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