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) : OS(&OS) {} 102 103 void EmitBytes(StringRef Data) { OS->EmitBytes(Data); } 104 105 void EmitIntValue(uint64_t Value, unsigned Size) { 106 OS->EmitIntValue(Value, Size); 107 } 108 109 void EmitBinaryData(StringRef Data) { OS->EmitBinaryData(Data); } 110 111 private: 112 MCStreamer *OS = nullptr; 113 }; 114 } // namespace 115 116 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 117 switch (Type) { 118 case Triple::ArchType::x86: 119 return CPUType::Pentium3; 120 case Triple::ArchType::x86_64: 121 return CPUType::X64; 122 case Triple::ArchType::thumb: 123 return CPUType::Thumb; 124 case Triple::ArchType::aarch64: 125 return CPUType::ARM64; 126 default: 127 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 128 } 129 } 130 131 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 132 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 133 // If module doesn't have named metadata anchors or COFF debug section 134 // is not available, skip any debug info related stuff. 135 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 136 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 137 Asm = nullptr; 138 MMI->setDebugInfoAvailability(false); 139 return; 140 } 141 // Tell MMI that we have debug info. 142 MMI->setDebugInfoAvailability(true); 143 144 TheCPU = 145 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 146 147 collectGlobalVariableInfo(); 148 149 // Check if we should emit type record hashes. 150 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 151 MMI->getModule()->getModuleFlag("CodeViewGHash")); 152 EmitDebugGlobalHashes = GH && !GH->isZero(); 153 } 154 155 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 156 std::string &Filepath = FileToFilepathMap[File]; 157 if (!Filepath.empty()) 158 return Filepath; 159 160 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 161 162 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 163 // it textually because one of the path components could be a symlink. 164 if (Dir.startswith("/") || Filename.startswith("/")) { 165 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 166 return Filename; 167 Filepath = Dir; 168 if (Dir.back() != '/') 169 Filepath += '/'; 170 Filepath += Filename; 171 return Filepath; 172 } 173 174 // Clang emits directory and relative filename info into the IR, but CodeView 175 // operates on full paths. We could change Clang to emit full paths too, but 176 // that would increase the IR size and probably not needed for other users. 177 // For now, just concatenate and canonicalize the path here. 178 if (Filename.find(':') == 1) 179 Filepath = Filename; 180 else 181 Filepath = (Dir + "\\" + Filename).str(); 182 183 // Canonicalize the path. We have to do it textually because we may no longer 184 // have access the file in the filesystem. 185 // First, replace all slashes with backslashes. 186 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 187 188 // Remove all "\.\" with "\". 189 size_t Cursor = 0; 190 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 191 Filepath.erase(Cursor, 2); 192 193 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 194 // path should be well-formatted, e.g. start with a drive letter, etc. 195 Cursor = 0; 196 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 197 // Something's wrong if the path starts with "\..\", abort. 198 if (Cursor == 0) 199 break; 200 201 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 202 if (PrevSlash == std::string::npos) 203 // Something's wrong, abort. 204 break; 205 206 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 207 // The next ".." might be following the one we've just erased. 208 Cursor = PrevSlash; 209 } 210 211 // Remove all duplicate backslashes. 212 Cursor = 0; 213 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 214 Filepath.erase(Cursor, 1); 215 216 return Filepath; 217 } 218 219 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 220 StringRef FullPath = getFullFilepath(F); 221 unsigned NextId = FileIdMap.size() + 1; 222 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 223 if (Insertion.second) { 224 // We have to compute the full filepath and emit a .cv_file directive. 225 ArrayRef<uint8_t> ChecksumAsBytes; 226 FileChecksumKind CSKind = FileChecksumKind::None; 227 if (F->getChecksum()) { 228 std::string Checksum = fromHex(F->getChecksum()->Value); 229 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 230 memcpy(CKMem, Checksum.data(), Checksum.size()); 231 ChecksumAsBytes = ArrayRef<uint8_t>( 232 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 233 switch (F->getChecksum()->Kind) { 234 case DIFile::CSK_MD5: CSKind = FileChecksumKind::MD5; break; 235 case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break; 236 } 237 } 238 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 239 static_cast<unsigned>(CSKind)); 240 (void)Success; 241 assert(Success && ".cv_file directive failed"); 242 } 243 return Insertion.first->second; 244 } 245 246 CodeViewDebug::InlineSite & 247 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 248 const DISubprogram *Inlinee) { 249 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 250 InlineSite *Site = &SiteInsertion.first->second; 251 if (SiteInsertion.second) { 252 unsigned ParentFuncId = CurFn->FuncId; 253 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 254 ParentFuncId = 255 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 256 .SiteFuncId; 257 258 Site->SiteFuncId = NextFuncId++; 259 OS.EmitCVInlineSiteIdDirective( 260 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 261 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 262 Site->Inlinee = Inlinee; 263 InlinedSubprograms.insert(Inlinee); 264 getFuncIdForSubprogram(Inlinee); 265 } 266 return *Site; 267 } 268 269 static StringRef getPrettyScopeName(const DIScope *Scope) { 270 StringRef ScopeName = Scope->getName(); 271 if (!ScopeName.empty()) 272 return ScopeName; 273 274 switch (Scope->getTag()) { 275 case dwarf::DW_TAG_enumeration_type: 276 case dwarf::DW_TAG_class_type: 277 case dwarf::DW_TAG_structure_type: 278 case dwarf::DW_TAG_union_type: 279 return "<unnamed-tag>"; 280 case dwarf::DW_TAG_namespace: 281 return "`anonymous namespace'"; 282 } 283 284 return StringRef(); 285 } 286 287 static const DISubprogram *getQualifiedNameComponents( 288 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 289 const DISubprogram *ClosestSubprogram = nullptr; 290 while (Scope != nullptr) { 291 if (ClosestSubprogram == nullptr) 292 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 293 StringRef ScopeName = getPrettyScopeName(Scope); 294 if (!ScopeName.empty()) 295 QualifiedNameComponents.push_back(ScopeName); 296 Scope = Scope->getScope(); 297 } 298 return ClosestSubprogram; 299 } 300 301 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents, 302 StringRef TypeName) { 303 std::string FullyQualifiedName; 304 for (StringRef QualifiedNameComponent : 305 llvm::reverse(QualifiedNameComponents)) { 306 FullyQualifiedName.append(QualifiedNameComponent); 307 FullyQualifiedName.append("::"); 308 } 309 FullyQualifiedName.append(TypeName); 310 return FullyQualifiedName; 311 } 312 313 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) { 314 SmallVector<StringRef, 5> QualifiedNameComponents; 315 getQualifiedNameComponents(Scope, QualifiedNameComponents); 316 return getQualifiedName(QualifiedNameComponents, Name); 317 } 318 319 struct CodeViewDebug::TypeLoweringScope { 320 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 321 ~TypeLoweringScope() { 322 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 323 // inner TypeLoweringScopes don't attempt to emit deferred types. 324 if (CVD.TypeEmissionLevel == 1) 325 CVD.emitDeferredCompleteTypes(); 326 --CVD.TypeEmissionLevel; 327 } 328 CodeViewDebug &CVD; 329 }; 330 331 static std::string getFullyQualifiedName(const DIScope *Ty) { 332 const DIScope *Scope = Ty->getScope(); 333 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 334 } 335 336 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 337 // No scope means global scope and that uses the zero index. 338 if (!Scope || isa<DIFile>(Scope)) 339 return TypeIndex(); 340 341 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 342 343 // Check if we've already translated this scope. 344 auto I = TypeIndices.find({Scope, nullptr}); 345 if (I != TypeIndices.end()) 346 return I->second; 347 348 // Build the fully qualified name of the scope. 349 std::string ScopeName = getFullyQualifiedName(Scope); 350 StringIdRecord SID(TypeIndex(), ScopeName); 351 auto TI = TypeTable.writeLeafType(SID); 352 return recordTypeIndexForDINode(Scope, TI); 353 } 354 355 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 356 assert(SP); 357 358 // Check if we've already translated this subprogram. 359 auto I = TypeIndices.find({SP, nullptr}); 360 if (I != TypeIndices.end()) 361 return I->second; 362 363 // The display name includes function template arguments. Drop them to match 364 // MSVC. 365 StringRef DisplayName = SP->getName().split('<').first; 366 367 const DIScope *Scope = SP->getScope(); 368 TypeIndex TI; 369 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 370 // If the scope is a DICompositeType, then this must be a method. Member 371 // function types take some special handling, and require access to the 372 // subprogram. 373 TypeIndex ClassType = getTypeIndex(Class); 374 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 375 DisplayName); 376 TI = TypeTable.writeLeafType(MFuncId); 377 } else { 378 // Otherwise, this must be a free function. 379 TypeIndex ParentScope = getScopeIndex(Scope); 380 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 381 TI = TypeTable.writeLeafType(FuncId); 382 } 383 384 return recordTypeIndexForDINode(SP, TI); 385 } 386 387 static bool isNonTrivial(const DICompositeType *DCTy) { 388 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 389 } 390 391 static FunctionOptions 392 getFunctionOptions(const DISubroutineType *Ty, 393 const DICompositeType *ClassTy = nullptr, 394 StringRef SPName = StringRef("")) { 395 FunctionOptions FO = FunctionOptions::None; 396 const DIType *ReturnTy = nullptr; 397 if (auto TypeArray = Ty->getTypeArray()) { 398 if (TypeArray.size()) 399 ReturnTy = TypeArray[0]; 400 } 401 402 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) { 403 if (isNonTrivial(ReturnDCTy)) 404 FO |= FunctionOptions::CxxReturnUdt; 405 } 406 407 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 408 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 409 FO |= FunctionOptions::Constructor; 410 411 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 412 413 } 414 return FO; 415 } 416 417 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 418 const DICompositeType *Class) { 419 // Always use the method declaration as the key for the function type. The 420 // method declaration contains the this adjustment. 421 if (SP->getDeclaration()) 422 SP = SP->getDeclaration(); 423 assert(!SP->getDeclaration() && "should use declaration as key"); 424 425 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 426 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 427 auto I = TypeIndices.find({SP, Class}); 428 if (I != TypeIndices.end()) 429 return I->second; 430 431 // Make sure complete type info for the class is emitted *after* the member 432 // function type, as the complete class type is likely to reference this 433 // member function type. 434 TypeLoweringScope S(*this); 435 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 436 437 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 438 TypeIndex TI = lowerTypeMemberFunction( 439 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 440 return recordTypeIndexForDINode(SP, TI, Class); 441 } 442 443 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 444 TypeIndex TI, 445 const DIType *ClassTy) { 446 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 447 (void)InsertResult; 448 assert(InsertResult.second && "DINode was already assigned a type index"); 449 return TI; 450 } 451 452 unsigned CodeViewDebug::getPointerSizeInBytes() { 453 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 454 } 455 456 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 457 const LexicalScope *LS) { 458 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 459 // This variable was inlined. Associate it with the InlineSite. 460 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 461 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 462 Site.InlinedLocals.emplace_back(Var); 463 } else { 464 // This variable goes into the corresponding lexical scope. 465 ScopeVariables[LS].emplace_back(Var); 466 } 467 } 468 469 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 470 const DILocation *Loc) { 471 auto B = Locs.begin(), E = Locs.end(); 472 if (std::find(B, E, Loc) == E) 473 Locs.push_back(Loc); 474 } 475 476 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 477 const MachineFunction *MF) { 478 // Skip this instruction if it has the same location as the previous one. 479 if (!DL || DL == PrevInstLoc) 480 return; 481 482 const DIScope *Scope = DL.get()->getScope(); 483 if (!Scope) 484 return; 485 486 // Skip this line if it is longer than the maximum we can record. 487 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 488 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 489 LI.isNeverStepInto()) 490 return; 491 492 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 493 if (CI.getStartColumn() != DL.getCol()) 494 return; 495 496 if (!CurFn->HaveLineInfo) 497 CurFn->HaveLineInfo = true; 498 unsigned FileId = 0; 499 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 500 FileId = CurFn->LastFileId; 501 else 502 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 503 PrevInstLoc = DL; 504 505 unsigned FuncId = CurFn->FuncId; 506 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 507 const DILocation *Loc = DL.get(); 508 509 // If this location was actually inlined from somewhere else, give it the ID 510 // of the inline call site. 511 FuncId = 512 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 513 514 // Ensure we have links in the tree of inline call sites. 515 bool FirstLoc = true; 516 while ((SiteLoc = Loc->getInlinedAt())) { 517 InlineSite &Site = 518 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 519 if (!FirstLoc) 520 addLocIfNotPresent(Site.ChildSites, Loc); 521 FirstLoc = false; 522 Loc = SiteLoc; 523 } 524 addLocIfNotPresent(CurFn->ChildSites, Loc); 525 } 526 527 OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 528 /*PrologueEnd=*/false, /*IsStmt=*/false, 529 DL->getFilename(), SMLoc()); 530 } 531 532 void CodeViewDebug::emitCodeViewMagicVersion() { 533 OS.EmitValueToAlignment(4); 534 OS.AddComment("Debug section magic"); 535 OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4); 536 } 537 538 void CodeViewDebug::endModule() { 539 if (!Asm || !MMI->hasDebugInfo()) 540 return; 541 542 assert(Asm != nullptr); 543 544 // The COFF .debug$S section consists of several subsections, each starting 545 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 546 // of the payload followed by the payload itself. The subsections are 4-byte 547 // aligned. 548 549 // Use the generic .debug$S section, and make a subsection for all the inlined 550 // subprograms. 551 switchToDebugSectionForSymbol(nullptr); 552 553 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 554 emitCompilerInformation(); 555 endCVSubsection(CompilerInfo); 556 557 emitInlineeLinesSubsection(); 558 559 // Emit per-function debug information. 560 for (auto &P : FnDebugInfo) 561 if (!P.first->isDeclarationForLinker()) 562 emitDebugInfoForFunction(P.first, *P.second); 563 564 // Emit global variable debug information. 565 setCurrentSubprogram(nullptr); 566 emitDebugInfoForGlobals(); 567 568 // Emit retained types. 569 emitDebugInfoForRetainedTypes(); 570 571 // Switch back to the generic .debug$S section after potentially processing 572 // comdat symbol sections. 573 switchToDebugSectionForSymbol(nullptr); 574 575 // Emit UDT records for any types used by global variables. 576 if (!GlobalUDTs.empty()) { 577 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 578 emitDebugInfoForUDTs(GlobalUDTs); 579 endCVSubsection(SymbolsEnd); 580 } 581 582 // This subsection holds a file index to offset in string table table. 583 OS.AddComment("File index to string table offset subsection"); 584 OS.EmitCVFileChecksumsDirective(); 585 586 // This subsection holds the string table. 587 OS.AddComment("String table"); 588 OS.EmitCVStringTableDirective(); 589 590 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 591 // subsection in the generic .debug$S section at the end. There is no 592 // particular reason for this ordering other than to match MSVC. 593 emitBuildInfo(); 594 595 // Emit type information and hashes last, so that any types we translate while 596 // emitting function info are included. 597 emitTypeInformation(); 598 599 if (EmitDebugGlobalHashes) 600 emitTypeGlobalHashes(); 601 602 clear(); 603 } 604 605 static void 606 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 607 unsigned MaxFixedRecordLength = 0xF00) { 608 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 609 // after a fixed length portion of the record. The fixed length portion should 610 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 611 // overall record size is less than the maximum allowed. 612 SmallString<32> NullTerminatedString( 613 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 614 NullTerminatedString.push_back('\0'); 615 OS.EmitBytes(NullTerminatedString); 616 } 617 618 void CodeViewDebug::emitTypeInformation() { 619 if (TypeTable.empty()) 620 return; 621 622 // Start the .debug$T or .debug$P section with 0x4. 623 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 624 emitCodeViewMagicVersion(); 625 626 SmallString<8> CommentPrefix; 627 if (OS.isVerboseAsm()) { 628 CommentPrefix += '\t'; 629 CommentPrefix += Asm->MAI->getCommentString(); 630 CommentPrefix += ' '; 631 } 632 633 TypeTableCollection Table(TypeTable.records()); 634 SmallString<512> CommentBlock; 635 raw_svector_ostream CommentOS(CommentBlock); 636 std::unique_ptr<ScopedPrinter> SP; 637 std::unique_ptr<TypeDumpVisitor> TDV; 638 TypeVisitorCallbackPipeline Pipeline; 639 640 if (OS.isVerboseAsm()) { 641 // To construct block comment describing the type record for readability. 642 SP = llvm::make_unique<ScopedPrinter>(CommentOS); 643 SP->setPrefix(CommentPrefix); 644 TDV = llvm::make_unique<TypeDumpVisitor>(Table, SP.get(), false); 645 Pipeline.addCallbackToPipeline(*TDV); 646 } 647 648 // To emit type record using Codeview MCStreamer adapter 649 CVMCAdapter CVMCOS(OS); 650 TypeRecordMapping typeMapping(CVMCOS); 651 Pipeline.addCallbackToPipeline(typeMapping); 652 653 Optional<TypeIndex> B = Table.getFirst(); 654 while (B) { 655 // This will fail if the record data is invalid. 656 CVType Record = Table.getType(*B); 657 658 CommentBlock.clear(); 659 660 auto RecordLen = Record.length(); 661 auto RecordKind = Record.kind(); 662 OS.EmitIntValue(RecordLen - 2, 2); 663 OS.EmitIntValue(RecordKind, sizeof(RecordKind)); 664 665 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 666 667 if (E) { 668 logAllUnhandledErrors(std::move(E), errs(), "error: "); 669 llvm_unreachable("produced malformed type record"); 670 } 671 672 if (OS.isVerboseAsm()) { 673 // emitRawComment will insert its own tab and comment string before 674 // the first line, so strip off our first one. It also prints its own 675 // newline. 676 OS.emitRawComment( 677 CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim()); 678 } 679 B = Table.getNext(*B); 680 } 681 } 682 683 void CodeViewDebug::emitTypeGlobalHashes() { 684 if (TypeTable.empty()) 685 return; 686 687 // Start the .debug$H section with the version and hash algorithm, currently 688 // hardcoded to version 0, SHA1. 689 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 690 691 OS.EmitValueToAlignment(4); 692 OS.AddComment("Magic"); 693 OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4); 694 OS.AddComment("Section Version"); 695 OS.EmitIntValue(0, 2); 696 OS.AddComment("Hash Algorithm"); 697 OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2); 698 699 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 700 for (const auto &GHR : TypeTable.hashes()) { 701 if (OS.isVerboseAsm()) { 702 // Emit an EOL-comment describing which TypeIndex this hash corresponds 703 // to, as well as the stringified SHA1 hash. 704 SmallString<32> Comment; 705 raw_svector_ostream CommentOS(Comment); 706 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 707 OS.AddComment(Comment); 708 ++TI; 709 } 710 assert(GHR.Hash.size() == 8); 711 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 712 GHR.Hash.size()); 713 OS.EmitBinaryData(S); 714 } 715 } 716 717 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 718 switch (DWLang) { 719 case dwarf::DW_LANG_C: 720 case dwarf::DW_LANG_C89: 721 case dwarf::DW_LANG_C99: 722 case dwarf::DW_LANG_C11: 723 case dwarf::DW_LANG_ObjC: 724 return SourceLanguage::C; 725 case dwarf::DW_LANG_C_plus_plus: 726 case dwarf::DW_LANG_C_plus_plus_03: 727 case dwarf::DW_LANG_C_plus_plus_11: 728 case dwarf::DW_LANG_C_plus_plus_14: 729 return SourceLanguage::Cpp; 730 case dwarf::DW_LANG_Fortran77: 731 case dwarf::DW_LANG_Fortran90: 732 case dwarf::DW_LANG_Fortran03: 733 case dwarf::DW_LANG_Fortran08: 734 return SourceLanguage::Fortran; 735 case dwarf::DW_LANG_Pascal83: 736 return SourceLanguage::Pascal; 737 case dwarf::DW_LANG_Cobol74: 738 case dwarf::DW_LANG_Cobol85: 739 return SourceLanguage::Cobol; 740 case dwarf::DW_LANG_Java: 741 return SourceLanguage::Java; 742 case dwarf::DW_LANG_D: 743 return SourceLanguage::D; 744 case dwarf::DW_LANG_Swift: 745 return SourceLanguage::Swift; 746 default: 747 // There's no CodeView representation for this language, and CV doesn't 748 // have an "unknown" option for the language field, so we'll use MASM, 749 // as it's very low level. 750 return SourceLanguage::Masm; 751 } 752 } 753 754 namespace { 755 struct Version { 756 int Part[4]; 757 }; 758 } // end anonymous namespace 759 760 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 761 // the version number. 762 static Version parseVersion(StringRef Name) { 763 Version V = {{0}}; 764 int N = 0; 765 for (const char C : Name) { 766 if (isdigit(C)) { 767 V.Part[N] *= 10; 768 V.Part[N] += C - '0'; 769 } else if (C == '.') { 770 ++N; 771 if (N >= 4) 772 return V; 773 } else if (N > 0) 774 return V; 775 } 776 return V; 777 } 778 779 void CodeViewDebug::emitCompilerInformation() { 780 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 781 uint32_t Flags = 0; 782 783 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 784 const MDNode *Node = *CUs->operands().begin(); 785 const auto *CU = cast<DICompileUnit>(Node); 786 787 // The low byte of the flags indicates the source language. 788 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 789 // TODO: Figure out which other flags need to be set. 790 791 OS.AddComment("Flags and language"); 792 OS.EmitIntValue(Flags, 4); 793 794 OS.AddComment("CPUType"); 795 OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2); 796 797 StringRef CompilerVersion = CU->getProducer(); 798 Version FrontVer = parseVersion(CompilerVersion); 799 OS.AddComment("Frontend version"); 800 for (int N = 0; N < 4; ++N) 801 OS.EmitIntValue(FrontVer.Part[N], 2); 802 803 // Some Microsoft tools, like Binscope, expect a backend version number of at 804 // least 8.something, so we'll coerce the LLVM version into a form that 805 // guarantees it'll be big enough without really lying about the version. 806 int Major = 1000 * LLVM_VERSION_MAJOR + 807 10 * LLVM_VERSION_MINOR + 808 LLVM_VERSION_PATCH; 809 // Clamp it for builds that use unusually large version numbers. 810 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 811 Version BackVer = {{ Major, 0, 0, 0 }}; 812 OS.AddComment("Backend version"); 813 for (int N = 0; N < 4; ++N) 814 OS.EmitIntValue(BackVer.Part[N], 2); 815 816 OS.AddComment("Null-terminated compiler version string"); 817 emitNullTerminatedSymbolName(OS, CompilerVersion); 818 819 endSymbolRecord(CompilerEnd); 820 } 821 822 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 823 StringRef S) { 824 StringIdRecord SIR(TypeIndex(0x0), S); 825 return TypeTable.writeLeafType(SIR); 826 } 827 828 void CodeViewDebug::emitBuildInfo() { 829 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 830 // build info. The known prefix is: 831 // - Absolute path of current directory 832 // - Compiler path 833 // - Main source file path, relative to CWD or absolute 834 // - Type server PDB file 835 // - Canonical compiler command line 836 // If frontend and backend compilation are separated (think llc or LTO), it's 837 // not clear if the compiler path should refer to the executable for the 838 // frontend or the backend. Leave it blank for now. 839 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 840 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 841 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 842 const auto *CU = cast<DICompileUnit>(Node); 843 const DIFile *MainSourceFile = CU->getFile(); 844 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 845 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 846 BuildInfoArgs[BuildInfoRecord::SourceFile] = 847 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 848 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 849 // we implement /Zi type servers. 850 BuildInfoRecord BIR(BuildInfoArgs); 851 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 852 853 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 854 // from the module symbols into the type stream. 855 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 856 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 857 OS.AddComment("LF_BUILDINFO index"); 858 OS.EmitIntValue(BuildInfoIndex.getIndex(), 4); 859 endSymbolRecord(BIEnd); 860 endCVSubsection(BISubsecEnd); 861 } 862 863 void CodeViewDebug::emitInlineeLinesSubsection() { 864 if (InlinedSubprograms.empty()) 865 return; 866 867 OS.AddComment("Inlinee lines subsection"); 868 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 869 870 // We emit the checksum info for files. This is used by debuggers to 871 // determine if a pdb matches the source before loading it. Visual Studio, 872 // for instance, will display a warning that the breakpoints are not valid if 873 // the pdb does not match the source. 874 OS.AddComment("Inlinee lines signature"); 875 OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4); 876 877 for (const DISubprogram *SP : InlinedSubprograms) { 878 assert(TypeIndices.count({SP, nullptr})); 879 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 880 881 OS.AddBlankLine(); 882 unsigned FileId = maybeRecordFile(SP->getFile()); 883 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 884 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 885 OS.AddBlankLine(); 886 OS.AddComment("Type index of inlined function"); 887 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 888 OS.AddComment("Offset into filechecksum table"); 889 OS.EmitCVFileChecksumOffsetDirective(FileId); 890 OS.AddComment("Starting line number"); 891 OS.EmitIntValue(SP->getLine(), 4); 892 } 893 894 endCVSubsection(InlineEnd); 895 } 896 897 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 898 const DILocation *InlinedAt, 899 const InlineSite &Site) { 900 assert(TypeIndices.count({Site.Inlinee, nullptr})); 901 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 902 903 // SymbolRecord 904 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 905 906 OS.AddComment("PtrParent"); 907 OS.EmitIntValue(0, 4); 908 OS.AddComment("PtrEnd"); 909 OS.EmitIntValue(0, 4); 910 OS.AddComment("Inlinee type index"); 911 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 912 913 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 914 unsigned StartLineNum = Site.Inlinee->getLine(); 915 916 OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 917 FI.Begin, FI.End); 918 919 endSymbolRecord(InlineEnd); 920 921 emitLocalVariableList(FI, Site.InlinedLocals); 922 923 // Recurse on child inlined call sites before closing the scope. 924 for (const DILocation *ChildSite : Site.ChildSites) { 925 auto I = FI.InlineSites.find(ChildSite); 926 assert(I != FI.InlineSites.end() && 927 "child site not in function inline site map"); 928 emitInlinedCallSite(FI, ChildSite, I->second); 929 } 930 931 // Close the scope. 932 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 933 } 934 935 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 936 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 937 // comdat key. A section may be comdat because of -ffunction-sections or 938 // because it is comdat in the IR. 939 MCSectionCOFF *GVSec = 940 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 941 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 942 943 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 944 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 945 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 946 947 OS.SwitchSection(DebugSec); 948 949 // Emit the magic version number if this is the first time we've switched to 950 // this section. 951 if (ComdatDebugSections.insert(DebugSec).second) 952 emitCodeViewMagicVersion(); 953 } 954 955 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 956 // The only supported thunk ordinal is currently the standard type. 957 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 958 FunctionInfo &FI, 959 const MCSymbol *Fn) { 960 std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 961 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 962 963 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 964 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 965 966 // Emit S_THUNK32 967 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 968 OS.AddComment("PtrParent"); 969 OS.EmitIntValue(0, 4); 970 OS.AddComment("PtrEnd"); 971 OS.EmitIntValue(0, 4); 972 OS.AddComment("PtrNext"); 973 OS.EmitIntValue(0, 4); 974 OS.AddComment("Thunk section relative address"); 975 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 976 OS.AddComment("Thunk section index"); 977 OS.EmitCOFFSectionIndex(Fn); 978 OS.AddComment("Code size"); 979 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 980 OS.AddComment("Ordinal"); 981 OS.EmitIntValue(unsigned(ordinal), 1); 982 OS.AddComment("Function name"); 983 emitNullTerminatedSymbolName(OS, FuncName); 984 // Additional fields specific to the thunk ordinal would go here. 985 endSymbolRecord(ThunkRecordEnd); 986 987 // Local variables/inlined routines are purposely omitted here. The point of 988 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 989 990 // Emit S_PROC_ID_END 991 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 992 993 endCVSubsection(SymbolsEnd); 994 } 995 996 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 997 FunctionInfo &FI) { 998 // For each function there is a separate subsection which holds the PC to 999 // file:line table. 1000 const MCSymbol *Fn = Asm->getSymbol(GV); 1001 assert(Fn); 1002 1003 // Switch to the to a comdat section, if appropriate. 1004 switchToDebugSectionForSymbol(Fn); 1005 1006 std::string FuncName; 1007 auto *SP = GV->getSubprogram(); 1008 assert(SP); 1009 setCurrentSubprogram(SP); 1010 1011 if (SP->isThunk()) { 1012 emitDebugInfoForThunk(GV, FI, Fn); 1013 return; 1014 } 1015 1016 // If we have a display name, build the fully qualified name by walking the 1017 // chain of scopes. 1018 if (!SP->getName().empty()) 1019 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1020 1021 // If our DISubprogram name is empty, use the mangled name. 1022 if (FuncName.empty()) 1023 FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 1024 1025 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1026 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1027 OS.EmitCVFPOData(Fn); 1028 1029 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1030 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1031 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1032 { 1033 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1034 : SymbolKind::S_GPROC32_ID; 1035 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1036 1037 // These fields are filled in by tools like CVPACK which run after the fact. 1038 OS.AddComment("PtrParent"); 1039 OS.EmitIntValue(0, 4); 1040 OS.AddComment("PtrEnd"); 1041 OS.EmitIntValue(0, 4); 1042 OS.AddComment("PtrNext"); 1043 OS.EmitIntValue(0, 4); 1044 // This is the important bit that tells the debugger where the function 1045 // code is located and what's its size: 1046 OS.AddComment("Code size"); 1047 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1048 OS.AddComment("Offset after prologue"); 1049 OS.EmitIntValue(0, 4); 1050 OS.AddComment("Offset before epilogue"); 1051 OS.EmitIntValue(0, 4); 1052 OS.AddComment("Function type index"); 1053 OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4); 1054 OS.AddComment("Function section relative address"); 1055 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1056 OS.AddComment("Function section index"); 1057 OS.EmitCOFFSectionIndex(Fn); 1058 OS.AddComment("Flags"); 1059 OS.EmitIntValue(0, 1); 1060 // Emit the function display name as a null-terminated string. 1061 OS.AddComment("Function name"); 1062 // Truncate the name so we won't overflow the record length field. 1063 emitNullTerminatedSymbolName(OS, FuncName); 1064 endSymbolRecord(ProcRecordEnd); 1065 1066 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1067 // Subtract out the CSR size since MSVC excludes that and we include it. 1068 OS.AddComment("FrameSize"); 1069 OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4); 1070 OS.AddComment("Padding"); 1071 OS.EmitIntValue(0, 4); 1072 OS.AddComment("Offset of padding"); 1073 OS.EmitIntValue(0, 4); 1074 OS.AddComment("Bytes of callee saved registers"); 1075 OS.EmitIntValue(FI.CSRSize, 4); 1076 OS.AddComment("Exception handler offset"); 1077 OS.EmitIntValue(0, 4); 1078 OS.AddComment("Exception handler section"); 1079 OS.EmitIntValue(0, 2); 1080 OS.AddComment("Flags (defines frame register)"); 1081 OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4); 1082 endSymbolRecord(FrameProcEnd); 1083 1084 emitLocalVariableList(FI, FI.Locals); 1085 emitGlobalVariableList(FI.Globals); 1086 emitLexicalBlockList(FI.ChildBlocks, FI); 1087 1088 // Emit inlined call site information. Only emit functions inlined directly 1089 // into the parent function. We'll emit the other sites recursively as part 1090 // of their parent inline site. 1091 for (const DILocation *InlinedAt : FI.ChildSites) { 1092 auto I = FI.InlineSites.find(InlinedAt); 1093 assert(I != FI.InlineSites.end() && 1094 "child site not in function inline site map"); 1095 emitInlinedCallSite(FI, InlinedAt, I->second); 1096 } 1097 1098 for (auto Annot : FI.Annotations) { 1099 MCSymbol *Label = Annot.first; 1100 MDTuple *Strs = cast<MDTuple>(Annot.second); 1101 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1102 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1103 // FIXME: Make sure we don't overflow the max record size. 1104 OS.EmitCOFFSectionIndex(Label); 1105 OS.EmitIntValue(Strs->getNumOperands(), 2); 1106 for (Metadata *MD : Strs->operands()) { 1107 // MDStrings are null terminated, so we can do EmitBytes and get the 1108 // nice .asciz directive. 1109 StringRef Str = cast<MDString>(MD)->getString(); 1110 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1111 OS.EmitBytes(StringRef(Str.data(), Str.size() + 1)); 1112 } 1113 endSymbolRecord(AnnotEnd); 1114 } 1115 1116 for (auto HeapAllocSite : FI.HeapAllocSites) { 1117 MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1118 MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1119 1120 // The labels might not be defined if the instruction was replaced 1121 // somewhere in the codegen pipeline. 1122 if (!BeginLabel->isDefined() || !EndLabel->isDefined()) 1123 continue; 1124 1125 DIType *DITy = std::get<2>(HeapAllocSite); 1126 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1127 OS.AddComment("Call site offset"); 1128 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1129 OS.AddComment("Call site section index"); 1130 OS.EmitCOFFSectionIndex(BeginLabel); 1131 OS.AddComment("Call instruction length"); 1132 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1133 OS.AddComment("Type index"); 1134 OS.EmitIntValue(getCompleteTypeIndex(DITy).getIndex(), 4); 1135 endSymbolRecord(HeapAllocEnd); 1136 } 1137 1138 if (SP != nullptr) 1139 emitDebugInfoForUDTs(LocalUDTs); 1140 1141 // We're done with this function. 1142 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1143 } 1144 endCVSubsection(SymbolsEnd); 1145 1146 // We have an assembler directive that takes care of the whole line table. 1147 OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1148 } 1149 1150 CodeViewDebug::LocalVarDefRange 1151 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1152 LocalVarDefRange DR; 1153 DR.InMemory = -1; 1154 DR.DataOffset = Offset; 1155 assert(DR.DataOffset == Offset && "truncation"); 1156 DR.IsSubfield = 0; 1157 DR.StructOffset = 0; 1158 DR.CVRegister = CVRegister; 1159 return DR; 1160 } 1161 1162 void CodeViewDebug::collectVariableInfoFromMFTable( 1163 DenseSet<InlinedEntity> &Processed) { 1164 const MachineFunction &MF = *Asm->MF; 1165 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1166 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1167 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1168 1169 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1170 if (!VI.Var) 1171 continue; 1172 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1173 "Expected inlined-at fields to agree"); 1174 1175 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1176 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1177 1178 // If variable scope is not found then skip this variable. 1179 if (!Scope) 1180 continue; 1181 1182 // If the variable has an attached offset expression, extract it. 1183 // FIXME: Try to handle DW_OP_deref as well. 1184 int64_t ExprOffset = 0; 1185 bool Deref = false; 1186 if (VI.Expr) { 1187 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1188 if (VI.Expr->getNumElements() == 1 && 1189 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1190 Deref = true; 1191 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1192 continue; 1193 } 1194 1195 // Get the frame register used and the offset. 1196 unsigned FrameReg = 0; 1197 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1198 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1199 1200 // Calculate the label ranges. 1201 LocalVarDefRange DefRange = 1202 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1203 1204 for (const InsnRange &Range : Scope->getRanges()) { 1205 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1206 const MCSymbol *End = getLabelAfterInsn(Range.second); 1207 End = End ? End : Asm->getFunctionEnd(); 1208 DefRange.Ranges.emplace_back(Begin, End); 1209 } 1210 1211 LocalVariable Var; 1212 Var.DIVar = VI.Var; 1213 Var.DefRanges.emplace_back(std::move(DefRange)); 1214 if (Deref) 1215 Var.UseReferenceType = true; 1216 1217 recordLocalVariable(std::move(Var), Scope); 1218 } 1219 } 1220 1221 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1222 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1223 } 1224 1225 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1226 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1227 } 1228 1229 void CodeViewDebug::calculateRanges( 1230 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1231 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1232 1233 // Calculate the definition ranges. 1234 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1235 const auto &Entry = *I; 1236 if (!Entry.isDbgValue()) 1237 continue; 1238 const MachineInstr *DVInst = Entry.getInstr(); 1239 assert(DVInst->isDebugValue() && "Invalid History entry"); 1240 // FIXME: Find a way to represent constant variables, since they are 1241 // relatively common. 1242 Optional<DbgVariableLocation> Location = 1243 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1244 if (!Location) 1245 continue; 1246 1247 // CodeView can only express variables in register and variables in memory 1248 // at a constant offset from a register. However, for variables passed 1249 // indirectly by pointer, it is common for that pointer to be spilled to a 1250 // stack location. For the special case of one offseted load followed by a 1251 // zero offset load (a pointer spilled to the stack), we change the type of 1252 // the local variable from a value type to a reference type. This tricks the 1253 // debugger into doing the load for us. 1254 if (Var.UseReferenceType) { 1255 // We're using a reference type. Drop the last zero offset load. 1256 if (canUseReferenceType(*Location)) 1257 Location->LoadChain.pop_back(); 1258 else 1259 continue; 1260 } else if (needsReferenceType(*Location)) { 1261 // This location can't be expressed without switching to a reference type. 1262 // Start over using that. 1263 Var.UseReferenceType = true; 1264 Var.DefRanges.clear(); 1265 calculateRanges(Var, Entries); 1266 return; 1267 } 1268 1269 // We can only handle a register or an offseted load of a register. 1270 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1271 continue; 1272 { 1273 LocalVarDefRange DR; 1274 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1275 DR.InMemory = !Location->LoadChain.empty(); 1276 DR.DataOffset = 1277 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1278 if (Location->FragmentInfo) { 1279 DR.IsSubfield = true; 1280 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1281 } else { 1282 DR.IsSubfield = false; 1283 DR.StructOffset = 0; 1284 } 1285 1286 if (Var.DefRanges.empty() || 1287 Var.DefRanges.back().isDifferentLocation(DR)) { 1288 Var.DefRanges.emplace_back(std::move(DR)); 1289 } 1290 } 1291 1292 // Compute the label range. 1293 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1294 const MCSymbol *End; 1295 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1296 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1297 End = EndingEntry.isDbgValue() 1298 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1299 : getLabelAfterInsn(EndingEntry.getInstr()); 1300 } else 1301 End = Asm->getFunctionEnd(); 1302 1303 // If the last range end is our begin, just extend the last range. 1304 // Otherwise make a new range. 1305 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1306 Var.DefRanges.back().Ranges; 1307 if (!R.empty() && R.back().second == Begin) 1308 R.back().second = End; 1309 else 1310 R.emplace_back(Begin, End); 1311 1312 // FIXME: Do more range combining. 1313 } 1314 } 1315 1316 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1317 DenseSet<InlinedEntity> Processed; 1318 // Grab the variable info that was squirreled away in the MMI side-table. 1319 collectVariableInfoFromMFTable(Processed); 1320 1321 for (const auto &I : DbgValues) { 1322 InlinedEntity IV = I.first; 1323 if (Processed.count(IV)) 1324 continue; 1325 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1326 const DILocation *InlinedAt = IV.second; 1327 1328 // Instruction ranges, specifying where IV is accessible. 1329 const auto &Entries = I.second; 1330 1331 LexicalScope *Scope = nullptr; 1332 if (InlinedAt) 1333 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1334 else 1335 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1336 // If variable scope is not found then skip this variable. 1337 if (!Scope) 1338 continue; 1339 1340 LocalVariable Var; 1341 Var.DIVar = DIVar; 1342 1343 calculateRanges(Var, Entries); 1344 recordLocalVariable(std::move(Var), Scope); 1345 } 1346 } 1347 1348 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1349 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1350 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1351 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1352 const Function &GV = MF->getFunction(); 1353 auto Insertion = FnDebugInfo.insert({&GV, llvm::make_unique<FunctionInfo>()}); 1354 assert(Insertion.second && "function already has info"); 1355 CurFn = Insertion.first->second.get(); 1356 CurFn->FuncId = NextFuncId++; 1357 CurFn->Begin = Asm->getFunctionBegin(); 1358 1359 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1360 // callee-saved registers were used. For targets that don't use a PUSH 1361 // instruction (AArch64), this will be zero. 1362 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1363 CurFn->FrameSize = MFI.getStackSize(); 1364 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1365 CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF); 1366 1367 // For this function S_FRAMEPROC record, figure out which codeview register 1368 // will be the frame pointer. 1369 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1370 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1371 if (CurFn->FrameSize > 0) { 1372 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1373 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1374 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1375 } else { 1376 // If there is an FP, parameters are always relative to it. 1377 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1378 if (CurFn->HasStackRealignment) { 1379 // If the stack needs realignment, locals are relative to SP or VFRAME. 1380 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1381 } else { 1382 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1383 // other stack adjustments. 1384 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1385 } 1386 } 1387 } 1388 1389 // Compute other frame procedure options. 1390 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1391 if (MFI.hasVarSizedObjects()) 1392 FPO |= FrameProcedureOptions::HasAlloca; 1393 if (MF->exposesReturnsTwice()) 1394 FPO |= FrameProcedureOptions::HasSetJmp; 1395 // FIXME: Set HasLongJmp if we ever track that info. 1396 if (MF->hasInlineAsm()) 1397 FPO |= FrameProcedureOptions::HasInlineAssembly; 1398 if (GV.hasPersonalityFn()) { 1399 if (isAsynchronousEHPersonality( 1400 classifyEHPersonality(GV.getPersonalityFn()))) 1401 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1402 else 1403 FPO |= FrameProcedureOptions::HasExceptionHandling; 1404 } 1405 if (GV.hasFnAttribute(Attribute::InlineHint)) 1406 FPO |= FrameProcedureOptions::MarkedInline; 1407 if (GV.hasFnAttribute(Attribute::Naked)) 1408 FPO |= FrameProcedureOptions::Naked; 1409 if (MFI.hasStackProtectorIndex()) 1410 FPO |= FrameProcedureOptions::SecurityChecks; 1411 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1412 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1413 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1414 !GV.hasOptSize() && !GV.hasOptNone()) 1415 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1416 // FIXME: Set GuardCfg when it is implemented. 1417 CurFn->FrameProcOpts = FPO; 1418 1419 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1420 1421 // Find the end of the function prolog. First known non-DBG_VALUE and 1422 // non-frame setup location marks the beginning of the function body. 1423 // FIXME: is there a simpler a way to do this? Can we just search 1424 // for the first instruction of the function, not the last of the prolog? 1425 DebugLoc PrologEndLoc; 1426 bool EmptyPrologue = true; 1427 for (const auto &MBB : *MF) { 1428 for (const auto &MI : MBB) { 1429 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1430 MI.getDebugLoc()) { 1431 PrologEndLoc = MI.getDebugLoc(); 1432 break; 1433 } else if (!MI.isMetaInstruction()) { 1434 EmptyPrologue = false; 1435 } 1436 } 1437 } 1438 1439 // Record beginning of function if we have a non-empty prologue. 1440 if (PrologEndLoc && !EmptyPrologue) { 1441 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1442 maybeRecordLocation(FnStartDL, MF); 1443 } 1444 } 1445 1446 static bool shouldEmitUdt(const DIType *T) { 1447 if (!T) 1448 return false; 1449 1450 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1451 if (T->getTag() == dwarf::DW_TAG_typedef) { 1452 if (DIScope *Scope = T->getScope()) { 1453 switch (Scope->getTag()) { 1454 case dwarf::DW_TAG_structure_type: 1455 case dwarf::DW_TAG_class_type: 1456 case dwarf::DW_TAG_union_type: 1457 return false; 1458 } 1459 } 1460 } 1461 1462 while (true) { 1463 if (!T || T->isForwardDecl()) 1464 return false; 1465 1466 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1467 if (!DT) 1468 return true; 1469 T = DT->getBaseType(); 1470 } 1471 return true; 1472 } 1473 1474 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1475 // Don't record empty UDTs. 1476 if (Ty->getName().empty()) 1477 return; 1478 if (!shouldEmitUdt(Ty)) 1479 return; 1480 1481 SmallVector<StringRef, 5> QualifiedNameComponents; 1482 const DISubprogram *ClosestSubprogram = 1483 getQualifiedNameComponents(Ty->getScope(), QualifiedNameComponents); 1484 1485 std::string FullyQualifiedName = 1486 getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty)); 1487 1488 if (ClosestSubprogram == nullptr) { 1489 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1490 } else if (ClosestSubprogram == CurrentSubprogram) { 1491 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1492 } 1493 1494 // TODO: What if the ClosestSubprogram is neither null or the current 1495 // subprogram? Currently, the UDT just gets dropped on the floor. 1496 // 1497 // The current behavior is not desirable. To get maximal fidelity, we would 1498 // need to perform all type translation before beginning emission of .debug$S 1499 // and then make LocalUDTs a member of FunctionInfo 1500 } 1501 1502 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1503 // Generic dispatch for lowering an unknown type. 1504 switch (Ty->getTag()) { 1505 case dwarf::DW_TAG_array_type: 1506 return lowerTypeArray(cast<DICompositeType>(Ty)); 1507 case dwarf::DW_TAG_typedef: 1508 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1509 case dwarf::DW_TAG_base_type: 1510 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1511 case dwarf::DW_TAG_pointer_type: 1512 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1513 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1514 LLVM_FALLTHROUGH; 1515 case dwarf::DW_TAG_reference_type: 1516 case dwarf::DW_TAG_rvalue_reference_type: 1517 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1518 case dwarf::DW_TAG_ptr_to_member_type: 1519 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1520 case dwarf::DW_TAG_restrict_type: 1521 case dwarf::DW_TAG_const_type: 1522 case dwarf::DW_TAG_volatile_type: 1523 // TODO: add support for DW_TAG_atomic_type here 1524 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1525 case dwarf::DW_TAG_subroutine_type: 1526 if (ClassTy) { 1527 // The member function type of a member function pointer has no 1528 // ThisAdjustment. 1529 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1530 /*ThisAdjustment=*/0, 1531 /*IsStaticMethod=*/false); 1532 } 1533 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1534 case dwarf::DW_TAG_enumeration_type: 1535 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1536 case dwarf::DW_TAG_class_type: 1537 case dwarf::DW_TAG_structure_type: 1538 return lowerTypeClass(cast<DICompositeType>(Ty)); 1539 case dwarf::DW_TAG_union_type: 1540 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1541 case dwarf::DW_TAG_unspecified_type: 1542 if (Ty->getName() == "decltype(nullptr)") 1543 return TypeIndex::NullptrT(); 1544 return TypeIndex::None(); 1545 default: 1546 // Use the null type index. 1547 return TypeIndex(); 1548 } 1549 } 1550 1551 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1552 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1553 StringRef TypeName = Ty->getName(); 1554 1555 addToUDTs(Ty); 1556 1557 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1558 TypeName == "HRESULT") 1559 return TypeIndex(SimpleTypeKind::HResult); 1560 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1561 TypeName == "wchar_t") 1562 return TypeIndex(SimpleTypeKind::WideCharacter); 1563 1564 return UnderlyingTypeIndex; 1565 } 1566 1567 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1568 const DIType *ElementType = Ty->getBaseType(); 1569 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1570 // IndexType is size_t, which depends on the bitness of the target. 1571 TypeIndex IndexType = getPointerSizeInBytes() == 8 1572 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1573 : TypeIndex(SimpleTypeKind::UInt32Long); 1574 1575 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1576 1577 // Add subranges to array type. 1578 DINodeArray Elements = Ty->getElements(); 1579 for (int i = Elements.size() - 1; i >= 0; --i) { 1580 const DINode *Element = Elements[i]; 1581 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1582 1583 const DISubrange *Subrange = cast<DISubrange>(Element); 1584 assert(Subrange->getLowerBound() == 0 && 1585 "codeview doesn't support subranges with lower bounds"); 1586 int64_t Count = -1; 1587 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>()) 1588 Count = CI->getSExtValue(); 1589 1590 // Forward declarations of arrays without a size and VLAs use a count of -1. 1591 // Emit a count of zero in these cases to match what MSVC does for arrays 1592 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1593 // should do for them even if we could distinguish them. 1594 if (Count == -1) 1595 Count = 0; 1596 1597 // Update the element size and element type index for subsequent subranges. 1598 ElementSize *= Count; 1599 1600 // If this is the outermost array, use the size from the array. It will be 1601 // more accurate if we had a VLA or an incomplete element type size. 1602 uint64_t ArraySize = 1603 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1604 1605 StringRef Name = (i == 0) ? Ty->getName() : ""; 1606 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1607 ElementTypeIndex = TypeTable.writeLeafType(AR); 1608 } 1609 1610 return ElementTypeIndex; 1611 } 1612 1613 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1614 TypeIndex Index; 1615 dwarf::TypeKind Kind; 1616 uint32_t ByteSize; 1617 1618 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1619 ByteSize = Ty->getSizeInBits() / 8; 1620 1621 SimpleTypeKind STK = SimpleTypeKind::None; 1622 switch (Kind) { 1623 case dwarf::DW_ATE_address: 1624 // FIXME: Translate 1625 break; 1626 case dwarf::DW_ATE_boolean: 1627 switch (ByteSize) { 1628 case 1: STK = SimpleTypeKind::Boolean8; break; 1629 case 2: STK = SimpleTypeKind::Boolean16; break; 1630 case 4: STK = SimpleTypeKind::Boolean32; break; 1631 case 8: STK = SimpleTypeKind::Boolean64; break; 1632 case 16: STK = SimpleTypeKind::Boolean128; break; 1633 } 1634 break; 1635 case dwarf::DW_ATE_complex_float: 1636 switch (ByteSize) { 1637 case 2: STK = SimpleTypeKind::Complex16; break; 1638 case 4: STK = SimpleTypeKind::Complex32; break; 1639 case 8: STK = SimpleTypeKind::Complex64; break; 1640 case 10: STK = SimpleTypeKind::Complex80; break; 1641 case 16: STK = SimpleTypeKind::Complex128; break; 1642 } 1643 break; 1644 case dwarf::DW_ATE_float: 1645 switch (ByteSize) { 1646 case 2: STK = SimpleTypeKind::Float16; break; 1647 case 4: STK = SimpleTypeKind::Float32; break; 1648 case 6: STK = SimpleTypeKind::Float48; break; 1649 case 8: STK = SimpleTypeKind::Float64; break; 1650 case 10: STK = SimpleTypeKind::Float80; break; 1651 case 16: STK = SimpleTypeKind::Float128; break; 1652 } 1653 break; 1654 case dwarf::DW_ATE_signed: 1655 switch (ByteSize) { 1656 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1657 case 2: STK = SimpleTypeKind::Int16Short; break; 1658 case 4: STK = SimpleTypeKind::Int32; break; 1659 case 8: STK = SimpleTypeKind::Int64Quad; break; 1660 case 16: STK = SimpleTypeKind::Int128Oct; break; 1661 } 1662 break; 1663 case dwarf::DW_ATE_unsigned: 1664 switch (ByteSize) { 1665 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1666 case 2: STK = SimpleTypeKind::UInt16Short; break; 1667 case 4: STK = SimpleTypeKind::UInt32; break; 1668 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1669 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1670 } 1671 break; 1672 case dwarf::DW_ATE_UTF: 1673 switch (ByteSize) { 1674 case 2: STK = SimpleTypeKind::Character16; break; 1675 case 4: STK = SimpleTypeKind::Character32; break; 1676 } 1677 break; 1678 case dwarf::DW_ATE_signed_char: 1679 if (ByteSize == 1) 1680 STK = SimpleTypeKind::SignedCharacter; 1681 break; 1682 case dwarf::DW_ATE_unsigned_char: 1683 if (ByteSize == 1) 1684 STK = SimpleTypeKind::UnsignedCharacter; 1685 break; 1686 default: 1687 break; 1688 } 1689 1690 // Apply some fixups based on the source-level type name. 1691 if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int") 1692 STK = SimpleTypeKind::Int32Long; 1693 if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int") 1694 STK = SimpleTypeKind::UInt32Long; 1695 if (STK == SimpleTypeKind::UInt16Short && 1696 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1697 STK = SimpleTypeKind::WideCharacter; 1698 if ((STK == SimpleTypeKind::SignedCharacter || 1699 STK == SimpleTypeKind::UnsignedCharacter) && 1700 Ty->getName() == "char") 1701 STK = SimpleTypeKind::NarrowCharacter; 1702 1703 return TypeIndex(STK); 1704 } 1705 1706 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1707 PointerOptions PO) { 1708 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1709 1710 // Pointers to simple types without any options can use SimpleTypeMode, rather 1711 // than having a dedicated pointer type record. 1712 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1713 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1714 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1715 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1716 ? SimpleTypeMode::NearPointer64 1717 : SimpleTypeMode::NearPointer32; 1718 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1719 } 1720 1721 PointerKind PK = 1722 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1723 PointerMode PM = PointerMode::Pointer; 1724 switch (Ty->getTag()) { 1725 default: llvm_unreachable("not a pointer tag type"); 1726 case dwarf::DW_TAG_pointer_type: 1727 PM = PointerMode::Pointer; 1728 break; 1729 case dwarf::DW_TAG_reference_type: 1730 PM = PointerMode::LValueReference; 1731 break; 1732 case dwarf::DW_TAG_rvalue_reference_type: 1733 PM = PointerMode::RValueReference; 1734 break; 1735 } 1736 1737 if (Ty->isObjectPointer()) 1738 PO |= PointerOptions::Const; 1739 1740 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1741 return TypeTable.writeLeafType(PR); 1742 } 1743 1744 static PointerToMemberRepresentation 1745 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1746 // SizeInBytes being zero generally implies that the member pointer type was 1747 // incomplete, which can happen if it is part of a function prototype. In this 1748 // case, use the unknown model instead of the general model. 1749 if (IsPMF) { 1750 switch (Flags & DINode::FlagPtrToMemberRep) { 1751 case 0: 1752 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1753 : PointerToMemberRepresentation::GeneralFunction; 1754 case DINode::FlagSingleInheritance: 1755 return PointerToMemberRepresentation::SingleInheritanceFunction; 1756 case DINode::FlagMultipleInheritance: 1757 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1758 case DINode::FlagVirtualInheritance: 1759 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1760 } 1761 } else { 1762 switch (Flags & DINode::FlagPtrToMemberRep) { 1763 case 0: 1764 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1765 : PointerToMemberRepresentation::GeneralData; 1766 case DINode::FlagSingleInheritance: 1767 return PointerToMemberRepresentation::SingleInheritanceData; 1768 case DINode::FlagMultipleInheritance: 1769 return PointerToMemberRepresentation::MultipleInheritanceData; 1770 case DINode::FlagVirtualInheritance: 1771 return PointerToMemberRepresentation::VirtualInheritanceData; 1772 } 1773 } 1774 llvm_unreachable("invalid ptr to member representation"); 1775 } 1776 1777 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1778 PointerOptions PO) { 1779 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1780 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1781 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType()); 1782 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1783 : PointerKind::Near32; 1784 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1785 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1786 : PointerMode::PointerToDataMember; 1787 1788 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1789 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1790 MemberPointerInfo MPI( 1791 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1792 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1793 return TypeTable.writeLeafType(PR); 1794 } 1795 1796 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1797 /// have a translation, use the NearC convention. 1798 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1799 switch (DwarfCC) { 1800 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1801 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1802 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1803 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1804 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1805 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1806 } 1807 return CallingConvention::NearC; 1808 } 1809 1810 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1811 ModifierOptions Mods = ModifierOptions::None; 1812 PointerOptions PO = PointerOptions::None; 1813 bool IsModifier = true; 1814 const DIType *BaseTy = Ty; 1815 while (IsModifier && BaseTy) { 1816 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1817 switch (BaseTy->getTag()) { 1818 case dwarf::DW_TAG_const_type: 1819 Mods |= ModifierOptions::Const; 1820 PO |= PointerOptions::Const; 1821 break; 1822 case dwarf::DW_TAG_volatile_type: 1823 Mods |= ModifierOptions::Volatile; 1824 PO |= PointerOptions::Volatile; 1825 break; 1826 case dwarf::DW_TAG_restrict_type: 1827 // Only pointer types be marked with __restrict. There is no known flag 1828 // for __restrict in LF_MODIFIER records. 1829 PO |= PointerOptions::Restrict; 1830 break; 1831 default: 1832 IsModifier = false; 1833 break; 1834 } 1835 if (IsModifier) 1836 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1837 } 1838 1839 // Check if the inner type will use an LF_POINTER record. If so, the 1840 // qualifiers will go in the LF_POINTER record. This comes up for types like 1841 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1842 // char *'. 1843 if (BaseTy) { 1844 switch (BaseTy->getTag()) { 1845 case dwarf::DW_TAG_pointer_type: 1846 case dwarf::DW_TAG_reference_type: 1847 case dwarf::DW_TAG_rvalue_reference_type: 1848 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1849 case dwarf::DW_TAG_ptr_to_member_type: 1850 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1851 default: 1852 break; 1853 } 1854 } 1855 1856 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1857 1858 // Return the base type index if there aren't any modifiers. For example, the 1859 // metadata could contain restrict wrappers around non-pointer types. 1860 if (Mods == ModifierOptions::None) 1861 return ModifiedTI; 1862 1863 ModifierRecord MR(ModifiedTI, Mods); 1864 return TypeTable.writeLeafType(MR); 1865 } 1866 1867 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1868 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1869 for (const DIType *ArgType : Ty->getTypeArray()) 1870 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1871 1872 // MSVC uses type none for variadic argument. 1873 if (ReturnAndArgTypeIndices.size() > 1 && 1874 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1875 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1876 } 1877 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1878 ArrayRef<TypeIndex> ArgTypeIndices = None; 1879 if (!ReturnAndArgTypeIndices.empty()) { 1880 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1881 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1882 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1883 } 1884 1885 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1886 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1887 1888 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1889 1890 FunctionOptions FO = getFunctionOptions(Ty); 1891 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 1892 ArgListIndex); 1893 return TypeTable.writeLeafType(Procedure); 1894 } 1895 1896 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 1897 const DIType *ClassTy, 1898 int ThisAdjustment, 1899 bool IsStaticMethod, 1900 FunctionOptions FO) { 1901 // Lower the containing class type. 1902 TypeIndex ClassType = getTypeIndex(ClassTy); 1903 1904 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 1905 1906 unsigned Index = 0; 1907 SmallVector<TypeIndex, 8> ArgTypeIndices; 1908 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1909 if (ReturnAndArgs.size() > Index) { 1910 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 1911 } 1912 1913 // If the first argument is a pointer type and this isn't a static method, 1914 // treat it as the special 'this' parameter, which is encoded separately from 1915 // the arguments. 1916 TypeIndex ThisTypeIndex; 1917 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 1918 if (const DIDerivedType *PtrTy = 1919 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 1920 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 1921 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 1922 Index++; 1923 } 1924 } 1925 } 1926 1927 while (Index < ReturnAndArgs.size()) 1928 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 1929 1930 // MSVC uses type none for variadic argument. 1931 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 1932 ArgTypeIndices.back() = TypeIndex::None(); 1933 1934 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1935 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1936 1937 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1938 1939 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1940 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1941 return TypeTable.writeLeafType(MFR); 1942 } 1943 1944 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1945 unsigned VSlotCount = 1946 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1947 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1948 1949 VFTableShapeRecord VFTSR(Slots); 1950 return TypeTable.writeLeafType(VFTSR); 1951 } 1952 1953 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1954 switch (Flags & DINode::FlagAccessibility) { 1955 case DINode::FlagPrivate: return MemberAccess::Private; 1956 case DINode::FlagPublic: return MemberAccess::Public; 1957 case DINode::FlagProtected: return MemberAccess::Protected; 1958 case 0: 1959 // If there was no explicit access control, provide the default for the tag. 1960 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1961 : MemberAccess::Public; 1962 } 1963 llvm_unreachable("access flags are exclusive"); 1964 } 1965 1966 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1967 if (SP->isArtificial()) 1968 return MethodOptions::CompilerGenerated; 1969 1970 // FIXME: Handle other MethodOptions. 1971 1972 return MethodOptions::None; 1973 } 1974 1975 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 1976 bool Introduced) { 1977 if (SP->getFlags() & DINode::FlagStaticMember) 1978 return MethodKind::Static; 1979 1980 switch (SP->getVirtuality()) { 1981 case dwarf::DW_VIRTUALITY_none: 1982 break; 1983 case dwarf::DW_VIRTUALITY_virtual: 1984 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 1985 case dwarf::DW_VIRTUALITY_pure_virtual: 1986 return Introduced ? MethodKind::PureIntroducingVirtual 1987 : MethodKind::PureVirtual; 1988 default: 1989 llvm_unreachable("unhandled virtuality case"); 1990 } 1991 1992 return MethodKind::Vanilla; 1993 } 1994 1995 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 1996 switch (Ty->getTag()) { 1997 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 1998 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 1999 } 2000 llvm_unreachable("unexpected tag"); 2001 } 2002 2003 /// Return ClassOptions that should be present on both the forward declaration 2004 /// and the defintion of a tag type. 2005 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2006 ClassOptions CO = ClassOptions::None; 2007 2008 // MSVC always sets this flag, even for local types. Clang doesn't always 2009 // appear to give every type a linkage name, which may be problematic for us. 2010 // FIXME: Investigate the consequences of not following them here. 2011 if (!Ty->getIdentifier().empty()) 2012 CO |= ClassOptions::HasUniqueName; 2013 2014 // Put the Nested flag on a type if it appears immediately inside a tag type. 2015 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2016 // here. That flag is only set on definitions, and not forward declarations. 2017 const DIScope *ImmediateScope = Ty->getScope(); 2018 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2019 CO |= ClassOptions::Nested; 2020 2021 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2022 // type only when it has an immediate function scope. Clang never puts enums 2023 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2024 // always in function, class, or file scopes. 2025 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2026 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2027 CO |= ClassOptions::Scoped; 2028 } else { 2029 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2030 Scope = Scope->getScope()) { 2031 if (isa<DISubprogram>(Scope)) { 2032 CO |= ClassOptions::Scoped; 2033 break; 2034 } 2035 } 2036 } 2037 2038 return CO; 2039 } 2040 2041 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2042 switch (Ty->getTag()) { 2043 case dwarf::DW_TAG_class_type: 2044 case dwarf::DW_TAG_structure_type: 2045 case dwarf::DW_TAG_union_type: 2046 case dwarf::DW_TAG_enumeration_type: 2047 break; 2048 default: 2049 return; 2050 } 2051 2052 if (const auto *File = Ty->getFile()) { 2053 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2054 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2055 2056 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2057 TypeTable.writeLeafType(USLR); 2058 } 2059 } 2060 2061 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2062 ClassOptions CO = getCommonClassOptions(Ty); 2063 TypeIndex FTI; 2064 unsigned EnumeratorCount = 0; 2065 2066 if (Ty->isForwardDecl()) { 2067 CO |= ClassOptions::ForwardReference; 2068 } else { 2069 ContinuationRecordBuilder ContinuationBuilder; 2070 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2071 for (const DINode *Element : Ty->getElements()) { 2072 // We assume that the frontend provides all members in source declaration 2073 // order, which is what MSVC does. 2074 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2075 EnumeratorRecord ER(MemberAccess::Public, 2076 APSInt::getUnsigned(Enumerator->getValue()), 2077 Enumerator->getName()); 2078 ContinuationBuilder.writeMemberType(ER); 2079 EnumeratorCount++; 2080 } 2081 } 2082 FTI = TypeTable.insertRecord(ContinuationBuilder); 2083 } 2084 2085 std::string FullName = getFullyQualifiedName(Ty); 2086 2087 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2088 getTypeIndex(Ty->getBaseType())); 2089 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2090 2091 addUDTSrcLine(Ty, EnumTI); 2092 2093 return EnumTI; 2094 } 2095 2096 //===----------------------------------------------------------------------===// 2097 // ClassInfo 2098 //===----------------------------------------------------------------------===// 2099 2100 struct llvm::ClassInfo { 2101 struct MemberInfo { 2102 const DIDerivedType *MemberTypeNode; 2103 uint64_t BaseOffset; 2104 }; 2105 // [MemberInfo] 2106 using MemberList = std::vector<MemberInfo>; 2107 2108 using MethodsList = TinyPtrVector<const DISubprogram *>; 2109 // MethodName -> MethodsList 2110 using MethodsMap = MapVector<MDString *, MethodsList>; 2111 2112 /// Base classes. 2113 std::vector<const DIDerivedType *> Inheritance; 2114 2115 /// Direct members. 2116 MemberList Members; 2117 // Direct overloaded methods gathered by name. 2118 MethodsMap Methods; 2119 2120 TypeIndex VShapeTI; 2121 2122 std::vector<const DIType *> NestedTypes; 2123 }; 2124 2125 void CodeViewDebug::clear() { 2126 assert(CurFn == nullptr); 2127 FileIdMap.clear(); 2128 FnDebugInfo.clear(); 2129 FileToFilepathMap.clear(); 2130 LocalUDTs.clear(); 2131 GlobalUDTs.clear(); 2132 TypeIndices.clear(); 2133 CompleteTypeIndices.clear(); 2134 ScopeGlobals.clear(); 2135 } 2136 2137 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2138 const DIDerivedType *DDTy) { 2139 if (!DDTy->getName().empty()) { 2140 Info.Members.push_back({DDTy, 0}); 2141 return; 2142 } 2143 2144 // An unnamed member may represent a nested struct or union. Attempt to 2145 // interpret the unnamed member as a DICompositeType possibly wrapped in 2146 // qualifier types. Add all the indirect fields to the current record if that 2147 // succeeds, and drop the member if that fails. 2148 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2149 uint64_t Offset = DDTy->getOffsetInBits(); 2150 const DIType *Ty = DDTy->getBaseType(); 2151 bool FullyResolved = false; 2152 while (!FullyResolved) { 2153 switch (Ty->getTag()) { 2154 case dwarf::DW_TAG_const_type: 2155 case dwarf::DW_TAG_volatile_type: 2156 // FIXME: we should apply the qualifier types to the indirect fields 2157 // rather than dropping them. 2158 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2159 break; 2160 default: 2161 FullyResolved = true; 2162 break; 2163 } 2164 } 2165 2166 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2167 if (!DCTy) 2168 return; 2169 2170 ClassInfo NestedInfo = collectClassInfo(DCTy); 2171 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2172 Info.Members.push_back( 2173 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2174 } 2175 2176 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2177 ClassInfo Info; 2178 // Add elements to structure type. 2179 DINodeArray Elements = Ty->getElements(); 2180 for (auto *Element : Elements) { 2181 // We assume that the frontend provides all members in source declaration 2182 // order, which is what MSVC does. 2183 if (!Element) 2184 continue; 2185 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2186 Info.Methods[SP->getRawName()].push_back(SP); 2187 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2188 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2189 collectMemberInfo(Info, DDTy); 2190 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2191 Info.Inheritance.push_back(DDTy); 2192 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2193 DDTy->getName() == "__vtbl_ptr_type") { 2194 Info.VShapeTI = getTypeIndex(DDTy); 2195 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2196 Info.NestedTypes.push_back(DDTy); 2197 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2198 // Ignore friend members. It appears that MSVC emitted info about 2199 // friends in the past, but modern versions do not. 2200 } 2201 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2202 Info.NestedTypes.push_back(Composite); 2203 } 2204 // Skip other unrecognized kinds of elements. 2205 } 2206 return Info; 2207 } 2208 2209 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2210 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2211 // if a complete type should be emitted instead of a forward reference. 2212 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2213 !Ty->isForwardDecl(); 2214 } 2215 2216 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2217 // Emit the complete type for unnamed structs. C++ classes with methods 2218 // which have a circular reference back to the class type are expected to 2219 // be named by the front-end and should not be "unnamed". C unnamed 2220 // structs should not have circular references. 2221 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2222 // If this unnamed complete type is already in the process of being defined 2223 // then the description of the type is malformed and cannot be emitted 2224 // into CodeView correctly so report a fatal error. 2225 auto I = CompleteTypeIndices.find(Ty); 2226 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2227 report_fatal_error("cannot debug circular reference to unnamed type"); 2228 return getCompleteTypeIndex(Ty); 2229 } 2230 2231 // First, construct the forward decl. Don't look into Ty to compute the 2232 // forward decl options, since it might not be available in all TUs. 2233 TypeRecordKind Kind = getRecordKind(Ty); 2234 ClassOptions CO = 2235 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2236 std::string FullName = getFullyQualifiedName(Ty); 2237 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2238 FullName, Ty->getIdentifier()); 2239 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2240 if (!Ty->isForwardDecl()) 2241 DeferredCompleteTypes.push_back(Ty); 2242 return FwdDeclTI; 2243 } 2244 2245 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2246 // Construct the field list and complete type record. 2247 TypeRecordKind Kind = getRecordKind(Ty); 2248 ClassOptions CO = getCommonClassOptions(Ty); 2249 TypeIndex FieldTI; 2250 TypeIndex VShapeTI; 2251 unsigned FieldCount; 2252 bool ContainsNestedClass; 2253 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2254 lowerRecordFieldList(Ty); 2255 2256 if (ContainsNestedClass) 2257 CO |= ClassOptions::ContainsNestedClass; 2258 2259 // MSVC appears to set this flag by searching any destructor or method with 2260 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2261 // the members, however special member functions are not yet emitted into 2262 // debug information. For now checking a class's non-triviality seems enough. 2263 // FIXME: not true for a nested unnamed struct. 2264 if (isNonTrivial(Ty)) 2265 CO |= ClassOptions::HasConstructorOrDestructor; 2266 2267 std::string FullName = getFullyQualifiedName(Ty); 2268 2269 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2270 2271 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2272 SizeInBytes, FullName, Ty->getIdentifier()); 2273 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2274 2275 addUDTSrcLine(Ty, ClassTI); 2276 2277 addToUDTs(Ty); 2278 2279 return ClassTI; 2280 } 2281 2282 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2283 // Emit the complete type for unnamed unions. 2284 if (shouldAlwaysEmitCompleteClassType(Ty)) 2285 return getCompleteTypeIndex(Ty); 2286 2287 ClassOptions CO = 2288 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2289 std::string FullName = getFullyQualifiedName(Ty); 2290 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2291 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2292 if (!Ty->isForwardDecl()) 2293 DeferredCompleteTypes.push_back(Ty); 2294 return FwdDeclTI; 2295 } 2296 2297 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2298 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2299 TypeIndex FieldTI; 2300 unsigned FieldCount; 2301 bool ContainsNestedClass; 2302 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2303 lowerRecordFieldList(Ty); 2304 2305 if (ContainsNestedClass) 2306 CO |= ClassOptions::ContainsNestedClass; 2307 2308 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2309 std::string FullName = getFullyQualifiedName(Ty); 2310 2311 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2312 Ty->getIdentifier()); 2313 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2314 2315 addUDTSrcLine(Ty, UnionTI); 2316 2317 addToUDTs(Ty); 2318 2319 return UnionTI; 2320 } 2321 2322 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2323 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2324 // Manually count members. MSVC appears to count everything that generates a 2325 // field list record. Each individual overload in a method overload group 2326 // contributes to this count, even though the overload group is a single field 2327 // list record. 2328 unsigned MemberCount = 0; 2329 ClassInfo Info = collectClassInfo(Ty); 2330 ContinuationRecordBuilder ContinuationBuilder; 2331 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2332 2333 // Create base classes. 2334 for (const DIDerivedType *I : Info.Inheritance) { 2335 if (I->getFlags() & DINode::FlagVirtual) { 2336 // Virtual base. 2337 unsigned VBPtrOffset = I->getVBPtrOffset(); 2338 // FIXME: Despite the accessor name, the offset is really in bytes. 2339 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2340 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2341 ? TypeRecordKind::IndirectVirtualBaseClass 2342 : TypeRecordKind::VirtualBaseClass; 2343 VirtualBaseClassRecord VBCR( 2344 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2345 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2346 VBTableIndex); 2347 2348 ContinuationBuilder.writeMemberType(VBCR); 2349 MemberCount++; 2350 } else { 2351 assert(I->getOffsetInBits() % 8 == 0 && 2352 "bases must be on byte boundaries"); 2353 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2354 getTypeIndex(I->getBaseType()), 2355 I->getOffsetInBits() / 8); 2356 ContinuationBuilder.writeMemberType(BCR); 2357 MemberCount++; 2358 } 2359 } 2360 2361 // Create members. 2362 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2363 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2364 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2365 StringRef MemberName = Member->getName(); 2366 MemberAccess Access = 2367 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2368 2369 if (Member->isStaticMember()) { 2370 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2371 ContinuationBuilder.writeMemberType(SDMR); 2372 MemberCount++; 2373 continue; 2374 } 2375 2376 // Virtual function pointer member. 2377 if ((Member->getFlags() & DINode::FlagArtificial) && 2378 Member->getName().startswith("_vptr$")) { 2379 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2380 ContinuationBuilder.writeMemberType(VFPR); 2381 MemberCount++; 2382 continue; 2383 } 2384 2385 // Data member. 2386 uint64_t MemberOffsetInBits = 2387 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2388 if (Member->isBitField()) { 2389 uint64_t StartBitOffset = MemberOffsetInBits; 2390 if (const auto *CI = 2391 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2392 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2393 } 2394 StartBitOffset -= MemberOffsetInBits; 2395 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2396 StartBitOffset); 2397 MemberBaseType = TypeTable.writeLeafType(BFR); 2398 } 2399 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2400 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2401 MemberName); 2402 ContinuationBuilder.writeMemberType(DMR); 2403 MemberCount++; 2404 } 2405 2406 // Create methods 2407 for (auto &MethodItr : Info.Methods) { 2408 StringRef Name = MethodItr.first->getString(); 2409 2410 std::vector<OneMethodRecord> Methods; 2411 for (const DISubprogram *SP : MethodItr.second) { 2412 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2413 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2414 2415 unsigned VFTableOffset = -1; 2416 if (Introduced) 2417 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2418 2419 Methods.push_back(OneMethodRecord( 2420 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2421 translateMethodKindFlags(SP, Introduced), 2422 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2423 MemberCount++; 2424 } 2425 assert(!Methods.empty() && "Empty methods map entry"); 2426 if (Methods.size() == 1) 2427 ContinuationBuilder.writeMemberType(Methods[0]); 2428 else { 2429 // FIXME: Make this use its own ContinuationBuilder so that 2430 // MethodOverloadList can be split correctly. 2431 MethodOverloadListRecord MOLR(Methods); 2432 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2433 2434 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2435 ContinuationBuilder.writeMemberType(OMR); 2436 } 2437 } 2438 2439 // Create nested classes. 2440 for (const DIType *Nested : Info.NestedTypes) { 2441 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2442 ContinuationBuilder.writeMemberType(R); 2443 MemberCount++; 2444 } 2445 2446 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2447 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2448 !Info.NestedTypes.empty()); 2449 } 2450 2451 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2452 if (!VBPType.getIndex()) { 2453 // Make a 'const int *' type. 2454 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2455 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2456 2457 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2458 : PointerKind::Near32; 2459 PointerMode PM = PointerMode::Pointer; 2460 PointerOptions PO = PointerOptions::None; 2461 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2462 VBPType = TypeTable.writeLeafType(PR); 2463 } 2464 2465 return VBPType; 2466 } 2467 2468 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2469 // The null DIType is the void type. Don't try to hash it. 2470 if (!Ty) 2471 return TypeIndex::Void(); 2472 2473 // Check if we've already translated this type. Don't try to do a 2474 // get-or-create style insertion that caches the hash lookup across the 2475 // lowerType call. It will update the TypeIndices map. 2476 auto I = TypeIndices.find({Ty, ClassTy}); 2477 if (I != TypeIndices.end()) 2478 return I->second; 2479 2480 TypeLoweringScope S(*this); 2481 TypeIndex TI = lowerType(Ty, ClassTy); 2482 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2483 } 2484 2485 codeview::TypeIndex 2486 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2487 const DISubroutineType *SubroutineTy) { 2488 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2489 "this type must be a pointer type"); 2490 2491 PointerOptions Options = PointerOptions::None; 2492 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2493 Options = PointerOptions::LValueRefThisPointer; 2494 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2495 Options = PointerOptions::RValueRefThisPointer; 2496 2497 // Check if we've already translated this type. If there is no ref qualifier 2498 // on the function then we look up this pointer type with no associated class 2499 // so that the TypeIndex for the this pointer can be shared with the type 2500 // index for other pointers to this class type. If there is a ref qualifier 2501 // then we lookup the pointer using the subroutine as the parent type. 2502 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2503 if (I != TypeIndices.end()) 2504 return I->second; 2505 2506 TypeLoweringScope S(*this); 2507 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2508 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2509 } 2510 2511 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2512 PointerRecord PR(getTypeIndex(Ty), 2513 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2514 : PointerKind::Near32, 2515 PointerMode::LValueReference, PointerOptions::None, 2516 Ty->getSizeInBits() / 8); 2517 return TypeTable.writeLeafType(PR); 2518 } 2519 2520 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2521 // The null DIType is the void type. Don't try to hash it. 2522 if (!Ty) 2523 return TypeIndex::Void(); 2524 2525 // Look through typedefs when getting the complete type index. Call 2526 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2527 // emitted only once. 2528 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2529 (void)getTypeIndex(Ty); 2530 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2531 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2532 2533 // If this is a non-record type, the complete type index is the same as the 2534 // normal type index. Just call getTypeIndex. 2535 switch (Ty->getTag()) { 2536 case dwarf::DW_TAG_class_type: 2537 case dwarf::DW_TAG_structure_type: 2538 case dwarf::DW_TAG_union_type: 2539 break; 2540 default: 2541 return getTypeIndex(Ty); 2542 } 2543 2544 const auto *CTy = cast<DICompositeType>(Ty); 2545 2546 TypeLoweringScope S(*this); 2547 2548 // Make sure the forward declaration is emitted first. It's unclear if this 2549 // is necessary, but MSVC does it, and we should follow suit until we can show 2550 // otherwise. 2551 // We only emit a forward declaration for named types. 2552 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2553 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2554 2555 // Just use the forward decl if we don't have complete type info. This 2556 // might happen if the frontend is using modules and expects the complete 2557 // definition to be emitted elsewhere. 2558 if (CTy->isForwardDecl()) 2559 return FwdDeclTI; 2560 } 2561 2562 // Check if we've already translated the complete record type. 2563 // Insert the type with a null TypeIndex to signify that the type is currently 2564 // being lowered. 2565 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2566 if (!InsertResult.second) 2567 return InsertResult.first->second; 2568 2569 TypeIndex TI; 2570 switch (CTy->getTag()) { 2571 case dwarf::DW_TAG_class_type: 2572 case dwarf::DW_TAG_structure_type: 2573 TI = lowerCompleteTypeClass(CTy); 2574 break; 2575 case dwarf::DW_TAG_union_type: 2576 TI = lowerCompleteTypeUnion(CTy); 2577 break; 2578 default: 2579 llvm_unreachable("not a record"); 2580 } 2581 2582 // Update the type index associated with this CompositeType. This cannot 2583 // use the 'InsertResult' iterator above because it is potentially 2584 // invalidated by map insertions which can occur while lowering the class 2585 // type above. 2586 CompleteTypeIndices[CTy] = TI; 2587 return TI; 2588 } 2589 2590 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2591 /// and do this until fixpoint, as each complete record type typically 2592 /// references 2593 /// many other record types. 2594 void CodeViewDebug::emitDeferredCompleteTypes() { 2595 SmallVector<const DICompositeType *, 4> TypesToEmit; 2596 while (!DeferredCompleteTypes.empty()) { 2597 std::swap(DeferredCompleteTypes, TypesToEmit); 2598 for (const DICompositeType *RecordTy : TypesToEmit) 2599 getCompleteTypeIndex(RecordTy); 2600 TypesToEmit.clear(); 2601 } 2602 } 2603 2604 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2605 ArrayRef<LocalVariable> Locals) { 2606 // Get the sorted list of parameters and emit them first. 2607 SmallVector<const LocalVariable *, 6> Params; 2608 for (const LocalVariable &L : Locals) 2609 if (L.DIVar->isParameter()) 2610 Params.push_back(&L); 2611 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2612 return L->DIVar->getArg() < R->DIVar->getArg(); 2613 }); 2614 for (const LocalVariable *L : Params) 2615 emitLocalVariable(FI, *L); 2616 2617 // Next emit all non-parameters in the order that we found them. 2618 for (const LocalVariable &L : Locals) 2619 if (!L.DIVar->isParameter()) 2620 emitLocalVariable(FI, L); 2621 } 2622 2623 /// Only call this on endian-specific types like ulittle16_t and little32_t, or 2624 /// structs composed of them. 2625 template <typename T> 2626 static void copyBytesForDefRange(SmallString<20> &BytePrefix, 2627 SymbolKind SymKind, const T &DefRangeHeader) { 2628 BytePrefix.resize(2 + sizeof(T)); 2629 ulittle16_t SymKindLE = ulittle16_t(SymKind); 2630 memcpy(&BytePrefix[0], &SymKindLE, 2); 2631 memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T)); 2632 } 2633 2634 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2635 const LocalVariable &Var) { 2636 // LocalSym record, see SymbolRecord.h for more info. 2637 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2638 2639 LocalSymFlags Flags = LocalSymFlags::None; 2640 if (Var.DIVar->isParameter()) 2641 Flags |= LocalSymFlags::IsParameter; 2642 if (Var.DefRanges.empty()) 2643 Flags |= LocalSymFlags::IsOptimizedOut; 2644 2645 OS.AddComment("TypeIndex"); 2646 TypeIndex TI = Var.UseReferenceType 2647 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2648 : getCompleteTypeIndex(Var.DIVar->getType()); 2649 OS.EmitIntValue(TI.getIndex(), 4); 2650 OS.AddComment("Flags"); 2651 OS.EmitIntValue(static_cast<uint16_t>(Flags), 2); 2652 // Truncate the name so we won't overflow the record length field. 2653 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2654 endSymbolRecord(LocalEnd); 2655 2656 // Calculate the on disk prefix of the appropriate def range record. The 2657 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2658 // should be big enough to hold all forms without memory allocation. 2659 SmallString<20> BytePrefix; 2660 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2661 BytePrefix.clear(); 2662 if (DefRange.InMemory) { 2663 int Offset = DefRange.DataOffset; 2664 unsigned Reg = DefRange.CVRegister; 2665 2666 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2667 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2668 // instead. In frames without stack realignment, $T0 will be the CFA. 2669 if (RegisterId(Reg) == RegisterId::ESP) { 2670 Reg = unsigned(RegisterId::VFRAME); 2671 Offset += FI.OffsetAdjustment; 2672 } 2673 2674 // If we can use the chosen frame pointer for the frame and this isn't a 2675 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2676 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2677 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2678 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2679 (bool(Flags & LocalSymFlags::IsParameter) 2680 ? (EncFP == FI.EncodedParamFramePtrReg) 2681 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2682 little32_t FPOffset = little32_t(Offset); 2683 copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset); 2684 } else { 2685 uint16_t RegRelFlags = 0; 2686 if (DefRange.IsSubfield) { 2687 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2688 (DefRange.StructOffset 2689 << DefRangeRegisterRelSym::OffsetInParentShift); 2690 } 2691 DefRangeRegisterRelSym::Header DRHdr; 2692 DRHdr.Register = Reg; 2693 DRHdr.Flags = RegRelFlags; 2694 DRHdr.BasePointerOffset = Offset; 2695 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr); 2696 } 2697 } else { 2698 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2699 if (DefRange.IsSubfield) { 2700 DefRangeSubfieldRegisterSym::Header DRHdr; 2701 DRHdr.Register = DefRange.CVRegister; 2702 DRHdr.MayHaveNoName = 0; 2703 DRHdr.OffsetInParent = DefRange.StructOffset; 2704 copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr); 2705 } else { 2706 DefRangeRegisterSym::Header DRHdr; 2707 DRHdr.Register = DefRange.CVRegister; 2708 DRHdr.MayHaveNoName = 0; 2709 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr); 2710 } 2711 } 2712 OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix); 2713 } 2714 } 2715 2716 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2717 const FunctionInfo& FI) { 2718 for (LexicalBlock *Block : Blocks) 2719 emitLexicalBlock(*Block, FI); 2720 } 2721 2722 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2723 /// lexical block scope. 2724 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2725 const FunctionInfo& FI) { 2726 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2727 OS.AddComment("PtrParent"); 2728 OS.EmitIntValue(0, 4); // PtrParent 2729 OS.AddComment("PtrEnd"); 2730 OS.EmitIntValue(0, 4); // PtrEnd 2731 OS.AddComment("Code size"); 2732 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2733 OS.AddComment("Function section relative address"); 2734 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2735 OS.AddComment("Function section index"); 2736 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2737 OS.AddComment("Lexical block name"); 2738 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2739 endSymbolRecord(RecordEnd); 2740 2741 // Emit variables local to this lexical block. 2742 emitLocalVariableList(FI, Block.Locals); 2743 emitGlobalVariableList(Block.Globals); 2744 2745 // Emit lexical blocks contained within this block. 2746 emitLexicalBlockList(Block.Children, FI); 2747 2748 // Close the lexical block scope. 2749 emitEndSymbolRecord(SymbolKind::S_END); 2750 } 2751 2752 /// Convenience routine for collecting lexical block information for a list 2753 /// of lexical scopes. 2754 void CodeViewDebug::collectLexicalBlockInfo( 2755 SmallVectorImpl<LexicalScope *> &Scopes, 2756 SmallVectorImpl<LexicalBlock *> &Blocks, 2757 SmallVectorImpl<LocalVariable> &Locals, 2758 SmallVectorImpl<CVGlobalVariable> &Globals) { 2759 for (LexicalScope *Scope : Scopes) 2760 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2761 } 2762 2763 /// Populate the lexical blocks and local variable lists of the parent with 2764 /// information about the specified lexical scope. 2765 void CodeViewDebug::collectLexicalBlockInfo( 2766 LexicalScope &Scope, 2767 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2768 SmallVectorImpl<LocalVariable> &ParentLocals, 2769 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2770 if (Scope.isAbstractScope()) 2771 return; 2772 2773 // Gather information about the lexical scope including local variables, 2774 // global variables, and address ranges. 2775 bool IgnoreScope = false; 2776 auto LI = ScopeVariables.find(&Scope); 2777 SmallVectorImpl<LocalVariable> *Locals = 2778 LI != ScopeVariables.end() ? &LI->second : nullptr; 2779 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2780 SmallVectorImpl<CVGlobalVariable> *Globals = 2781 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2782 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2783 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2784 2785 // Ignore lexical scopes which do not contain variables. 2786 if (!Locals && !Globals) 2787 IgnoreScope = true; 2788 2789 // Ignore lexical scopes which are not lexical blocks. 2790 if (!DILB) 2791 IgnoreScope = true; 2792 2793 // Ignore scopes which have too many address ranges to represent in the 2794 // current CodeView format or do not have a valid address range. 2795 // 2796 // For lexical scopes with multiple address ranges you may be tempted to 2797 // construct a single range covering every instruction where the block is 2798 // live and everything in between. Unfortunately, Visual Studio only 2799 // displays variables from the first matching lexical block scope. If the 2800 // first lexical block contains exception handling code or cold code which 2801 // is moved to the bottom of the routine creating a single range covering 2802 // nearly the entire routine, then it will hide all other lexical blocks 2803 // and the variables they contain. 2804 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2805 IgnoreScope = true; 2806 2807 if (IgnoreScope) { 2808 // This scope can be safely ignored and eliminating it will reduce the 2809 // size of the debug information. Be sure to collect any variable and scope 2810 // information from the this scope or any of its children and collapse them 2811 // into the parent scope. 2812 if (Locals) 2813 ParentLocals.append(Locals->begin(), Locals->end()); 2814 if (Globals) 2815 ParentGlobals.append(Globals->begin(), Globals->end()); 2816 collectLexicalBlockInfo(Scope.getChildren(), 2817 ParentBlocks, 2818 ParentLocals, 2819 ParentGlobals); 2820 return; 2821 } 2822 2823 // Create a new CodeView lexical block for this lexical scope. If we've 2824 // seen this DILexicalBlock before then the scope tree is malformed and 2825 // we can handle this gracefully by not processing it a second time. 2826 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2827 if (!BlockInsertion.second) 2828 return; 2829 2830 // Create a lexical block containing the variables and collect the the 2831 // lexical block information for the children. 2832 const InsnRange &Range = Ranges.front(); 2833 assert(Range.first && Range.second); 2834 LexicalBlock &Block = BlockInsertion.first->second; 2835 Block.Begin = getLabelBeforeInsn(Range.first); 2836 Block.End = getLabelAfterInsn(Range.second); 2837 assert(Block.Begin && "missing label for scope begin"); 2838 assert(Block.End && "missing label for scope end"); 2839 Block.Name = DILB->getName(); 2840 if (Locals) 2841 Block.Locals = std::move(*Locals); 2842 if (Globals) 2843 Block.Globals = std::move(*Globals); 2844 ParentBlocks.push_back(&Block); 2845 collectLexicalBlockInfo(Scope.getChildren(), 2846 Block.Children, 2847 Block.Locals, 2848 Block.Globals); 2849 } 2850 2851 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2852 const Function &GV = MF->getFunction(); 2853 assert(FnDebugInfo.count(&GV)); 2854 assert(CurFn == FnDebugInfo[&GV].get()); 2855 2856 collectVariableInfo(GV.getSubprogram()); 2857 2858 // Build the lexical block structure to emit for this routine. 2859 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2860 collectLexicalBlockInfo(*CFS, 2861 CurFn->ChildBlocks, 2862 CurFn->Locals, 2863 CurFn->Globals); 2864 2865 // Clear the scope and variable information from the map which will not be 2866 // valid after we have finished processing this routine. This also prepares 2867 // the map for the subsequent routine. 2868 ScopeVariables.clear(); 2869 2870 // Don't emit anything if we don't have any line tables. 2871 // Thunks are compiler-generated and probably won't have source correlation. 2872 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2873 FnDebugInfo.erase(&GV); 2874 CurFn = nullptr; 2875 return; 2876 } 2877 2878 CurFn->Annotations = MF->getCodeViewAnnotations(); 2879 CurFn->HeapAllocSites = MF->getCodeViewHeapAllocSites(); 2880 2881 CurFn->End = Asm->getFunctionEnd(); 2882 2883 CurFn = nullptr; 2884 } 2885 2886 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2887 DebugHandlerBase::beginInstruction(MI); 2888 2889 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2890 if (!Asm || !CurFn || MI->isDebugInstr() || 2891 MI->getFlag(MachineInstr::FrameSetup)) 2892 return; 2893 2894 // If the first instruction of a new MBB has no location, find the first 2895 // instruction with a location and use that. 2896 DebugLoc DL = MI->getDebugLoc(); 2897 if (!DL && MI->getParent() != PrevInstBB) { 2898 for (const auto &NextMI : *MI->getParent()) { 2899 if (NextMI.isDebugInstr()) 2900 continue; 2901 DL = NextMI.getDebugLoc(); 2902 if (DL) 2903 break; 2904 } 2905 } 2906 PrevInstBB = MI->getParent(); 2907 2908 // If we still don't have a debug location, don't record a location. 2909 if (!DL) 2910 return; 2911 2912 maybeRecordLocation(DL, Asm->MF); 2913 } 2914 2915 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2916 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2917 *EndLabel = MMI->getContext().createTempSymbol(); 2918 OS.EmitIntValue(unsigned(Kind), 4); 2919 OS.AddComment("Subsection size"); 2920 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2921 OS.EmitLabel(BeginLabel); 2922 return EndLabel; 2923 } 2924 2925 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2926 OS.EmitLabel(EndLabel); 2927 // Every subsection must be aligned to a 4-byte boundary. 2928 OS.EmitValueToAlignment(4); 2929 } 2930 2931 static StringRef getSymbolName(SymbolKind SymKind) { 2932 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 2933 if (EE.Value == SymKind) 2934 return EE.Name; 2935 return ""; 2936 } 2937 2938 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 2939 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2940 *EndLabel = MMI->getContext().createTempSymbol(); 2941 OS.AddComment("Record length"); 2942 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 2943 OS.EmitLabel(BeginLabel); 2944 if (OS.isVerboseAsm()) 2945 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 2946 OS.EmitIntValue(unsigned(SymKind), 2); 2947 return EndLabel; 2948 } 2949 2950 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 2951 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 2952 // an extra copy of every symbol record in LLD. This increases object file 2953 // size by less than 1% in the clang build, and is compatible with the Visual 2954 // C++ linker. 2955 OS.EmitValueToAlignment(4); 2956 OS.EmitLabel(SymEnd); 2957 } 2958 2959 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 2960 OS.AddComment("Record length"); 2961 OS.EmitIntValue(2, 2); 2962 if (OS.isVerboseAsm()) 2963 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 2964 OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind 2965 } 2966 2967 void CodeViewDebug::emitDebugInfoForUDTs( 2968 ArrayRef<std::pair<std::string, const DIType *>> UDTs) { 2969 for (const auto &UDT : UDTs) { 2970 const DIType *T = UDT.second; 2971 assert(shouldEmitUdt(T)); 2972 2973 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 2974 OS.AddComment("Type"); 2975 OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4); 2976 emitNullTerminatedSymbolName(OS, UDT.first); 2977 endSymbolRecord(UDTRecordEnd); 2978 } 2979 } 2980 2981 void CodeViewDebug::collectGlobalVariableInfo() { 2982 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 2983 GlobalMap; 2984 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 2985 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 2986 GV.getDebugInfo(GVEs); 2987 for (const auto *GVE : GVEs) 2988 GlobalMap[GVE] = &GV; 2989 } 2990 2991 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2992 for (const MDNode *Node : CUs->operands()) { 2993 const auto *CU = cast<DICompileUnit>(Node); 2994 for (const auto *GVE : CU->getGlobalVariables()) { 2995 const DIGlobalVariable *DIGV = GVE->getVariable(); 2996 const DIExpression *DIE = GVE->getExpression(); 2997 2998 // Emit constant global variables in a global symbol section. 2999 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3000 CVGlobalVariable CVGV = {DIGV, DIE}; 3001 GlobalVariables.emplace_back(std::move(CVGV)); 3002 } 3003 3004 const auto *GV = GlobalMap.lookup(GVE); 3005 if (!GV || GV->isDeclarationForLinker()) 3006 continue; 3007 3008 DIScope *Scope = DIGV->getScope(); 3009 SmallVector<CVGlobalVariable, 1> *VariableList; 3010 if (Scope && isa<DILocalScope>(Scope)) { 3011 // Locate a global variable list for this scope, creating one if 3012 // necessary. 3013 auto Insertion = ScopeGlobals.insert( 3014 {Scope, std::unique_ptr<GlobalVariableList>()}); 3015 if (Insertion.second) 3016 Insertion.first->second = llvm::make_unique<GlobalVariableList>(); 3017 VariableList = Insertion.first->second.get(); 3018 } else if (GV->hasComdat()) 3019 // Emit this global variable into a COMDAT section. 3020 VariableList = &ComdatVariables; 3021 else 3022 // Emit this global variable in a single global symbol section. 3023 VariableList = &GlobalVariables; 3024 CVGlobalVariable CVGV = {DIGV, GV}; 3025 VariableList->emplace_back(std::move(CVGV)); 3026 } 3027 } 3028 } 3029 3030 void CodeViewDebug::emitDebugInfoForGlobals() { 3031 // First, emit all globals that are not in a comdat in a single symbol 3032 // substream. MSVC doesn't like it if the substream is empty, so only open 3033 // it if we have at least one global to emit. 3034 switchToDebugSectionForSymbol(nullptr); 3035 if (!GlobalVariables.empty()) { 3036 OS.AddComment("Symbol subsection for globals"); 3037 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3038 emitGlobalVariableList(GlobalVariables); 3039 endCVSubsection(EndLabel); 3040 } 3041 3042 // Second, emit each global that is in a comdat into its own .debug$S 3043 // section along with its own symbol substream. 3044 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3045 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3046 MCSymbol *GVSym = Asm->getSymbol(GV); 3047 OS.AddComment("Symbol subsection for " + 3048 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3049 switchToDebugSectionForSymbol(GVSym); 3050 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3051 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3052 emitDebugInfoForGlobal(CVGV); 3053 endCVSubsection(EndLabel); 3054 } 3055 } 3056 3057 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3058 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3059 for (const MDNode *Node : CUs->operands()) { 3060 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3061 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3062 getTypeIndex(RT); 3063 // FIXME: Add to global/local DTU list. 3064 } 3065 } 3066 } 3067 } 3068 3069 // Emit each global variable in the specified array. 3070 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3071 for (const CVGlobalVariable &CVGV : Globals) { 3072 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3073 emitDebugInfoForGlobal(CVGV); 3074 } 3075 } 3076 3077 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3078 const DIGlobalVariable *DIGV = CVGV.DIGV; 3079 if (const GlobalVariable *GV = 3080 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3081 // DataSym record, see SymbolRecord.h for more info. Thread local data 3082 // happens to have the same format as global data. 3083 MCSymbol *GVSym = Asm->getSymbol(GV); 3084 SymbolKind DataSym = GV->isThreadLocal() 3085 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3086 : SymbolKind::S_GTHREAD32) 3087 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3088 : SymbolKind::S_GDATA32); 3089 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3090 OS.AddComment("Type"); 3091 OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4); 3092 OS.AddComment("DataOffset"); 3093 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 3094 OS.AddComment("Segment"); 3095 OS.EmitCOFFSectionIndex(GVSym); 3096 OS.AddComment("Name"); 3097 const unsigned LengthOfDataRecord = 12; 3098 emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord); 3099 endSymbolRecord(DataEnd); 3100 } else { 3101 // FIXME: Currently this only emits the global variables in the IR metadata. 3102 // This should also emit enums and static data members. 3103 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3104 assert(DIE->isConstant() && 3105 "Global constant variables must contain a constant expression."); 3106 uint64_t Val = DIE->getElement(1); 3107 3108 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3109 OS.AddComment("Type"); 3110 OS.EmitIntValue(getTypeIndex(DIGV->getType()).getIndex(), 4); 3111 OS.AddComment("Value"); 3112 3113 // Encoded integers shouldn't need more than 10 bytes. 3114 uint8_t data[10]; 3115 BinaryStreamWriter Writer(data, llvm::support::endianness::little); 3116 CodeViewRecordIO IO(Writer); 3117 cantFail(IO.mapEncodedInteger(Val)); 3118 StringRef SRef((char *)data, Writer.getOffset()); 3119 OS.EmitBinaryData(SRef); 3120 3121 OS.AddComment("Name"); 3122 const DIScope *Scope = DIGV->getScope(); 3123 // For static data members, get the scope from the declaration. 3124 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3125 DIGV->getRawStaticDataMemberDeclaration())) 3126 Scope = MemberDecl->getScope(); 3127 emitNullTerminatedSymbolName(OS, 3128 getFullyQualifiedName(Scope, DIGV->getName())); 3129 endSymbolRecord(SConstantEnd); 3130 } 3131 } 3132