1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 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 contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "DwarfExpression.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/ArrayRef.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/DenseSet.h" 19 #include "llvm/ADT/MapVector.h" 20 #include "llvm/ADT/None.h" 21 #include "llvm/ADT/Optional.h" 22 #include "llvm/ADT/STLExtras.h" 23 #include "llvm/ADT/SmallString.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/StringRef.h" 26 #include "llvm/ADT/TinyPtrVector.h" 27 #include "llvm/ADT/Triple.h" 28 #include "llvm/ADT/Twine.h" 29 #include "llvm/BinaryFormat/COFF.h" 30 #include "llvm/BinaryFormat/Dwarf.h" 31 #include "llvm/CodeGen/AsmPrinter.h" 32 #include "llvm/CodeGen/LexicalScopes.h" 33 #include "llvm/CodeGen/MachineFrameInfo.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineInstr.h" 36 #include "llvm/CodeGen/MachineModuleInfo.h" 37 #include "llvm/CodeGen/MachineOperand.h" 38 #include "llvm/CodeGen/TargetFrameLowering.h" 39 #include "llvm/CodeGen/TargetRegisterInfo.h" 40 #include "llvm/CodeGen/TargetSubtargetInfo.h" 41 #include "llvm/Config/llvm-config.h" 42 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 43 #include "llvm/DebugInfo/CodeView/CodeView.h" 44 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 45 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 46 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 47 #include "llvm/DebugInfo/CodeView/EnumTables.h" 48 #include "llvm/DebugInfo/CodeView/Line.h" 49 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 50 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 51 #include "llvm/DebugInfo/CodeView/TypeIndex.h" 52 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 53 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 54 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 55 #include "llvm/IR/Constants.h" 56 #include "llvm/IR/DataLayout.h" 57 #include "llvm/IR/DebugInfoMetadata.h" 58 #include "llvm/IR/DebugLoc.h" 59 #include "llvm/IR/Function.h" 60 #include "llvm/IR/GlobalValue.h" 61 #include "llvm/IR/GlobalVariable.h" 62 #include "llvm/IR/Metadata.h" 63 #include "llvm/IR/Module.h" 64 #include "llvm/MC/MCAsmInfo.h" 65 #include "llvm/MC/MCContext.h" 66 #include "llvm/MC/MCSectionCOFF.h" 67 #include "llvm/MC/MCStreamer.h" 68 #include "llvm/MC/MCSymbol.h" 69 #include "llvm/Support/BinaryByteStream.h" 70 #include "llvm/Support/BinaryStreamReader.h" 71 #include "llvm/Support/BinaryStreamWriter.h" 72 #include "llvm/Support/Casting.h" 73 #include "llvm/Support/CommandLine.h" 74 #include "llvm/Support/Compiler.h" 75 #include "llvm/Support/Endian.h" 76 #include "llvm/Support/Error.h" 77 #include "llvm/Support/ErrorHandling.h" 78 #include "llvm/Support/FormatVariadic.h" 79 #include "llvm/Support/Path.h" 80 #include "llvm/Support/SMLoc.h" 81 #include "llvm/Support/ScopedPrinter.h" 82 #include "llvm/Target/TargetLoweringObjectFile.h" 83 #include "llvm/Target/TargetMachine.h" 84 #include <algorithm> 85 #include <cassert> 86 #include <cctype> 87 #include <cstddef> 88 #include <cstdint> 89 #include <iterator> 90 #include <limits> 91 #include <string> 92 #include <utility> 93 #include <vector> 94 95 using namespace llvm; 96 using namespace llvm::codeview; 97 98 namespace { 99 class CVMCAdapter : public CodeViewRecordStreamer { 100 public: 101 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 102 : OS(&OS), TypeTable(TypeTable) {} 103 104 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 105 106 void emitIntValue(uint64_t Value, unsigned Size) override { 107 OS->emitIntValueInHex(Value, Size); 108 } 109 110 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 111 112 void AddComment(const Twine &T) override { OS->AddComment(T); } 113 114 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 115 116 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 117 118 std::string getTypeName(TypeIndex TI) override { 119 std::string TypeName; 120 if (!TI.isNoneType()) { 121 if (TI.isSimple()) 122 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 123 else 124 TypeName = std::string(TypeTable.getTypeName(TI)); 125 } 126 return TypeName; 127 } 128 129 private: 130 MCStreamer *OS = nullptr; 131 TypeCollection &TypeTable; 132 }; 133 } // namespace 134 135 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 136 switch (Type) { 137 case Triple::ArchType::x86: 138 return CPUType::Pentium3; 139 case Triple::ArchType::x86_64: 140 return CPUType::X64; 141 case Triple::ArchType::thumb: 142 return CPUType::Thumb; 143 case Triple::ArchType::aarch64: 144 return CPUType::ARM64; 145 default: 146 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 147 } 148 } 149 150 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 151 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 152 // If module doesn't have named metadata anchors or COFF debug section 153 // is not available, skip any debug info related stuff. 154 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 155 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 156 Asm = nullptr; 157 MMI->setDebugInfoAvailability(false); 158 return; 159 } 160 // Tell MMI that we have debug info. 161 MMI->setDebugInfoAvailability(true); 162 163 TheCPU = 164 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 165 166 collectGlobalVariableInfo(); 167 168 // Check if we should emit type record hashes. 169 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 170 MMI->getModule()->getModuleFlag("CodeViewGHash")); 171 EmitDebugGlobalHashes = GH && !GH->isZero(); 172 } 173 174 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 175 std::string &Filepath = FileToFilepathMap[File]; 176 if (!Filepath.empty()) 177 return Filepath; 178 179 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 180 181 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 182 // it textually because one of the path components could be a symlink. 183 if (Dir.startswith("/") || Filename.startswith("/")) { 184 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 185 return Filename; 186 Filepath = std::string(Dir); 187 if (Dir.back() != '/') 188 Filepath += '/'; 189 Filepath += Filename; 190 return Filepath; 191 } 192 193 // Clang emits directory and relative filename info into the IR, but CodeView 194 // operates on full paths. We could change Clang to emit full paths too, but 195 // that would increase the IR size and probably not needed for other users. 196 // For now, just concatenate and canonicalize the path here. 197 if (Filename.find(':') == 1) 198 Filepath = std::string(Filename); 199 else 200 Filepath = (Dir + "\\" + Filename).str(); 201 202 // Canonicalize the path. We have to do it textually because we may no longer 203 // have access the file in the filesystem. 204 // First, replace all slashes with backslashes. 205 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 206 207 // Remove all "\.\" with "\". 208 size_t Cursor = 0; 209 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 210 Filepath.erase(Cursor, 2); 211 212 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 213 // path should be well-formatted, e.g. start with a drive letter, etc. 214 Cursor = 0; 215 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 216 // Something's wrong if the path starts with "\..\", abort. 217 if (Cursor == 0) 218 break; 219 220 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 221 if (PrevSlash == std::string::npos) 222 // Something's wrong, abort. 223 break; 224 225 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 226 // The next ".." might be following the one we've just erased. 227 Cursor = PrevSlash; 228 } 229 230 // Remove all duplicate backslashes. 231 Cursor = 0; 232 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 233 Filepath.erase(Cursor, 1); 234 235 return Filepath; 236 } 237 238 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 239 StringRef FullPath = getFullFilepath(F); 240 unsigned NextId = FileIdMap.size() + 1; 241 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 242 if (Insertion.second) { 243 // We have to compute the full filepath and emit a .cv_file directive. 244 ArrayRef<uint8_t> ChecksumAsBytes; 245 FileChecksumKind CSKind = FileChecksumKind::None; 246 if (F->getChecksum()) { 247 std::string Checksum = fromHex(F->getChecksum()->Value); 248 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 249 memcpy(CKMem, Checksum.data(), Checksum.size()); 250 ChecksumAsBytes = ArrayRef<uint8_t>( 251 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 252 switch (F->getChecksum()->Kind) { 253 case DIFile::CSK_MD5: 254 CSKind = FileChecksumKind::MD5; 255 break; 256 case DIFile::CSK_SHA1: 257 CSKind = FileChecksumKind::SHA1; 258 break; 259 case DIFile::CSK_SHA256: 260 CSKind = FileChecksumKind::SHA256; 261 break; 262 } 263 } 264 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 265 static_cast<unsigned>(CSKind)); 266 (void)Success; 267 assert(Success && ".cv_file directive failed"); 268 } 269 return Insertion.first->second; 270 } 271 272 CodeViewDebug::InlineSite & 273 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 274 const DISubprogram *Inlinee) { 275 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 276 InlineSite *Site = &SiteInsertion.first->second; 277 if (SiteInsertion.second) { 278 unsigned ParentFuncId = CurFn->FuncId; 279 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 280 ParentFuncId = 281 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 282 .SiteFuncId; 283 284 Site->SiteFuncId = NextFuncId++; 285 OS.EmitCVInlineSiteIdDirective( 286 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 287 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 288 Site->Inlinee = Inlinee; 289 InlinedSubprograms.insert(Inlinee); 290 getFuncIdForSubprogram(Inlinee); 291 } 292 return *Site; 293 } 294 295 static StringRef getPrettyScopeName(const DIScope *Scope) { 296 StringRef ScopeName = Scope->getName(); 297 if (!ScopeName.empty()) 298 return ScopeName; 299 300 switch (Scope->getTag()) { 301 case dwarf::DW_TAG_enumeration_type: 302 case dwarf::DW_TAG_class_type: 303 case dwarf::DW_TAG_structure_type: 304 case dwarf::DW_TAG_union_type: 305 return "<unnamed-tag>"; 306 case dwarf::DW_TAG_namespace: 307 return "`anonymous namespace'"; 308 } 309 310 return StringRef(); 311 } 312 313 const DISubprogram *CodeViewDebug::collectParentScopeNames( 314 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 315 const DISubprogram *ClosestSubprogram = nullptr; 316 while (Scope != nullptr) { 317 if (ClosestSubprogram == nullptr) 318 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 319 320 // If a type appears in a scope chain, make sure it gets emitted. The 321 // frontend will be responsible for deciding if this should be a forward 322 // declaration or a complete type. 323 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 324 DeferredCompleteTypes.push_back(Ty); 325 326 StringRef ScopeName = getPrettyScopeName(Scope); 327 if (!ScopeName.empty()) 328 QualifiedNameComponents.push_back(ScopeName); 329 Scope = Scope->getScope(); 330 } 331 return ClosestSubprogram; 332 } 333 334 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 335 StringRef TypeName) { 336 std::string FullyQualifiedName; 337 for (StringRef QualifiedNameComponent : 338 llvm::reverse(QualifiedNameComponents)) { 339 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 340 FullyQualifiedName.append("::"); 341 } 342 FullyQualifiedName.append(std::string(TypeName)); 343 return FullyQualifiedName; 344 } 345 346 struct CodeViewDebug::TypeLoweringScope { 347 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 348 ~TypeLoweringScope() { 349 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 350 // inner TypeLoweringScopes don't attempt to emit deferred types. 351 if (CVD.TypeEmissionLevel == 1) 352 CVD.emitDeferredCompleteTypes(); 353 --CVD.TypeEmissionLevel; 354 } 355 CodeViewDebug &CVD; 356 }; 357 358 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 359 StringRef Name) { 360 // Ensure types in the scope chain are emitted as soon as possible. 361 // This can create otherwise a situation where S_UDTs are emitted while 362 // looping in emitDebugInfoForUDTs. 363 TypeLoweringScope S(*this); 364 SmallVector<StringRef, 5> QualifiedNameComponents; 365 collectParentScopeNames(Scope, QualifiedNameComponents); 366 return formatNestedName(QualifiedNameComponents, Name); 367 } 368 369 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 370 const DIScope *Scope = Ty->getScope(); 371 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 372 } 373 374 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 375 // No scope means global scope and that uses the zero index. 376 if (!Scope || isa<DIFile>(Scope)) 377 return TypeIndex(); 378 379 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 380 381 // Check if we've already translated this scope. 382 auto I = TypeIndices.find({Scope, nullptr}); 383 if (I != TypeIndices.end()) 384 return I->second; 385 386 // Build the fully qualified name of the scope. 387 std::string ScopeName = getFullyQualifiedName(Scope); 388 StringIdRecord SID(TypeIndex(), ScopeName); 389 auto TI = TypeTable.writeLeafType(SID); 390 return recordTypeIndexForDINode(Scope, TI); 391 } 392 393 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 394 assert(SP); 395 396 // Check if we've already translated this subprogram. 397 auto I = TypeIndices.find({SP, nullptr}); 398 if (I != TypeIndices.end()) 399 return I->second; 400 401 // The display name includes function template arguments. Drop them to match 402 // MSVC. 403 StringRef DisplayName = SP->getName().split('<').first; 404 405 const DIScope *Scope = SP->getScope(); 406 TypeIndex TI; 407 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 408 // If the scope is a DICompositeType, then this must be a method. Member 409 // function types take some special handling, and require access to the 410 // subprogram. 411 TypeIndex ClassType = getTypeIndex(Class); 412 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 413 DisplayName); 414 TI = TypeTable.writeLeafType(MFuncId); 415 } else { 416 // Otherwise, this must be a free function. 417 TypeIndex ParentScope = getScopeIndex(Scope); 418 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 419 TI = TypeTable.writeLeafType(FuncId); 420 } 421 422 return recordTypeIndexForDINode(SP, TI); 423 } 424 425 static bool isNonTrivial(const DICompositeType *DCTy) { 426 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 427 } 428 429 static FunctionOptions 430 getFunctionOptions(const DISubroutineType *Ty, 431 const DICompositeType *ClassTy = nullptr, 432 StringRef SPName = StringRef("")) { 433 FunctionOptions FO = FunctionOptions::None; 434 const DIType *ReturnTy = nullptr; 435 if (auto TypeArray = Ty->getTypeArray()) { 436 if (TypeArray.size()) 437 ReturnTy = TypeArray[0]; 438 } 439 440 // Add CxxReturnUdt option to functions that return nontrivial record types 441 // or methods that return record types. 442 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 443 if (isNonTrivial(ReturnDCTy) || ClassTy) 444 FO |= FunctionOptions::CxxReturnUdt; 445 446 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 447 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 448 FO |= FunctionOptions::Constructor; 449 450 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 451 452 } 453 return FO; 454 } 455 456 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 457 const DICompositeType *Class) { 458 // Always use the method declaration as the key for the function type. The 459 // method declaration contains the this adjustment. 460 if (SP->getDeclaration()) 461 SP = SP->getDeclaration(); 462 assert(!SP->getDeclaration() && "should use declaration as key"); 463 464 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 465 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 466 auto I = TypeIndices.find({SP, Class}); 467 if (I != TypeIndices.end()) 468 return I->second; 469 470 // Make sure complete type info for the class is emitted *after* the member 471 // function type, as the complete class type is likely to reference this 472 // member function type. 473 TypeLoweringScope S(*this); 474 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 475 476 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 477 TypeIndex TI = lowerTypeMemberFunction( 478 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 479 return recordTypeIndexForDINode(SP, TI, Class); 480 } 481 482 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 483 TypeIndex TI, 484 const DIType *ClassTy) { 485 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 486 (void)InsertResult; 487 assert(InsertResult.second && "DINode was already assigned a type index"); 488 return TI; 489 } 490 491 unsigned CodeViewDebug::getPointerSizeInBytes() { 492 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 493 } 494 495 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 496 const LexicalScope *LS) { 497 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 498 // This variable was inlined. Associate it with the InlineSite. 499 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 500 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 501 Site.InlinedLocals.emplace_back(Var); 502 } else { 503 // This variable goes into the corresponding lexical scope. 504 ScopeVariables[LS].emplace_back(Var); 505 } 506 } 507 508 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 509 const DILocation *Loc) { 510 auto B = Locs.begin(), E = Locs.end(); 511 if (std::find(B, E, Loc) == E) 512 Locs.push_back(Loc); 513 } 514 515 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 516 const MachineFunction *MF) { 517 // Skip this instruction if it has the same location as the previous one. 518 if (!DL || DL == PrevInstLoc) 519 return; 520 521 const DIScope *Scope = DL.get()->getScope(); 522 if (!Scope) 523 return; 524 525 // Skip this line if it is longer than the maximum we can record. 526 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 527 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 528 LI.isNeverStepInto()) 529 return; 530 531 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 532 if (CI.getStartColumn() != DL.getCol()) 533 return; 534 535 if (!CurFn->HaveLineInfo) 536 CurFn->HaveLineInfo = true; 537 unsigned FileId = 0; 538 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 539 FileId = CurFn->LastFileId; 540 else 541 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 542 PrevInstLoc = DL; 543 544 unsigned FuncId = CurFn->FuncId; 545 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 546 const DILocation *Loc = DL.get(); 547 548 // If this location was actually inlined from somewhere else, give it the ID 549 // of the inline call site. 550 FuncId = 551 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 552 553 // Ensure we have links in the tree of inline call sites. 554 bool FirstLoc = true; 555 while ((SiteLoc = Loc->getInlinedAt())) { 556 InlineSite &Site = 557 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 558 if (!FirstLoc) 559 addLocIfNotPresent(Site.ChildSites, Loc); 560 FirstLoc = false; 561 Loc = SiteLoc; 562 } 563 addLocIfNotPresent(CurFn->ChildSites, Loc); 564 } 565 566 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 567 /*PrologueEnd=*/false, /*IsStmt=*/false, 568 DL->getFilename(), SMLoc()); 569 } 570 571 void CodeViewDebug::emitCodeViewMagicVersion() { 572 OS.emitValueToAlignment(4); 573 OS.AddComment("Debug section magic"); 574 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 575 } 576 577 void CodeViewDebug::endModule() { 578 if (!Asm || !MMI->hasDebugInfo()) 579 return; 580 581 assert(Asm != nullptr); 582 583 // The COFF .debug$S section consists of several subsections, each starting 584 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 585 // of the payload followed by the payload itself. The subsections are 4-byte 586 // aligned. 587 588 // Use the generic .debug$S section, and make a subsection for all the inlined 589 // subprograms. 590 switchToDebugSectionForSymbol(nullptr); 591 592 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 593 emitCompilerInformation(); 594 endCVSubsection(CompilerInfo); 595 596 emitInlineeLinesSubsection(); 597 598 // Emit per-function debug information. 599 for (auto &P : FnDebugInfo) 600 if (!P.first->isDeclarationForLinker()) 601 emitDebugInfoForFunction(P.first, *P.second); 602 603 // Emit global variable debug information. 604 setCurrentSubprogram(nullptr); 605 emitDebugInfoForGlobals(); 606 607 // Emit retained types. 608 emitDebugInfoForRetainedTypes(); 609 610 // Switch back to the generic .debug$S section after potentially processing 611 // comdat symbol sections. 612 switchToDebugSectionForSymbol(nullptr); 613 614 // Emit UDT records for any types used by global variables. 615 if (!GlobalUDTs.empty()) { 616 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 617 emitDebugInfoForUDTs(GlobalUDTs); 618 endCVSubsection(SymbolsEnd); 619 } 620 621 // This subsection holds a file index to offset in string table table. 622 OS.AddComment("File index to string table offset subsection"); 623 OS.emitCVFileChecksumsDirective(); 624 625 // This subsection holds the string table. 626 OS.AddComment("String table"); 627 OS.emitCVStringTableDirective(); 628 629 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 630 // subsection in the generic .debug$S section at the end. There is no 631 // particular reason for this ordering other than to match MSVC. 632 emitBuildInfo(); 633 634 // Emit type information and hashes last, so that any types we translate while 635 // emitting function info are included. 636 emitTypeInformation(); 637 638 if (EmitDebugGlobalHashes) 639 emitTypeGlobalHashes(); 640 641 clear(); 642 } 643 644 static void 645 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 646 unsigned MaxFixedRecordLength = 0xF00) { 647 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 648 // after a fixed length portion of the record. The fixed length portion should 649 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 650 // overall record size is less than the maximum allowed. 651 SmallString<32> NullTerminatedString( 652 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 653 NullTerminatedString.push_back('\0'); 654 OS.emitBytes(NullTerminatedString); 655 } 656 657 void CodeViewDebug::emitTypeInformation() { 658 if (TypeTable.empty()) 659 return; 660 661 // Start the .debug$T or .debug$P section with 0x4. 662 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 663 emitCodeViewMagicVersion(); 664 665 TypeTableCollection Table(TypeTable.records()); 666 TypeVisitorCallbackPipeline Pipeline; 667 668 // To emit type record using Codeview MCStreamer adapter 669 CVMCAdapter CVMCOS(OS, Table); 670 TypeRecordMapping typeMapping(CVMCOS); 671 Pipeline.addCallbackToPipeline(typeMapping); 672 673 Optional<TypeIndex> B = Table.getFirst(); 674 while (B) { 675 // This will fail if the record data is invalid. 676 CVType Record = Table.getType(*B); 677 678 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 679 680 if (E) { 681 logAllUnhandledErrors(std::move(E), errs(), "error: "); 682 llvm_unreachable("produced malformed type record"); 683 } 684 685 B = Table.getNext(*B); 686 } 687 } 688 689 void CodeViewDebug::emitTypeGlobalHashes() { 690 if (TypeTable.empty()) 691 return; 692 693 // Start the .debug$H section with the version and hash algorithm, currently 694 // hardcoded to version 0, SHA1. 695 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 696 697 OS.emitValueToAlignment(4); 698 OS.AddComment("Magic"); 699 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 700 OS.AddComment("Section Version"); 701 OS.emitInt16(0); 702 OS.AddComment("Hash Algorithm"); 703 OS.emitInt16(uint16_t(GlobalTypeHashAlg::SHA1_8)); 704 705 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 706 for (const auto &GHR : TypeTable.hashes()) { 707 if (OS.isVerboseAsm()) { 708 // Emit an EOL-comment describing which TypeIndex this hash corresponds 709 // to, as well as the stringified SHA1 hash. 710 SmallString<32> Comment; 711 raw_svector_ostream CommentOS(Comment); 712 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 713 OS.AddComment(Comment); 714 ++TI; 715 } 716 assert(GHR.Hash.size() == 8); 717 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 718 GHR.Hash.size()); 719 OS.emitBinaryData(S); 720 } 721 } 722 723 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 724 switch (DWLang) { 725 case dwarf::DW_LANG_C: 726 case dwarf::DW_LANG_C89: 727 case dwarf::DW_LANG_C99: 728 case dwarf::DW_LANG_C11: 729 case dwarf::DW_LANG_ObjC: 730 return SourceLanguage::C; 731 case dwarf::DW_LANG_C_plus_plus: 732 case dwarf::DW_LANG_C_plus_plus_03: 733 case dwarf::DW_LANG_C_plus_plus_11: 734 case dwarf::DW_LANG_C_plus_plus_14: 735 return SourceLanguage::Cpp; 736 case dwarf::DW_LANG_Fortran77: 737 case dwarf::DW_LANG_Fortran90: 738 case dwarf::DW_LANG_Fortran03: 739 case dwarf::DW_LANG_Fortran08: 740 return SourceLanguage::Fortran; 741 case dwarf::DW_LANG_Pascal83: 742 return SourceLanguage::Pascal; 743 case dwarf::DW_LANG_Cobol74: 744 case dwarf::DW_LANG_Cobol85: 745 return SourceLanguage::Cobol; 746 case dwarf::DW_LANG_Java: 747 return SourceLanguage::Java; 748 case dwarf::DW_LANG_D: 749 return SourceLanguage::D; 750 case dwarf::DW_LANG_Swift: 751 return SourceLanguage::Swift; 752 default: 753 // There's no CodeView representation for this language, and CV doesn't 754 // have an "unknown" option for the language field, so we'll use MASM, 755 // as it's very low level. 756 return SourceLanguage::Masm; 757 } 758 } 759 760 namespace { 761 struct Version { 762 int Part[4]; 763 }; 764 } // end anonymous namespace 765 766 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 767 // the version number. 768 static Version parseVersion(StringRef Name) { 769 Version V = {{0}}; 770 int N = 0; 771 for (const char C : Name) { 772 if (isdigit(C)) { 773 V.Part[N] *= 10; 774 V.Part[N] += C - '0'; 775 } else if (C == '.') { 776 ++N; 777 if (N >= 4) 778 return V; 779 } else if (N > 0) 780 return V; 781 } 782 return V; 783 } 784 785 void CodeViewDebug::emitCompilerInformation() { 786 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 787 uint32_t Flags = 0; 788 789 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 790 const MDNode *Node = *CUs->operands().begin(); 791 const auto *CU = cast<DICompileUnit>(Node); 792 793 // The low byte of the flags indicates the source language. 794 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 795 // TODO: Figure out which other flags need to be set. 796 797 OS.AddComment("Flags and language"); 798 OS.emitInt32(Flags); 799 800 OS.AddComment("CPUType"); 801 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 802 803 StringRef CompilerVersion = CU->getProducer(); 804 Version FrontVer = parseVersion(CompilerVersion); 805 OS.AddComment("Frontend version"); 806 for (int N = 0; N < 4; ++N) 807 OS.emitInt16(FrontVer.Part[N]); 808 809 // Some Microsoft tools, like Binscope, expect a backend version number of at 810 // least 8.something, so we'll coerce the LLVM version into a form that 811 // guarantees it'll be big enough without really lying about the version. 812 int Major = 1000 * LLVM_VERSION_MAJOR + 813 10 * LLVM_VERSION_MINOR + 814 LLVM_VERSION_PATCH; 815 // Clamp it for builds that use unusually large version numbers. 816 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 817 Version BackVer = {{ Major, 0, 0, 0 }}; 818 OS.AddComment("Backend version"); 819 for (int N = 0; N < 4; ++N) 820 OS.emitInt16(BackVer.Part[N]); 821 822 OS.AddComment("Null-terminated compiler version string"); 823 emitNullTerminatedSymbolName(OS, CompilerVersion); 824 825 endSymbolRecord(CompilerEnd); 826 } 827 828 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 829 StringRef S) { 830 StringIdRecord SIR(TypeIndex(0x0), S); 831 return TypeTable.writeLeafType(SIR); 832 } 833 834 void CodeViewDebug::emitBuildInfo() { 835 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 836 // build info. The known prefix is: 837 // - Absolute path of current directory 838 // - Compiler path 839 // - Main source file path, relative to CWD or absolute 840 // - Type server PDB file 841 // - Canonical compiler command line 842 // If frontend and backend compilation are separated (think llc or LTO), it's 843 // not clear if the compiler path should refer to the executable for the 844 // frontend or the backend. Leave it blank for now. 845 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 846 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 847 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 848 const auto *CU = cast<DICompileUnit>(Node); 849 const DIFile *MainSourceFile = CU->getFile(); 850 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 851 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 852 BuildInfoArgs[BuildInfoRecord::SourceFile] = 853 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 854 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 855 // we implement /Zi type servers. 856 BuildInfoRecord BIR(BuildInfoArgs); 857 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 858 859 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 860 // from the module symbols into the type stream. 861 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 862 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 863 OS.AddComment("LF_BUILDINFO index"); 864 OS.emitInt32(BuildInfoIndex.getIndex()); 865 endSymbolRecord(BIEnd); 866 endCVSubsection(BISubsecEnd); 867 } 868 869 void CodeViewDebug::emitInlineeLinesSubsection() { 870 if (InlinedSubprograms.empty()) 871 return; 872 873 OS.AddComment("Inlinee lines subsection"); 874 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 875 876 // We emit the checksum info for files. This is used by debuggers to 877 // determine if a pdb matches the source before loading it. Visual Studio, 878 // for instance, will display a warning that the breakpoints are not valid if 879 // the pdb does not match the source. 880 OS.AddComment("Inlinee lines signature"); 881 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 882 883 for (const DISubprogram *SP : InlinedSubprograms) { 884 assert(TypeIndices.count({SP, nullptr})); 885 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 886 887 OS.AddBlankLine(); 888 unsigned FileId = maybeRecordFile(SP->getFile()); 889 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 890 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 891 OS.AddBlankLine(); 892 OS.AddComment("Type index of inlined function"); 893 OS.emitInt32(InlineeIdx.getIndex()); 894 OS.AddComment("Offset into filechecksum table"); 895 OS.emitCVFileChecksumOffsetDirective(FileId); 896 OS.AddComment("Starting line number"); 897 OS.emitInt32(SP->getLine()); 898 } 899 900 endCVSubsection(InlineEnd); 901 } 902 903 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 904 const DILocation *InlinedAt, 905 const InlineSite &Site) { 906 assert(TypeIndices.count({Site.Inlinee, nullptr})); 907 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 908 909 // SymbolRecord 910 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 911 912 OS.AddComment("PtrParent"); 913 OS.emitInt32(0); 914 OS.AddComment("PtrEnd"); 915 OS.emitInt32(0); 916 OS.AddComment("Inlinee type index"); 917 OS.emitInt32(InlineeIdx.getIndex()); 918 919 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 920 unsigned StartLineNum = Site.Inlinee->getLine(); 921 922 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 923 FI.Begin, FI.End); 924 925 endSymbolRecord(InlineEnd); 926 927 emitLocalVariableList(FI, Site.InlinedLocals); 928 929 // Recurse on child inlined call sites before closing the scope. 930 for (const DILocation *ChildSite : Site.ChildSites) { 931 auto I = FI.InlineSites.find(ChildSite); 932 assert(I != FI.InlineSites.end() && 933 "child site not in function inline site map"); 934 emitInlinedCallSite(FI, ChildSite, I->second); 935 } 936 937 // Close the scope. 938 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 939 } 940 941 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 942 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 943 // comdat key. A section may be comdat because of -ffunction-sections or 944 // because it is comdat in the IR. 945 MCSectionCOFF *GVSec = 946 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 947 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 948 949 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 950 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 951 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 952 953 OS.SwitchSection(DebugSec); 954 955 // Emit the magic version number if this is the first time we've switched to 956 // this section. 957 if (ComdatDebugSections.insert(DebugSec).second) 958 emitCodeViewMagicVersion(); 959 } 960 961 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 962 // The only supported thunk ordinal is currently the standard type. 963 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 964 FunctionInfo &FI, 965 const MCSymbol *Fn) { 966 std::string FuncName = 967 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 968 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 969 970 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 971 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 972 973 // Emit S_THUNK32 974 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 975 OS.AddComment("PtrParent"); 976 OS.emitInt32(0); 977 OS.AddComment("PtrEnd"); 978 OS.emitInt32(0); 979 OS.AddComment("PtrNext"); 980 OS.emitInt32(0); 981 OS.AddComment("Thunk section relative address"); 982 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 983 OS.AddComment("Thunk section index"); 984 OS.EmitCOFFSectionIndex(Fn); 985 OS.AddComment("Code size"); 986 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 987 OS.AddComment("Ordinal"); 988 OS.emitInt8(unsigned(ordinal)); 989 OS.AddComment("Function name"); 990 emitNullTerminatedSymbolName(OS, FuncName); 991 // Additional fields specific to the thunk ordinal would go here. 992 endSymbolRecord(ThunkRecordEnd); 993 994 // Local variables/inlined routines are purposely omitted here. The point of 995 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 996 997 // Emit S_PROC_ID_END 998 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 999 1000 endCVSubsection(SymbolsEnd); 1001 } 1002 1003 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1004 FunctionInfo &FI) { 1005 // For each function there is a separate subsection which holds the PC to 1006 // file:line table. 1007 const MCSymbol *Fn = Asm->getSymbol(GV); 1008 assert(Fn); 1009 1010 // Switch to the to a comdat section, if appropriate. 1011 switchToDebugSectionForSymbol(Fn); 1012 1013 std::string FuncName; 1014 auto *SP = GV->getSubprogram(); 1015 assert(SP); 1016 setCurrentSubprogram(SP); 1017 1018 if (SP->isThunk()) { 1019 emitDebugInfoForThunk(GV, FI, Fn); 1020 return; 1021 } 1022 1023 // If we have a display name, build the fully qualified name by walking the 1024 // chain of scopes. 1025 if (!SP->getName().empty()) 1026 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1027 1028 // If our DISubprogram name is empty, use the mangled name. 1029 if (FuncName.empty()) 1030 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1031 1032 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1033 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1034 OS.EmitCVFPOData(Fn); 1035 1036 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1037 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1038 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1039 { 1040 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1041 : SymbolKind::S_GPROC32_ID; 1042 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1043 1044 // These fields are filled in by tools like CVPACK which run after the fact. 1045 OS.AddComment("PtrParent"); 1046 OS.emitInt32(0); 1047 OS.AddComment("PtrEnd"); 1048 OS.emitInt32(0); 1049 OS.AddComment("PtrNext"); 1050 OS.emitInt32(0); 1051 // This is the important bit that tells the debugger where the function 1052 // code is located and what's its size: 1053 OS.AddComment("Code size"); 1054 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1055 OS.AddComment("Offset after prologue"); 1056 OS.emitInt32(0); 1057 OS.AddComment("Offset before epilogue"); 1058 OS.emitInt32(0); 1059 OS.AddComment("Function type index"); 1060 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1061 OS.AddComment("Function section relative address"); 1062 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1063 OS.AddComment("Function section index"); 1064 OS.EmitCOFFSectionIndex(Fn); 1065 OS.AddComment("Flags"); 1066 OS.emitInt8(0); 1067 // Emit the function display name as a null-terminated string. 1068 OS.AddComment("Function name"); 1069 // Truncate the name so we won't overflow the record length field. 1070 emitNullTerminatedSymbolName(OS, FuncName); 1071 endSymbolRecord(ProcRecordEnd); 1072 1073 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1074 // Subtract out the CSR size since MSVC excludes that and we include it. 1075 OS.AddComment("FrameSize"); 1076 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1077 OS.AddComment("Padding"); 1078 OS.emitInt32(0); 1079 OS.AddComment("Offset of padding"); 1080 OS.emitInt32(0); 1081 OS.AddComment("Bytes of callee saved registers"); 1082 OS.emitInt32(FI.CSRSize); 1083 OS.AddComment("Exception handler offset"); 1084 OS.emitInt32(0); 1085 OS.AddComment("Exception handler section"); 1086 OS.emitInt16(0); 1087 OS.AddComment("Flags (defines frame register)"); 1088 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1089 endSymbolRecord(FrameProcEnd); 1090 1091 emitLocalVariableList(FI, FI.Locals); 1092 emitGlobalVariableList(FI.Globals); 1093 emitLexicalBlockList(FI.ChildBlocks, FI); 1094 1095 // Emit inlined call site information. Only emit functions inlined directly 1096 // into the parent function. We'll emit the other sites recursively as part 1097 // of their parent inline site. 1098 for (const DILocation *InlinedAt : FI.ChildSites) { 1099 auto I = FI.InlineSites.find(InlinedAt); 1100 assert(I != FI.InlineSites.end() && 1101 "child site not in function inline site map"); 1102 emitInlinedCallSite(FI, InlinedAt, I->second); 1103 } 1104 1105 for (auto Annot : FI.Annotations) { 1106 MCSymbol *Label = Annot.first; 1107 MDTuple *Strs = cast<MDTuple>(Annot.second); 1108 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1109 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1110 // FIXME: Make sure we don't overflow the max record size. 1111 OS.EmitCOFFSectionIndex(Label); 1112 OS.emitInt16(Strs->getNumOperands()); 1113 for (Metadata *MD : Strs->operands()) { 1114 // MDStrings are null terminated, so we can do EmitBytes and get the 1115 // nice .asciz directive. 1116 StringRef Str = cast<MDString>(MD)->getString(); 1117 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1118 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1119 } 1120 endSymbolRecord(AnnotEnd); 1121 } 1122 1123 for (auto HeapAllocSite : FI.HeapAllocSites) { 1124 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1125 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1126 const DIType *DITy = std::get<2>(HeapAllocSite); 1127 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1128 OS.AddComment("Call site offset"); 1129 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1130 OS.AddComment("Call site section index"); 1131 OS.EmitCOFFSectionIndex(BeginLabel); 1132 OS.AddComment("Call instruction length"); 1133 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1134 OS.AddComment("Type index"); 1135 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1136 endSymbolRecord(HeapAllocEnd); 1137 } 1138 1139 if (SP != nullptr) 1140 emitDebugInfoForUDTs(LocalUDTs); 1141 1142 // We're done with this function. 1143 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1144 } 1145 endCVSubsection(SymbolsEnd); 1146 1147 // We have an assembler directive that takes care of the whole line table. 1148 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1149 } 1150 1151 CodeViewDebug::LocalVarDefRange 1152 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1153 LocalVarDefRange DR; 1154 DR.InMemory = -1; 1155 DR.DataOffset = Offset; 1156 assert(DR.DataOffset == Offset && "truncation"); 1157 DR.IsSubfield = 0; 1158 DR.StructOffset = 0; 1159 DR.CVRegister = CVRegister; 1160 return DR; 1161 } 1162 1163 void CodeViewDebug::collectVariableInfoFromMFTable( 1164 DenseSet<InlinedEntity> &Processed) { 1165 const MachineFunction &MF = *Asm->MF; 1166 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1167 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1168 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1169 1170 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1171 if (!VI.Var) 1172 continue; 1173 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1174 "Expected inlined-at fields to agree"); 1175 1176 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1177 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1178 1179 // If variable scope is not found then skip this variable. 1180 if (!Scope) 1181 continue; 1182 1183 // If the variable has an attached offset expression, extract it. 1184 // FIXME: Try to handle DW_OP_deref as well. 1185 int64_t ExprOffset = 0; 1186 bool Deref = false; 1187 if (VI.Expr) { 1188 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1189 if (VI.Expr->getNumElements() == 1 && 1190 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1191 Deref = true; 1192 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1193 continue; 1194 } 1195 1196 // Get the frame register used and the offset. 1197 Register FrameReg; 1198 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1199 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1200 1201 // Calculate the label ranges. 1202 LocalVarDefRange DefRange = 1203 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1204 1205 for (const InsnRange &Range : Scope->getRanges()) { 1206 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1207 const MCSymbol *End = getLabelAfterInsn(Range.second); 1208 End = End ? End : Asm->getFunctionEnd(); 1209 DefRange.Ranges.emplace_back(Begin, End); 1210 } 1211 1212 LocalVariable Var; 1213 Var.DIVar = VI.Var; 1214 Var.DefRanges.emplace_back(std::move(DefRange)); 1215 if (Deref) 1216 Var.UseReferenceType = true; 1217 1218 recordLocalVariable(std::move(Var), Scope); 1219 } 1220 } 1221 1222 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1223 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1224 } 1225 1226 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1227 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1228 } 1229 1230 void CodeViewDebug::calculateRanges( 1231 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1232 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1233 1234 // Calculate the definition ranges. 1235 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1236 const auto &Entry = *I; 1237 if (!Entry.isDbgValue()) 1238 continue; 1239 const MachineInstr *DVInst = Entry.getInstr(); 1240 assert(DVInst->isDebugValue() && "Invalid History entry"); 1241 // FIXME: Find a way to represent constant variables, since they are 1242 // relatively common. 1243 Optional<DbgVariableLocation> Location = 1244 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1245 if (!Location) 1246 continue; 1247 1248 // CodeView can only express variables in register and variables in memory 1249 // at a constant offset from a register. However, for variables passed 1250 // indirectly by pointer, it is common for that pointer to be spilled to a 1251 // stack location. For the special case of one offseted load followed by a 1252 // zero offset load (a pointer spilled to the stack), we change the type of 1253 // the local variable from a value type to a reference type. This tricks the 1254 // debugger into doing the load for us. 1255 if (Var.UseReferenceType) { 1256 // We're using a reference type. Drop the last zero offset load. 1257 if (canUseReferenceType(*Location)) 1258 Location->LoadChain.pop_back(); 1259 else 1260 continue; 1261 } else if (needsReferenceType(*Location)) { 1262 // This location can't be expressed without switching to a reference type. 1263 // Start over using that. 1264 Var.UseReferenceType = true; 1265 Var.DefRanges.clear(); 1266 calculateRanges(Var, Entries); 1267 return; 1268 } 1269 1270 // We can only handle a register or an offseted load of a register. 1271 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1272 continue; 1273 { 1274 LocalVarDefRange DR; 1275 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1276 DR.InMemory = !Location->LoadChain.empty(); 1277 DR.DataOffset = 1278 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1279 if (Location->FragmentInfo) { 1280 DR.IsSubfield = true; 1281 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1282 } else { 1283 DR.IsSubfield = false; 1284 DR.StructOffset = 0; 1285 } 1286 1287 if (Var.DefRanges.empty() || 1288 Var.DefRanges.back().isDifferentLocation(DR)) { 1289 Var.DefRanges.emplace_back(std::move(DR)); 1290 } 1291 } 1292 1293 // Compute the label range. 1294 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1295 const MCSymbol *End; 1296 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1297 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1298 End = EndingEntry.isDbgValue() 1299 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1300 : getLabelAfterInsn(EndingEntry.getInstr()); 1301 } else 1302 End = Asm->getFunctionEnd(); 1303 1304 // If the last range end is our begin, just extend the last range. 1305 // Otherwise make a new range. 1306 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1307 Var.DefRanges.back().Ranges; 1308 if (!R.empty() && R.back().second == Begin) 1309 R.back().second = End; 1310 else 1311 R.emplace_back(Begin, End); 1312 1313 // FIXME: Do more range combining. 1314 } 1315 } 1316 1317 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1318 DenseSet<InlinedEntity> Processed; 1319 // Grab the variable info that was squirreled away in the MMI side-table. 1320 collectVariableInfoFromMFTable(Processed); 1321 1322 for (const auto &I : DbgValues) { 1323 InlinedEntity IV = I.first; 1324 if (Processed.count(IV)) 1325 continue; 1326 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1327 const DILocation *InlinedAt = IV.second; 1328 1329 // Instruction ranges, specifying where IV is accessible. 1330 const auto &Entries = I.second; 1331 1332 LexicalScope *Scope = nullptr; 1333 if (InlinedAt) 1334 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1335 else 1336 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1337 // If variable scope is not found then skip this variable. 1338 if (!Scope) 1339 continue; 1340 1341 LocalVariable Var; 1342 Var.DIVar = DIVar; 1343 1344 calculateRanges(Var, Entries); 1345 recordLocalVariable(std::move(Var), Scope); 1346 } 1347 } 1348 1349 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1350 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1351 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1352 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1353 const Function &GV = MF->getFunction(); 1354 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1355 assert(Insertion.second && "function already has info"); 1356 CurFn = Insertion.first->second.get(); 1357 CurFn->FuncId = NextFuncId++; 1358 CurFn->Begin = Asm->getFunctionBegin(); 1359 1360 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1361 // callee-saved registers were used. For targets that don't use a PUSH 1362 // instruction (AArch64), this will be zero. 1363 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1364 CurFn->FrameSize = MFI.getStackSize(); 1365 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1366 CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF); 1367 1368 // For this function S_FRAMEPROC record, figure out which codeview register 1369 // will be the frame pointer. 1370 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1371 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1372 if (CurFn->FrameSize > 0) { 1373 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1374 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1375 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1376 } else { 1377 // If there is an FP, parameters are always relative to it. 1378 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1379 if (CurFn->HasStackRealignment) { 1380 // If the stack needs realignment, locals are relative to SP or VFRAME. 1381 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1382 } else { 1383 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1384 // other stack adjustments. 1385 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1386 } 1387 } 1388 } 1389 1390 // Compute other frame procedure options. 1391 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1392 if (MFI.hasVarSizedObjects()) 1393 FPO |= FrameProcedureOptions::HasAlloca; 1394 if (MF->exposesReturnsTwice()) 1395 FPO |= FrameProcedureOptions::HasSetJmp; 1396 // FIXME: Set HasLongJmp if we ever track that info. 1397 if (MF->hasInlineAsm()) 1398 FPO |= FrameProcedureOptions::HasInlineAssembly; 1399 if (GV.hasPersonalityFn()) { 1400 if (isAsynchronousEHPersonality( 1401 classifyEHPersonality(GV.getPersonalityFn()))) 1402 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1403 else 1404 FPO |= FrameProcedureOptions::HasExceptionHandling; 1405 } 1406 if (GV.hasFnAttribute(Attribute::InlineHint)) 1407 FPO |= FrameProcedureOptions::MarkedInline; 1408 if (GV.hasFnAttribute(Attribute::Naked)) 1409 FPO |= FrameProcedureOptions::Naked; 1410 if (MFI.hasStackProtectorIndex()) 1411 FPO |= FrameProcedureOptions::SecurityChecks; 1412 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1413 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1414 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1415 !GV.hasOptSize() && !GV.hasOptNone()) 1416 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1417 // FIXME: Set GuardCfg when it is implemented. 1418 CurFn->FrameProcOpts = FPO; 1419 1420 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1421 1422 // Find the end of the function prolog. First known non-DBG_VALUE and 1423 // non-frame setup location marks the beginning of the function body. 1424 // FIXME: is there a simpler a way to do this? Can we just search 1425 // for the first instruction of the function, not the last of the prolog? 1426 DebugLoc PrologEndLoc; 1427 bool EmptyPrologue = true; 1428 for (const auto &MBB : *MF) { 1429 for (const auto &MI : MBB) { 1430 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1431 MI.getDebugLoc()) { 1432 PrologEndLoc = MI.getDebugLoc(); 1433 break; 1434 } else if (!MI.isMetaInstruction()) { 1435 EmptyPrologue = false; 1436 } 1437 } 1438 } 1439 1440 // Record beginning of function if we have a non-empty prologue. 1441 if (PrologEndLoc && !EmptyPrologue) { 1442 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1443 maybeRecordLocation(FnStartDL, MF); 1444 } 1445 1446 // Find heap alloc sites and emit labels around them. 1447 for (const auto &MBB : *MF) { 1448 for (const auto &MI : MBB) { 1449 if (MI.getHeapAllocMarker()) { 1450 requestLabelBeforeInsn(&MI); 1451 requestLabelAfterInsn(&MI); 1452 } 1453 } 1454 } 1455 } 1456 1457 static bool shouldEmitUdt(const DIType *T) { 1458 if (!T) 1459 return false; 1460 1461 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1462 if (T->getTag() == dwarf::DW_TAG_typedef) { 1463 if (DIScope *Scope = T->getScope()) { 1464 switch (Scope->getTag()) { 1465 case dwarf::DW_TAG_structure_type: 1466 case dwarf::DW_TAG_class_type: 1467 case dwarf::DW_TAG_union_type: 1468 return false; 1469 } 1470 } 1471 } 1472 1473 while (true) { 1474 if (!T || T->isForwardDecl()) 1475 return false; 1476 1477 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1478 if (!DT) 1479 return true; 1480 T = DT->getBaseType(); 1481 } 1482 return true; 1483 } 1484 1485 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1486 // Don't record empty UDTs. 1487 if (Ty->getName().empty()) 1488 return; 1489 if (!shouldEmitUdt(Ty)) 1490 return; 1491 1492 SmallVector<StringRef, 5> ParentScopeNames; 1493 const DISubprogram *ClosestSubprogram = 1494 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1495 1496 std::string FullyQualifiedName = 1497 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1498 1499 if (ClosestSubprogram == nullptr) { 1500 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1501 } else if (ClosestSubprogram == CurrentSubprogram) { 1502 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1503 } 1504 1505 // TODO: What if the ClosestSubprogram is neither null or the current 1506 // subprogram? Currently, the UDT just gets dropped on the floor. 1507 // 1508 // The current behavior is not desirable. To get maximal fidelity, we would 1509 // need to perform all type translation before beginning emission of .debug$S 1510 // and then make LocalUDTs a member of FunctionInfo 1511 } 1512 1513 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1514 // Generic dispatch for lowering an unknown type. 1515 switch (Ty->getTag()) { 1516 case dwarf::DW_TAG_array_type: 1517 return lowerTypeArray(cast<DICompositeType>(Ty)); 1518 case dwarf::DW_TAG_typedef: 1519 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1520 case dwarf::DW_TAG_base_type: 1521 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1522 case dwarf::DW_TAG_pointer_type: 1523 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1524 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1525 LLVM_FALLTHROUGH; 1526 case dwarf::DW_TAG_reference_type: 1527 case dwarf::DW_TAG_rvalue_reference_type: 1528 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1529 case dwarf::DW_TAG_ptr_to_member_type: 1530 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1531 case dwarf::DW_TAG_restrict_type: 1532 case dwarf::DW_TAG_const_type: 1533 case dwarf::DW_TAG_volatile_type: 1534 // TODO: add support for DW_TAG_atomic_type here 1535 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1536 case dwarf::DW_TAG_subroutine_type: 1537 if (ClassTy) { 1538 // The member function type of a member function pointer has no 1539 // ThisAdjustment. 1540 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1541 /*ThisAdjustment=*/0, 1542 /*IsStaticMethod=*/false); 1543 } 1544 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1545 case dwarf::DW_TAG_enumeration_type: 1546 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1547 case dwarf::DW_TAG_class_type: 1548 case dwarf::DW_TAG_structure_type: 1549 return lowerTypeClass(cast<DICompositeType>(Ty)); 1550 case dwarf::DW_TAG_union_type: 1551 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1552 case dwarf::DW_TAG_unspecified_type: 1553 if (Ty->getName() == "decltype(nullptr)") 1554 return TypeIndex::NullptrT(); 1555 return TypeIndex::None(); 1556 default: 1557 // Use the null type index. 1558 return TypeIndex(); 1559 } 1560 } 1561 1562 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1563 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1564 StringRef TypeName = Ty->getName(); 1565 1566 addToUDTs(Ty); 1567 1568 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1569 TypeName == "HRESULT") 1570 return TypeIndex(SimpleTypeKind::HResult); 1571 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1572 TypeName == "wchar_t") 1573 return TypeIndex(SimpleTypeKind::WideCharacter); 1574 1575 return UnderlyingTypeIndex; 1576 } 1577 1578 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1579 const DIType *ElementType = Ty->getBaseType(); 1580 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1581 // IndexType is size_t, which depends on the bitness of the target. 1582 TypeIndex IndexType = getPointerSizeInBytes() == 8 1583 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1584 : TypeIndex(SimpleTypeKind::UInt32Long); 1585 1586 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1587 1588 // Add subranges to array type. 1589 DINodeArray Elements = Ty->getElements(); 1590 for (int i = Elements.size() - 1; i >= 0; --i) { 1591 const DINode *Element = Elements[i]; 1592 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1593 1594 const DISubrange *Subrange = cast<DISubrange>(Element); 1595 int64_t Count = -1; 1596 // Calculate the count if either LowerBound is absent or is zero and 1597 // either of Count or UpperBound are constant. 1598 auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>(); 1599 if (!Subrange->getRawLowerBound() || (LI && (LI->getSExtValue() == 0))) { 1600 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>()) 1601 Count = CI->getSExtValue(); 1602 else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt*>()) 1603 Count = UI->getSExtValue() + 1; // LowerBound is zero 1604 } 1605 1606 // Forward declarations of arrays without a size and VLAs use a count of -1. 1607 // Emit a count of zero in these cases to match what MSVC does for arrays 1608 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1609 // should do for them even if we could distinguish them. 1610 if (Count == -1) 1611 Count = 0; 1612 1613 // Update the element size and element type index for subsequent subranges. 1614 ElementSize *= Count; 1615 1616 // If this is the outermost array, use the size from the array. It will be 1617 // more accurate if we had a VLA or an incomplete element type size. 1618 uint64_t ArraySize = 1619 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1620 1621 StringRef Name = (i == 0) ? Ty->getName() : ""; 1622 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1623 ElementTypeIndex = TypeTable.writeLeafType(AR); 1624 } 1625 1626 return ElementTypeIndex; 1627 } 1628 1629 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1630 TypeIndex Index; 1631 dwarf::TypeKind Kind; 1632 uint32_t ByteSize; 1633 1634 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1635 ByteSize = Ty->getSizeInBits() / 8; 1636 1637 SimpleTypeKind STK = SimpleTypeKind::None; 1638 switch (Kind) { 1639 case dwarf::DW_ATE_address: 1640 // FIXME: Translate 1641 break; 1642 case dwarf::DW_ATE_boolean: 1643 switch (ByteSize) { 1644 case 1: STK = SimpleTypeKind::Boolean8; break; 1645 case 2: STK = SimpleTypeKind::Boolean16; break; 1646 case 4: STK = SimpleTypeKind::Boolean32; break; 1647 case 8: STK = SimpleTypeKind::Boolean64; break; 1648 case 16: STK = SimpleTypeKind::Boolean128; break; 1649 } 1650 break; 1651 case dwarf::DW_ATE_complex_float: 1652 switch (ByteSize) { 1653 case 2: STK = SimpleTypeKind::Complex16; break; 1654 case 4: STK = SimpleTypeKind::Complex32; break; 1655 case 8: STK = SimpleTypeKind::Complex64; break; 1656 case 10: STK = SimpleTypeKind::Complex80; break; 1657 case 16: STK = SimpleTypeKind::Complex128; break; 1658 } 1659 break; 1660 case dwarf::DW_ATE_float: 1661 switch (ByteSize) { 1662 case 2: STK = SimpleTypeKind::Float16; break; 1663 case 4: STK = SimpleTypeKind::Float32; break; 1664 case 6: STK = SimpleTypeKind::Float48; break; 1665 case 8: STK = SimpleTypeKind::Float64; break; 1666 case 10: STK = SimpleTypeKind::Float80; break; 1667 case 16: STK = SimpleTypeKind::Float128; break; 1668 } 1669 break; 1670 case dwarf::DW_ATE_signed: 1671 switch (ByteSize) { 1672 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1673 case 2: STK = SimpleTypeKind::Int16Short; break; 1674 case 4: STK = SimpleTypeKind::Int32; break; 1675 case 8: STK = SimpleTypeKind::Int64Quad; break; 1676 case 16: STK = SimpleTypeKind::Int128Oct; break; 1677 } 1678 break; 1679 case dwarf::DW_ATE_unsigned: 1680 switch (ByteSize) { 1681 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1682 case 2: STK = SimpleTypeKind::UInt16Short; break; 1683 case 4: STK = SimpleTypeKind::UInt32; break; 1684 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1685 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1686 } 1687 break; 1688 case dwarf::DW_ATE_UTF: 1689 switch (ByteSize) { 1690 case 2: STK = SimpleTypeKind::Character16; break; 1691 case 4: STK = SimpleTypeKind::Character32; break; 1692 } 1693 break; 1694 case dwarf::DW_ATE_signed_char: 1695 if (ByteSize == 1) 1696 STK = SimpleTypeKind::SignedCharacter; 1697 break; 1698 case dwarf::DW_ATE_unsigned_char: 1699 if (ByteSize == 1) 1700 STK = SimpleTypeKind::UnsignedCharacter; 1701 break; 1702 default: 1703 break; 1704 } 1705 1706 // Apply some fixups based on the source-level type name. 1707 if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int") 1708 STK = SimpleTypeKind::Int32Long; 1709 if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int") 1710 STK = SimpleTypeKind::UInt32Long; 1711 if (STK == SimpleTypeKind::UInt16Short && 1712 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1713 STK = SimpleTypeKind::WideCharacter; 1714 if ((STK == SimpleTypeKind::SignedCharacter || 1715 STK == SimpleTypeKind::UnsignedCharacter) && 1716 Ty->getName() == "char") 1717 STK = SimpleTypeKind::NarrowCharacter; 1718 1719 return TypeIndex(STK); 1720 } 1721 1722 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1723 PointerOptions PO) { 1724 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1725 1726 // Pointers to simple types without any options can use SimpleTypeMode, rather 1727 // than having a dedicated pointer type record. 1728 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1729 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1730 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1731 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1732 ? SimpleTypeMode::NearPointer64 1733 : SimpleTypeMode::NearPointer32; 1734 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1735 } 1736 1737 PointerKind PK = 1738 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1739 PointerMode PM = PointerMode::Pointer; 1740 switch (Ty->getTag()) { 1741 default: llvm_unreachable("not a pointer tag type"); 1742 case dwarf::DW_TAG_pointer_type: 1743 PM = PointerMode::Pointer; 1744 break; 1745 case dwarf::DW_TAG_reference_type: 1746 PM = PointerMode::LValueReference; 1747 break; 1748 case dwarf::DW_TAG_rvalue_reference_type: 1749 PM = PointerMode::RValueReference; 1750 break; 1751 } 1752 1753 if (Ty->isObjectPointer()) 1754 PO |= PointerOptions::Const; 1755 1756 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1757 return TypeTable.writeLeafType(PR); 1758 } 1759 1760 static PointerToMemberRepresentation 1761 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1762 // SizeInBytes being zero generally implies that the member pointer type was 1763 // incomplete, which can happen if it is part of a function prototype. In this 1764 // case, use the unknown model instead of the general model. 1765 if (IsPMF) { 1766 switch (Flags & DINode::FlagPtrToMemberRep) { 1767 case 0: 1768 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1769 : PointerToMemberRepresentation::GeneralFunction; 1770 case DINode::FlagSingleInheritance: 1771 return PointerToMemberRepresentation::SingleInheritanceFunction; 1772 case DINode::FlagMultipleInheritance: 1773 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1774 case DINode::FlagVirtualInheritance: 1775 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1776 } 1777 } else { 1778 switch (Flags & DINode::FlagPtrToMemberRep) { 1779 case 0: 1780 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1781 : PointerToMemberRepresentation::GeneralData; 1782 case DINode::FlagSingleInheritance: 1783 return PointerToMemberRepresentation::SingleInheritanceData; 1784 case DINode::FlagMultipleInheritance: 1785 return PointerToMemberRepresentation::MultipleInheritanceData; 1786 case DINode::FlagVirtualInheritance: 1787 return PointerToMemberRepresentation::VirtualInheritanceData; 1788 } 1789 } 1790 llvm_unreachable("invalid ptr to member representation"); 1791 } 1792 1793 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1794 PointerOptions PO) { 1795 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1796 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1797 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1798 TypeIndex PointeeTI = 1799 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1800 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1801 : PointerKind::Near32; 1802 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1803 : PointerMode::PointerToDataMember; 1804 1805 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1806 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1807 MemberPointerInfo MPI( 1808 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1809 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1810 return TypeTable.writeLeafType(PR); 1811 } 1812 1813 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1814 /// have a translation, use the NearC convention. 1815 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1816 switch (DwarfCC) { 1817 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1818 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1819 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1820 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1821 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1822 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1823 } 1824 return CallingConvention::NearC; 1825 } 1826 1827 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1828 ModifierOptions Mods = ModifierOptions::None; 1829 PointerOptions PO = PointerOptions::None; 1830 bool IsModifier = true; 1831 const DIType *BaseTy = Ty; 1832 while (IsModifier && BaseTy) { 1833 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1834 switch (BaseTy->getTag()) { 1835 case dwarf::DW_TAG_const_type: 1836 Mods |= ModifierOptions::Const; 1837 PO |= PointerOptions::Const; 1838 break; 1839 case dwarf::DW_TAG_volatile_type: 1840 Mods |= ModifierOptions::Volatile; 1841 PO |= PointerOptions::Volatile; 1842 break; 1843 case dwarf::DW_TAG_restrict_type: 1844 // Only pointer types be marked with __restrict. There is no known flag 1845 // for __restrict in LF_MODIFIER records. 1846 PO |= PointerOptions::Restrict; 1847 break; 1848 default: 1849 IsModifier = false; 1850 break; 1851 } 1852 if (IsModifier) 1853 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1854 } 1855 1856 // Check if the inner type will use an LF_POINTER record. If so, the 1857 // qualifiers will go in the LF_POINTER record. This comes up for types like 1858 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1859 // char *'. 1860 if (BaseTy) { 1861 switch (BaseTy->getTag()) { 1862 case dwarf::DW_TAG_pointer_type: 1863 case dwarf::DW_TAG_reference_type: 1864 case dwarf::DW_TAG_rvalue_reference_type: 1865 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1866 case dwarf::DW_TAG_ptr_to_member_type: 1867 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1868 default: 1869 break; 1870 } 1871 } 1872 1873 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1874 1875 // Return the base type index if there aren't any modifiers. For example, the 1876 // metadata could contain restrict wrappers around non-pointer types. 1877 if (Mods == ModifierOptions::None) 1878 return ModifiedTI; 1879 1880 ModifierRecord MR(ModifiedTI, Mods); 1881 return TypeTable.writeLeafType(MR); 1882 } 1883 1884 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1885 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1886 for (const DIType *ArgType : Ty->getTypeArray()) 1887 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1888 1889 // MSVC uses type none for variadic argument. 1890 if (ReturnAndArgTypeIndices.size() > 1 && 1891 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1892 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1893 } 1894 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1895 ArrayRef<TypeIndex> ArgTypeIndices = None; 1896 if (!ReturnAndArgTypeIndices.empty()) { 1897 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1898 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1899 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1900 } 1901 1902 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1903 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1904 1905 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1906 1907 FunctionOptions FO = getFunctionOptions(Ty); 1908 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 1909 ArgListIndex); 1910 return TypeTable.writeLeafType(Procedure); 1911 } 1912 1913 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 1914 const DIType *ClassTy, 1915 int ThisAdjustment, 1916 bool IsStaticMethod, 1917 FunctionOptions FO) { 1918 // Lower the containing class type. 1919 TypeIndex ClassType = getTypeIndex(ClassTy); 1920 1921 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 1922 1923 unsigned Index = 0; 1924 SmallVector<TypeIndex, 8> ArgTypeIndices; 1925 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1926 if (ReturnAndArgs.size() > Index) { 1927 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 1928 } 1929 1930 // If the first argument is a pointer type and this isn't a static method, 1931 // treat it as the special 'this' parameter, which is encoded separately from 1932 // the arguments. 1933 TypeIndex ThisTypeIndex; 1934 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 1935 if (const DIDerivedType *PtrTy = 1936 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 1937 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 1938 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 1939 Index++; 1940 } 1941 } 1942 } 1943 1944 while (Index < ReturnAndArgs.size()) 1945 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 1946 1947 // MSVC uses type none for variadic argument. 1948 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 1949 ArgTypeIndices.back() = TypeIndex::None(); 1950 1951 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1952 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1953 1954 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1955 1956 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1957 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1958 return TypeTable.writeLeafType(MFR); 1959 } 1960 1961 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1962 unsigned VSlotCount = 1963 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1964 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1965 1966 VFTableShapeRecord VFTSR(Slots); 1967 return TypeTable.writeLeafType(VFTSR); 1968 } 1969 1970 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1971 switch (Flags & DINode::FlagAccessibility) { 1972 case DINode::FlagPrivate: return MemberAccess::Private; 1973 case DINode::FlagPublic: return MemberAccess::Public; 1974 case DINode::FlagProtected: return MemberAccess::Protected; 1975 case 0: 1976 // If there was no explicit access control, provide the default for the tag. 1977 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1978 : MemberAccess::Public; 1979 } 1980 llvm_unreachable("access flags are exclusive"); 1981 } 1982 1983 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1984 if (SP->isArtificial()) 1985 return MethodOptions::CompilerGenerated; 1986 1987 // FIXME: Handle other MethodOptions. 1988 1989 return MethodOptions::None; 1990 } 1991 1992 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 1993 bool Introduced) { 1994 if (SP->getFlags() & DINode::FlagStaticMember) 1995 return MethodKind::Static; 1996 1997 switch (SP->getVirtuality()) { 1998 case dwarf::DW_VIRTUALITY_none: 1999 break; 2000 case dwarf::DW_VIRTUALITY_virtual: 2001 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2002 case dwarf::DW_VIRTUALITY_pure_virtual: 2003 return Introduced ? MethodKind::PureIntroducingVirtual 2004 : MethodKind::PureVirtual; 2005 default: 2006 llvm_unreachable("unhandled virtuality case"); 2007 } 2008 2009 return MethodKind::Vanilla; 2010 } 2011 2012 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2013 switch (Ty->getTag()) { 2014 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 2015 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 2016 } 2017 llvm_unreachable("unexpected tag"); 2018 } 2019 2020 /// Return ClassOptions that should be present on both the forward declaration 2021 /// and the defintion of a tag type. 2022 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2023 ClassOptions CO = ClassOptions::None; 2024 2025 // MSVC always sets this flag, even for local types. Clang doesn't always 2026 // appear to give every type a linkage name, which may be problematic for us. 2027 // FIXME: Investigate the consequences of not following them here. 2028 if (!Ty->getIdentifier().empty()) 2029 CO |= ClassOptions::HasUniqueName; 2030 2031 // Put the Nested flag on a type if it appears immediately inside a tag type. 2032 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2033 // here. That flag is only set on definitions, and not forward declarations. 2034 const DIScope *ImmediateScope = Ty->getScope(); 2035 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2036 CO |= ClassOptions::Nested; 2037 2038 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2039 // type only when it has an immediate function scope. Clang never puts enums 2040 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2041 // always in function, class, or file scopes. 2042 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2043 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2044 CO |= ClassOptions::Scoped; 2045 } else { 2046 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2047 Scope = Scope->getScope()) { 2048 if (isa<DISubprogram>(Scope)) { 2049 CO |= ClassOptions::Scoped; 2050 break; 2051 } 2052 } 2053 } 2054 2055 return CO; 2056 } 2057 2058 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2059 switch (Ty->getTag()) { 2060 case dwarf::DW_TAG_class_type: 2061 case dwarf::DW_TAG_structure_type: 2062 case dwarf::DW_TAG_union_type: 2063 case dwarf::DW_TAG_enumeration_type: 2064 break; 2065 default: 2066 return; 2067 } 2068 2069 if (const auto *File = Ty->getFile()) { 2070 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2071 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2072 2073 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2074 TypeTable.writeLeafType(USLR); 2075 } 2076 } 2077 2078 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2079 ClassOptions CO = getCommonClassOptions(Ty); 2080 TypeIndex FTI; 2081 unsigned EnumeratorCount = 0; 2082 2083 if (Ty->isForwardDecl()) { 2084 CO |= ClassOptions::ForwardReference; 2085 } else { 2086 ContinuationRecordBuilder ContinuationBuilder; 2087 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2088 for (const DINode *Element : Ty->getElements()) { 2089 // We assume that the frontend provides all members in source declaration 2090 // order, which is what MSVC does. 2091 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2092 EnumeratorRecord ER(MemberAccess::Public, 2093 APSInt(Enumerator->getValue(), true), 2094 Enumerator->getName()); 2095 ContinuationBuilder.writeMemberType(ER); 2096 EnumeratorCount++; 2097 } 2098 } 2099 FTI = TypeTable.insertRecord(ContinuationBuilder); 2100 } 2101 2102 std::string FullName = getFullyQualifiedName(Ty); 2103 2104 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2105 getTypeIndex(Ty->getBaseType())); 2106 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2107 2108 addUDTSrcLine(Ty, EnumTI); 2109 2110 return EnumTI; 2111 } 2112 2113 //===----------------------------------------------------------------------===// 2114 // ClassInfo 2115 //===----------------------------------------------------------------------===// 2116 2117 struct llvm::ClassInfo { 2118 struct MemberInfo { 2119 const DIDerivedType *MemberTypeNode; 2120 uint64_t BaseOffset; 2121 }; 2122 // [MemberInfo] 2123 using MemberList = std::vector<MemberInfo>; 2124 2125 using MethodsList = TinyPtrVector<const DISubprogram *>; 2126 // MethodName -> MethodsList 2127 using MethodsMap = MapVector<MDString *, MethodsList>; 2128 2129 /// Base classes. 2130 std::vector<const DIDerivedType *> Inheritance; 2131 2132 /// Direct members. 2133 MemberList Members; 2134 // Direct overloaded methods gathered by name. 2135 MethodsMap Methods; 2136 2137 TypeIndex VShapeTI; 2138 2139 std::vector<const DIType *> NestedTypes; 2140 }; 2141 2142 void CodeViewDebug::clear() { 2143 assert(CurFn == nullptr); 2144 FileIdMap.clear(); 2145 FnDebugInfo.clear(); 2146 FileToFilepathMap.clear(); 2147 LocalUDTs.clear(); 2148 GlobalUDTs.clear(); 2149 TypeIndices.clear(); 2150 CompleteTypeIndices.clear(); 2151 ScopeGlobals.clear(); 2152 } 2153 2154 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2155 const DIDerivedType *DDTy) { 2156 if (!DDTy->getName().empty()) { 2157 Info.Members.push_back({DDTy, 0}); 2158 return; 2159 } 2160 2161 // An unnamed member may represent a nested struct or union. Attempt to 2162 // interpret the unnamed member as a DICompositeType possibly wrapped in 2163 // qualifier types. Add all the indirect fields to the current record if that 2164 // succeeds, and drop the member if that fails. 2165 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2166 uint64_t Offset = DDTy->getOffsetInBits(); 2167 const DIType *Ty = DDTy->getBaseType(); 2168 bool FullyResolved = false; 2169 while (!FullyResolved) { 2170 switch (Ty->getTag()) { 2171 case dwarf::DW_TAG_const_type: 2172 case dwarf::DW_TAG_volatile_type: 2173 // FIXME: we should apply the qualifier types to the indirect fields 2174 // rather than dropping them. 2175 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2176 break; 2177 default: 2178 FullyResolved = true; 2179 break; 2180 } 2181 } 2182 2183 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2184 if (!DCTy) 2185 return; 2186 2187 ClassInfo NestedInfo = collectClassInfo(DCTy); 2188 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2189 Info.Members.push_back( 2190 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2191 } 2192 2193 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2194 ClassInfo Info; 2195 // Add elements to structure type. 2196 DINodeArray Elements = Ty->getElements(); 2197 for (auto *Element : Elements) { 2198 // We assume that the frontend provides all members in source declaration 2199 // order, which is what MSVC does. 2200 if (!Element) 2201 continue; 2202 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2203 Info.Methods[SP->getRawName()].push_back(SP); 2204 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2205 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2206 collectMemberInfo(Info, DDTy); 2207 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2208 Info.Inheritance.push_back(DDTy); 2209 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2210 DDTy->getName() == "__vtbl_ptr_type") { 2211 Info.VShapeTI = getTypeIndex(DDTy); 2212 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2213 Info.NestedTypes.push_back(DDTy); 2214 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2215 // Ignore friend members. It appears that MSVC emitted info about 2216 // friends in the past, but modern versions do not. 2217 } 2218 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2219 Info.NestedTypes.push_back(Composite); 2220 } 2221 // Skip other unrecognized kinds of elements. 2222 } 2223 return Info; 2224 } 2225 2226 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2227 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2228 // if a complete type should be emitted instead of a forward reference. 2229 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2230 !Ty->isForwardDecl(); 2231 } 2232 2233 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2234 // Emit the complete type for unnamed structs. C++ classes with methods 2235 // which have a circular reference back to the class type are expected to 2236 // be named by the front-end and should not be "unnamed". C unnamed 2237 // structs should not have circular references. 2238 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2239 // If this unnamed complete type is already in the process of being defined 2240 // then the description of the type is malformed and cannot be emitted 2241 // into CodeView correctly so report a fatal error. 2242 auto I = CompleteTypeIndices.find(Ty); 2243 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2244 report_fatal_error("cannot debug circular reference to unnamed type"); 2245 return getCompleteTypeIndex(Ty); 2246 } 2247 2248 // First, construct the forward decl. Don't look into Ty to compute the 2249 // forward decl options, since it might not be available in all TUs. 2250 TypeRecordKind Kind = getRecordKind(Ty); 2251 ClassOptions CO = 2252 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2253 std::string FullName = getFullyQualifiedName(Ty); 2254 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2255 FullName, Ty->getIdentifier()); 2256 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2257 if (!Ty->isForwardDecl()) 2258 DeferredCompleteTypes.push_back(Ty); 2259 return FwdDeclTI; 2260 } 2261 2262 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2263 // Construct the field list and complete type record. 2264 TypeRecordKind Kind = getRecordKind(Ty); 2265 ClassOptions CO = getCommonClassOptions(Ty); 2266 TypeIndex FieldTI; 2267 TypeIndex VShapeTI; 2268 unsigned FieldCount; 2269 bool ContainsNestedClass; 2270 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2271 lowerRecordFieldList(Ty); 2272 2273 if (ContainsNestedClass) 2274 CO |= ClassOptions::ContainsNestedClass; 2275 2276 // MSVC appears to set this flag by searching any destructor or method with 2277 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2278 // the members, however special member functions are not yet emitted into 2279 // debug information. For now checking a class's non-triviality seems enough. 2280 // FIXME: not true for a nested unnamed struct. 2281 if (isNonTrivial(Ty)) 2282 CO |= ClassOptions::HasConstructorOrDestructor; 2283 2284 std::string FullName = getFullyQualifiedName(Ty); 2285 2286 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2287 2288 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2289 SizeInBytes, FullName, Ty->getIdentifier()); 2290 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2291 2292 addUDTSrcLine(Ty, ClassTI); 2293 2294 addToUDTs(Ty); 2295 2296 return ClassTI; 2297 } 2298 2299 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2300 // Emit the complete type for unnamed unions. 2301 if (shouldAlwaysEmitCompleteClassType(Ty)) 2302 return getCompleteTypeIndex(Ty); 2303 2304 ClassOptions CO = 2305 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2306 std::string FullName = getFullyQualifiedName(Ty); 2307 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2308 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2309 if (!Ty->isForwardDecl()) 2310 DeferredCompleteTypes.push_back(Ty); 2311 return FwdDeclTI; 2312 } 2313 2314 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2315 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2316 TypeIndex FieldTI; 2317 unsigned FieldCount; 2318 bool ContainsNestedClass; 2319 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2320 lowerRecordFieldList(Ty); 2321 2322 if (ContainsNestedClass) 2323 CO |= ClassOptions::ContainsNestedClass; 2324 2325 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2326 std::string FullName = getFullyQualifiedName(Ty); 2327 2328 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2329 Ty->getIdentifier()); 2330 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2331 2332 addUDTSrcLine(Ty, UnionTI); 2333 2334 addToUDTs(Ty); 2335 2336 return UnionTI; 2337 } 2338 2339 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2340 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2341 // Manually count members. MSVC appears to count everything that generates a 2342 // field list record. Each individual overload in a method overload group 2343 // contributes to this count, even though the overload group is a single field 2344 // list record. 2345 unsigned MemberCount = 0; 2346 ClassInfo Info = collectClassInfo(Ty); 2347 ContinuationRecordBuilder ContinuationBuilder; 2348 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2349 2350 // Create base classes. 2351 for (const DIDerivedType *I : Info.Inheritance) { 2352 if (I->getFlags() & DINode::FlagVirtual) { 2353 // Virtual base. 2354 unsigned VBPtrOffset = I->getVBPtrOffset(); 2355 // FIXME: Despite the accessor name, the offset is really in bytes. 2356 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2357 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2358 ? TypeRecordKind::IndirectVirtualBaseClass 2359 : TypeRecordKind::VirtualBaseClass; 2360 VirtualBaseClassRecord VBCR( 2361 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2362 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2363 VBTableIndex); 2364 2365 ContinuationBuilder.writeMemberType(VBCR); 2366 MemberCount++; 2367 } else { 2368 assert(I->getOffsetInBits() % 8 == 0 && 2369 "bases must be on byte boundaries"); 2370 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2371 getTypeIndex(I->getBaseType()), 2372 I->getOffsetInBits() / 8); 2373 ContinuationBuilder.writeMemberType(BCR); 2374 MemberCount++; 2375 } 2376 } 2377 2378 // Create members. 2379 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2380 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2381 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2382 StringRef MemberName = Member->getName(); 2383 MemberAccess Access = 2384 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2385 2386 if (Member->isStaticMember()) { 2387 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2388 ContinuationBuilder.writeMemberType(SDMR); 2389 MemberCount++; 2390 continue; 2391 } 2392 2393 // Virtual function pointer member. 2394 if ((Member->getFlags() & DINode::FlagArtificial) && 2395 Member->getName().startswith("_vptr$")) { 2396 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2397 ContinuationBuilder.writeMemberType(VFPR); 2398 MemberCount++; 2399 continue; 2400 } 2401 2402 // Data member. 2403 uint64_t MemberOffsetInBits = 2404 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2405 if (Member->isBitField()) { 2406 uint64_t StartBitOffset = MemberOffsetInBits; 2407 if (const auto *CI = 2408 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2409 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2410 } 2411 StartBitOffset -= MemberOffsetInBits; 2412 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2413 StartBitOffset); 2414 MemberBaseType = TypeTable.writeLeafType(BFR); 2415 } 2416 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2417 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2418 MemberName); 2419 ContinuationBuilder.writeMemberType(DMR); 2420 MemberCount++; 2421 } 2422 2423 // Create methods 2424 for (auto &MethodItr : Info.Methods) { 2425 StringRef Name = MethodItr.first->getString(); 2426 2427 std::vector<OneMethodRecord> Methods; 2428 for (const DISubprogram *SP : MethodItr.second) { 2429 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2430 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2431 2432 unsigned VFTableOffset = -1; 2433 if (Introduced) 2434 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2435 2436 Methods.push_back(OneMethodRecord( 2437 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2438 translateMethodKindFlags(SP, Introduced), 2439 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2440 MemberCount++; 2441 } 2442 assert(!Methods.empty() && "Empty methods map entry"); 2443 if (Methods.size() == 1) 2444 ContinuationBuilder.writeMemberType(Methods[0]); 2445 else { 2446 // FIXME: Make this use its own ContinuationBuilder so that 2447 // MethodOverloadList can be split correctly. 2448 MethodOverloadListRecord MOLR(Methods); 2449 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2450 2451 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2452 ContinuationBuilder.writeMemberType(OMR); 2453 } 2454 } 2455 2456 // Create nested classes. 2457 for (const DIType *Nested : Info.NestedTypes) { 2458 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2459 ContinuationBuilder.writeMemberType(R); 2460 MemberCount++; 2461 } 2462 2463 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2464 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2465 !Info.NestedTypes.empty()); 2466 } 2467 2468 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2469 if (!VBPType.getIndex()) { 2470 // Make a 'const int *' type. 2471 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2472 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2473 2474 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2475 : PointerKind::Near32; 2476 PointerMode PM = PointerMode::Pointer; 2477 PointerOptions PO = PointerOptions::None; 2478 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2479 VBPType = TypeTable.writeLeafType(PR); 2480 } 2481 2482 return VBPType; 2483 } 2484 2485 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2486 // The null DIType is the void type. Don't try to hash it. 2487 if (!Ty) 2488 return TypeIndex::Void(); 2489 2490 // Check if we've already translated this type. Don't try to do a 2491 // get-or-create style insertion that caches the hash lookup across the 2492 // lowerType call. It will update the TypeIndices map. 2493 auto I = TypeIndices.find({Ty, ClassTy}); 2494 if (I != TypeIndices.end()) 2495 return I->second; 2496 2497 TypeLoweringScope S(*this); 2498 TypeIndex TI = lowerType(Ty, ClassTy); 2499 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2500 } 2501 2502 codeview::TypeIndex 2503 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2504 const DISubroutineType *SubroutineTy) { 2505 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2506 "this type must be a pointer type"); 2507 2508 PointerOptions Options = PointerOptions::None; 2509 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2510 Options = PointerOptions::LValueRefThisPointer; 2511 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2512 Options = PointerOptions::RValueRefThisPointer; 2513 2514 // Check if we've already translated this type. If there is no ref qualifier 2515 // on the function then we look up this pointer type with no associated class 2516 // so that the TypeIndex for the this pointer can be shared with the type 2517 // index for other pointers to this class type. If there is a ref qualifier 2518 // then we lookup the pointer using the subroutine as the parent type. 2519 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2520 if (I != TypeIndices.end()) 2521 return I->second; 2522 2523 TypeLoweringScope S(*this); 2524 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2525 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2526 } 2527 2528 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2529 PointerRecord PR(getTypeIndex(Ty), 2530 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2531 : PointerKind::Near32, 2532 PointerMode::LValueReference, PointerOptions::None, 2533 Ty->getSizeInBits() / 8); 2534 return TypeTable.writeLeafType(PR); 2535 } 2536 2537 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2538 // The null DIType is the void type. Don't try to hash it. 2539 if (!Ty) 2540 return TypeIndex::Void(); 2541 2542 // Look through typedefs when getting the complete type index. Call 2543 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2544 // emitted only once. 2545 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2546 (void)getTypeIndex(Ty); 2547 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2548 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2549 2550 // If this is a non-record type, the complete type index is the same as the 2551 // normal type index. Just call getTypeIndex. 2552 switch (Ty->getTag()) { 2553 case dwarf::DW_TAG_class_type: 2554 case dwarf::DW_TAG_structure_type: 2555 case dwarf::DW_TAG_union_type: 2556 break; 2557 default: 2558 return getTypeIndex(Ty); 2559 } 2560 2561 const auto *CTy = cast<DICompositeType>(Ty); 2562 2563 TypeLoweringScope S(*this); 2564 2565 // Make sure the forward declaration is emitted first. It's unclear if this 2566 // is necessary, but MSVC does it, and we should follow suit until we can show 2567 // otherwise. 2568 // We only emit a forward declaration for named types. 2569 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2570 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2571 2572 // Just use the forward decl if we don't have complete type info. This 2573 // might happen if the frontend is using modules and expects the complete 2574 // definition to be emitted elsewhere. 2575 if (CTy->isForwardDecl()) 2576 return FwdDeclTI; 2577 } 2578 2579 // Check if we've already translated the complete record type. 2580 // Insert the type with a null TypeIndex to signify that the type is currently 2581 // being lowered. 2582 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2583 if (!InsertResult.second) 2584 return InsertResult.first->second; 2585 2586 TypeIndex TI; 2587 switch (CTy->getTag()) { 2588 case dwarf::DW_TAG_class_type: 2589 case dwarf::DW_TAG_structure_type: 2590 TI = lowerCompleteTypeClass(CTy); 2591 break; 2592 case dwarf::DW_TAG_union_type: 2593 TI = lowerCompleteTypeUnion(CTy); 2594 break; 2595 default: 2596 llvm_unreachable("not a record"); 2597 } 2598 2599 // Update the type index associated with this CompositeType. This cannot 2600 // use the 'InsertResult' iterator above because it is potentially 2601 // invalidated by map insertions which can occur while lowering the class 2602 // type above. 2603 CompleteTypeIndices[CTy] = TI; 2604 return TI; 2605 } 2606 2607 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2608 /// and do this until fixpoint, as each complete record type typically 2609 /// references 2610 /// many other record types. 2611 void CodeViewDebug::emitDeferredCompleteTypes() { 2612 SmallVector<const DICompositeType *, 4> TypesToEmit; 2613 while (!DeferredCompleteTypes.empty()) { 2614 std::swap(DeferredCompleteTypes, TypesToEmit); 2615 for (const DICompositeType *RecordTy : TypesToEmit) 2616 getCompleteTypeIndex(RecordTy); 2617 TypesToEmit.clear(); 2618 } 2619 } 2620 2621 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2622 ArrayRef<LocalVariable> Locals) { 2623 // Get the sorted list of parameters and emit them first. 2624 SmallVector<const LocalVariable *, 6> Params; 2625 for (const LocalVariable &L : Locals) 2626 if (L.DIVar->isParameter()) 2627 Params.push_back(&L); 2628 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2629 return L->DIVar->getArg() < R->DIVar->getArg(); 2630 }); 2631 for (const LocalVariable *L : Params) 2632 emitLocalVariable(FI, *L); 2633 2634 // Next emit all non-parameters in the order that we found them. 2635 for (const LocalVariable &L : Locals) 2636 if (!L.DIVar->isParameter()) 2637 emitLocalVariable(FI, L); 2638 } 2639 2640 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2641 const LocalVariable &Var) { 2642 // LocalSym record, see SymbolRecord.h for more info. 2643 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2644 2645 LocalSymFlags Flags = LocalSymFlags::None; 2646 if (Var.DIVar->isParameter()) 2647 Flags |= LocalSymFlags::IsParameter; 2648 if (Var.DefRanges.empty()) 2649 Flags |= LocalSymFlags::IsOptimizedOut; 2650 2651 OS.AddComment("TypeIndex"); 2652 TypeIndex TI = Var.UseReferenceType 2653 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2654 : getCompleteTypeIndex(Var.DIVar->getType()); 2655 OS.emitInt32(TI.getIndex()); 2656 OS.AddComment("Flags"); 2657 OS.emitInt16(static_cast<uint16_t>(Flags)); 2658 // Truncate the name so we won't overflow the record length field. 2659 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2660 endSymbolRecord(LocalEnd); 2661 2662 // Calculate the on disk prefix of the appropriate def range record. The 2663 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2664 // should be big enough to hold all forms without memory allocation. 2665 SmallString<20> BytePrefix; 2666 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2667 BytePrefix.clear(); 2668 if (DefRange.InMemory) { 2669 int Offset = DefRange.DataOffset; 2670 unsigned Reg = DefRange.CVRegister; 2671 2672 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2673 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2674 // instead. In frames without stack realignment, $T0 will be the CFA. 2675 if (RegisterId(Reg) == RegisterId::ESP) { 2676 Reg = unsigned(RegisterId::VFRAME); 2677 Offset += FI.OffsetAdjustment; 2678 } 2679 2680 // If we can use the chosen frame pointer for the frame and this isn't a 2681 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2682 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2683 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2684 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2685 (bool(Flags & LocalSymFlags::IsParameter) 2686 ? (EncFP == FI.EncodedParamFramePtrReg) 2687 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2688 DefRangeFramePointerRelHeader DRHdr; 2689 DRHdr.Offset = Offset; 2690 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2691 } else { 2692 uint16_t RegRelFlags = 0; 2693 if (DefRange.IsSubfield) { 2694 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2695 (DefRange.StructOffset 2696 << DefRangeRegisterRelSym::OffsetInParentShift); 2697 } 2698 DefRangeRegisterRelHeader DRHdr; 2699 DRHdr.Register = Reg; 2700 DRHdr.Flags = RegRelFlags; 2701 DRHdr.BasePointerOffset = Offset; 2702 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2703 } 2704 } else { 2705 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2706 if (DefRange.IsSubfield) { 2707 DefRangeSubfieldRegisterHeader DRHdr; 2708 DRHdr.Register = DefRange.CVRegister; 2709 DRHdr.MayHaveNoName = 0; 2710 DRHdr.OffsetInParent = DefRange.StructOffset; 2711 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2712 } else { 2713 DefRangeRegisterHeader DRHdr; 2714 DRHdr.Register = DefRange.CVRegister; 2715 DRHdr.MayHaveNoName = 0; 2716 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2717 } 2718 } 2719 } 2720 } 2721 2722 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2723 const FunctionInfo& FI) { 2724 for (LexicalBlock *Block : Blocks) 2725 emitLexicalBlock(*Block, FI); 2726 } 2727 2728 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2729 /// lexical block scope. 2730 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2731 const FunctionInfo& FI) { 2732 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2733 OS.AddComment("PtrParent"); 2734 OS.emitInt32(0); // PtrParent 2735 OS.AddComment("PtrEnd"); 2736 OS.emitInt32(0); // PtrEnd 2737 OS.AddComment("Code size"); 2738 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2739 OS.AddComment("Function section relative address"); 2740 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2741 OS.AddComment("Function section index"); 2742 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2743 OS.AddComment("Lexical block name"); 2744 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2745 endSymbolRecord(RecordEnd); 2746 2747 // Emit variables local to this lexical block. 2748 emitLocalVariableList(FI, Block.Locals); 2749 emitGlobalVariableList(Block.Globals); 2750 2751 // Emit lexical blocks contained within this block. 2752 emitLexicalBlockList(Block.Children, FI); 2753 2754 // Close the lexical block scope. 2755 emitEndSymbolRecord(SymbolKind::S_END); 2756 } 2757 2758 /// Convenience routine for collecting lexical block information for a list 2759 /// of lexical scopes. 2760 void CodeViewDebug::collectLexicalBlockInfo( 2761 SmallVectorImpl<LexicalScope *> &Scopes, 2762 SmallVectorImpl<LexicalBlock *> &Blocks, 2763 SmallVectorImpl<LocalVariable> &Locals, 2764 SmallVectorImpl<CVGlobalVariable> &Globals) { 2765 for (LexicalScope *Scope : Scopes) 2766 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2767 } 2768 2769 /// Populate the lexical blocks and local variable lists of the parent with 2770 /// information about the specified lexical scope. 2771 void CodeViewDebug::collectLexicalBlockInfo( 2772 LexicalScope &Scope, 2773 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2774 SmallVectorImpl<LocalVariable> &ParentLocals, 2775 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2776 if (Scope.isAbstractScope()) 2777 return; 2778 2779 // Gather information about the lexical scope including local variables, 2780 // global variables, and address ranges. 2781 bool IgnoreScope = false; 2782 auto LI = ScopeVariables.find(&Scope); 2783 SmallVectorImpl<LocalVariable> *Locals = 2784 LI != ScopeVariables.end() ? &LI->second : nullptr; 2785 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2786 SmallVectorImpl<CVGlobalVariable> *Globals = 2787 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2788 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2789 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2790 2791 // Ignore lexical scopes which do not contain variables. 2792 if (!Locals && !Globals) 2793 IgnoreScope = true; 2794 2795 // Ignore lexical scopes which are not lexical blocks. 2796 if (!DILB) 2797 IgnoreScope = true; 2798 2799 // Ignore scopes which have too many address ranges to represent in the 2800 // current CodeView format or do not have a valid address range. 2801 // 2802 // For lexical scopes with multiple address ranges you may be tempted to 2803 // construct a single range covering every instruction where the block is 2804 // live and everything in between. Unfortunately, Visual Studio only 2805 // displays variables from the first matching lexical block scope. If the 2806 // first lexical block contains exception handling code or cold code which 2807 // is moved to the bottom of the routine creating a single range covering 2808 // nearly the entire routine, then it will hide all other lexical blocks 2809 // and the variables they contain. 2810 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2811 IgnoreScope = true; 2812 2813 if (IgnoreScope) { 2814 // This scope can be safely ignored and eliminating it will reduce the 2815 // size of the debug information. Be sure to collect any variable and scope 2816 // information from the this scope or any of its children and collapse them 2817 // into the parent scope. 2818 if (Locals) 2819 ParentLocals.append(Locals->begin(), Locals->end()); 2820 if (Globals) 2821 ParentGlobals.append(Globals->begin(), Globals->end()); 2822 collectLexicalBlockInfo(Scope.getChildren(), 2823 ParentBlocks, 2824 ParentLocals, 2825 ParentGlobals); 2826 return; 2827 } 2828 2829 // Create a new CodeView lexical block for this lexical scope. If we've 2830 // seen this DILexicalBlock before then the scope tree is malformed and 2831 // we can handle this gracefully by not processing it a second time. 2832 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2833 if (!BlockInsertion.second) 2834 return; 2835 2836 // Create a lexical block containing the variables and collect the the 2837 // lexical block information for the children. 2838 const InsnRange &Range = Ranges.front(); 2839 assert(Range.first && Range.second); 2840 LexicalBlock &Block = BlockInsertion.first->second; 2841 Block.Begin = getLabelBeforeInsn(Range.first); 2842 Block.End = getLabelAfterInsn(Range.second); 2843 assert(Block.Begin && "missing label for scope begin"); 2844 assert(Block.End && "missing label for scope end"); 2845 Block.Name = DILB->getName(); 2846 if (Locals) 2847 Block.Locals = std::move(*Locals); 2848 if (Globals) 2849 Block.Globals = std::move(*Globals); 2850 ParentBlocks.push_back(&Block); 2851 collectLexicalBlockInfo(Scope.getChildren(), 2852 Block.Children, 2853 Block.Locals, 2854 Block.Globals); 2855 } 2856 2857 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2858 const Function &GV = MF->getFunction(); 2859 assert(FnDebugInfo.count(&GV)); 2860 assert(CurFn == FnDebugInfo[&GV].get()); 2861 2862 collectVariableInfo(GV.getSubprogram()); 2863 2864 // Build the lexical block structure to emit for this routine. 2865 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2866 collectLexicalBlockInfo(*CFS, 2867 CurFn->ChildBlocks, 2868 CurFn->Locals, 2869 CurFn->Globals); 2870 2871 // Clear the scope and variable information from the map which will not be 2872 // valid after we have finished processing this routine. This also prepares 2873 // the map for the subsequent routine. 2874 ScopeVariables.clear(); 2875 2876 // Don't emit anything if we don't have any line tables. 2877 // Thunks are compiler-generated and probably won't have source correlation. 2878 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2879 FnDebugInfo.erase(&GV); 2880 CurFn = nullptr; 2881 return; 2882 } 2883 2884 // Find heap alloc sites and add to list. 2885 for (const auto &MBB : *MF) { 2886 for (const auto &MI : MBB) { 2887 if (MDNode *MD = MI.getHeapAllocMarker()) { 2888 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 2889 getLabelAfterInsn(&MI), 2890 dyn_cast<DIType>(MD))); 2891 } 2892 } 2893 } 2894 2895 CurFn->Annotations = MF->getCodeViewAnnotations(); 2896 2897 CurFn->End = Asm->getFunctionEnd(); 2898 2899 CurFn = nullptr; 2900 } 2901 2902 // Usable locations are valid with non-zero line numbers. A line number of zero 2903 // corresponds to optimized code that doesn't have a distinct source location. 2904 // In this case, we try to use the previous or next source location depending on 2905 // the context. 2906 static bool isUsableDebugLoc(DebugLoc DL) { 2907 return DL && DL.getLine() != 0; 2908 } 2909 2910 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2911 DebugHandlerBase::beginInstruction(MI); 2912 2913 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2914 if (!Asm || !CurFn || MI->isDebugInstr() || 2915 MI->getFlag(MachineInstr::FrameSetup)) 2916 return; 2917 2918 // If the first instruction of a new MBB has no location, find the first 2919 // instruction with a location and use that. 2920 DebugLoc DL = MI->getDebugLoc(); 2921 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 2922 for (const auto &NextMI : *MI->getParent()) { 2923 if (NextMI.isDebugInstr()) 2924 continue; 2925 DL = NextMI.getDebugLoc(); 2926 if (isUsableDebugLoc(DL)) 2927 break; 2928 } 2929 // FIXME: Handle the case where the BB has no valid locations. This would 2930 // probably require doing a real dataflow analysis. 2931 } 2932 PrevInstBB = MI->getParent(); 2933 2934 // If we still don't have a debug location, don't record a location. 2935 if (!isUsableDebugLoc(DL)) 2936 return; 2937 2938 maybeRecordLocation(DL, Asm->MF); 2939 } 2940 2941 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2942 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2943 *EndLabel = MMI->getContext().createTempSymbol(); 2944 OS.emitInt32(unsigned(Kind)); 2945 OS.AddComment("Subsection size"); 2946 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2947 OS.emitLabel(BeginLabel); 2948 return EndLabel; 2949 } 2950 2951 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2952 OS.emitLabel(EndLabel); 2953 // Every subsection must be aligned to a 4-byte boundary. 2954 OS.emitValueToAlignment(4); 2955 } 2956 2957 static StringRef getSymbolName(SymbolKind SymKind) { 2958 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 2959 if (EE.Value == SymKind) 2960 return EE.Name; 2961 return ""; 2962 } 2963 2964 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 2965 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2966 *EndLabel = MMI->getContext().createTempSymbol(); 2967 OS.AddComment("Record length"); 2968 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 2969 OS.emitLabel(BeginLabel); 2970 if (OS.isVerboseAsm()) 2971 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 2972 OS.emitInt16(unsigned(SymKind)); 2973 return EndLabel; 2974 } 2975 2976 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 2977 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 2978 // an extra copy of every symbol record in LLD. This increases object file 2979 // size by less than 1% in the clang build, and is compatible with the Visual 2980 // C++ linker. 2981 OS.emitValueToAlignment(4); 2982 OS.emitLabel(SymEnd); 2983 } 2984 2985 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 2986 OS.AddComment("Record length"); 2987 OS.emitInt16(2); 2988 if (OS.isVerboseAsm()) 2989 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 2990 OS.emitInt16(uint16_t(EndKind)); // Record Kind 2991 } 2992 2993 void CodeViewDebug::emitDebugInfoForUDTs( 2994 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 2995 #ifndef NDEBUG 2996 size_t OriginalSize = UDTs.size(); 2997 #endif 2998 for (const auto &UDT : UDTs) { 2999 const DIType *T = UDT.second; 3000 assert(shouldEmitUdt(T)); 3001 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3002 OS.AddComment("Type"); 3003 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3004 assert(OriginalSize == UDTs.size() && 3005 "getCompleteTypeIndex found new UDTs!"); 3006 emitNullTerminatedSymbolName(OS, UDT.first); 3007 endSymbolRecord(UDTRecordEnd); 3008 } 3009 } 3010 3011 void CodeViewDebug::collectGlobalVariableInfo() { 3012 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3013 GlobalMap; 3014 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3015 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3016 GV.getDebugInfo(GVEs); 3017 for (const auto *GVE : GVEs) 3018 GlobalMap[GVE] = &GV; 3019 } 3020 3021 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3022 for (const MDNode *Node : CUs->operands()) { 3023 const auto *CU = cast<DICompileUnit>(Node); 3024 for (const auto *GVE : CU->getGlobalVariables()) { 3025 const DIGlobalVariable *DIGV = GVE->getVariable(); 3026 const DIExpression *DIE = GVE->getExpression(); 3027 3028 // Emit constant global variables in a global symbol section. 3029 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3030 CVGlobalVariable CVGV = {DIGV, DIE}; 3031 GlobalVariables.emplace_back(std::move(CVGV)); 3032 } 3033 3034 const auto *GV = GlobalMap.lookup(GVE); 3035 if (!GV || GV->isDeclarationForLinker()) 3036 continue; 3037 3038 DIScope *Scope = DIGV->getScope(); 3039 SmallVector<CVGlobalVariable, 1> *VariableList; 3040 if (Scope && isa<DILocalScope>(Scope)) { 3041 // Locate a global variable list for this scope, creating one if 3042 // necessary. 3043 auto Insertion = ScopeGlobals.insert( 3044 {Scope, std::unique_ptr<GlobalVariableList>()}); 3045 if (Insertion.second) 3046 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3047 VariableList = Insertion.first->second.get(); 3048 } else if (GV->hasComdat()) 3049 // Emit this global variable into a COMDAT section. 3050 VariableList = &ComdatVariables; 3051 else 3052 // Emit this global variable in a single global symbol section. 3053 VariableList = &GlobalVariables; 3054 CVGlobalVariable CVGV = {DIGV, GV}; 3055 VariableList->emplace_back(std::move(CVGV)); 3056 } 3057 } 3058 } 3059 3060 void CodeViewDebug::emitDebugInfoForGlobals() { 3061 // First, emit all globals that are not in a comdat in a single symbol 3062 // substream. MSVC doesn't like it if the substream is empty, so only open 3063 // it if we have at least one global to emit. 3064 switchToDebugSectionForSymbol(nullptr); 3065 if (!GlobalVariables.empty()) { 3066 OS.AddComment("Symbol subsection for globals"); 3067 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3068 emitGlobalVariableList(GlobalVariables); 3069 endCVSubsection(EndLabel); 3070 } 3071 3072 // Second, emit each global that is in a comdat into its own .debug$S 3073 // section along with its own symbol substream. 3074 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3075 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3076 MCSymbol *GVSym = Asm->getSymbol(GV); 3077 OS.AddComment("Symbol subsection for " + 3078 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3079 switchToDebugSectionForSymbol(GVSym); 3080 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3081 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3082 emitDebugInfoForGlobal(CVGV); 3083 endCVSubsection(EndLabel); 3084 } 3085 } 3086 3087 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3088 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3089 for (const MDNode *Node : CUs->operands()) { 3090 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3091 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3092 getTypeIndex(RT); 3093 // FIXME: Add to global/local DTU list. 3094 } 3095 } 3096 } 3097 } 3098 3099 // Emit each global variable in the specified array. 3100 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3101 for (const CVGlobalVariable &CVGV : Globals) { 3102 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3103 emitDebugInfoForGlobal(CVGV); 3104 } 3105 } 3106 3107 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3108 const DIGlobalVariable *DIGV = CVGV.DIGV; 3109 3110 const DIScope *Scope = DIGV->getScope(); 3111 // For static data members, get the scope from the declaration. 3112 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3113 DIGV->getRawStaticDataMemberDeclaration())) 3114 Scope = MemberDecl->getScope(); 3115 std::string QualifiedName = getFullyQualifiedName(Scope, DIGV->getName()); 3116 3117 if (const GlobalVariable *GV = 3118 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3119 // DataSym record, see SymbolRecord.h for more info. Thread local data 3120 // happens to have the same format as global data. 3121 MCSymbol *GVSym = Asm->getSymbol(GV); 3122 SymbolKind DataSym = GV->isThreadLocal() 3123 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3124 : SymbolKind::S_GTHREAD32) 3125 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3126 : SymbolKind::S_GDATA32); 3127 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3128 OS.AddComment("Type"); 3129 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3130 OS.AddComment("DataOffset"); 3131 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 3132 OS.AddComment("Segment"); 3133 OS.EmitCOFFSectionIndex(GVSym); 3134 OS.AddComment("Name"); 3135 const unsigned LengthOfDataRecord = 12; 3136 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3137 endSymbolRecord(DataEnd); 3138 } else { 3139 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3140 assert(DIE->isConstant() && 3141 "Global constant variables must contain a constant expression."); 3142 uint64_t Val = DIE->getElement(1); 3143 3144 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3145 OS.AddComment("Type"); 3146 OS.emitInt32(getTypeIndex(DIGV->getType()).getIndex()); 3147 OS.AddComment("Value"); 3148 3149 // Encoded integers shouldn't need more than 10 bytes. 3150 uint8_t data[10]; 3151 BinaryStreamWriter Writer(data, llvm::support::endianness::little); 3152 CodeViewRecordIO IO(Writer); 3153 cantFail(IO.mapEncodedInteger(Val)); 3154 StringRef SRef((char *)data, Writer.getOffset()); 3155 OS.emitBinaryData(SRef); 3156 3157 OS.AddComment("Name"); 3158 emitNullTerminatedSymbolName(OS, QualifiedName); 3159 endSymbolRecord(SConstantEnd); 3160 } 3161 } 3162