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