEnumerator(const Module &M, Type *PrefixType) { 363 EnumerateType(PrefixType); 364 365 UseListOrders = predictUseListOrder(M); 366 367 // Enumerate the global variables. 368 for (const GlobalVariable &GV : M.globals()) { 369 EnumerateValue(&GV); 370 EnumerateType(GV.getValueType()); 371 } 372 373 // Enumerate the functions. 374 for (const Function &F : M) { 375 EnumerateValue(&F); 376 EnumerateType(F.getValueType()); 377 EnumerateType( 378 dxil::TypedPointerType::get(F.getFunctionType(), F.getAddressSpace())); 379 EnumerateAttributes(F.getAttributes()); 380 } 381 382 // Enumerate the aliases. 383 for (const GlobalAlias &GA : M.aliases()) { 384 EnumerateValue(&GA); 385 EnumerateType(GA.getValueType()); 386 } 387 388 // Enumerate the ifuncs. 389 for (const GlobalIFunc &GIF : M.ifuncs()) { 390 EnumerateValue(&GIF); 391 EnumerateType(GIF.getValueType()); 392 } 393 394 // Enumerate the global variable initializers and attributes. 395 for (const GlobalVariable &GV : M.globals()) { 396 if (GV.hasInitializer()) 397 EnumerateValue(GV.getInitializer()); 398 EnumerateType( 399 dxil::TypedPointerType::get(GV.getValueType(), GV.getAddressSpace())); 400 if (GV.hasAttributes()) 401 EnumerateAttributes(GV.getAttributesAsList(AttributeList::FunctionIndex)); 402 } 403 404 // Enumerate the aliasees. 405 for (const GlobalAlias &GA : M.aliases()) 406 EnumerateValue(GA.getAliasee()); 407 408 // Enumerate the ifunc resolvers. 409 for (const GlobalIFunc &GIF : M.ifuncs()) 410 EnumerateValue(GIF.getResolver()); 411 412 // Enumerate any optional Function data. 413 for (const Function &F : M) 414 for (const Use &U : F.operands()) 415 EnumerateValue(U.get()); 416 417 // Enumerate the metadata type. 418 // 419 // TODO: Move this to ValueEnumerator::EnumerateOperandType() once bitcode 420 // only encodes the metadata type when it's used as a value. 421 EnumerateType(Type::getMetadataTy(M.getContext())); 422 423 // Insert constants and metadata that are named at module level into the slot 424 // pool so that the module symbol table can refer to them... 425 EnumerateValueSymbolTable(M.getValueSymbolTable()); 426 EnumerateNamedMetadata(M); 427 428 SmallVector<std::pair<unsigned, MDNode *>, 8> MDs; 429 for (const GlobalVariable &GV : M.globals()) { 430 MDs.clear(); 431 GV.getAllMetadata(MDs); 432 for (const auto &I : MDs) 433 // FIXME: Pass GV to EnumerateMetadata and arrange for the bitcode writer 434 // to write metadata to the global variable's own metadata block 435 // (PR28134). 436 EnumerateMetadata(nullptr, I.second); 437 } 438 439 // Enumerate types used by function bodies and argument lists. 440 for (const Function &F : M) { 441 for (const Argument &A : F.args()) 442 EnumerateType(A.getType()); 443 444 // Enumerate metadata attached to this function. 445 MDs.clear(); 446 F.getAllMetadata(MDs); 447 for (const auto &I : MDs) 448 EnumerateMetadata(F.isDeclaration() ? nullptr : &F, I.second); 449 450 for (const BasicBlock &BB : F) 451 for (const Instruction &I : BB) { 452 for (const Use &Op : I.operands()) { 453 auto *MD = dyn_cast<MetadataAsValue>(&Op); 454 if (!MD) { 455 EnumerateOperandType(Op); 456 continue; 457 } 458 459 // Local metadata is enumerated during function-incorporation, but 460 // any ConstantAsMetadata arguments in a DIArgList should be examined 461 // now. 462 if (isa<LocalAsMetadata>(MD->getMetadata())) 463 continue; 464 if (auto *AL = dyn_cast<DIArgList>(MD->getMetadata())) { 465 for (auto *VAM : AL->getArgs()) 466 if (isa<ConstantAsMetadata>(VAM)) 467 EnumerateMetadata(&F, VAM); 468 continue; 469 } 470 471 EnumerateMetadata(&F, MD->getMetadata()); 472 } 473 if (auto *SVI = dyn_cast<ShuffleVectorInst>(&I)) 474 EnumerateType(SVI->getShuffleMaskForBitcode()->getType()); 475 if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) 476 EnumerateType(GEP->getSourceElementType()); 477 if (auto *AI = dyn_cast<AllocaInst>(&I)) 478 EnumerateType(AI->getAllocatedType()); 479 EnumerateType(I.getType()); 480 if (const auto *Call = dyn_cast<CallBase>(&I)) { 481 EnumerateAttributes(Call->getAttributes()); 482 EnumerateType(Call->getFunctionType()); 483 } 484 485 // Enumerate metadata attached with this instruction. 486 MDs.clear(); 487 I.getAllMetadataOtherThanDebugLoc(MDs); 488 for (unsigned i = 0, e = MDs.size(); i != e; ++i) 489 EnumerateMetadata(&F, MDs[i].second); 490 491 // Don't enumerate the location directly -- it has a special record 492 // type -- but enumerate its operands. 493 if (DILocation *L = I.getDebugLoc()) 494 for (const Metadata *Op : L->operands()) 495 EnumerateMetadata(&F, Op); 496 } 497 } 498 499 // Organize metadata ordering. 500 organizeMetadata(); 501 } 502 503 unsigned ValueEnumerator::getInstructionID(const Instruction *Inst) const { 504 InstructionMapType::const_iterator I = InstructionMap.find(Inst); 505 assert(I != InstructionMap.end() && "Instruction is not mapped!"); 506 return I->second; 507 } 508 509 unsigned ValueEnumerator::getComdatID(const Comdat *C) const { 510 unsigned ComdatID = Comdats.idFor(C); 511 assert(ComdatID && "Comdat not found!"); 512 return ComdatID; 513 } 514 515 void ValueEnumerator::setInstructionID(const Instruction *I) { 516 InstructionMap[I] = InstructionCount++; 517 } 518 519 unsigned ValueEnumerator::getValueID(const Value *V) const { 520 if (auto *MD = dyn_cast<MetadataAsValue>(V)) 521 return getMetadataID(MD->getMetadata()); 522 523 ValueMapType::const_iterator I = ValueMap.find(V); 524 assert(I != ValueMap.end() && "Value not in slotcalculator!"); 525 return I->second - 1; 526 } 527 528 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 529 LLVM_DUMP_METHOD void ValueEnumerator::dump() const { 530 print(dbgs(), ValueMap, "Default"); 531 dbgs() << '\n'; 532 print(dbgs(), MetadataMap, "MetaData"); 533 dbgs() << '\n'; 534 } 535 #endif 536 537 void ValueEnumerator::print(raw_ostream &OS, const ValueMapType &Map, 538 const char *Name) const { 539 OS << "Map Name: " << Name << "\n"; 540 OS << "Size: " << Map.size() << "\n"; 541 for (const auto &I : Map) { 542 const Value *V = I.first; 543 if (V->hasName()) 544 OS << "Value: " << V->getName(); 545 else 546 OS << "Value: [null]\n"; 547 V->print(errs()); 548 errs() << '\n'; 549 550 OS << " Uses(" << V->getNumUses() << "):"; 551 for (const Use &U : V->uses()) { 552 if (&U != &*V->use_begin()) 553 OS << ","; 554 if (U->hasName()) 555 OS << " " << U->getName(); 556 else 557 OS << " [null]"; 558 } 559 OS << "\n\n"; 560 } 561 } 562 563 void ValueEnumerator::print(raw_ostream &OS, const MetadataMapType &Map, 564 const char *Name) const { 565 OS << "Map Name: " << Name << "\n"; 566 OS << "Size: " << Map.size() << "\n"; 567 for (const auto &I : Map) { 568 const Metadata *MD = I.first; 569 OS << "Metadata: slot = " << I.second.ID << "\n"; 570 OS << "Metadata: function = " << I.second.F << "\n"; 571 MD->print(OS); 572 OS << "\n"; 573 } 574 } 575 576 /// EnumerateValueSymbolTable - Insert all of the values in the specified symbol 577 /// table into the values table. 578 void ValueEnumerator::EnumerateValueSymbolTable(const ValueSymbolTable &VST) { 579 for (ValueSymbolTable::const_iterator VI = VST.begin(), VE = VST.end(); 580 VI != VE; ++VI) 581 EnumerateValue(VI->getValue()); 582 } 583 584 /// Insert all of the values referenced by named metadata in the specified 585 /// module. 586 void ValueEnumerator::EnumerateNamedMetadata(const Module &M) { 587 for (const auto &I : M.named_metadata()) 588 EnumerateNamedMDNode(&I); 589 } 590 591 void ValueEnumerator::EnumerateNamedMDNode(const NamedMDNode *MD) { 592 for (unsigned i = 0, e = MD->getNumOperands(); i != e; ++i) 593 EnumerateMetadata(nullptr, MD->getOperand(i)); 594 } 595 596 unsigned ValueEnumerator::getMetadataFunctionID(const Function *F) const { 597 return F ? getValueID(F) + 1 : 0; 598 } 599 600 void ValueEnumerator::EnumerateMetadata(const Function *F, const Metadata *MD) { 601 EnumerateMetadata(getMetadataFunctionID(F), MD); 602 } 603 604 void ValueEnumerator::EnumerateFunctionLocalMetadata( 605 const Function &F, const LocalAsMetadata *Local) { 606 EnumerateFunctionLocalMetadata(getMetadataFunctionID(&F), Local); 607 } 608 609 void ValueEnumerator::EnumerateFunctionLocalListMetadata( 610 const Function &F, const DIArgList *ArgList) { 611 EnumerateFunctionLocalListMetadata(getMetadataFunctionID(&F), ArgList); 612 } 613 614 void ValueEnumerator::dropFunctionFromMetadata( 615 MetadataMapType::value_type &FirstMD) { 616 SmallVector<const MDNode *, 64> Worklist; 617 auto push = [&Worklist](MetadataMapType::value_type &MD) { 618 auto &Entry = MD.second; 619 620 // Nothing to do if this metadata isn't tagged. 621 if (!Entry.F) 622 return; 623 624 // Drop the function tag. 625 Entry.F = 0; 626 627 // If this is has an ID and is an MDNode, then its operands have entries as 628 // well. We need to drop the function from them too. 629 if (Entry.ID) 630 if (auto *N = dyn_cast<MDNode>(MD.first)) 631 Worklist.push_back(N); 632 }; 633 push(FirstMD); 634 while (!Worklist.empty()) 635 for (const Metadata *Op : Worklist.pop_back_val()->operands()) { 636 if (!Op) 637 continue; 638 auto MD = MetadataMap.find(Op); 639 if (MD != MetadataMap.end()) 640 push(*MD); 641 } 642 } 643 644 void ValueEnumerator::EnumerateMetadata(unsigned F, const Metadata *MD) { 645 // It's vital for reader efficiency that uniqued subgraphs are done in 646 // post-order; it's expensive when their operands have forward references. 647 // If a distinct node is referenced from a uniqued node, it'll be delayed 648 // until the uniqued subgraph has been completely traversed. 649 SmallVector<const MDNode *, 32> DelayedDistinctNodes; 650 651 // Start by enumerating MD, and then work through its transitive operands in 652 // post-order. This requires a depth-first search. 653 SmallVector<std::pair<const MDNode *, MDNode::op_iterator>, 32> Worklist; 654 if (const MDNode *N = enumerateMetadataImpl(F, MD)) 655 Worklist.push_back(std::make_pair(N, N->op_begin())); 656 657 while (!Worklist.empty()) { 658 const MDNode *N = Worklist.back().first; 659 660 // Enumerate operands until we hit a new node. We need to traverse these 661 // nodes' operands before visiting the rest of N's operands. 662 MDNode::op_iterator I = std::find_if( 663 Worklist.back().second, N->op_end(), 664 [&](const Metadata *MD) { return enumerateMetadataImpl(F, MD); }); 665 if (I != N->op_end()) { 666 auto *Op = cast<MDNode>(*I); 667 Worklist.back().second = ++I; 668 669 // Delay traversing Op if it's a distinct node and N is uniqued. 670 if (Op->isDistinct() && !N->isDistinct()) 671 DelayedDistinctNodes.push_back(Op); 672 else 673 Worklist.push_back(std::make_pair(Op, Op->op_begin())); 674 continue; 675 } 676 677 // All the operands have been visited. Now assign an ID. 678 Worklist.pop_back(); 679 MDs.push_back(N); 680 MetadataMap[N].ID = MDs.size(); 681 682 // Flush out any delayed distinct nodes; these are all the distinct nodes 683 // that are leaves in last uniqued subgraph. 684 if (Worklist.empty() || Worklist.back().first->isDistinct()) { 685 for (const MDNode *N : DelayedDistinctNodes) 686 Worklist.push_back(std::make_pair(N, N->op_begin())); 687 DelayedDistinctNodes.clear(); 688 } 689 } 690 } 691 692 const MDNode *ValueEnumerator::enumerateMetadataImpl(unsigned F, 693 const Metadata *MD) { 694 if (!MD) 695 return nullptr; 696 697 assert( 698 (isa<MDNode>(MD) || isa<MDString>(MD) || isa<ConstantAsMetadata>(MD)) && 699 "Invalid metadata kind"); 700 701 auto Insertion = MetadataMap.insert(std::make_pair(MD, MDIndex(F))); 702 MDIndex &Entry = Insertion.first->second; 703 if (!Insertion.second) { 704 // Already mapped. If F doesn't match the function tag, drop it. 705 if (Entry.hasDifferentFunction(F)) 706 dropFunctionFromMetadata(*Insertion.first); 707 return nullptr; 708 } 709 710 // Don't assign IDs to metadata nodes. 711 if (auto *N = dyn_cast<MDNode>(MD)) 712 return N; 713 714 // Save the metadata. 715 MDs.push_back(MD); 716 Entry.ID = MDs.size(); 717 718 // Enumerate the constant, if any. 719 if (auto *C = dyn_cast<ConstantAsMetadata>(MD)) 720 EnumerateValue(C->getValue()); 721 722 return nullptr; 723 } 724 725 /// EnumerateFunctionLocalMetadata - Incorporate function-local metadata 726 /// information reachable from the metadata. 727 void ValueEnumerator::EnumerateFunctionLocalMetadata( 728 unsigned F, const LocalAsMetadata *Local) { 729 assert(F && "Expected a function"); 730 731 // Check to see if it's already in! 732 MDIndex &Index = MetadataMap[Local]; 733 if (Index.ID) { 734 assert(Index.F == F && "Expected the same function"); 735 return; 736 } 737 738 MDs.push_back(Local); 739 Index.F = F; 740 Index.ID = MDs.size(); 741 742 EnumerateValue(Local->getValue()); 743 } 744 745 /// EnumerateFunctionLocalListMetadata - Incorporate function-local metadata 746 /// information reachable from the metadata. 747 void ValueEnumerator::EnumerateFunctionLocalListMetadata( 748 unsigned F, const DIArgList *ArgList) { 749 assert(F && "Expected a function"); 750 751 // Check to see if it's already in! 752 MDIndex &Index = MetadataMap[ArgList]; 753 if (Index.ID) { 754 assert(Index.F == F && "Expected the same function"); 755 return; 756 } 757 758 for (ValueAsMetadata *VAM : ArgList->getArgs()) { 759 if (isa<LocalAsMetadata>(VAM)) { 760 assert(MetadataMap.count(VAM) && 761 "LocalAsMetadata should be enumerated before DIArgList"); 762 assert(MetadataMap[VAM].F == F && 763 "Expected LocalAsMetadata in the same function"); 764 } else { 765 assert(isa<ConstantAsMetadata>(VAM) && 766 "Expected LocalAsMetadata or ConstantAsMetadata"); 767 assert(ValueMap.count(VAM->getValue()) && 768 "Constant should be enumerated beforeDIArgList"); 769 EnumerateMetadata(F, VAM); 770 } 771 } 772 773 MDs.push_back(ArgList); 774 Index.F = F; 775 Index.ID = MDs.size(); 776 } 777 778 static unsigned getMetadataTypeOrder(const Metadata *MD) { 779 // Strings are emitted in bulk and must come first. 780 if (isa<MDString>(MD)) 781 return 0; 782 783 // ConstantAsMetadata doesn't reference anything. We may as well shuffle it 784 // to the front since we can detect it. 785 auto *N = dyn_cast<MDNode>(MD); 786 if (!N) 787 return 1; 788 789 // The reader is fast forward references for distinct node operands, but slow 790 // when uniqued operands are unresolved. 791 return N->isDistinct() ? 2 : 3; 792 } 793 794 void ValueEnumerator::organizeMetadata() { 795 assert(MetadataMap.size() == MDs.size() && 796 "Metadata map and vector out of sync"); 797 798 if (MDs.empty()) 799 return; 800 801 // Copy out the index information from MetadataMap in order to choose a new 802 // order. 803 SmallVector<MDIndex, 64> Order; 804 Order.reserve(MetadataMap.size()); 805 for (const Metadata *MD : MDs) 806 Order.push_back(MetadataMap.lookup(MD)); 807 808 // Partition: 809 // - by function, then 810 // - by isa<MDString> 811 // and then sort by the original/current ID. Since the IDs are guaranteed to 812 // be unique, the result of std::sort will be deterministic. There's no need 813 // for std::stable_sort. 814 llvm::sort(Order, [this](MDIndex LHS, MDIndex RHS) { 815 return std::make_tuple(LHS.F, getMetadataTypeOrder(LHS.get(MDs)), LHS.ID) < 816 std::make_tuple(RHS.F, getMetadataTypeOrder(RHS.get(MDs)), RHS.ID); 817 }); 818 819 // Rebuild MDs, index the metadata ranges for each function in FunctionMDs, 820 // and fix up MetadataMap. 821 std::vector<const Metadata *> OldMDs; 822 MDs.swap(OldMDs); 823 MDs.reserve(OldMDs.size()); 824 for (unsigned I = 0, E = Order.size(); I != E && !Order[I].F; ++I) { 825 auto *MD = Order[I].get(OldMDs); 826 MDs.push_back(MD); 827 MetadataMap[MD].ID = I + 1; 828 if (isa<MDString>(MD)) 829 ++NumMDStrings; 830 } 831 832 // Return early if there's nothing for the functions. 833 if (MDs.size() == Order.size()) 834 return; 835 836 // Build the function metadata ranges. 837 MDRange R; 838 FunctionMDs.reserve(OldMDs.size()); 839 unsigned PrevF = 0; 840 for (unsigned I = MDs.size(), E = Order.size(), ID = MDs.size(); I != E; 841 ++I) { 842 unsigned F = Order[I].F; 843 if (!PrevF) { 844 PrevF = F; 845 } else if (PrevF != F) { 846 R.Last = FunctionMDs.size(); 847 std::swap(R, FunctionMDInfo[PrevF]); 848 R.First = FunctionMDs.size(); 849 850 ID = MDs.size(); 851 PrevF = F; 852 } 853 854 auto *MD = Order[I].get(OldMDs); 855 FunctionMDs.push_back(MD); 856 MetadataMap[MD].ID = ++ID; 857 if (isa<MDString>(MD)) 858 ++R.NumStrings; 859 } 860 R.Last = FunctionMDs.size(); 861 FunctionMDInfo[PrevF] = R; 862 } 863 864 void ValueEnumerator::incorporateFunctionMetadata(const Function &F) { 865 NumModuleMDs = MDs.size(); 866 867 auto R = FunctionMDInfo.lookup(getValueID(&F) + 1); 868 NumMDStrings = R.NumStrings; 869 MDs.insert(MDs.end(), FunctionMDs.begin() + R.First, 870 FunctionMDs.begin() + R.Last); 871 } 872 873 void ValueEnumerator::EnumerateValue(const Value *V) { 874 assert(!V->getType()->isVoidTy() && "Can't insert void values!"); 875 assert(!isa<MetadataAsValue>(V) && "EnumerateValue doesn't handle Metadata!"); 876 877 // Check to see if it's already in! 878 unsigned &ValueID = ValueMap[V]; 879 if (ValueID) { 880 // Increment use count. 881 Values[ValueID - 1].second++; 882 return; 883 } 884 885 if (auto *GO = dyn_cast<GlobalObject>(V)) 886 if (const Comdat *C = GO->getComdat()) 887 Comdats.insert(C); 888 889 // Enumerate the type of this value. 890 EnumerateType(V->getType()); 891 892 if (const Constant *C = dyn_cast<Constant>(V)) { 893 if (isa<GlobalValue>(C)) { 894 // Initializers for globals are handled explicitly elsewhere. 895 } else if (C->getNumOperands()) { 896 // If a constant has operands, enumerate them. This makes sure that if a 897 // constant has uses (for example an array of const ints), that they are 898 // inserted also. 899 900 // We prefer to enumerate them with values before we enumerate the user 901 // itself. This makes it more likely that we can avoid forward references 902 // in the reader. We know that there can be no cycles in the constants 903 // graph that don't go through a global variable. 904 for (User::const_op_iterator I = C->op_begin(), E = C->op_end(); I != E; 905 ++I) 906 if (!isa<BasicBlock>(*I)) // Don't enumerate BB operand to BlockAddress. 907 EnumerateValue(*I); 908 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 909 if (CE->getOpcode() == Instruction::ShuffleVector) 910 EnumerateValue(CE->getShuffleMaskForBitcode()); 911 if (auto *GEP = dyn_cast<GEPOperator>(CE)) 912 EnumerateType(GEP->getSourceElementType()); 913 } 914 915 // Finally, add the value. Doing this could make the ValueID reference be 916 // dangling, don't reuse it. 917 Values.push_back(std::make_pair(V, 1U)); 918 ValueMap[V] = Values.size(); 919 return; 920 } 921 } 922 923 // Add the value. 924 Values.push_back(std::make_pair(V, 1U)); 925 ValueID = Values.size(); 926 } 927 928 void ValueEnumerator::EnumerateType(Type *Ty) { 929 unsigned *TypeID = &TypeMap[Ty]; 930 931 // We've already seen this type. 932 if (*TypeID) 933 return; 934 935 // If it is a non-anonymous struct, mark the type as being visited so that we 936 // don't recursively visit it. This is safe because we allow forward 937 // references of these in the bitcode reader. 938 if (StructType *STy = dyn_cast<StructType>(Ty)) 939 if (!STy->isLiteral()) 940 *TypeID = ~0U; 941 942 // Enumerate all of the subtypes before we enumerate this type. This ensures 943 // that the type will be enumerated in an order that can be directly built. 944 for (Type *SubTy : Ty->subtypes()) 945 EnumerateType(SubTy); 946 947 // Refresh the TypeID pointer in case the table rehashed. 948 TypeID = &TypeMap[Ty]; 949 950 // Check to see if we got the pointer another way. This can happen when 951 // enumerating recursive types that hit the base case deeper than they start. 952 // 953 // If this is actually a struct that we are treating as forward ref'able, 954 // then emit the definition now that all of its contents are available. 955 if (*TypeID && *TypeID != ~0U) 956 return; 957 958 // Add this type now that its contents are all happily enumerated. 959 Types.push_back(Ty); 960 961 *TypeID = Types.size(); 962 } 963 964 // Enumerate the types for the specified value. If the value is a constant, 965 // walk through it, enumerating the types of the constant. 966 void ValueEnumerator::EnumerateOperandType(const Value *V) { 967 EnumerateType(V->getType()); 968 969 assert(!isa<MetadataAsValue>(V) && "Unexpected metadata operand"); 970 971 const Constant *C = dyn_cast<Constant>(V); 972 if (!C) 973 return; 974 975 // If this constant is already enumerated, ignore it, we know its type must 976 // be enumerated. 977 if (ValueMap.count(C)) 978 return; 979 980 // This constant may have operands, make sure to enumerate the types in 981 // them. 982 for (const Value *Op : C->operands()) { 983 // Don't enumerate basic blocks here, this happens as operands to 984 // blockaddress. 985 if (isa<BasicBlock>(Op)) 986 continue; 987 988 EnumerateOperandType(Op); 989 } 990 if (auto *CE = dyn_cast<ConstantExpr>(C)) { 991 if (CE->getOpcode() == Instruction::ShuffleVector) 992 EnumerateOperandType(CE->getShuffleMaskForBitcode()); 993 if (CE->getOpcode() == Instruction::GetElementPtr) 994 EnumerateType(cast<GEPOperator>(CE)->getSourceElementType()); 995 } 996 } 997 998 void ValueEnumerator::EnumerateAttributes(AttributeList PAL) { 999 if (PAL.isEmpty()) 1000 return; // null is always 0. 1001 1002 // Do a lookup. 1003 unsigned &Entry = AttributeListMap[PAL]; 1004 if (Entry == 0) { 1005 // Never saw this before, add it. 1006 AttributeLists.push_back(PAL); 1007 Entry = AttributeLists.size(); 1008 } 1009 1010 // Do lookups for all attribute groups. 1011 for (unsigned i : PAL.indexes()) { 1012 AttributeSet AS = PAL.getAttributes(i); 1013 if (!AS.hasAttributes()) 1014 continue; 1015 IndexAndAttrSet Pair = {i, AS}; 1016 unsigned &Entry = AttributeGroupMap[Pair]; 1017 if (Entry == 0) { 1018 AttributeGroups.push_back(Pair); 1019 Entry = AttributeGroups.size(); 1020 1021 for (Attribute Attr : AS) { 1022 if (Attr.isTypeAttribute()) 1023 EnumerateType(Attr.getValueAsType()); 1024 } 1025 } 1026 } 1027 } 1028 1029 void ValueEnumerator::incorporateFunction(const Function &F) { 1030 InstructionCount = 0; 1031 NumModuleValues = Values.size(); 1032 1033 // Add global metadata to the function block. This doesn't include 1034 // LocalAsMetadata. 1035 incorporateFunctionMetadata(F); 1036 1037 // Adding function arguments to the value table. 1038 for (const auto &I : F.args()) { 1039 EnumerateValue(&I); 1040 if (I.hasAttribute(Attribute::ByVal)) 1041 EnumerateType(I.getParamByValType()); 1042 else if (I.hasAttribute(Attribute::StructRet)) 1043 EnumerateType(I.getParamStructRetType()); 1044 else if (I.hasAttribute(Attribute::ByRef)) 1045 EnumerateType(I.getParamByRefType()); 1046 } 1047 FirstFuncConstantID = Values.size(); 1048 1049 // Add all function-level constants to the value table. 1050 for (const BasicBlock &BB : F) { 1051 for (const Instruction &I : BB) { 1052 for (const Use &OI : I.operands()) { 1053 if ((isa<Constant>(OI) && !isa<GlobalValue>(OI)) || isa<InlineAsm>(OI)) 1054 EnumerateValue(OI); 1055 } 1056 if (auto *SVI = dyn_cast<ShuffleVectorInst>(&I)) 1057 EnumerateValue(SVI->getShuffleMaskForBitcode()); 1058 } 1059 BasicBlocks.push_back(&BB); 1060 ValueMap[&BB] = BasicBlocks.size(); 1061 } 1062 1063 // Add the function's parameter attributes so they are available for use in 1064 // the function's instruction. 1065 EnumerateAttributes(F.getAttributes()); 1066 1067 FirstInstID = Values.size(); 1068 1069 SmallVector<LocalAsMetadata *, 8> FnLocalMDVector; 1070 SmallVector<DIArgList *, 8> ArgListMDVector; 1071 // Add all of the instructions. 1072 for (const BasicBlock &BB : F) { 1073 for (const Instruction &I : BB) { 1074 for (const Use &OI : I.operands()) { 1075 if (auto *MD = dyn_cast<MetadataAsValue>(&OI)) { 1076 if (auto *Local = dyn_cast<LocalAsMetadata>(MD->getMetadata())) { 1077 // Enumerate metadata after the instructions they might refer to. 1078 FnLocalMDVector.push_back(Local); 1079 } else if (auto *ArgList = dyn_cast<DIArgList>(MD->getMetadata())) { 1080 ArgListMDVector.push_back(ArgList); 1081 for (ValueAsMetadata *VMD : ArgList->getArgs()) { 1082 if (auto *Local = dyn_cast<LocalAsMetadata>(VMD)) { 1083 // Enumerate metadata after the instructions they might refer 1084 // to. 1085 FnLocalMDVector.push_back(Local); 1086 } 1087 } 1088 } 1089 } 1090 } 1091 1092 if (!I.getType()->isVoidTy()) 1093 EnumerateValue(&I); 1094 } 1095 } 1096 1097 // Add all of the function-local metadata. 1098 for (unsigned i = 0, e = FnLocalMDVector.size(); i != e; ++i) { 1099 // At this point, every local values have been incorporated, we shouldn't 1100 // have a metadata operand that references a value that hasn't been seen. 1101 assert(ValueMap.count(FnLocalMDVector[i]->getValue()) && 1102 "Missing value for metadata operand"); 1103 EnumerateFunctionLocalMetadata(F, FnLocalMDVector[i]); 1104 } 1105 // DIArgList entries must come after function-local metadata, as it is not 1106 // possible to forward-reference them. 1107 for (const DIArgList *ArgList : ArgListMDVector) 1108 EnumerateFunctionLocalListMetadata(F, ArgList); 1109 } 1110 1111 void ValueEnumerator::purgeFunction() { 1112 /// Remove purged values from the ValueMap. 1113 for (unsigned i = NumModuleValues, e = Values.size(); i != e; ++i) 1114 ValueMap.erase(Values[i].first); 1115 for (unsigned i = NumModuleMDs, e = MDs.size(); i != e; ++i) 1116 MetadataMap.erase(MDs[i]); 1117 for (const BasicBlock *BB : BasicBlocks) 1118 ValueMap.erase(BB); 1119 1120 Values.resize(NumModuleValues); 1121 MDs.resize(NumModuleMDs); 1122 BasicBlocks.clear(); 1123 NumMDStrings = 0; 1124 } 1125 1126 static void IncorporateFunctionInfoGlobalBBIDs( 1127 const Function *F, DenseMap<const BasicBlock *, unsigned> &IDMap) { 1128 unsigned Counter = 0; 1129 for (const BasicBlock &BB : *F) 1130 IDMap[&BB] = ++Counter; 1131 } 1132 1133 /// getGlobalBasicBlockID - This returns the function-specific ID for the 1134 /// specified basic block. This is relatively expensive information, so it 1135 /// should only be used by rare constructs such as address-of-label. 1136 unsigned ValueEnumerator::getGlobalBasicBlockID(const BasicBlock *BB) const { 1137 unsigned &Idx = GlobalBasicBlockIDs[BB]; 1138 if (Idx != 0) 1139 return Idx - 1; 1140 1141 IncorporateFunctionInfoGlobalBBIDs(BB->getParent(), GlobalBasicBlockIDs); 1142 return getGlobalBasicBlockID(BB); 1143 } 1144 1145 uint64_t ValueEnumerator::computeBitsRequiredForTypeIndicies() const { 1146 return Log2_32_Ceil(getTypes().size() + 1); 1147 } 1148