PDB.cpp revision 327952
1//===- PDB.cpp ------------------------------------------------------------===// 2// 3// The LLVM Linker 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9 10#include "PDB.h" 11#include "Chunks.h" 12#include "Config.h" 13#include "Driver.h" 14#include "SymbolTable.h" 15#include "Symbols.h" 16#include "Writer.h" 17#include "lld/Common/ErrorHandler.h" 18#include "llvm/DebugInfo/CodeView/CVDebugRecord.h" 19#include "llvm/DebugInfo/CodeView/DebugSubsectionRecord.h" 20#include "llvm/DebugInfo/CodeView/GlobalTypeTableBuilder.h" 21#include "llvm/DebugInfo/CodeView/LazyRandomTypeCollection.h" 22#include "llvm/DebugInfo/CodeView/MergingTypeTableBuilder.h" 23#include "llvm/DebugInfo/CodeView/RecordName.h" 24#include "llvm/DebugInfo/CodeView/SymbolDeserializer.h" 25#include "llvm/DebugInfo/CodeView/SymbolSerializer.h" 26#include "llvm/DebugInfo/CodeView/TypeDeserializer.h" 27#include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 28#include "llvm/DebugInfo/CodeView/TypeIndexDiscovery.h" 29#include "llvm/DebugInfo/CodeView/TypeStreamMerger.h" 30#include "llvm/DebugInfo/MSF/MSFBuilder.h" 31#include "llvm/DebugInfo/MSF/MSFCommon.h" 32#include "llvm/DebugInfo/PDB/GenericError.h" 33#include "llvm/DebugInfo/PDB/Native/DbiModuleDescriptorBuilder.h" 34#include "llvm/DebugInfo/PDB/Native/DbiStream.h" 35#include "llvm/DebugInfo/PDB/Native/DbiStreamBuilder.h" 36#include "llvm/DebugInfo/PDB/Native/GSIStreamBuilder.h" 37#include "llvm/DebugInfo/PDB/Native/InfoStream.h" 38#include "llvm/DebugInfo/PDB/Native/InfoStreamBuilder.h" 39#include "llvm/DebugInfo/PDB/Native/NativeSession.h" 40#include "llvm/DebugInfo/PDB/Native/PDBFile.h" 41#include "llvm/DebugInfo/PDB/Native/PDBFileBuilder.h" 42#include "llvm/DebugInfo/PDB/Native/PDBStringTableBuilder.h" 43#include "llvm/DebugInfo/PDB/Native/TpiHashing.h" 44#include "llvm/DebugInfo/PDB/Native/TpiStream.h" 45#include "llvm/DebugInfo/PDB/Native/TpiStreamBuilder.h" 46#include "llvm/DebugInfo/PDB/PDB.h" 47#include "llvm/Object/COFF.h" 48#include "llvm/Support/BinaryByteStream.h" 49#include "llvm/Support/Endian.h" 50#include "llvm/Support/JamCRC.h" 51#include "llvm/Support/Path.h" 52#include "llvm/Support/ScopedPrinter.h" 53#include <memory> 54 55using namespace lld; 56using namespace lld::coff; 57using namespace llvm; 58using namespace llvm::codeview; 59 60using llvm::object::coff_section; 61 62static ExitOnError ExitOnErr; 63 64namespace { 65/// Map from type index and item index in a type server PDB to the 66/// corresponding index in the destination PDB. 67struct CVIndexMap { 68 SmallVector<TypeIndex, 0> TPIMap; 69 SmallVector<TypeIndex, 0> IPIMap; 70 bool IsTypeServerMap = false; 71}; 72 73class PDBLinker { 74public: 75 PDBLinker(SymbolTable *Symtab) 76 : Alloc(), Symtab(Symtab), Builder(Alloc), TypeTable(Alloc), 77 IDTable(Alloc), GlobalTypeTable(Alloc), GlobalIDTable(Alloc) {} 78 79 /// Emit the basic PDB structure: initial streams, headers, etc. 80 void initialize(const llvm::codeview::DebugInfo &BuildId); 81 82 /// Link CodeView from each object file in the symbol table into the PDB. 83 void addObjectsToPDB(); 84 85 /// Link CodeView from a single object file into the PDB. 86 void addObjFile(ObjFile *File); 87 88 /// Produce a mapping from the type and item indices used in the object 89 /// file to those in the destination PDB. 90 /// 91 /// If the object file uses a type server PDB (compiled with /Zi), merge TPI 92 /// and IPI from the type server PDB and return a map for it. Each unique type 93 /// server PDB is merged at most once, so this may return an existing index 94 /// mapping. 95 /// 96 /// If the object does not use a type server PDB (compiled with /Z7), we merge 97 /// all the type and item records from the .debug$S stream and fill in the 98 /// caller-provided ObjectIndexMap. 99 const CVIndexMap &mergeDebugT(ObjFile *File, CVIndexMap &ObjectIndexMap); 100 101 const CVIndexMap &maybeMergeTypeServerPDB(ObjFile *File, 102 TypeServer2Record &TS); 103 104 /// Add the section map and section contributions to the PDB. 105 void addSections(ArrayRef<OutputSection *> OutputSections, 106 ArrayRef<uint8_t> SectionTable); 107 108 void addSectionContrib(pdb::DbiModuleDescriptorBuilder &LinkerModule, 109 OutputSection *OS, Chunk *C); 110 111 /// Write the PDB to disk. 112 void commit(); 113 114private: 115 BumpPtrAllocator Alloc; 116 117 SymbolTable *Symtab; 118 119 pdb::PDBFileBuilder Builder; 120 121 /// Type records that will go into the PDB TPI stream. 122 MergingTypeTableBuilder TypeTable; 123 124 /// Item records that will go into the PDB IPI stream. 125 MergingTypeTableBuilder IDTable; 126 127 /// Type records that will go into the PDB TPI stream (for /DEBUG:GHASH) 128 GlobalTypeTableBuilder GlobalTypeTable; 129 130 /// Item records that will go into the PDB IPI stream (for /DEBUG:GHASH) 131 GlobalTypeTableBuilder GlobalIDTable; 132 133 /// PDBs use a single global string table for filenames in the file checksum 134 /// table. 135 DebugStringTableSubsection PDBStrTab; 136 137 llvm::SmallString<128> NativePath; 138 139 std::vector<pdb::SecMapEntry> SectionMap; 140 141 /// Type index mappings of type server PDBs that we've loaded so far. 142 std::map<GUID, CVIndexMap> TypeServerIndexMappings; 143}; 144} 145 146static SectionChunk *findByName(ArrayRef<SectionChunk *> Sections, 147 StringRef Name) { 148 for (SectionChunk *C : Sections) 149 if (C->getSectionName() == Name) 150 return C; 151 return nullptr; 152} 153 154static ArrayRef<uint8_t> consumeDebugMagic(ArrayRef<uint8_t> Data, 155 StringRef SecName) { 156 // First 4 bytes are section magic. 157 if (Data.size() < 4) 158 fatal(SecName + " too short"); 159 if (support::endian::read32le(Data.data()) != COFF::DEBUG_SECTION_MAGIC) 160 fatal(SecName + " has an invalid magic"); 161 return Data.slice(4); 162} 163 164static ArrayRef<uint8_t> getDebugSection(ObjFile *File, StringRef SecName) { 165 if (SectionChunk *Sec = findByName(File->getDebugChunks(), SecName)) 166 return consumeDebugMagic(Sec->getContents(), SecName); 167 return {}; 168} 169 170// A COFF .debug$H section is currently a clang extension. This function checks 171// if a .debug$H section is in a format that we expect / understand, so that we 172// can ignore any sections which are coincidentally also named .debug$H but do 173// not contain a format we recognize. 174static bool canUseDebugH(ArrayRef<uint8_t> DebugH) { 175 if (DebugH.size() < sizeof(object::debug_h_header)) 176 return false; 177 auto *Header = 178 reinterpret_cast<const object::debug_h_header *>(DebugH.data()); 179 DebugH = DebugH.drop_front(sizeof(object::debug_h_header)); 180 return Header->Magic == COFF::DEBUG_HASHES_SECTION_MAGIC && 181 Header->Version == 0 && 182 Header->HashAlgorithm == uint16_t(GlobalTypeHashAlg::SHA1) && 183 (DebugH.size() % 20 == 0); 184} 185 186static Optional<ArrayRef<uint8_t>> getDebugH(ObjFile *File) { 187 SectionChunk *Sec = findByName(File->getDebugChunks(), ".debug$H"); 188 if (!Sec) 189 return llvm::None; 190 ArrayRef<uint8_t> Contents = Sec->getContents(); 191 if (!canUseDebugH(Contents)) 192 return None; 193 return Contents; 194} 195 196static ArrayRef<GloballyHashedType> 197getHashesFromDebugH(ArrayRef<uint8_t> DebugH) { 198 assert(canUseDebugH(DebugH)); 199 200 DebugH = DebugH.drop_front(sizeof(object::debug_h_header)); 201 uint32_t Count = DebugH.size() / sizeof(GloballyHashedType); 202 return {reinterpret_cast<const GloballyHashedType *>(DebugH.data()), Count}; 203} 204 205static void addTypeInfo(pdb::TpiStreamBuilder &TpiBuilder, 206 TypeCollection &TypeTable) { 207 // Start the TPI or IPI stream header. 208 TpiBuilder.setVersionHeader(pdb::PdbTpiV80); 209 210 // Flatten the in memory type table and hash each type. 211 TypeTable.ForEachRecord([&](TypeIndex TI, const CVType &Type) { 212 auto Hash = pdb::hashTypeRecord(Type); 213 if (auto E = Hash.takeError()) 214 fatal("type hashing error"); 215 TpiBuilder.addTypeRecord(Type.RecordData, *Hash); 216 }); 217} 218 219static Optional<TypeServer2Record> 220maybeReadTypeServerRecord(CVTypeArray &Types) { 221 auto I = Types.begin(); 222 if (I == Types.end()) 223 return None; 224 const CVType &Type = *I; 225 if (Type.kind() != LF_TYPESERVER2) 226 return None; 227 TypeServer2Record TS; 228 if (auto EC = TypeDeserializer::deserializeAs(const_cast<CVType &>(Type), TS)) 229 fatal("error reading type server record: " + toString(std::move(EC))); 230 return std::move(TS); 231} 232 233const CVIndexMap &PDBLinker::mergeDebugT(ObjFile *File, 234 CVIndexMap &ObjectIndexMap) { 235 ArrayRef<uint8_t> Data = getDebugSection(File, ".debug$T"); 236 if (Data.empty()) 237 return ObjectIndexMap; 238 239 BinaryByteStream Stream(Data, support::little); 240 CVTypeArray Types; 241 BinaryStreamReader Reader(Stream); 242 if (auto EC = Reader.readArray(Types, Reader.getLength())) 243 fatal("Reader::readArray failed: " + toString(std::move(EC))); 244 245 // Look through type servers. If we've already seen this type server, don't 246 // merge any type information. 247 if (Optional<TypeServer2Record> TS = maybeReadTypeServerRecord(Types)) 248 return maybeMergeTypeServerPDB(File, *TS); 249 250 // This is a /Z7 object. Fill in the temporary, caller-provided 251 // ObjectIndexMap. 252 if (Config->DebugGHashes) { 253 ArrayRef<GloballyHashedType> Hashes; 254 std::vector<GloballyHashedType> OwnedHashes; 255 if (Optional<ArrayRef<uint8_t>> DebugH = getDebugH(File)) 256 Hashes = getHashesFromDebugH(*DebugH); 257 else { 258 OwnedHashes = GloballyHashedType::hashTypes(Types); 259 Hashes = OwnedHashes; 260 } 261 262 if (auto Err = mergeTypeAndIdRecords(GlobalIDTable, GlobalTypeTable, 263 ObjectIndexMap.TPIMap, Types, Hashes)) 264 fatal("codeview::mergeTypeAndIdRecords failed: " + 265 toString(std::move(Err))); 266 } else { 267 if (auto Err = mergeTypeAndIdRecords(IDTable, TypeTable, 268 ObjectIndexMap.TPIMap, Types)) 269 fatal("codeview::mergeTypeAndIdRecords failed: " + 270 toString(std::move(Err))); 271 } 272 return ObjectIndexMap; 273} 274 275static Expected<std::unique_ptr<pdb::NativeSession>> 276tryToLoadPDB(const GUID &GuidFromObj, StringRef TSPath) { 277 ErrorOr<std::unique_ptr<MemoryBuffer>> MBOrErr = MemoryBuffer::getFile( 278 TSPath, /*FileSize=*/-1, /*RequiresNullTerminator=*/false); 279 if (!MBOrErr) 280 return errorCodeToError(MBOrErr.getError()); 281 282 std::unique_ptr<pdb::IPDBSession> ThisSession; 283 if (auto EC = pdb::NativeSession::createFromPdb( 284 MemoryBuffer::getMemBuffer(Driver->takeBuffer(std::move(*MBOrErr)), 285 /*RequiresNullTerminator=*/false), 286 ThisSession)) 287 return std::move(EC); 288 289 std::unique_ptr<pdb::NativeSession> NS( 290 static_cast<pdb::NativeSession *>(ThisSession.release())); 291 pdb::PDBFile &File = NS->getPDBFile(); 292 auto ExpectedInfo = File.getPDBInfoStream(); 293 // All PDB Files should have an Info stream. 294 if (!ExpectedInfo) 295 return ExpectedInfo.takeError(); 296 297 // Just because a file with a matching name was found and it was an actual 298 // PDB file doesn't mean it matches. For it to match the InfoStream's GUID 299 // must match the GUID specified in the TypeServer2 record. 300 if (ExpectedInfo->getGuid() != GuidFromObj) 301 return make_error<pdb::GenericError>( 302 pdb::generic_error_code::type_server_not_found, TSPath); 303 304 return std::move(NS); 305} 306 307const CVIndexMap &PDBLinker::maybeMergeTypeServerPDB(ObjFile *File, 308 TypeServer2Record &TS) { 309 // First, check if we already loaded a PDB with this GUID. Return the type 310 // index mapping if we have it. 311 auto Insertion = TypeServerIndexMappings.insert({TS.getGuid(), CVIndexMap()}); 312 CVIndexMap &IndexMap = Insertion.first->second; 313 if (!Insertion.second) 314 return IndexMap; 315 316 // Mark this map as a type server map. 317 IndexMap.IsTypeServerMap = true; 318 319 // Check for a PDB at: 320 // 1. The given file path 321 // 2. Next to the object file or archive file 322 auto ExpectedSession = tryToLoadPDB(TS.getGuid(), TS.getName()); 323 if (!ExpectedSession) { 324 consumeError(ExpectedSession.takeError()); 325 StringRef LocalPath = 326 !File->ParentName.empty() ? File->ParentName : File->getName(); 327 SmallString<128> Path = sys::path::parent_path(LocalPath); 328 sys::path::append( 329 Path, sys::path::filename(TS.getName(), sys::path::Style::windows)); 330 ExpectedSession = tryToLoadPDB(TS.getGuid(), Path); 331 } 332 if (auto E = ExpectedSession.takeError()) 333 fatal("Type server PDB was not found: " + toString(std::move(E))); 334 335 auto ExpectedTpi = (*ExpectedSession)->getPDBFile().getPDBTpiStream(); 336 if (auto E = ExpectedTpi.takeError()) 337 fatal("Type server does not have TPI stream: " + toString(std::move(E))); 338 auto ExpectedIpi = (*ExpectedSession)->getPDBFile().getPDBIpiStream(); 339 if (auto E = ExpectedIpi.takeError()) 340 fatal("Type server does not have TPI stream: " + toString(std::move(E))); 341 342 if (Config->DebugGHashes) { 343 // PDBs do not actually store global hashes, so when merging a type server 344 // PDB we have to synthesize global hashes. To do this, we first synthesize 345 // global hashes for the TPI stream, since it is independent, then we 346 // synthesize hashes for the IPI stream, using the hashes for the TPI stream 347 // as inputs. 348 auto TpiHashes = GloballyHashedType::hashTypes(ExpectedTpi->typeArray()); 349 auto IpiHashes = 350 GloballyHashedType::hashIds(ExpectedIpi->typeArray(), TpiHashes); 351 352 // Merge TPI first, because the IPI stream will reference type indices. 353 if (auto Err = mergeTypeRecords(GlobalTypeTable, IndexMap.TPIMap, 354 ExpectedTpi->typeArray(), TpiHashes)) 355 fatal("codeview::mergeTypeRecords failed: " + toString(std::move(Err))); 356 357 // Merge IPI. 358 if (auto Err = 359 mergeIdRecords(GlobalIDTable, IndexMap.TPIMap, IndexMap.IPIMap, 360 ExpectedIpi->typeArray(), IpiHashes)) 361 fatal("codeview::mergeIdRecords failed: " + toString(std::move(Err))); 362 } else { 363 // Merge TPI first, because the IPI stream will reference type indices. 364 if (auto Err = mergeTypeRecords(TypeTable, IndexMap.TPIMap, 365 ExpectedTpi->typeArray())) 366 fatal("codeview::mergeTypeRecords failed: " + toString(std::move(Err))); 367 368 // Merge IPI. 369 if (auto Err = mergeIdRecords(IDTable, IndexMap.TPIMap, IndexMap.IPIMap, 370 ExpectedIpi->typeArray())) 371 fatal("codeview::mergeIdRecords failed: " + toString(std::move(Err))); 372 } 373 374 return IndexMap; 375} 376 377static bool remapTypeIndex(TypeIndex &TI, ArrayRef<TypeIndex> TypeIndexMap) { 378 if (TI.isSimple()) 379 return true; 380 if (TI.toArrayIndex() >= TypeIndexMap.size()) 381 return false; 382 TI = TypeIndexMap[TI.toArrayIndex()]; 383 return true; 384} 385 386static void remapTypesInSymbolRecord(ObjFile *File, SymbolKind SymKind, 387 MutableArrayRef<uint8_t> Contents, 388 const CVIndexMap &IndexMap, 389 ArrayRef<TiReference> TypeRefs) { 390 for (const TiReference &Ref : TypeRefs) { 391 unsigned ByteSize = Ref.Count * sizeof(TypeIndex); 392 if (Contents.size() < Ref.Offset + ByteSize) 393 fatal("symbol record too short"); 394 395 // This can be an item index or a type index. Choose the appropriate map. 396 ArrayRef<TypeIndex> TypeOrItemMap = IndexMap.TPIMap; 397 bool IsItemIndex = Ref.Kind == TiRefKind::IndexRef; 398 if (IsItemIndex && IndexMap.IsTypeServerMap) 399 TypeOrItemMap = IndexMap.IPIMap; 400 401 MutableArrayRef<TypeIndex> TIs( 402 reinterpret_cast<TypeIndex *>(Contents.data() + Ref.Offset), Ref.Count); 403 for (TypeIndex &TI : TIs) { 404 if (!remapTypeIndex(TI, TypeOrItemMap)) { 405 log("ignoring symbol record of kind 0x" + utohexstr(SymKind) + " in " + 406 File->getName() + " with bad " + (IsItemIndex ? "item" : "type") + 407 " index 0x" + utohexstr(TI.getIndex())); 408 TI = TypeIndex(SimpleTypeKind::NotTranslated); 409 continue; 410 } 411 } 412 } 413} 414 415static SymbolKind symbolKind(ArrayRef<uint8_t> RecordData) { 416 const RecordPrefix *Prefix = 417 reinterpret_cast<const RecordPrefix *>(RecordData.data()); 418 return static_cast<SymbolKind>(uint16_t(Prefix->RecordKind)); 419} 420 421/// MSVC translates S_PROC_ID_END to S_END, and S_[LG]PROC32_ID to S_[LG]PROC32 422static void translateIdSymbols(MutableArrayRef<uint8_t> &RecordData, 423 TypeCollection &IDTable) { 424 RecordPrefix *Prefix = reinterpret_cast<RecordPrefix *>(RecordData.data()); 425 426 SymbolKind Kind = symbolKind(RecordData); 427 428 if (Kind == SymbolKind::S_PROC_ID_END) { 429 Prefix->RecordKind = SymbolKind::S_END; 430 return; 431 } 432 433 // In an object file, GPROC32_ID has an embedded reference which refers to the 434 // single object file type index namespace. This has already been translated 435 // to the PDB file's ID stream index space, but we need to convert this to a 436 // symbol that refers to the type stream index space. So we remap again from 437 // ID index space to type index space. 438 if (Kind == SymbolKind::S_GPROC32_ID || Kind == SymbolKind::S_LPROC32_ID) { 439 SmallVector<TiReference, 1> Refs; 440 auto Content = RecordData.drop_front(sizeof(RecordPrefix)); 441 CVSymbol Sym(Kind, RecordData); 442 discoverTypeIndicesInSymbol(Sym, Refs); 443 assert(Refs.size() == 1); 444 assert(Refs.front().Count == 1); 445 446 TypeIndex *TI = 447 reinterpret_cast<TypeIndex *>(Content.data() + Refs[0].Offset); 448 // `TI` is the index of a FuncIdRecord or MemberFuncIdRecord which lives in 449 // the IPI stream, whose `FunctionType` member refers to the TPI stream. 450 // Note that LF_FUNC_ID and LF_MEMFUNC_ID have the same record layout, and 451 // in both cases we just need the second type index. 452 if (!TI->isSimple() && !TI->isNoneType()) { 453 CVType FuncIdData = IDTable.getType(*TI); 454 SmallVector<TypeIndex, 2> Indices; 455 discoverTypeIndices(FuncIdData, Indices); 456 assert(Indices.size() == 2); 457 *TI = Indices[1]; 458 } 459 460 Kind = (Kind == SymbolKind::S_GPROC32_ID) ? SymbolKind::S_GPROC32 461 : SymbolKind::S_LPROC32; 462 Prefix->RecordKind = uint16_t(Kind); 463 } 464} 465 466/// Copy the symbol record. In a PDB, symbol records must be 4 byte aligned. 467/// The object file may not be aligned. 468static MutableArrayRef<uint8_t> copySymbolForPdb(const CVSymbol &Sym, 469 BumpPtrAllocator &Alloc) { 470 size_t Size = alignTo(Sym.length(), alignOf(CodeViewContainer::Pdb)); 471 assert(Size >= 4 && "record too short"); 472 assert(Size <= MaxRecordLength && "record too long"); 473 void *Mem = Alloc.Allocate(Size, 4); 474 475 // Copy the symbol record and zero out any padding bytes. 476 MutableArrayRef<uint8_t> NewData(reinterpret_cast<uint8_t *>(Mem), Size); 477 memcpy(NewData.data(), Sym.data().data(), Sym.length()); 478 memset(NewData.data() + Sym.length(), 0, Size - Sym.length()); 479 480 // Update the record prefix length. It should point to the beginning of the 481 // next record. 482 auto *Prefix = reinterpret_cast<RecordPrefix *>(Mem); 483 Prefix->RecordLen = Size - 2; 484 return NewData; 485} 486 487/// Return true if this symbol opens a scope. This implies that the symbol has 488/// "parent" and "end" fields, which contain the offset of the S_END or 489/// S_INLINESITE_END record. 490static bool symbolOpensScope(SymbolKind Kind) { 491 switch (Kind) { 492 case SymbolKind::S_GPROC32: 493 case SymbolKind::S_LPROC32: 494 case SymbolKind::S_LPROC32_ID: 495 case SymbolKind::S_GPROC32_ID: 496 case SymbolKind::S_BLOCK32: 497 case SymbolKind::S_SEPCODE: 498 case SymbolKind::S_THUNK32: 499 case SymbolKind::S_INLINESITE: 500 case SymbolKind::S_INLINESITE2: 501 return true; 502 default: 503 break; 504 } 505 return false; 506} 507 508static bool symbolEndsScope(SymbolKind Kind) { 509 switch (Kind) { 510 case SymbolKind::S_END: 511 case SymbolKind::S_PROC_ID_END: 512 case SymbolKind::S_INLINESITE_END: 513 return true; 514 default: 515 break; 516 } 517 return false; 518} 519 520struct ScopeRecord { 521 ulittle32_t PtrParent; 522 ulittle32_t PtrEnd; 523}; 524 525struct SymbolScope { 526 ScopeRecord *OpeningRecord; 527 uint32_t ScopeOffset; 528}; 529 530static void scopeStackOpen(SmallVectorImpl<SymbolScope> &Stack, 531 uint32_t CurOffset, CVSymbol &Sym) { 532 assert(symbolOpensScope(Sym.kind())); 533 SymbolScope S; 534 S.ScopeOffset = CurOffset; 535 S.OpeningRecord = const_cast<ScopeRecord *>( 536 reinterpret_cast<const ScopeRecord *>(Sym.content().data())); 537 S.OpeningRecord->PtrParent = Stack.empty() ? 0 : Stack.back().ScopeOffset; 538 Stack.push_back(S); 539} 540 541static void scopeStackClose(SmallVectorImpl<SymbolScope> &Stack, 542 uint32_t CurOffset, ObjFile *File) { 543 if (Stack.empty()) { 544 warn("symbol scopes are not balanced in " + File->getName()); 545 return; 546 } 547 SymbolScope S = Stack.pop_back_val(); 548 S.OpeningRecord->PtrEnd = CurOffset; 549} 550 551static bool symbolGoesInModuleStream(const CVSymbol &Sym) { 552 switch (Sym.kind()) { 553 case SymbolKind::S_GDATA32: 554 case SymbolKind::S_CONSTANT: 555 case SymbolKind::S_UDT: 556 // We really should not be seeing S_PROCREF and S_LPROCREF in the first place 557 // since they are synthesized by the linker in response to S_GPROC32 and 558 // S_LPROC32, but if we do see them, don't put them in the module stream I 559 // guess. 560 case SymbolKind::S_PROCREF: 561 case SymbolKind::S_LPROCREF: 562 return false; 563 // S_GDATA32 does not go in the module stream, but S_LDATA32 does. 564 case SymbolKind::S_LDATA32: 565 default: 566 return true; 567 } 568} 569 570static bool symbolGoesInGlobalsStream(const CVSymbol &Sym) { 571 switch (Sym.kind()) { 572 case SymbolKind::S_CONSTANT: 573 case SymbolKind::S_GDATA32: 574 // S_LDATA32 goes in both the module stream and the globals stream. 575 case SymbolKind::S_LDATA32: 576 case SymbolKind::S_GPROC32: 577 case SymbolKind::S_LPROC32: 578 // We really should not be seeing S_PROCREF and S_LPROCREF in the first place 579 // since they are synthesized by the linker in response to S_GPROC32 and 580 // S_LPROC32, but if we do see them, copy them straight through. 581 case SymbolKind::S_PROCREF: 582 case SymbolKind::S_LPROCREF: 583 return true; 584 // FIXME: For now, we drop all S_UDT symbols (i.e. they don't go in the 585 // globals stream or the modules stream). These have special handling which 586 // needs more investigation before we can get right, but by putting them all 587 // into the globals stream WinDbg fails to display local variables of class 588 // types saying that it cannot find the type Foo *. So as a stopgap just to 589 // keep things working, we drop them. 590 case SymbolKind::S_UDT: 591 default: 592 return false; 593 } 594} 595 596static void addGlobalSymbol(pdb::GSIStreamBuilder &Builder, ObjFile &File, 597 const CVSymbol &Sym) { 598 switch (Sym.kind()) { 599 case SymbolKind::S_CONSTANT: 600 case SymbolKind::S_UDT: 601 case SymbolKind::S_GDATA32: 602 case SymbolKind::S_LDATA32: 603 case SymbolKind::S_PROCREF: 604 case SymbolKind::S_LPROCREF: 605 Builder.addGlobalSymbol(Sym); 606 break; 607 case SymbolKind::S_GPROC32: 608 case SymbolKind::S_LPROC32: { 609 SymbolRecordKind K = SymbolRecordKind::ProcRefSym; 610 if (Sym.kind() == SymbolKind::S_LPROC32) 611 K = SymbolRecordKind::LocalProcRef; 612 ProcRefSym PS(K); 613 PS.Module = static_cast<uint16_t>(File.ModuleDBI->getModuleIndex()); 614 // For some reason, MSVC seems to add one to this value. 615 ++PS.Module; 616 PS.Name = getSymbolName(Sym); 617 PS.SumName = 0; 618 PS.SymOffset = File.ModuleDBI->getNextSymbolOffset(); 619 Builder.addGlobalSymbol(PS); 620 break; 621 } 622 default: 623 llvm_unreachable("Invalid symbol kind!"); 624 } 625} 626 627static void mergeSymbolRecords(BumpPtrAllocator &Alloc, ObjFile *File, 628 pdb::GSIStreamBuilder &GsiBuilder, 629 const CVIndexMap &IndexMap, 630 TypeCollection &IDTable, 631 BinaryStreamRef SymData) { 632 // FIXME: Improve error recovery by warning and skipping records when 633 // possible. 634 CVSymbolArray Syms; 635 BinaryStreamReader Reader(SymData); 636 ExitOnErr(Reader.readArray(Syms, Reader.getLength())); 637 SmallVector<SymbolScope, 4> Scopes; 638 for (CVSymbol Sym : Syms) { 639 // Discover type index references in the record. Skip it if we don't know 640 // where they are. 641 SmallVector<TiReference, 32> TypeRefs; 642 if (!discoverTypeIndicesInSymbol(Sym, TypeRefs)) { 643 log("ignoring unknown symbol record with kind 0x" + utohexstr(Sym.kind())); 644 continue; 645 } 646 647 // Copy the symbol record so we can mutate it. 648 MutableArrayRef<uint8_t> NewData = copySymbolForPdb(Sym, Alloc); 649 650 // Re-map all the type index references. 651 MutableArrayRef<uint8_t> Contents = 652 NewData.drop_front(sizeof(RecordPrefix)); 653 remapTypesInSymbolRecord(File, Sym.kind(), Contents, IndexMap, TypeRefs); 654 655 // An object file may have S_xxx_ID symbols, but these get converted to 656 // "real" symbols in a PDB. 657 translateIdSymbols(NewData, IDTable); 658 659 SymbolKind NewKind = symbolKind(NewData); 660 661 // Fill in "Parent" and "End" fields by maintaining a stack of scopes. 662 CVSymbol NewSym(NewKind, NewData); 663 if (symbolOpensScope(NewKind)) 664 scopeStackOpen(Scopes, File->ModuleDBI->getNextSymbolOffset(), NewSym); 665 else if (symbolEndsScope(NewKind)) 666 scopeStackClose(Scopes, File->ModuleDBI->getNextSymbolOffset(), File); 667 668 // Add the symbol to the globals stream if necessary. Do this before adding 669 // the symbol to the module since we may need to get the next symbol offset, 670 // and writing to the module's symbol stream will update that offset. 671 if (symbolGoesInGlobalsStream(NewSym)) 672 addGlobalSymbol(GsiBuilder, *File, NewSym); 673 674 // Add the symbol to the module. 675 if (symbolGoesInModuleStream(NewSym)) 676 File->ModuleDBI->addSymbol(NewSym); 677 } 678} 679 680// Allocate memory for a .debug$S section and relocate it. 681static ArrayRef<uint8_t> relocateDebugChunk(BumpPtrAllocator &Alloc, 682 SectionChunk *DebugChunk) { 683 uint8_t *Buffer = Alloc.Allocate<uint8_t>(DebugChunk->getSize()); 684 assert(DebugChunk->OutputSectionOff == 0 && 685 "debug sections should not be in output sections"); 686 DebugChunk->writeTo(Buffer); 687 return consumeDebugMagic(makeArrayRef(Buffer, DebugChunk->getSize()), 688 ".debug$S"); 689} 690 691void PDBLinker::addObjFile(ObjFile *File) { 692 // Add a module descriptor for every object file. We need to put an absolute 693 // path to the object into the PDB. If this is a plain object, we make its 694 // path absolute. If it's an object in an archive, we make the archive path 695 // absolute. 696 bool InArchive = !File->ParentName.empty(); 697 SmallString<128> Path = InArchive ? File->ParentName : File->getName(); 698 sys::fs::make_absolute(Path); 699 sys::path::native(Path, sys::path::Style::windows); 700 StringRef Name = InArchive ? File->getName() : StringRef(Path); 701 702 File->ModuleDBI = &ExitOnErr(Builder.getDbiBuilder().addModuleInfo(Name)); 703 File->ModuleDBI->setObjFileName(Path); 704 705 // Before we can process symbol substreams from .debug$S, we need to process 706 // type information, file checksums, and the string table. Add type info to 707 // the PDB first, so that we can get the map from object file type and item 708 // indices to PDB type and item indices. 709 CVIndexMap ObjectIndexMap; 710 const CVIndexMap &IndexMap = mergeDebugT(File, ObjectIndexMap); 711 712 // Now do all live .debug$S sections. 713 for (SectionChunk *DebugChunk : File->getDebugChunks()) { 714 if (!DebugChunk->isLive() || DebugChunk->getSectionName() != ".debug$S") 715 continue; 716 717 ArrayRef<uint8_t> RelocatedDebugContents = 718 relocateDebugChunk(Alloc, DebugChunk); 719 if (RelocatedDebugContents.empty()) 720 continue; 721 722 DebugSubsectionArray Subsections; 723 BinaryStreamReader Reader(RelocatedDebugContents, support::little); 724 ExitOnErr(Reader.readArray(Subsections, RelocatedDebugContents.size())); 725 726 DebugStringTableSubsectionRef CVStrTab; 727 DebugChecksumsSubsectionRef Checksums; 728 for (const DebugSubsectionRecord &SS : Subsections) { 729 switch (SS.kind()) { 730 case DebugSubsectionKind::StringTable: 731 ExitOnErr(CVStrTab.initialize(SS.getRecordData())); 732 break; 733 case DebugSubsectionKind::FileChecksums: 734 ExitOnErr(Checksums.initialize(SS.getRecordData())); 735 break; 736 case DebugSubsectionKind::Lines: 737 // We can add the relocated line table directly to the PDB without 738 // modification because the file checksum offsets will stay the same. 739 File->ModuleDBI->addDebugSubsection(SS); 740 break; 741 case DebugSubsectionKind::Symbols: 742 if (Config->DebugGHashes) { 743 mergeSymbolRecords(Alloc, File, Builder.getGsiBuilder(), IndexMap, 744 GlobalIDTable, SS.getRecordData()); 745 } else { 746 mergeSymbolRecords(Alloc, File, Builder.getGsiBuilder(), IndexMap, 747 IDTable, SS.getRecordData()); 748 } 749 break; 750 default: 751 // FIXME: Process the rest of the subsections. 752 break; 753 } 754 } 755 756 if (Checksums.valid()) { 757 // Make a new file checksum table that refers to offsets in the PDB-wide 758 // string table. Generally the string table subsection appears after the 759 // checksum table, so we have to do this after looping over all the 760 // subsections. 761 if (!CVStrTab.valid()) 762 fatal(".debug$S sections must have both a string table subsection " 763 "and a checksum subsection table or neither"); 764 auto NewChecksums = make_unique<DebugChecksumsSubsection>(PDBStrTab); 765 for (FileChecksumEntry &FC : Checksums) { 766 StringRef FileName = ExitOnErr(CVStrTab.getString(FC.FileNameOffset)); 767 ExitOnErr(Builder.getDbiBuilder().addModuleSourceFile(*File->ModuleDBI, 768 FileName)); 769 NewChecksums->addChecksum(FileName, FC.Kind, FC.Checksum); 770 } 771 File->ModuleDBI->addDebugSubsection(std::move(NewChecksums)); 772 } 773 } 774} 775 776static PublicSym32 createPublic(Defined *Def) { 777 PublicSym32 Pub(SymbolKind::S_PUB32); 778 Pub.Name = Def->getName(); 779 if (auto *D = dyn_cast<DefinedCOFF>(Def)) { 780 if (D->getCOFFSymbol().isFunctionDefinition()) 781 Pub.Flags = PublicSymFlags::Function; 782 } else if (isa<DefinedImportThunk>(Def)) { 783 Pub.Flags = PublicSymFlags::Function; 784 } 785 786 OutputSection *OS = Def->getChunk()->getOutputSection(); 787 assert(OS && "all publics should be in final image"); 788 Pub.Offset = Def->getRVA() - OS->getRVA(); 789 Pub.Segment = OS->SectionIndex; 790 return Pub; 791} 792 793// Add all object files to the PDB. Merge .debug$T sections into IpiData and 794// TpiData. 795void PDBLinker::addObjectsToPDB() { 796 for (ObjFile *File : ObjFile::Instances) 797 addObjFile(File); 798 799 Builder.getStringTableBuilder().setStrings(PDBStrTab); 800 801 // Construct TPI and IPI stream contents. 802 if (Config->DebugGHashes) { 803 addTypeInfo(Builder.getTpiBuilder(), GlobalTypeTable); 804 addTypeInfo(Builder.getIpiBuilder(), GlobalIDTable); 805 } else { 806 addTypeInfo(Builder.getTpiBuilder(), TypeTable); 807 addTypeInfo(Builder.getIpiBuilder(), IDTable); 808 } 809 810 // Compute the public and global symbols. 811 auto &GsiBuilder = Builder.getGsiBuilder(); 812 std::vector<PublicSym32> Publics; 813 Symtab->forEachSymbol([&Publics](Symbol *S) { 814 // Only emit defined, live symbols that have a chunk. 815 auto *Def = dyn_cast<Defined>(S); 816 if (Def && Def->isLive() && Def->getChunk()) 817 Publics.push_back(createPublic(Def)); 818 }); 819 820 if (!Publics.empty()) { 821 // Sort the public symbols and add them to the stream. 822 std::sort(Publics.begin(), Publics.end(), 823 [](const PublicSym32 &L, const PublicSym32 &R) { 824 return L.Name < R.Name; 825 }); 826 for (const PublicSym32 &Pub : Publics) 827 GsiBuilder.addPublicSymbol(Pub); 828 } 829} 830 831static void addCommonLinkerModuleSymbols(StringRef Path, 832 pdb::DbiModuleDescriptorBuilder &Mod, 833 BumpPtrAllocator &Allocator) { 834 ObjNameSym ONS(SymbolRecordKind::ObjNameSym); 835 Compile3Sym CS(SymbolRecordKind::Compile3Sym); 836 EnvBlockSym EBS(SymbolRecordKind::EnvBlockSym); 837 838 ONS.Name = "* Linker *"; 839 ONS.Signature = 0; 840 841 CS.Machine = Config->is64() ? CPUType::X64 : CPUType::Intel80386; 842 // Interestingly, if we set the string to 0.0.0.0, then when trying to view 843 // local variables WinDbg emits an error that private symbols are not present. 844 // By setting this to a valid MSVC linker version string, local variables are 845 // displayed properly. As such, even though it is not representative of 846 // LLVM's version information, we need this for compatibility. 847 CS.Flags = CompileSym3Flags::None; 848 CS.VersionBackendBuild = 25019; 849 CS.VersionBackendMajor = 14; 850 CS.VersionBackendMinor = 10; 851 CS.VersionBackendQFE = 0; 852 853 // MSVC also sets the frontend to 0.0.0.0 since this is specifically for the 854 // linker module (which is by definition a backend), so we don't need to do 855 // anything here. Also, it seems we can use "LLVM Linker" for the linker name 856 // without any problems. Only the backend version has to be hardcoded to a 857 // magic number. 858 CS.VersionFrontendBuild = 0; 859 CS.VersionFrontendMajor = 0; 860 CS.VersionFrontendMinor = 0; 861 CS.VersionFrontendQFE = 0; 862 CS.Version = "LLVM Linker"; 863 CS.setLanguage(SourceLanguage::Link); 864 865 ArrayRef<StringRef> Args = makeArrayRef(Config->Argv).drop_front(); 866 std::string ArgStr = llvm::join(Args, " "); 867 EBS.Fields.push_back("cwd"); 868 SmallString<64> cwd; 869 sys::fs::current_path(cwd); 870 EBS.Fields.push_back(cwd); 871 EBS.Fields.push_back("exe"); 872 SmallString<64> exe = Config->Argv[0]; 873 llvm::sys::fs::make_absolute(exe); 874 EBS.Fields.push_back(exe); 875 EBS.Fields.push_back("pdb"); 876 EBS.Fields.push_back(Path); 877 EBS.Fields.push_back("cmd"); 878 EBS.Fields.push_back(ArgStr); 879 Mod.addSymbol(codeview::SymbolSerializer::writeOneSymbol( 880 ONS, Allocator, CodeViewContainer::Pdb)); 881 Mod.addSymbol(codeview::SymbolSerializer::writeOneSymbol( 882 CS, Allocator, CodeViewContainer::Pdb)); 883 Mod.addSymbol(codeview::SymbolSerializer::writeOneSymbol( 884 EBS, Allocator, CodeViewContainer::Pdb)); 885} 886 887static void addLinkerModuleSectionSymbol(pdb::DbiModuleDescriptorBuilder &Mod, 888 OutputSection &OS, 889 BumpPtrAllocator &Allocator) { 890 SectionSym Sym(SymbolRecordKind::SectionSym); 891 Sym.Alignment = 12; // 2^12 = 4KB 892 Sym.Characteristics = OS.getCharacteristics(); 893 Sym.Length = OS.getVirtualSize(); 894 Sym.Name = OS.getName(); 895 Sym.Rva = OS.getRVA(); 896 Sym.SectionNumber = OS.SectionIndex; 897 Mod.addSymbol(codeview::SymbolSerializer::writeOneSymbol( 898 Sym, Allocator, CodeViewContainer::Pdb)); 899} 900 901// Creates a PDB file. 902void coff::createPDB(SymbolTable *Symtab, 903 ArrayRef<OutputSection *> OutputSections, 904 ArrayRef<uint8_t> SectionTable, 905 const llvm::codeview::DebugInfo &BuildId) { 906 PDBLinker PDB(Symtab); 907 PDB.initialize(BuildId); 908 PDB.addObjectsToPDB(); 909 PDB.addSections(OutputSections, SectionTable); 910 PDB.commit(); 911} 912 913void PDBLinker::initialize(const llvm::codeview::DebugInfo &BuildId) { 914 ExitOnErr(Builder.initialize(4096)); // 4096 is blocksize 915 916 // Create streams in MSF for predefined streams, namely 917 // PDB, TPI, DBI and IPI. 918 for (int I = 0; I < (int)pdb::kSpecialStreamCount; ++I) 919 ExitOnErr(Builder.getMsfBuilder().addStream(0)); 920 921 // Add an Info stream. 922 auto &InfoBuilder = Builder.getInfoBuilder(); 923 InfoBuilder.setAge(BuildId.PDB70.Age); 924 925 GUID uuid; 926 memcpy(&uuid, &BuildId.PDB70.Signature, sizeof(uuid)); 927 InfoBuilder.setGuid(uuid); 928 InfoBuilder.setSignature(time(nullptr)); 929 InfoBuilder.setVersion(pdb::PdbRaw_ImplVer::PdbImplVC70); 930 931 // Add an empty DBI stream. 932 pdb::DbiStreamBuilder &DbiBuilder = Builder.getDbiBuilder(); 933 DbiBuilder.setAge(BuildId.PDB70.Age); 934 DbiBuilder.setVersionHeader(pdb::PdbDbiV70); 935 ExitOnErr(DbiBuilder.addDbgStream(pdb::DbgHeaderType::NewFPO, {})); 936} 937 938void PDBLinker::addSectionContrib(pdb::DbiModuleDescriptorBuilder &LinkerModule, 939 OutputSection *OS, Chunk *C) { 940 pdb::SectionContrib SC; 941 memset(&SC, 0, sizeof(SC)); 942 SC.ISect = OS->SectionIndex; 943 SC.Off = C->getRVA() - OS->getRVA(); 944 SC.Size = C->getSize(); 945 if (auto *SecChunk = dyn_cast<SectionChunk>(C)) { 946 SC.Characteristics = SecChunk->Header->Characteristics; 947 SC.Imod = SecChunk->File->ModuleDBI->getModuleIndex(); 948 ArrayRef<uint8_t> Contents = SecChunk->getContents(); 949 JamCRC CRC(0); 950 ArrayRef<char> CharContents = makeArrayRef( 951 reinterpret_cast<const char *>(Contents.data()), Contents.size()); 952 CRC.update(CharContents); 953 SC.DataCrc = CRC.getCRC(); 954 } else { 955 SC.Characteristics = OS->getCharacteristics(); 956 // FIXME: When we start creating DBI for import libraries, use those here. 957 SC.Imod = LinkerModule.getModuleIndex(); 958 } 959 SC.RelocCrc = 0; // FIXME 960 Builder.getDbiBuilder().addSectionContrib(SC); 961} 962 963void PDBLinker::addSections(ArrayRef<OutputSection *> OutputSections, 964 ArrayRef<uint8_t> SectionTable) { 965 // It's not entirely clear what this is, but the * Linker * module uses it. 966 pdb::DbiStreamBuilder &DbiBuilder = Builder.getDbiBuilder(); 967 NativePath = Config->PDBPath; 968 sys::fs::make_absolute(NativePath); 969 sys::path::native(NativePath, sys::path::Style::windows); 970 uint32_t PdbFilePathNI = DbiBuilder.addECName(NativePath); 971 auto &LinkerModule = ExitOnErr(DbiBuilder.addModuleInfo("* Linker *")); 972 LinkerModule.setPdbFilePathNI(PdbFilePathNI); 973 addCommonLinkerModuleSymbols(NativePath, LinkerModule, Alloc); 974 975 // Add section contributions. They must be ordered by ascending RVA. 976 for (OutputSection *OS : OutputSections) { 977 addLinkerModuleSectionSymbol(LinkerModule, *OS, Alloc); 978 for (Chunk *C : OS->getChunks()) 979 addSectionContrib(LinkerModule, OS, C); 980 } 981 982 // Add Section Map stream. 983 ArrayRef<object::coff_section> Sections = { 984 (const object::coff_section *)SectionTable.data(), 985 SectionTable.size() / sizeof(object::coff_section)}; 986 SectionMap = pdb::DbiStreamBuilder::createSectionMap(Sections); 987 DbiBuilder.setSectionMap(SectionMap); 988 989 // Add COFF section header stream. 990 ExitOnErr( 991 DbiBuilder.addDbgStream(pdb::DbgHeaderType::SectionHdr, SectionTable)); 992} 993 994void PDBLinker::commit() { 995 // Write to a file. 996 ExitOnErr(Builder.commit(Config->PDBPath)); 997} 998