1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains support for writing Microsoft CodeView debug info. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeViewDebug.h" 15 #include "DwarfExpression.h" 16 #include "llvm/ADT/APSInt.h" 17 #include "llvm/ADT/ArrayRef.h" 18 #include "llvm/ADT/DenseMap.h" 19 #include "llvm/ADT/DenseSet.h" 20 #include "llvm/ADT/MapVector.h" 21 #include "llvm/ADT/None.h" 22 #include "llvm/ADT/Optional.h" 23 #include "llvm/ADT/STLExtras.h" 24 #include "llvm/ADT/SmallString.h" 25 #include "llvm/ADT/SmallVector.h" 26 #include "llvm/ADT/StringRef.h" 27 #include "llvm/ADT/TinyPtrVector.h" 28 #include "llvm/ADT/Triple.h" 29 #include "llvm/ADT/Twine.h" 30 #include "llvm/BinaryFormat/COFF.h" 31 #include "llvm/BinaryFormat/Dwarf.h" 32 #include "llvm/CodeGen/AsmPrinter.h" 33 #include "llvm/CodeGen/LexicalScopes.h" 34 #include "llvm/CodeGen/MachineFrameInfo.h" 35 #include "llvm/CodeGen/MachineFunction.h" 36 #include "llvm/CodeGen/MachineInstr.h" 37 #include "llvm/CodeGen/MachineModuleInfo.h" 38 #include "llvm/CodeGen/MachineOperand.h" 39 #include "llvm/CodeGen/TargetFrameLowering.h" 40 #include "llvm/CodeGen/TargetRegisterInfo.h" 41 #include "llvm/CodeGen/TargetSubtargetInfo.h" 42 #include "llvm/Config/llvm-config.h" 43 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 44 #include "llvm/DebugInfo/CodeView/CodeView.h" 45 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 46 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 47 #include "llvm/DebugInfo/CodeView/Line.h" 48 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 49 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 50 #include "llvm/DebugInfo/CodeView/TypeIndex.h" 51 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 52 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 53 #include "llvm/IR/Constants.h" 54 #include "llvm/IR/DataLayout.h" 55 #include "llvm/IR/DebugInfoMetadata.h" 56 #include "llvm/IR/DebugLoc.h" 57 #include "llvm/IR/Function.h" 58 #include "llvm/IR/GlobalValue.h" 59 #include "llvm/IR/GlobalVariable.h" 60 #include "llvm/IR/Metadata.h" 61 #include "llvm/IR/Module.h" 62 #include "llvm/MC/MCAsmInfo.h" 63 #include "llvm/MC/MCContext.h" 64 #include "llvm/MC/MCSectionCOFF.h" 65 #include "llvm/MC/MCStreamer.h" 66 #include "llvm/MC/MCSymbol.h" 67 #include "llvm/Support/BinaryByteStream.h" 68 #include "llvm/Support/BinaryStreamReader.h" 69 #include "llvm/Support/Casting.h" 70 #include "llvm/Support/CommandLine.h" 71 #include "llvm/Support/Compiler.h" 72 #include "llvm/Support/Endian.h" 73 #include "llvm/Support/Error.h" 74 #include "llvm/Support/ErrorHandling.h" 75 #include "llvm/Support/FormatVariadic.h" 76 #include "llvm/Support/Path.h" 77 #include "llvm/Support/SMLoc.h" 78 #include "llvm/Support/ScopedPrinter.h" 79 #include "llvm/Target/TargetLoweringObjectFile.h" 80 #include "llvm/Target/TargetMachine.h" 81 #include <algorithm> 82 #include <cassert> 83 #include <cctype> 84 #include <cstddef> 85 #include <cstdint> 86 #include <iterator> 87 #include <limits> 88 #include <string> 89 #include <utility> 90 #include <vector> 91 92 using namespace llvm; 93 using namespace llvm::codeview; 94 95 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 96 switch (Type) { 97 case Triple::ArchType::x86: 98 return CPUType::Pentium3; 99 case Triple::ArchType::x86_64: 100 return CPUType::X64; 101 case Triple::ArchType::thumb: 102 return CPUType::Thumb; 103 case Triple::ArchType::aarch64: 104 return CPUType::ARM64; 105 default: 106 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 107 } 108 } 109 110 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 111 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 112 // If module doesn't have named metadata anchors or COFF debug section 113 // is not available, skip any debug info related stuff. 114 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 115 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 116 Asm = nullptr; 117 MMI->setDebugInfoAvailability(false); 118 return; 119 } 120 // Tell MMI that we have debug info. 121 MMI->setDebugInfoAvailability(true); 122 123 TheCPU = 124 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 125 126 // Check if we should emit type record hashes. 127 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 128 MMI->getModule()->getModuleFlag("CodeViewGHash")); 129 EmitDebugGlobalHashes = GH && !GH->isZero(); 130 } 131 132 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 133 std::string &Filepath = FileToFilepathMap[File]; 134 if (!Filepath.empty()) 135 return Filepath; 136 137 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 138 139 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 140 // it textually because one of the path components could be a symlink. 141 if (Dir.startswith("/") || Filename.startswith("/")) { 142 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 143 return Filename; 144 Filepath = Dir; 145 if (Dir.back() != '/') 146 Filepath += '/'; 147 Filepath += Filename; 148 return Filepath; 149 } 150 151 // Clang emits directory and relative filename info into the IR, but CodeView 152 // operates on full paths. We could change Clang to emit full paths too, but 153 // that would increase the IR size and probably not needed for other users. 154 // For now, just concatenate and canonicalize the path here. 155 if (Filename.find(':') == 1) 156 Filepath = Filename; 157 else 158 Filepath = (Dir + "\\" + Filename).str(); 159 160 // Canonicalize the path. We have to do it textually because we may no longer 161 // have access the file in the filesystem. 162 // First, replace all slashes with backslashes. 163 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 164 165 // Remove all "\.\" with "\". 166 size_t Cursor = 0; 167 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 168 Filepath.erase(Cursor, 2); 169 170 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 171 // path should be well-formatted, e.g. start with a drive letter, etc. 172 Cursor = 0; 173 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 174 // Something's wrong if the path starts with "\..\", abort. 175 if (Cursor == 0) 176 break; 177 178 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 179 if (PrevSlash == std::string::npos) 180 // Something's wrong, abort. 181 break; 182 183 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 184 // The next ".." might be following the one we've just erased. 185 Cursor = PrevSlash; 186 } 187 188 // Remove all duplicate backslashes. 189 Cursor = 0; 190 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 191 Filepath.erase(Cursor, 1); 192 193 return Filepath; 194 } 195 196 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 197 StringRef FullPath = getFullFilepath(F); 198 unsigned NextId = FileIdMap.size() + 1; 199 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 200 if (Insertion.second) { 201 // We have to compute the full filepath and emit a .cv_file directive. 202 ArrayRef<uint8_t> ChecksumAsBytes; 203 FileChecksumKind CSKind = FileChecksumKind::None; 204 if (F->getChecksum()) { 205 std::string Checksum = fromHex(F->getChecksum()->Value); 206 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 207 memcpy(CKMem, Checksum.data(), Checksum.size()); 208 ChecksumAsBytes = ArrayRef<uint8_t>( 209 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 210 switch (F->getChecksum()->Kind) { 211 case DIFile::CSK_MD5: CSKind = FileChecksumKind::MD5; break; 212 case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break; 213 } 214 } 215 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 216 static_cast<unsigned>(CSKind)); 217 (void)Success; 218 assert(Success && ".cv_file directive failed"); 219 } 220 return Insertion.first->second; 221 } 222 223 CodeViewDebug::InlineSite & 224 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 225 const DISubprogram *Inlinee) { 226 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 227 InlineSite *Site = &SiteInsertion.first->second; 228 if (SiteInsertion.second) { 229 unsigned ParentFuncId = CurFn->FuncId; 230 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 231 ParentFuncId = 232 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 233 .SiteFuncId; 234 235 Site->SiteFuncId = NextFuncId++; 236 OS.EmitCVInlineSiteIdDirective( 237 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 238 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 239 Site->Inlinee = Inlinee; 240 InlinedSubprograms.insert(Inlinee); 241 getFuncIdForSubprogram(Inlinee); 242 } 243 return *Site; 244 } 245 246 static StringRef getPrettyScopeName(const DIScope *Scope) { 247 StringRef ScopeName = Scope->getName(); 248 if (!ScopeName.empty()) 249 return ScopeName; 250 251 switch (Scope->getTag()) { 252 case dwarf::DW_TAG_enumeration_type: 253 case dwarf::DW_TAG_class_type: 254 case dwarf::DW_TAG_structure_type: 255 case dwarf::DW_TAG_union_type: 256 return "<unnamed-tag>"; 257 case dwarf::DW_TAG_namespace: 258 return "`anonymous namespace'"; 259 } 260 261 return StringRef(); 262 } 263 264 static const DISubprogram *getQualifiedNameComponents( 265 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 266 const DISubprogram *ClosestSubprogram = nullptr; 267 while (Scope != nullptr) { 268 if (ClosestSubprogram == nullptr) 269 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 270 StringRef ScopeName = getPrettyScopeName(Scope); 271 if (!ScopeName.empty()) 272 QualifiedNameComponents.push_back(ScopeName); 273 Scope = Scope->getScope().resolve(); 274 } 275 return ClosestSubprogram; 276 } 277 278 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents, 279 StringRef TypeName) { 280 std::string FullyQualifiedName; 281 for (StringRef QualifiedNameComponent : 282 llvm::reverse(QualifiedNameComponents)) { 283 FullyQualifiedName.append(QualifiedNameComponent); 284 FullyQualifiedName.append("::"); 285 } 286 FullyQualifiedName.append(TypeName); 287 return FullyQualifiedName; 288 } 289 290 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) { 291 SmallVector<StringRef, 5> QualifiedNameComponents; 292 getQualifiedNameComponents(Scope, QualifiedNameComponents); 293 return getQualifiedName(QualifiedNameComponents, Name); 294 } 295 296 struct CodeViewDebug::TypeLoweringScope { 297 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 298 ~TypeLoweringScope() { 299 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 300 // inner TypeLoweringScopes don't attempt to emit deferred types. 301 if (CVD.TypeEmissionLevel == 1) 302 CVD.emitDeferredCompleteTypes(); 303 --CVD.TypeEmissionLevel; 304 } 305 CodeViewDebug &CVD; 306 }; 307 308 static std::string getFullyQualifiedName(const DIScope *Ty) { 309 const DIScope *Scope = Ty->getScope().resolve(); 310 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 311 } 312 313 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 314 // No scope means global scope and that uses the zero index. 315 if (!Scope || isa<DIFile>(Scope)) 316 return TypeIndex(); 317 318 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 319 320 // Check if we've already translated this scope. 321 auto I = TypeIndices.find({Scope, nullptr}); 322 if (I != TypeIndices.end()) 323 return I->second; 324 325 // Build the fully qualified name of the scope. 326 std::string ScopeName = getFullyQualifiedName(Scope); 327 StringIdRecord SID(TypeIndex(), ScopeName); 328 auto TI = TypeTable.writeLeafType(SID); 329 return recordTypeIndexForDINode(Scope, TI); 330 } 331 332 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 333 assert(SP); 334 335 // Check if we've already translated this subprogram. 336 auto I = TypeIndices.find({SP, nullptr}); 337 if (I != TypeIndices.end()) 338 return I->second; 339 340 // The display name includes function template arguments. Drop them to match 341 // MSVC. 342 StringRef DisplayName = SP->getName().split('<').first; 343 344 const DIScope *Scope = SP->getScope().resolve(); 345 TypeIndex TI; 346 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 347 // If the scope is a DICompositeType, then this must be a method. Member 348 // function types take some special handling, and require access to the 349 // subprogram. 350 TypeIndex ClassType = getTypeIndex(Class); 351 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 352 DisplayName); 353 TI = TypeTable.writeLeafType(MFuncId); 354 } else { 355 // Otherwise, this must be a free function. 356 TypeIndex ParentScope = getScopeIndex(Scope); 357 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 358 TI = TypeTable.writeLeafType(FuncId); 359 } 360 361 return recordTypeIndexForDINode(SP, TI); 362 } 363 364 static bool isTrivial(const DICompositeType *DCTy) { 365 return ((DCTy->getFlags() & DINode::FlagTrivial) == DINode::FlagTrivial); 366 } 367 368 static FunctionOptions 369 getFunctionOptions(const DISubroutineType *Ty, 370 const DICompositeType *ClassTy = nullptr, 371 StringRef SPName = StringRef("")) { 372 FunctionOptions FO = FunctionOptions::None; 373 const DIType *ReturnTy = nullptr; 374 if (auto TypeArray = Ty->getTypeArray()) { 375 if (TypeArray.size()) 376 ReturnTy = TypeArray[0].resolve(); 377 } 378 379 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) { 380 if (!isTrivial(ReturnDCTy)) 381 FO |= FunctionOptions::CxxReturnUdt; 382 } 383 384 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 385 if (ClassTy && !isTrivial(ClassTy) && SPName == ClassTy->getName()) { 386 FO |= FunctionOptions::Constructor; 387 388 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 389 390 } 391 return FO; 392 } 393 394 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 395 const DICompositeType *Class) { 396 // Always use the method declaration as the key for the function type. The 397 // method declaration contains the this adjustment. 398 if (SP->getDeclaration()) 399 SP = SP->getDeclaration(); 400 assert(!SP->getDeclaration() && "should use declaration as key"); 401 402 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 403 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 404 auto I = TypeIndices.find({SP, Class}); 405 if (I != TypeIndices.end()) 406 return I->second; 407 408 // Make sure complete type info for the class is emitted *after* the member 409 // function type, as the complete class type is likely to reference this 410 // member function type. 411 TypeLoweringScope S(*this); 412 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 413 414 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 415 TypeIndex TI = lowerTypeMemberFunction( 416 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 417 return recordTypeIndexForDINode(SP, TI, Class); 418 } 419 420 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 421 TypeIndex TI, 422 const DIType *ClassTy) { 423 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 424 (void)InsertResult; 425 assert(InsertResult.second && "DINode was already assigned a type index"); 426 return TI; 427 } 428 429 unsigned CodeViewDebug::getPointerSizeInBytes() { 430 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 431 } 432 433 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 434 const LexicalScope *LS) { 435 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 436 // This variable was inlined. Associate it with the InlineSite. 437 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 438 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 439 Site.InlinedLocals.emplace_back(Var); 440 } else { 441 // This variable goes into the corresponding lexical scope. 442 ScopeVariables[LS].emplace_back(Var); 443 } 444 } 445 446 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 447 const DILocation *Loc) { 448 auto B = Locs.begin(), E = Locs.end(); 449 if (std::find(B, E, Loc) == E) 450 Locs.push_back(Loc); 451 } 452 453 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 454 const MachineFunction *MF) { 455 // Skip this instruction if it has the same location as the previous one. 456 if (!DL || DL == PrevInstLoc) 457 return; 458 459 const DIScope *Scope = DL.get()->getScope(); 460 if (!Scope) 461 return; 462 463 // Skip this line if it is longer than the maximum we can record. 464 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 465 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 466 LI.isNeverStepInto()) 467 return; 468 469 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 470 if (CI.getStartColumn() != DL.getCol()) 471 return; 472 473 if (!CurFn->HaveLineInfo) 474 CurFn->HaveLineInfo = true; 475 unsigned FileId = 0; 476 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 477 FileId = CurFn->LastFileId; 478 else 479 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 480 PrevInstLoc = DL; 481 482 unsigned FuncId = CurFn->FuncId; 483 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 484 const DILocation *Loc = DL.get(); 485 486 // If this location was actually inlined from somewhere else, give it the ID 487 // of the inline call site. 488 FuncId = 489 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 490 491 // Ensure we have links in the tree of inline call sites. 492 bool FirstLoc = true; 493 while ((SiteLoc = Loc->getInlinedAt())) { 494 InlineSite &Site = 495 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 496 if (!FirstLoc) 497 addLocIfNotPresent(Site.ChildSites, Loc); 498 FirstLoc = false; 499 Loc = SiteLoc; 500 } 501 addLocIfNotPresent(CurFn->ChildSites, Loc); 502 } 503 504 OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 505 /*PrologueEnd=*/false, /*IsStmt=*/false, 506 DL->getFilename(), SMLoc()); 507 } 508 509 void CodeViewDebug::emitCodeViewMagicVersion() { 510 OS.EmitValueToAlignment(4); 511 OS.AddComment("Debug section magic"); 512 OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4); 513 } 514 515 void CodeViewDebug::endModule() { 516 if (!Asm || !MMI->hasDebugInfo()) 517 return; 518 519 assert(Asm != nullptr); 520 521 // The COFF .debug$S section consists of several subsections, each starting 522 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 523 // of the payload followed by the payload itself. The subsections are 4-byte 524 // aligned. 525 526 // Use the generic .debug$S section, and make a subsection for all the inlined 527 // subprograms. 528 switchToDebugSectionForSymbol(nullptr); 529 530 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 531 emitCompilerInformation(); 532 endCVSubsection(CompilerInfo); 533 534 emitInlineeLinesSubsection(); 535 536 // Emit per-function debug information. 537 for (auto &P : FnDebugInfo) 538 if (!P.first->isDeclarationForLinker()) 539 emitDebugInfoForFunction(P.first, *P.second); 540 541 // Emit global variable debug information. 542 setCurrentSubprogram(nullptr); 543 emitDebugInfoForGlobals(); 544 545 // Emit retained types. 546 emitDebugInfoForRetainedTypes(); 547 548 // Switch back to the generic .debug$S section after potentially processing 549 // comdat symbol sections. 550 switchToDebugSectionForSymbol(nullptr); 551 552 // Emit UDT records for any types used by global variables. 553 if (!GlobalUDTs.empty()) { 554 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 555 emitDebugInfoForUDTs(GlobalUDTs); 556 endCVSubsection(SymbolsEnd); 557 } 558 559 // This subsection holds a file index to offset in string table table. 560 OS.AddComment("File index to string table offset subsection"); 561 OS.EmitCVFileChecksumsDirective(); 562 563 // This subsection holds the string table. 564 OS.AddComment("String table"); 565 OS.EmitCVStringTableDirective(); 566 567 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 568 // subsection in the generic .debug$S section at the end. There is no 569 // particular reason for this ordering other than to match MSVC. 570 emitBuildInfo(); 571 572 // Emit type information and hashes last, so that any types we translate while 573 // emitting function info are included. 574 emitTypeInformation(); 575 576 if (EmitDebugGlobalHashes) 577 emitTypeGlobalHashes(); 578 579 clear(); 580 } 581 582 static void emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 583 unsigned MaxFixedRecordLength = 0xF00) { 584 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 585 // after a fixed length portion of the record. The fixed length portion should 586 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 587 // overall record size is less than the maximum allowed. 588 SmallString<32> NullTerminatedString( 589 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 590 NullTerminatedString.push_back('\0'); 591 OS.EmitBytes(NullTerminatedString); 592 } 593 594 void CodeViewDebug::emitTypeInformation() { 595 if (TypeTable.empty()) 596 return; 597 598 // Start the .debug$T or .debug$P section with 0x4. 599 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 600 emitCodeViewMagicVersion(); 601 602 SmallString<8> CommentPrefix; 603 if (OS.isVerboseAsm()) { 604 CommentPrefix += '\t'; 605 CommentPrefix += Asm->MAI->getCommentString(); 606 CommentPrefix += ' '; 607 } 608 609 TypeTableCollection Table(TypeTable.records()); 610 Optional<TypeIndex> B = Table.getFirst(); 611 while (B) { 612 // This will fail if the record data is invalid. 613 CVType Record = Table.getType(*B); 614 615 if (OS.isVerboseAsm()) { 616 // Emit a block comment describing the type record for readability. 617 SmallString<512> CommentBlock; 618 raw_svector_ostream CommentOS(CommentBlock); 619 ScopedPrinter SP(CommentOS); 620 SP.setPrefix(CommentPrefix); 621 TypeDumpVisitor TDV(Table, &SP, false); 622 623 Error E = codeview::visitTypeRecord(Record, *B, TDV); 624 if (E) { 625 logAllUnhandledErrors(std::move(E), errs(), "error: "); 626 llvm_unreachable("produced malformed type record"); 627 } 628 // emitRawComment will insert its own tab and comment string before 629 // the first line, so strip off our first one. It also prints its own 630 // newline. 631 OS.emitRawComment( 632 CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim()); 633 } 634 OS.EmitBinaryData(Record.str_data()); 635 B = Table.getNext(*B); 636 } 637 } 638 639 void CodeViewDebug::emitTypeGlobalHashes() { 640 if (TypeTable.empty()) 641 return; 642 643 // Start the .debug$H section with the version and hash algorithm, currently 644 // hardcoded to version 0, SHA1. 645 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 646 647 OS.EmitValueToAlignment(4); 648 OS.AddComment("Magic"); 649 OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4); 650 OS.AddComment("Section Version"); 651 OS.EmitIntValue(0, 2); 652 OS.AddComment("Hash Algorithm"); 653 OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2); 654 655 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 656 for (const auto &GHR : TypeTable.hashes()) { 657 if (OS.isVerboseAsm()) { 658 // Emit an EOL-comment describing which TypeIndex this hash corresponds 659 // to, as well as the stringified SHA1 hash. 660 SmallString<32> Comment; 661 raw_svector_ostream CommentOS(Comment); 662 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 663 OS.AddComment(Comment); 664 ++TI; 665 } 666 assert(GHR.Hash.size() == 8); 667 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 668 GHR.Hash.size()); 669 OS.EmitBinaryData(S); 670 } 671 } 672 673 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 674 switch (DWLang) { 675 case dwarf::DW_LANG_C: 676 case dwarf::DW_LANG_C89: 677 case dwarf::DW_LANG_C99: 678 case dwarf::DW_LANG_C11: 679 case dwarf::DW_LANG_ObjC: 680 return SourceLanguage::C; 681 case dwarf::DW_LANG_C_plus_plus: 682 case dwarf::DW_LANG_C_plus_plus_03: 683 case dwarf::DW_LANG_C_plus_plus_11: 684 case dwarf::DW_LANG_C_plus_plus_14: 685 return SourceLanguage::Cpp; 686 case dwarf::DW_LANG_Fortran77: 687 case dwarf::DW_LANG_Fortran90: 688 case dwarf::DW_LANG_Fortran03: 689 case dwarf::DW_LANG_Fortran08: 690 return SourceLanguage::Fortran; 691 case dwarf::DW_LANG_Pascal83: 692 return SourceLanguage::Pascal; 693 case dwarf::DW_LANG_Cobol74: 694 case dwarf::DW_LANG_Cobol85: 695 return SourceLanguage::Cobol; 696 case dwarf::DW_LANG_Java: 697 return SourceLanguage::Java; 698 case dwarf::DW_LANG_D: 699 return SourceLanguage::D; 700 default: 701 // There's no CodeView representation for this language, and CV doesn't 702 // have an "unknown" option for the language field, so we'll use MASM, 703 // as it's very low level. 704 return SourceLanguage::Masm; 705 } 706 } 707 708 namespace { 709 struct Version { 710 int Part[4]; 711 }; 712 } // end anonymous namespace 713 714 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 715 // the version number. 716 static Version parseVersion(StringRef Name) { 717 Version V = {{0}}; 718 int N = 0; 719 for (const char C : Name) { 720 if (isdigit(C)) { 721 V.Part[N] *= 10; 722 V.Part[N] += C - '0'; 723 } else if (C == '.') { 724 ++N; 725 if (N >= 4) 726 return V; 727 } else if (N > 0) 728 return V; 729 } 730 return V; 731 } 732 733 void CodeViewDebug::emitCompilerInformation() { 734 MCContext &Context = MMI->getContext(); 735 MCSymbol *CompilerBegin = Context.createTempSymbol(), 736 *CompilerEnd = Context.createTempSymbol(); 737 OS.AddComment("Record length"); 738 OS.emitAbsoluteSymbolDiff(CompilerEnd, CompilerBegin, 2); 739 OS.EmitLabel(CompilerBegin); 740 OS.AddComment("Record kind: S_COMPILE3"); 741 OS.EmitIntValue(SymbolKind::S_COMPILE3, 2); 742 uint32_t Flags = 0; 743 744 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 745 const MDNode *Node = *CUs->operands().begin(); 746 const auto *CU = cast<DICompileUnit>(Node); 747 748 // The low byte of the flags indicates the source language. 749 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 750 // TODO: Figure out which other flags need to be set. 751 752 OS.AddComment("Flags and language"); 753 OS.EmitIntValue(Flags, 4); 754 755 OS.AddComment("CPUType"); 756 OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2); 757 758 StringRef CompilerVersion = CU->getProducer(); 759 Version FrontVer = parseVersion(CompilerVersion); 760 OS.AddComment("Frontend version"); 761 for (int N = 0; N < 4; ++N) 762 OS.EmitIntValue(FrontVer.Part[N], 2); 763 764 // Some Microsoft tools, like Binscope, expect a backend version number of at 765 // least 8.something, so we'll coerce the LLVM version into a form that 766 // guarantees it'll be big enough without really lying about the version. 767 int Major = 1000 * LLVM_VERSION_MAJOR + 768 10 * LLVM_VERSION_MINOR + 769 LLVM_VERSION_PATCH; 770 // Clamp it for builds that use unusually large version numbers. 771 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 772 Version BackVer = {{ Major, 0, 0, 0 }}; 773 OS.AddComment("Backend version"); 774 for (int N = 0; N < 4; ++N) 775 OS.EmitIntValue(BackVer.Part[N], 2); 776 777 OS.AddComment("Null-terminated compiler version string"); 778 emitNullTerminatedSymbolName(OS, CompilerVersion); 779 780 OS.EmitLabel(CompilerEnd); 781 } 782 783 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 784 StringRef S) { 785 StringIdRecord SIR(TypeIndex(0x0), S); 786 return TypeTable.writeLeafType(SIR); 787 } 788 789 void CodeViewDebug::emitBuildInfo() { 790 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 791 // build info. The known prefix is: 792 // - Absolute path of current directory 793 // - Compiler path 794 // - Main source file path, relative to CWD or absolute 795 // - Type server PDB file 796 // - Canonical compiler command line 797 // If frontend and backend compilation are separated (think llc or LTO), it's 798 // not clear if the compiler path should refer to the executable for the 799 // frontend or the backend. Leave it blank for now. 800 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 801 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 802 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 803 const auto *CU = cast<DICompileUnit>(Node); 804 const DIFile *MainSourceFile = CU->getFile(); 805 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 806 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 807 BuildInfoArgs[BuildInfoRecord::SourceFile] = 808 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 809 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 810 // we implement /Zi type servers. 811 BuildInfoRecord BIR(BuildInfoArgs); 812 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 813 814 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 815 // from the module symbols into the type stream. 816 MCSymbol *BuildInfoEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 817 OS.AddComment("Record length"); 818 OS.EmitIntValue(6, 2); 819 OS.AddComment("Record kind: S_BUILDINFO"); 820 OS.EmitIntValue(unsigned(SymbolKind::S_BUILDINFO), 2); 821 OS.AddComment("LF_BUILDINFO index"); 822 OS.EmitIntValue(BuildInfoIndex.getIndex(), 4); 823 endCVSubsection(BuildInfoEnd); 824 } 825 826 void CodeViewDebug::emitInlineeLinesSubsection() { 827 if (InlinedSubprograms.empty()) 828 return; 829 830 OS.AddComment("Inlinee lines subsection"); 831 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 832 833 // We emit the checksum info for files. This is used by debuggers to 834 // determine if a pdb matches the source before loading it. Visual Studio, 835 // for instance, will display a warning that the breakpoints are not valid if 836 // the pdb does not match the source. 837 OS.AddComment("Inlinee lines signature"); 838 OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4); 839 840 for (const DISubprogram *SP : InlinedSubprograms) { 841 assert(TypeIndices.count({SP, nullptr})); 842 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 843 844 OS.AddBlankLine(); 845 unsigned FileId = maybeRecordFile(SP->getFile()); 846 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 847 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 848 OS.AddBlankLine(); 849 OS.AddComment("Type index of inlined function"); 850 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 851 OS.AddComment("Offset into filechecksum table"); 852 OS.EmitCVFileChecksumOffsetDirective(FileId); 853 OS.AddComment("Starting line number"); 854 OS.EmitIntValue(SP->getLine(), 4); 855 } 856 857 endCVSubsection(InlineEnd); 858 } 859 860 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 861 const DILocation *InlinedAt, 862 const InlineSite &Site) { 863 MCSymbol *InlineBegin = MMI->getContext().createTempSymbol(), 864 *InlineEnd = MMI->getContext().createTempSymbol(); 865 866 assert(TypeIndices.count({Site.Inlinee, nullptr})); 867 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 868 869 // SymbolRecord 870 OS.AddComment("Record length"); 871 OS.emitAbsoluteSymbolDiff(InlineEnd, InlineBegin, 2); // RecordLength 872 OS.EmitLabel(InlineBegin); 873 OS.AddComment("Record kind: S_INLINESITE"); 874 OS.EmitIntValue(SymbolKind::S_INLINESITE, 2); // RecordKind 875 876 OS.AddComment("PtrParent"); 877 OS.EmitIntValue(0, 4); 878 OS.AddComment("PtrEnd"); 879 OS.EmitIntValue(0, 4); 880 OS.AddComment("Inlinee type index"); 881 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 882 883 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 884 unsigned StartLineNum = Site.Inlinee->getLine(); 885 886 OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 887 FI.Begin, FI.End); 888 889 OS.EmitLabel(InlineEnd); 890 891 emitLocalVariableList(FI, Site.InlinedLocals); 892 893 // Recurse on child inlined call sites before closing the scope. 894 for (const DILocation *ChildSite : Site.ChildSites) { 895 auto I = FI.InlineSites.find(ChildSite); 896 assert(I != FI.InlineSites.end() && 897 "child site not in function inline site map"); 898 emitInlinedCallSite(FI, ChildSite, I->second); 899 } 900 901 // Close the scope. 902 OS.AddComment("Record length"); 903 OS.EmitIntValue(2, 2); // RecordLength 904 OS.AddComment("Record kind: S_INLINESITE_END"); 905 OS.EmitIntValue(SymbolKind::S_INLINESITE_END, 2); // RecordKind 906 } 907 908 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 909 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 910 // comdat key. A section may be comdat because of -ffunction-sections or 911 // because it is comdat in the IR. 912 MCSectionCOFF *GVSec = 913 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 914 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 915 916 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 917 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 918 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 919 920 OS.SwitchSection(DebugSec); 921 922 // Emit the magic version number if this is the first time we've switched to 923 // this section. 924 if (ComdatDebugSections.insert(DebugSec).second) 925 emitCodeViewMagicVersion(); 926 } 927 928 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 929 // The only supported thunk ordinal is currently the standard type. 930 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 931 FunctionInfo &FI, 932 const MCSymbol *Fn) { 933 std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 934 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 935 936 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 937 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 938 939 // Emit S_THUNK32 940 MCSymbol *ThunkRecordBegin = MMI->getContext().createTempSymbol(), 941 *ThunkRecordEnd = MMI->getContext().createTempSymbol(); 942 OS.AddComment("Record length"); 943 OS.emitAbsoluteSymbolDiff(ThunkRecordEnd, ThunkRecordBegin, 2); 944 OS.EmitLabel(ThunkRecordBegin); 945 OS.AddComment("Record kind: S_THUNK32"); 946 OS.EmitIntValue(unsigned(SymbolKind::S_THUNK32), 2); 947 OS.AddComment("PtrParent"); 948 OS.EmitIntValue(0, 4); 949 OS.AddComment("PtrEnd"); 950 OS.EmitIntValue(0, 4); 951 OS.AddComment("PtrNext"); 952 OS.EmitIntValue(0, 4); 953 OS.AddComment("Thunk section relative address"); 954 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 955 OS.AddComment("Thunk section index"); 956 OS.EmitCOFFSectionIndex(Fn); 957 OS.AddComment("Code size"); 958 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 959 OS.AddComment("Ordinal"); 960 OS.EmitIntValue(unsigned(ordinal), 1); 961 OS.AddComment("Function name"); 962 emitNullTerminatedSymbolName(OS, FuncName); 963 // Additional fields specific to the thunk ordinal would go here. 964 OS.EmitLabel(ThunkRecordEnd); 965 966 // Local variables/inlined routines are purposely omitted here. The point of 967 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 968 969 // Emit S_PROC_ID_END 970 const unsigned RecordLengthForSymbolEnd = 2; 971 OS.AddComment("Record length"); 972 OS.EmitIntValue(RecordLengthForSymbolEnd, 2); 973 OS.AddComment("Record kind: S_PROC_ID_END"); 974 OS.EmitIntValue(unsigned(SymbolKind::S_PROC_ID_END), 2); 975 976 endCVSubsection(SymbolsEnd); 977 } 978 979 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 980 FunctionInfo &FI) { 981 // For each function there is a separate subsection which holds the PC to 982 // file:line table. 983 const MCSymbol *Fn = Asm->getSymbol(GV); 984 assert(Fn); 985 986 // Switch to the to a comdat section, if appropriate. 987 switchToDebugSectionForSymbol(Fn); 988 989 std::string FuncName; 990 auto *SP = GV->getSubprogram(); 991 assert(SP); 992 setCurrentSubprogram(SP); 993 994 if (SP->isThunk()) { 995 emitDebugInfoForThunk(GV, FI, Fn); 996 return; 997 } 998 999 // If we have a display name, build the fully qualified name by walking the 1000 // chain of scopes. 1001 if (!SP->getName().empty()) 1002 FuncName = 1003 getFullyQualifiedName(SP->getScope().resolve(), SP->getName()); 1004 1005 // If our DISubprogram name is empty, use the mangled name. 1006 if (FuncName.empty()) 1007 FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 1008 1009 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1010 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1011 OS.EmitCVFPOData(Fn); 1012 1013 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1014 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1015 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1016 { 1017 MCSymbol *ProcRecordBegin = MMI->getContext().createTempSymbol(), 1018 *ProcRecordEnd = MMI->getContext().createTempSymbol(); 1019 OS.AddComment("Record length"); 1020 OS.emitAbsoluteSymbolDiff(ProcRecordEnd, ProcRecordBegin, 2); 1021 OS.EmitLabel(ProcRecordBegin); 1022 1023 if (GV->hasLocalLinkage()) { 1024 OS.AddComment("Record kind: S_LPROC32_ID"); 1025 OS.EmitIntValue(unsigned(SymbolKind::S_LPROC32_ID), 2); 1026 } else { 1027 OS.AddComment("Record kind: S_GPROC32_ID"); 1028 OS.EmitIntValue(unsigned(SymbolKind::S_GPROC32_ID), 2); 1029 } 1030 1031 // These fields are filled in by tools like CVPACK which run after the fact. 1032 OS.AddComment("PtrParent"); 1033 OS.EmitIntValue(0, 4); 1034 OS.AddComment("PtrEnd"); 1035 OS.EmitIntValue(0, 4); 1036 OS.AddComment("PtrNext"); 1037 OS.EmitIntValue(0, 4); 1038 // This is the important bit that tells the debugger where the function 1039 // code is located and what's its size: 1040 OS.AddComment("Code size"); 1041 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1042 OS.AddComment("Offset after prologue"); 1043 OS.EmitIntValue(0, 4); 1044 OS.AddComment("Offset before epilogue"); 1045 OS.EmitIntValue(0, 4); 1046 OS.AddComment("Function type index"); 1047 OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4); 1048 OS.AddComment("Function section relative address"); 1049 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1050 OS.AddComment("Function section index"); 1051 OS.EmitCOFFSectionIndex(Fn); 1052 OS.AddComment("Flags"); 1053 OS.EmitIntValue(0, 1); 1054 // Emit the function display name as a null-terminated string. 1055 OS.AddComment("Function name"); 1056 // Truncate the name so we won't overflow the record length field. 1057 emitNullTerminatedSymbolName(OS, FuncName); 1058 OS.EmitLabel(ProcRecordEnd); 1059 1060 MCSymbol *FrameProcBegin = MMI->getContext().createTempSymbol(), 1061 *FrameProcEnd = MMI->getContext().createTempSymbol(); 1062 OS.AddComment("Record length"); 1063 OS.emitAbsoluteSymbolDiff(FrameProcEnd, FrameProcBegin, 2); 1064 OS.EmitLabel(FrameProcBegin); 1065 OS.AddComment("Record kind: S_FRAMEPROC"); 1066 OS.EmitIntValue(unsigned(SymbolKind::S_FRAMEPROC), 2); 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 OS.EmitLabel(FrameProcEnd); 1083 1084 emitLocalVariableList(FI, FI.Locals); 1085 emitLexicalBlockList(FI.ChildBlocks, FI); 1086 1087 // Emit inlined call site information. Only emit functions inlined directly 1088 // into the parent function. We'll emit the other sites recursively as part 1089 // of their parent inline site. 1090 for (const DILocation *InlinedAt : FI.ChildSites) { 1091 auto I = FI.InlineSites.find(InlinedAt); 1092 assert(I != FI.InlineSites.end() && 1093 "child site not in function inline site map"); 1094 emitInlinedCallSite(FI, InlinedAt, I->second); 1095 } 1096 1097 for (auto Annot : FI.Annotations) { 1098 MCSymbol *Label = Annot.first; 1099 MDTuple *Strs = cast<MDTuple>(Annot.second); 1100 MCSymbol *AnnotBegin = MMI->getContext().createTempSymbol(), 1101 *AnnotEnd = MMI->getContext().createTempSymbol(); 1102 OS.AddComment("Record length"); 1103 OS.emitAbsoluteSymbolDiff(AnnotEnd, AnnotBegin, 2); 1104 OS.EmitLabel(AnnotBegin); 1105 OS.AddComment("Record kind: S_ANNOTATION"); 1106 OS.EmitIntValue(SymbolKind::S_ANNOTATION, 2); 1107 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1108 // FIXME: Make sure we don't overflow the max record size. 1109 OS.EmitCOFFSectionIndex(Label); 1110 OS.EmitIntValue(Strs->getNumOperands(), 2); 1111 for (Metadata *MD : Strs->operands()) { 1112 // MDStrings are null terminated, so we can do EmitBytes and get the 1113 // nice .asciz directive. 1114 StringRef Str = cast<MDString>(MD)->getString(); 1115 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1116 OS.EmitBytes(StringRef(Str.data(), Str.size() + 1)); 1117 } 1118 OS.EmitLabel(AnnotEnd); 1119 } 1120 1121 if (SP != nullptr) 1122 emitDebugInfoForUDTs(LocalUDTs); 1123 1124 // We're done with this function. 1125 OS.AddComment("Record length"); 1126 OS.EmitIntValue(0x0002, 2); 1127 OS.AddComment("Record kind: S_PROC_ID_END"); 1128 OS.EmitIntValue(unsigned(SymbolKind::S_PROC_ID_END), 2); 1129 } 1130 endCVSubsection(SymbolsEnd); 1131 1132 // We have an assembler directive that takes care of the whole line table. 1133 OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1134 } 1135 1136 CodeViewDebug::LocalVarDefRange 1137 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1138 LocalVarDefRange DR; 1139 DR.InMemory = -1; 1140 DR.DataOffset = Offset; 1141 assert(DR.DataOffset == Offset && "truncation"); 1142 DR.IsSubfield = 0; 1143 DR.StructOffset = 0; 1144 DR.CVRegister = CVRegister; 1145 return DR; 1146 } 1147 1148 void CodeViewDebug::collectVariableInfoFromMFTable( 1149 DenseSet<InlinedEntity> &Processed) { 1150 const MachineFunction &MF = *Asm->MF; 1151 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1152 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1153 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1154 1155 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1156 if (!VI.Var) 1157 continue; 1158 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1159 "Expected inlined-at fields to agree"); 1160 1161 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1162 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1163 1164 // If variable scope is not found then skip this variable. 1165 if (!Scope) 1166 continue; 1167 1168 // If the variable has an attached offset expression, extract it. 1169 // FIXME: Try to handle DW_OP_deref as well. 1170 int64_t ExprOffset = 0; 1171 if (VI.Expr) 1172 if (!VI.Expr->extractIfOffset(ExprOffset)) 1173 continue; 1174 1175 // Get the frame register used and the offset. 1176 unsigned FrameReg = 0; 1177 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1178 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1179 1180 // Calculate the label ranges. 1181 LocalVarDefRange DefRange = 1182 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1183 for (const InsnRange &Range : Scope->getRanges()) { 1184 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1185 const MCSymbol *End = getLabelAfterInsn(Range.second); 1186 End = End ? End : Asm->getFunctionEnd(); 1187 DefRange.Ranges.emplace_back(Begin, End); 1188 } 1189 1190 LocalVariable Var; 1191 Var.DIVar = VI.Var; 1192 Var.DefRanges.emplace_back(std::move(DefRange)); 1193 recordLocalVariable(std::move(Var), Scope); 1194 } 1195 } 1196 1197 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1198 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1199 } 1200 1201 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1202 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1203 } 1204 1205 void CodeViewDebug::calculateRanges( 1206 LocalVariable &Var, const DbgValueHistoryMap::InstrRanges &Ranges) { 1207 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1208 1209 // Calculate the definition ranges. 1210 for (auto I = Ranges.begin(), E = Ranges.end(); I != E; ++I) { 1211 const InsnRange &Range = *I; 1212 const MachineInstr *DVInst = Range.first; 1213 assert(DVInst->isDebugValue() && "Invalid History entry"); 1214 // FIXME: Find a way to represent constant variables, since they are 1215 // relatively common. 1216 Optional<DbgVariableLocation> Location = 1217 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1218 if (!Location) 1219 continue; 1220 1221 // CodeView can only express variables in register and variables in memory 1222 // at a constant offset from a register. However, for variables passed 1223 // indirectly by pointer, it is common for that pointer to be spilled to a 1224 // stack location. For the special case of one offseted load followed by a 1225 // zero offset load (a pointer spilled to the stack), we change the type of 1226 // the local variable from a value type to a reference type. This tricks the 1227 // debugger into doing the load for us. 1228 if (Var.UseReferenceType) { 1229 // We're using a reference type. Drop the last zero offset load. 1230 if (canUseReferenceType(*Location)) 1231 Location->LoadChain.pop_back(); 1232 else 1233 continue; 1234 } else if (needsReferenceType(*Location)) { 1235 // This location can't be expressed without switching to a reference type. 1236 // Start over using that. 1237 Var.UseReferenceType = true; 1238 Var.DefRanges.clear(); 1239 calculateRanges(Var, Ranges); 1240 return; 1241 } 1242 1243 // We can only handle a register or an offseted load of a register. 1244 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1245 continue; 1246 { 1247 LocalVarDefRange DR; 1248 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1249 DR.InMemory = !Location->LoadChain.empty(); 1250 DR.DataOffset = 1251 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1252 if (Location->FragmentInfo) { 1253 DR.IsSubfield = true; 1254 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1255 } else { 1256 DR.IsSubfield = false; 1257 DR.StructOffset = 0; 1258 } 1259 1260 if (Var.DefRanges.empty() || 1261 Var.DefRanges.back().isDifferentLocation(DR)) { 1262 Var.DefRanges.emplace_back(std::move(DR)); 1263 } 1264 } 1265 1266 // Compute the label range. 1267 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1268 const MCSymbol *End = getLabelAfterInsn(Range.second); 1269 if (!End) { 1270 // This range is valid until the next overlapping bitpiece. In the 1271 // common case, ranges will not be bitpieces, so they will overlap. 1272 auto J = std::next(I); 1273 const DIExpression *DIExpr = DVInst->getDebugExpression(); 1274 while (J != E && 1275 !DIExpr->fragmentsOverlap(J->first->getDebugExpression())) 1276 ++J; 1277 if (J != E) 1278 End = getLabelBeforeInsn(J->first); 1279 else 1280 End = Asm->getFunctionEnd(); 1281 } 1282 1283 // If the last range end is our begin, just extend the last range. 1284 // Otherwise make a new range. 1285 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1286 Var.DefRanges.back().Ranges; 1287 if (!R.empty() && R.back().second == Begin) 1288 R.back().second = End; 1289 else 1290 R.emplace_back(Begin, End); 1291 1292 // FIXME: Do more range combining. 1293 } 1294 } 1295 1296 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1297 DenseSet<InlinedEntity> Processed; 1298 // Grab the variable info that was squirreled away in the MMI side-table. 1299 collectVariableInfoFromMFTable(Processed); 1300 1301 for (const auto &I : DbgValues) { 1302 InlinedEntity IV = I.first; 1303 if (Processed.count(IV)) 1304 continue; 1305 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1306 const DILocation *InlinedAt = IV.second; 1307 1308 // Instruction ranges, specifying where IV is accessible. 1309 const auto &Ranges = I.second; 1310 1311 LexicalScope *Scope = nullptr; 1312 if (InlinedAt) 1313 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1314 else 1315 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1316 // If variable scope is not found then skip this variable. 1317 if (!Scope) 1318 continue; 1319 1320 LocalVariable Var; 1321 Var.DIVar = DIVar; 1322 1323 calculateRanges(Var, Ranges); 1324 recordLocalVariable(std::move(Var), Scope); 1325 } 1326 } 1327 1328 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1329 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1330 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1331 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1332 const Function &GV = MF->getFunction(); 1333 auto Insertion = FnDebugInfo.insert({&GV, llvm::make_unique<FunctionInfo>()}); 1334 assert(Insertion.second && "function already has info"); 1335 CurFn = Insertion.first->second.get(); 1336 CurFn->FuncId = NextFuncId++; 1337 CurFn->Begin = Asm->getFunctionBegin(); 1338 1339 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1340 // callee-saved registers were used. For targets that don't use a PUSH 1341 // instruction (AArch64), this will be zero. 1342 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1343 CurFn->FrameSize = MFI.getStackSize(); 1344 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1345 CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF); 1346 1347 // For this function S_FRAMEPROC record, figure out which codeview register 1348 // will be the frame pointer. 1349 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1350 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1351 if (CurFn->FrameSize > 0) { 1352 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1353 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1354 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1355 } else { 1356 // If there is an FP, parameters are always relative to it. 1357 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1358 if (CurFn->HasStackRealignment) { 1359 // If the stack needs realignment, locals are relative to SP or VFRAME. 1360 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1361 } else { 1362 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1363 // other stack adjustments. 1364 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1365 } 1366 } 1367 } 1368 1369 // Compute other frame procedure options. 1370 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1371 if (MFI.hasVarSizedObjects()) 1372 FPO |= FrameProcedureOptions::HasAlloca; 1373 if (MF->exposesReturnsTwice()) 1374 FPO |= FrameProcedureOptions::HasSetJmp; 1375 // FIXME: Set HasLongJmp if we ever track that info. 1376 if (MF->hasInlineAsm()) 1377 FPO |= FrameProcedureOptions::HasInlineAssembly; 1378 if (GV.hasPersonalityFn()) { 1379 if (isAsynchronousEHPersonality( 1380 classifyEHPersonality(GV.getPersonalityFn()))) 1381 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1382 else 1383 FPO |= FrameProcedureOptions::HasExceptionHandling; 1384 } 1385 if (GV.hasFnAttribute(Attribute::InlineHint)) 1386 FPO |= FrameProcedureOptions::MarkedInline; 1387 if (GV.hasFnAttribute(Attribute::Naked)) 1388 FPO |= FrameProcedureOptions::Naked; 1389 if (MFI.hasStackProtectorIndex()) 1390 FPO |= FrameProcedureOptions::SecurityChecks; 1391 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1392 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1393 if (Asm->TM.getOptLevel() != CodeGenOpt::None && !GV.optForSize() && 1394 !GV.hasFnAttribute(Attribute::OptimizeNone)) 1395 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1396 // FIXME: Set GuardCfg when it is implemented. 1397 CurFn->FrameProcOpts = FPO; 1398 1399 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1400 1401 // Find the end of the function prolog. First known non-DBG_VALUE and 1402 // non-frame setup location marks the beginning of the function body. 1403 // FIXME: is there a simpler a way to do this? Can we just search 1404 // for the first instruction of the function, not the last of the prolog? 1405 DebugLoc PrologEndLoc; 1406 bool EmptyPrologue = true; 1407 for (const auto &MBB : *MF) { 1408 for (const auto &MI : MBB) { 1409 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1410 MI.getDebugLoc()) { 1411 PrologEndLoc = MI.getDebugLoc(); 1412 break; 1413 } else if (!MI.isMetaInstruction()) { 1414 EmptyPrologue = false; 1415 } 1416 } 1417 } 1418 1419 // Record beginning of function if we have a non-empty prologue. 1420 if (PrologEndLoc && !EmptyPrologue) { 1421 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1422 maybeRecordLocation(FnStartDL, MF); 1423 } 1424 } 1425 1426 static bool shouldEmitUdt(const DIType *T) { 1427 if (!T) 1428 return false; 1429 1430 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1431 if (T->getTag() == dwarf::DW_TAG_typedef) { 1432 if (DIScope *Scope = T->getScope().resolve()) { 1433 switch (Scope->getTag()) { 1434 case dwarf::DW_TAG_structure_type: 1435 case dwarf::DW_TAG_class_type: 1436 case dwarf::DW_TAG_union_type: 1437 return false; 1438 } 1439 } 1440 } 1441 1442 while (true) { 1443 if (!T || T->isForwardDecl()) 1444 return false; 1445 1446 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1447 if (!DT) 1448 return true; 1449 T = DT->getBaseType().resolve(); 1450 } 1451 return true; 1452 } 1453 1454 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1455 // Don't record empty UDTs. 1456 if (Ty->getName().empty()) 1457 return; 1458 if (!shouldEmitUdt(Ty)) 1459 return; 1460 1461 SmallVector<StringRef, 5> QualifiedNameComponents; 1462 const DISubprogram *ClosestSubprogram = getQualifiedNameComponents( 1463 Ty->getScope().resolve(), QualifiedNameComponents); 1464 1465 std::string FullyQualifiedName = 1466 getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty)); 1467 1468 if (ClosestSubprogram == nullptr) { 1469 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1470 } else if (ClosestSubprogram == CurrentSubprogram) { 1471 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1472 } 1473 1474 // TODO: What if the ClosestSubprogram is neither null or the current 1475 // subprogram? Currently, the UDT just gets dropped on the floor. 1476 // 1477 // The current behavior is not desirable. To get maximal fidelity, we would 1478 // need to perform all type translation before beginning emission of .debug$S 1479 // and then make LocalUDTs a member of FunctionInfo 1480 } 1481 1482 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1483 // Generic dispatch for lowering an unknown type. 1484 switch (Ty->getTag()) { 1485 case dwarf::DW_TAG_array_type: 1486 return lowerTypeArray(cast<DICompositeType>(Ty)); 1487 case dwarf::DW_TAG_typedef: 1488 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1489 case dwarf::DW_TAG_base_type: 1490 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1491 case dwarf::DW_TAG_pointer_type: 1492 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1493 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1494 LLVM_FALLTHROUGH; 1495 case dwarf::DW_TAG_reference_type: 1496 case dwarf::DW_TAG_rvalue_reference_type: 1497 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1498 case dwarf::DW_TAG_ptr_to_member_type: 1499 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1500 case dwarf::DW_TAG_restrict_type: 1501 case dwarf::DW_TAG_const_type: 1502 case dwarf::DW_TAG_volatile_type: 1503 // TODO: add support for DW_TAG_atomic_type here 1504 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1505 case dwarf::DW_TAG_subroutine_type: 1506 if (ClassTy) { 1507 // The member function type of a member function pointer has no 1508 // ThisAdjustment. 1509 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1510 /*ThisAdjustment=*/0, 1511 /*IsStaticMethod=*/false); 1512 } 1513 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1514 case dwarf::DW_TAG_enumeration_type: 1515 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1516 case dwarf::DW_TAG_class_type: 1517 case dwarf::DW_TAG_structure_type: 1518 return lowerTypeClass(cast<DICompositeType>(Ty)); 1519 case dwarf::DW_TAG_union_type: 1520 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1521 case dwarf::DW_TAG_unspecified_type: 1522 if (Ty->getName() == "decltype(nullptr)") 1523 return TypeIndex::NullptrT(); 1524 return TypeIndex::None(); 1525 default: 1526 // Use the null type index. 1527 return TypeIndex(); 1528 } 1529 } 1530 1531 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1532 DITypeRef UnderlyingTypeRef = Ty->getBaseType(); 1533 TypeIndex UnderlyingTypeIndex = getTypeIndex(UnderlyingTypeRef); 1534 StringRef TypeName = Ty->getName(); 1535 1536 addToUDTs(Ty); 1537 1538 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1539 TypeName == "HRESULT") 1540 return TypeIndex(SimpleTypeKind::HResult); 1541 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1542 TypeName == "wchar_t") 1543 return TypeIndex(SimpleTypeKind::WideCharacter); 1544 1545 return UnderlyingTypeIndex; 1546 } 1547 1548 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1549 DITypeRef ElementTypeRef = Ty->getBaseType(); 1550 TypeIndex ElementTypeIndex = getTypeIndex(ElementTypeRef); 1551 // IndexType is size_t, which depends on the bitness of the target. 1552 TypeIndex IndexType = getPointerSizeInBytes() == 8 1553 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1554 : TypeIndex(SimpleTypeKind::UInt32Long); 1555 1556 uint64_t ElementSize = getBaseTypeSize(ElementTypeRef) / 8; 1557 1558 // Add subranges to array type. 1559 DINodeArray Elements = Ty->getElements(); 1560 for (int i = Elements.size() - 1; i >= 0; --i) { 1561 const DINode *Element = Elements[i]; 1562 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1563 1564 const DISubrange *Subrange = cast<DISubrange>(Element); 1565 assert(Subrange->getLowerBound() == 0 && 1566 "codeview doesn't support subranges with lower bounds"); 1567 int64_t Count = -1; 1568 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>()) 1569 Count = CI->getSExtValue(); 1570 1571 // Forward declarations of arrays without a size and VLAs use a count of -1. 1572 // Emit a count of zero in these cases to match what MSVC does for arrays 1573 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1574 // should do for them even if we could distinguish them. 1575 if (Count == -1) 1576 Count = 0; 1577 1578 // Update the element size and element type index for subsequent subranges. 1579 ElementSize *= Count; 1580 1581 // If this is the outermost array, use the size from the array. It will be 1582 // more accurate if we had a VLA or an incomplete element type size. 1583 uint64_t ArraySize = 1584 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1585 1586 StringRef Name = (i == 0) ? Ty->getName() : ""; 1587 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1588 ElementTypeIndex = TypeTable.writeLeafType(AR); 1589 } 1590 1591 return ElementTypeIndex; 1592 } 1593 1594 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1595 TypeIndex Index; 1596 dwarf::TypeKind Kind; 1597 uint32_t ByteSize; 1598 1599 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1600 ByteSize = Ty->getSizeInBits() / 8; 1601 1602 SimpleTypeKind STK = SimpleTypeKind::None; 1603 switch (Kind) { 1604 case dwarf::DW_ATE_address: 1605 // FIXME: Translate 1606 break; 1607 case dwarf::DW_ATE_boolean: 1608 switch (ByteSize) { 1609 case 1: STK = SimpleTypeKind::Boolean8; break; 1610 case 2: STK = SimpleTypeKind::Boolean16; break; 1611 case 4: STK = SimpleTypeKind::Boolean32; break; 1612 case 8: STK = SimpleTypeKind::Boolean64; break; 1613 case 16: STK = SimpleTypeKind::Boolean128; break; 1614 } 1615 break; 1616 case dwarf::DW_ATE_complex_float: 1617 switch (ByteSize) { 1618 case 2: STK = SimpleTypeKind::Complex16; break; 1619 case 4: STK = SimpleTypeKind::Complex32; break; 1620 case 8: STK = SimpleTypeKind::Complex64; break; 1621 case 10: STK = SimpleTypeKind::Complex80; break; 1622 case 16: STK = SimpleTypeKind::Complex128; break; 1623 } 1624 break; 1625 case dwarf::DW_ATE_float: 1626 switch (ByteSize) { 1627 case 2: STK = SimpleTypeKind::Float16; break; 1628 case 4: STK = SimpleTypeKind::Float32; break; 1629 case 6: STK = SimpleTypeKind::Float48; break; 1630 case 8: STK = SimpleTypeKind::Float64; break; 1631 case 10: STK = SimpleTypeKind::Float80; break; 1632 case 16: STK = SimpleTypeKind::Float128; break; 1633 } 1634 break; 1635 case dwarf::DW_ATE_signed: 1636 switch (ByteSize) { 1637 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1638 case 2: STK = SimpleTypeKind::Int16Short; break; 1639 case 4: STK = SimpleTypeKind::Int32; break; 1640 case 8: STK = SimpleTypeKind::Int64Quad; break; 1641 case 16: STK = SimpleTypeKind::Int128Oct; break; 1642 } 1643 break; 1644 case dwarf::DW_ATE_unsigned: 1645 switch (ByteSize) { 1646 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1647 case 2: STK = SimpleTypeKind::UInt16Short; break; 1648 case 4: STK = SimpleTypeKind::UInt32; break; 1649 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1650 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1651 } 1652 break; 1653 case dwarf::DW_ATE_UTF: 1654 switch (ByteSize) { 1655 case 2: STK = SimpleTypeKind::Character16; break; 1656 case 4: STK = SimpleTypeKind::Character32; break; 1657 } 1658 break; 1659 case dwarf::DW_ATE_signed_char: 1660 if (ByteSize == 1) 1661 STK = SimpleTypeKind::SignedCharacter; 1662 break; 1663 case dwarf::DW_ATE_unsigned_char: 1664 if (ByteSize == 1) 1665 STK = SimpleTypeKind::UnsignedCharacter; 1666 break; 1667 default: 1668 break; 1669 } 1670 1671 // Apply some fixups based on the source-level type name. 1672 if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int") 1673 STK = SimpleTypeKind::Int32Long; 1674 if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int") 1675 STK = SimpleTypeKind::UInt32Long; 1676 if (STK == SimpleTypeKind::UInt16Short && 1677 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1678 STK = SimpleTypeKind::WideCharacter; 1679 if ((STK == SimpleTypeKind::SignedCharacter || 1680 STK == SimpleTypeKind::UnsignedCharacter) && 1681 Ty->getName() == "char") 1682 STK = SimpleTypeKind::NarrowCharacter; 1683 1684 return TypeIndex(STK); 1685 } 1686 1687 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1688 PointerOptions PO) { 1689 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1690 1691 // Pointers to simple types without any options can use SimpleTypeMode, rather 1692 // than having a dedicated pointer type record. 1693 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1694 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1695 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1696 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1697 ? SimpleTypeMode::NearPointer64 1698 : SimpleTypeMode::NearPointer32; 1699 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1700 } 1701 1702 PointerKind PK = 1703 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1704 PointerMode PM = PointerMode::Pointer; 1705 switch (Ty->getTag()) { 1706 default: llvm_unreachable("not a pointer tag type"); 1707 case dwarf::DW_TAG_pointer_type: 1708 PM = PointerMode::Pointer; 1709 break; 1710 case dwarf::DW_TAG_reference_type: 1711 PM = PointerMode::LValueReference; 1712 break; 1713 case dwarf::DW_TAG_rvalue_reference_type: 1714 PM = PointerMode::RValueReference; 1715 break; 1716 } 1717 1718 if (Ty->isObjectPointer()) 1719 PO |= PointerOptions::Const; 1720 1721 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1722 return TypeTable.writeLeafType(PR); 1723 } 1724 1725 static PointerToMemberRepresentation 1726 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1727 // SizeInBytes being zero generally implies that the member pointer type was 1728 // incomplete, which can happen if it is part of a function prototype. In this 1729 // case, use the unknown model instead of the general model. 1730 if (IsPMF) { 1731 switch (Flags & DINode::FlagPtrToMemberRep) { 1732 case 0: 1733 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1734 : PointerToMemberRepresentation::GeneralFunction; 1735 case DINode::FlagSingleInheritance: 1736 return PointerToMemberRepresentation::SingleInheritanceFunction; 1737 case DINode::FlagMultipleInheritance: 1738 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1739 case DINode::FlagVirtualInheritance: 1740 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1741 } 1742 } else { 1743 switch (Flags & DINode::FlagPtrToMemberRep) { 1744 case 0: 1745 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1746 : PointerToMemberRepresentation::GeneralData; 1747 case DINode::FlagSingleInheritance: 1748 return PointerToMemberRepresentation::SingleInheritanceData; 1749 case DINode::FlagMultipleInheritance: 1750 return PointerToMemberRepresentation::MultipleInheritanceData; 1751 case DINode::FlagVirtualInheritance: 1752 return PointerToMemberRepresentation::VirtualInheritanceData; 1753 } 1754 } 1755 llvm_unreachable("invalid ptr to member representation"); 1756 } 1757 1758 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1759 PointerOptions PO) { 1760 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1761 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1762 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType()); 1763 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1764 : PointerKind::Near32; 1765 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1766 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1767 : PointerMode::PointerToDataMember; 1768 1769 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1770 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1771 MemberPointerInfo MPI( 1772 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1773 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1774 return TypeTable.writeLeafType(PR); 1775 } 1776 1777 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1778 /// have a translation, use the NearC convention. 1779 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1780 switch (DwarfCC) { 1781 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1782 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1783 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1784 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1785 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1786 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1787 } 1788 return CallingConvention::NearC; 1789 } 1790 1791 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1792 ModifierOptions Mods = ModifierOptions::None; 1793 PointerOptions PO = PointerOptions::None; 1794 bool IsModifier = true; 1795 const DIType *BaseTy = Ty; 1796 while (IsModifier && BaseTy) { 1797 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1798 switch (BaseTy->getTag()) { 1799 case dwarf::DW_TAG_const_type: 1800 Mods |= ModifierOptions::Const; 1801 PO |= PointerOptions::Const; 1802 break; 1803 case dwarf::DW_TAG_volatile_type: 1804 Mods |= ModifierOptions::Volatile; 1805 PO |= PointerOptions::Volatile; 1806 break; 1807 case dwarf::DW_TAG_restrict_type: 1808 // Only pointer types be marked with __restrict. There is no known flag 1809 // for __restrict in LF_MODIFIER records. 1810 PO |= PointerOptions::Restrict; 1811 break; 1812 default: 1813 IsModifier = false; 1814 break; 1815 } 1816 if (IsModifier) 1817 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType().resolve(); 1818 } 1819 1820 // Check if the inner type will use an LF_POINTER record. If so, the 1821 // qualifiers will go in the LF_POINTER record. This comes up for types like 1822 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1823 // char *'. 1824 if (BaseTy) { 1825 switch (BaseTy->getTag()) { 1826 case dwarf::DW_TAG_pointer_type: 1827 case dwarf::DW_TAG_reference_type: 1828 case dwarf::DW_TAG_rvalue_reference_type: 1829 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1830 case dwarf::DW_TAG_ptr_to_member_type: 1831 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1832 default: 1833 break; 1834 } 1835 } 1836 1837 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1838 1839 // Return the base type index if there aren't any modifiers. For example, the 1840 // metadata could contain restrict wrappers around non-pointer types. 1841 if (Mods == ModifierOptions::None) 1842 return ModifiedTI; 1843 1844 ModifierRecord MR(ModifiedTI, Mods); 1845 return TypeTable.writeLeafType(MR); 1846 } 1847 1848 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1849 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1850 for (DITypeRef ArgTypeRef : Ty->getTypeArray()) 1851 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgTypeRef)); 1852 1853 // MSVC uses type none for variadic argument. 1854 if (ReturnAndArgTypeIndices.size() > 1 && 1855 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1856 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1857 } 1858 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1859 ArrayRef<TypeIndex> ArgTypeIndices = None; 1860 if (!ReturnAndArgTypeIndices.empty()) { 1861 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1862 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1863 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1864 } 1865 1866 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1867 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1868 1869 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1870 1871 FunctionOptions FO = getFunctionOptions(Ty); 1872 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 1873 ArgListIndex); 1874 return TypeTable.writeLeafType(Procedure); 1875 } 1876 1877 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 1878 const DIType *ClassTy, 1879 int ThisAdjustment, 1880 bool IsStaticMethod, 1881 FunctionOptions FO) { 1882 // Lower the containing class type. 1883 TypeIndex ClassType = getTypeIndex(ClassTy); 1884 1885 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1886 for (DITypeRef ArgTypeRef : Ty->getTypeArray()) 1887 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgTypeRef)); 1888 1889 // MSVC uses type none for variadic argument. 1890 if (ReturnAndArgTypeIndices.size() > 1 && 1891 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1892 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1893 } 1894 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1895 ArrayRef<TypeIndex> ArgTypeIndices = None; 1896 if (!ReturnAndArgTypeIndices.empty()) { 1897 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1898 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1899 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1900 } 1901 TypeIndex ThisTypeIndex; 1902 if (!IsStaticMethod && !ArgTypeIndices.empty()) { 1903 ThisTypeIndex = ArgTypeIndices.front(); 1904 ArgTypeIndices = ArgTypeIndices.drop_front(); 1905 } 1906 1907 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1908 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1909 1910 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1911 1912 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1913 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1914 return TypeTable.writeLeafType(MFR); 1915 } 1916 1917 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1918 unsigned VSlotCount = 1919 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1920 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1921 1922 VFTableShapeRecord VFTSR(Slots); 1923 return TypeTable.writeLeafType(VFTSR); 1924 } 1925 1926 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1927 switch (Flags & DINode::FlagAccessibility) { 1928 case DINode::FlagPrivate: return MemberAccess::Private; 1929 case DINode::FlagPublic: return MemberAccess::Public; 1930 case DINode::FlagProtected: return MemberAccess::Protected; 1931 case 0: 1932 // If there was no explicit access control, provide the default for the tag. 1933 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1934 : MemberAccess::Public; 1935 } 1936 llvm_unreachable("access flags are exclusive"); 1937 } 1938 1939 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1940 if (SP->isArtificial()) 1941 return MethodOptions::CompilerGenerated; 1942 1943 // FIXME: Handle other MethodOptions. 1944 1945 return MethodOptions::None; 1946 } 1947 1948 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 1949 bool Introduced) { 1950 if (SP->getFlags() & DINode::FlagStaticMember) 1951 return MethodKind::Static; 1952 1953 switch (SP->getVirtuality()) { 1954 case dwarf::DW_VIRTUALITY_none: 1955 break; 1956 case dwarf::DW_VIRTUALITY_virtual: 1957 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 1958 case dwarf::DW_VIRTUALITY_pure_virtual: 1959 return Introduced ? MethodKind::PureIntroducingVirtual 1960 : MethodKind::PureVirtual; 1961 default: 1962 llvm_unreachable("unhandled virtuality case"); 1963 } 1964 1965 return MethodKind::Vanilla; 1966 } 1967 1968 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 1969 switch (Ty->getTag()) { 1970 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 1971 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 1972 } 1973 llvm_unreachable("unexpected tag"); 1974 } 1975 1976 /// Return ClassOptions that should be present on both the forward declaration 1977 /// and the defintion of a tag type. 1978 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 1979 ClassOptions CO = ClassOptions::None; 1980 1981 // MSVC always sets this flag, even for local types. Clang doesn't always 1982 // appear to give every type a linkage name, which may be problematic for us. 1983 // FIXME: Investigate the consequences of not following them here. 1984 if (!Ty->getIdentifier().empty()) 1985 CO |= ClassOptions::HasUniqueName; 1986 1987 // Put the Nested flag on a type if it appears immediately inside a tag type. 1988 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 1989 // here. That flag is only set on definitions, and not forward declarations. 1990 const DIScope *ImmediateScope = Ty->getScope().resolve(); 1991 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 1992 CO |= ClassOptions::Nested; 1993 1994 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 1995 // type only when it has an immediate function scope. Clang never puts enums 1996 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 1997 // always in function, class, or file scopes. 1998 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 1999 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2000 CO |= ClassOptions::Scoped; 2001 } else { 2002 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2003 Scope = Scope->getScope().resolve()) { 2004 if (isa<DISubprogram>(Scope)) { 2005 CO |= ClassOptions::Scoped; 2006 break; 2007 } 2008 } 2009 } 2010 2011 return CO; 2012 } 2013 2014 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2015 switch (Ty->getTag()) { 2016 case dwarf::DW_TAG_class_type: 2017 case dwarf::DW_TAG_structure_type: 2018 case dwarf::DW_TAG_union_type: 2019 case dwarf::DW_TAG_enumeration_type: 2020 break; 2021 default: 2022 return; 2023 } 2024 2025 if (const auto *File = Ty->getFile()) { 2026 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2027 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2028 2029 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2030 TypeTable.writeLeafType(USLR); 2031 } 2032 } 2033 2034 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2035 ClassOptions CO = getCommonClassOptions(Ty); 2036 TypeIndex FTI; 2037 unsigned EnumeratorCount = 0; 2038 2039 if (Ty->isForwardDecl()) { 2040 CO |= ClassOptions::ForwardReference; 2041 } else { 2042 ContinuationRecordBuilder ContinuationBuilder; 2043 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2044 for (const DINode *Element : Ty->getElements()) { 2045 // We assume that the frontend provides all members in source declaration 2046 // order, which is what MSVC does. 2047 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2048 EnumeratorRecord ER(MemberAccess::Public, 2049 APSInt::getUnsigned(Enumerator->getValue()), 2050 Enumerator->getName()); 2051 ContinuationBuilder.writeMemberType(ER); 2052 EnumeratorCount++; 2053 } 2054 } 2055 FTI = TypeTable.insertRecord(ContinuationBuilder); 2056 } 2057 2058 std::string FullName = getFullyQualifiedName(Ty); 2059 2060 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2061 getTypeIndex(Ty->getBaseType())); 2062 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2063 2064 addUDTSrcLine(Ty, EnumTI); 2065 2066 return EnumTI; 2067 } 2068 2069 //===----------------------------------------------------------------------===// 2070 // ClassInfo 2071 //===----------------------------------------------------------------------===// 2072 2073 struct llvm::ClassInfo { 2074 struct MemberInfo { 2075 const DIDerivedType *MemberTypeNode; 2076 uint64_t BaseOffset; 2077 }; 2078 // [MemberInfo] 2079 using MemberList = std::vector<MemberInfo>; 2080 2081 using MethodsList = TinyPtrVector<const DISubprogram *>; 2082 // MethodName -> MethodsList 2083 using MethodsMap = MapVector<MDString *, MethodsList>; 2084 2085 /// Base classes. 2086 std::vector<const DIDerivedType *> Inheritance; 2087 2088 /// Direct members. 2089 MemberList Members; 2090 // Direct overloaded methods gathered by name. 2091 MethodsMap Methods; 2092 2093 TypeIndex VShapeTI; 2094 2095 std::vector<const DIType *> NestedTypes; 2096 }; 2097 2098 void CodeViewDebug::clear() { 2099 assert(CurFn == nullptr); 2100 FileIdMap.clear(); 2101 FnDebugInfo.clear(); 2102 FileToFilepathMap.clear(); 2103 LocalUDTs.clear(); 2104 GlobalUDTs.clear(); 2105 TypeIndices.clear(); 2106 CompleteTypeIndices.clear(); 2107 } 2108 2109 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2110 const DIDerivedType *DDTy) { 2111 if (!DDTy->getName().empty()) { 2112 Info.Members.push_back({DDTy, 0}); 2113 return; 2114 } 2115 2116 // An unnamed member may represent a nested struct or union. Attempt to 2117 // interpret the unnamed member as a DICompositeType possibly wrapped in 2118 // qualifier types. Add all the indirect fields to the current record if that 2119 // succeeds, and drop the member if that fails. 2120 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2121 uint64_t Offset = DDTy->getOffsetInBits(); 2122 const DIType *Ty = DDTy->getBaseType().resolve(); 2123 bool FullyResolved = false; 2124 while (!FullyResolved) { 2125 switch (Ty->getTag()) { 2126 case dwarf::DW_TAG_const_type: 2127 case dwarf::DW_TAG_volatile_type: 2128 // FIXME: we should apply the qualifier types to the indirect fields 2129 // rather than dropping them. 2130 Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve(); 2131 break; 2132 default: 2133 FullyResolved = true; 2134 break; 2135 } 2136 } 2137 2138 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2139 if (!DCTy) 2140 return; 2141 2142 ClassInfo NestedInfo = collectClassInfo(DCTy); 2143 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2144 Info.Members.push_back( 2145 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2146 } 2147 2148 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2149 ClassInfo Info; 2150 // Add elements to structure type. 2151 DINodeArray Elements = Ty->getElements(); 2152 for (auto *Element : Elements) { 2153 // We assume that the frontend provides all members in source declaration 2154 // order, which is what MSVC does. 2155 if (!Element) 2156 continue; 2157 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2158 Info.Methods[SP->getRawName()].push_back(SP); 2159 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2160 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2161 collectMemberInfo(Info, DDTy); 2162 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2163 Info.Inheritance.push_back(DDTy); 2164 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2165 DDTy->getName() == "__vtbl_ptr_type") { 2166 Info.VShapeTI = getTypeIndex(DDTy); 2167 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2168 Info.NestedTypes.push_back(DDTy); 2169 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2170 // Ignore friend members. It appears that MSVC emitted info about 2171 // friends in the past, but modern versions do not. 2172 } 2173 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2174 Info.NestedTypes.push_back(Composite); 2175 } 2176 // Skip other unrecognized kinds of elements. 2177 } 2178 return Info; 2179 } 2180 2181 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2182 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2183 // if a complete type should be emitted instead of a forward reference. 2184 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2185 !Ty->isForwardDecl(); 2186 } 2187 2188 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2189 // Emit the complete type for unnamed structs. C++ classes with methods 2190 // which have a circular reference back to the class type are expected to 2191 // be named by the front-end and should not be "unnamed". C unnamed 2192 // structs should not have circular references. 2193 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2194 // If this unnamed complete type is already in the process of being defined 2195 // then the description of the type is malformed and cannot be emitted 2196 // into CodeView correctly so report a fatal error. 2197 auto I = CompleteTypeIndices.find(Ty); 2198 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2199 report_fatal_error("cannot debug circular reference to unnamed type"); 2200 return getCompleteTypeIndex(Ty); 2201 } 2202 2203 // First, construct the forward decl. Don't look into Ty to compute the 2204 // forward decl options, since it might not be available in all TUs. 2205 TypeRecordKind Kind = getRecordKind(Ty); 2206 ClassOptions CO = 2207 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2208 std::string FullName = getFullyQualifiedName(Ty); 2209 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2210 FullName, Ty->getIdentifier()); 2211 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2212 if (!Ty->isForwardDecl()) 2213 DeferredCompleteTypes.push_back(Ty); 2214 return FwdDeclTI; 2215 } 2216 2217 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2218 // Construct the field list and complete type record. 2219 TypeRecordKind Kind = getRecordKind(Ty); 2220 ClassOptions CO = getCommonClassOptions(Ty); 2221 TypeIndex FieldTI; 2222 TypeIndex VShapeTI; 2223 unsigned FieldCount; 2224 bool ContainsNestedClass; 2225 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2226 lowerRecordFieldList(Ty); 2227 2228 if (ContainsNestedClass) 2229 CO |= ClassOptions::ContainsNestedClass; 2230 2231 std::string FullName = getFullyQualifiedName(Ty); 2232 2233 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2234 2235 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2236 SizeInBytes, FullName, Ty->getIdentifier()); 2237 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2238 2239 addUDTSrcLine(Ty, ClassTI); 2240 2241 addToUDTs(Ty); 2242 2243 return ClassTI; 2244 } 2245 2246 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2247 // Emit the complete type for unnamed unions. 2248 if (shouldAlwaysEmitCompleteClassType(Ty)) 2249 return getCompleteTypeIndex(Ty); 2250 2251 ClassOptions CO = 2252 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2253 std::string FullName = getFullyQualifiedName(Ty); 2254 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2255 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2256 if (!Ty->isForwardDecl()) 2257 DeferredCompleteTypes.push_back(Ty); 2258 return FwdDeclTI; 2259 } 2260 2261 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2262 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2263 TypeIndex FieldTI; 2264 unsigned FieldCount; 2265 bool ContainsNestedClass; 2266 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2267 lowerRecordFieldList(Ty); 2268 2269 if (ContainsNestedClass) 2270 CO |= ClassOptions::ContainsNestedClass; 2271 2272 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2273 std::string FullName = getFullyQualifiedName(Ty); 2274 2275 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2276 Ty->getIdentifier()); 2277 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2278 2279 addUDTSrcLine(Ty, UnionTI); 2280 2281 addToUDTs(Ty); 2282 2283 return UnionTI; 2284 } 2285 2286 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2287 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2288 // Manually count members. MSVC appears to count everything that generates a 2289 // field list record. Each individual overload in a method overload group 2290 // contributes to this count, even though the overload group is a single field 2291 // list record. 2292 unsigned MemberCount = 0; 2293 ClassInfo Info = collectClassInfo(Ty); 2294 ContinuationRecordBuilder ContinuationBuilder; 2295 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2296 2297 // Create base classes. 2298 for (const DIDerivedType *I : Info.Inheritance) { 2299 if (I->getFlags() & DINode::FlagVirtual) { 2300 // Virtual base. 2301 unsigned VBPtrOffset = I->getVBPtrOffset(); 2302 // FIXME: Despite the accessor name, the offset is really in bytes. 2303 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2304 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2305 ? TypeRecordKind::IndirectVirtualBaseClass 2306 : TypeRecordKind::VirtualBaseClass; 2307 VirtualBaseClassRecord VBCR( 2308 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2309 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2310 VBTableIndex); 2311 2312 ContinuationBuilder.writeMemberType(VBCR); 2313 MemberCount++; 2314 } else { 2315 assert(I->getOffsetInBits() % 8 == 0 && 2316 "bases must be on byte boundaries"); 2317 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2318 getTypeIndex(I->getBaseType()), 2319 I->getOffsetInBits() / 8); 2320 ContinuationBuilder.writeMemberType(BCR); 2321 MemberCount++; 2322 } 2323 } 2324 2325 // Create members. 2326 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2327 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2328 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2329 StringRef MemberName = Member->getName(); 2330 MemberAccess Access = 2331 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2332 2333 if (Member->isStaticMember()) { 2334 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2335 ContinuationBuilder.writeMemberType(SDMR); 2336 MemberCount++; 2337 continue; 2338 } 2339 2340 // Virtual function pointer member. 2341 if ((Member->getFlags() & DINode::FlagArtificial) && 2342 Member->getName().startswith("_vptr$")) { 2343 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2344 ContinuationBuilder.writeMemberType(VFPR); 2345 MemberCount++; 2346 continue; 2347 } 2348 2349 // Data member. 2350 uint64_t MemberOffsetInBits = 2351 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2352 if (Member->isBitField()) { 2353 uint64_t StartBitOffset = MemberOffsetInBits; 2354 if (const auto *CI = 2355 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2356 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2357 } 2358 StartBitOffset -= MemberOffsetInBits; 2359 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2360 StartBitOffset); 2361 MemberBaseType = TypeTable.writeLeafType(BFR); 2362 } 2363 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2364 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2365 MemberName); 2366 ContinuationBuilder.writeMemberType(DMR); 2367 MemberCount++; 2368 } 2369 2370 // Create methods 2371 for (auto &MethodItr : Info.Methods) { 2372 StringRef Name = MethodItr.first->getString(); 2373 2374 std::vector<OneMethodRecord> Methods; 2375 for (const DISubprogram *SP : MethodItr.second) { 2376 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2377 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2378 2379 unsigned VFTableOffset = -1; 2380 if (Introduced) 2381 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2382 2383 Methods.push_back(OneMethodRecord( 2384 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2385 translateMethodKindFlags(SP, Introduced), 2386 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2387 MemberCount++; 2388 } 2389 assert(!Methods.empty() && "Empty methods map entry"); 2390 if (Methods.size() == 1) 2391 ContinuationBuilder.writeMemberType(Methods[0]); 2392 else { 2393 // FIXME: Make this use its own ContinuationBuilder so that 2394 // MethodOverloadList can be split correctly. 2395 MethodOverloadListRecord MOLR(Methods); 2396 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2397 2398 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2399 ContinuationBuilder.writeMemberType(OMR); 2400 } 2401 } 2402 2403 // Create nested classes. 2404 for (const DIType *Nested : Info.NestedTypes) { 2405 NestedTypeRecord R(getTypeIndex(DITypeRef(Nested)), Nested->getName()); 2406 ContinuationBuilder.writeMemberType(R); 2407 MemberCount++; 2408 } 2409 2410 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2411 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2412 !Info.NestedTypes.empty()); 2413 } 2414 2415 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2416 if (!VBPType.getIndex()) { 2417 // Make a 'const int *' type. 2418 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2419 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2420 2421 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2422 : PointerKind::Near32; 2423 PointerMode PM = PointerMode::Pointer; 2424 PointerOptions PO = PointerOptions::None; 2425 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2426 VBPType = TypeTable.writeLeafType(PR); 2427 } 2428 2429 return VBPType; 2430 } 2431 2432 TypeIndex CodeViewDebug::getTypeIndex(DITypeRef TypeRef, DITypeRef ClassTyRef) { 2433 const DIType *Ty = TypeRef.resolve(); 2434 const DIType *ClassTy = ClassTyRef.resolve(); 2435 2436 // The null DIType is the void type. Don't try to hash it. 2437 if (!Ty) 2438 return TypeIndex::Void(); 2439 2440 // Check if we've already translated this type. Don't try to do a 2441 // get-or-create style insertion that caches the hash lookup across the 2442 // lowerType call. It will update the TypeIndices map. 2443 auto I = TypeIndices.find({Ty, ClassTy}); 2444 if (I != TypeIndices.end()) 2445 return I->second; 2446 2447 TypeLoweringScope S(*this); 2448 TypeIndex TI = lowerType(Ty, ClassTy); 2449 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2450 } 2451 2452 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(DITypeRef TypeRef) { 2453 DIType *Ty = TypeRef.resolve(); 2454 PointerRecord PR(getTypeIndex(Ty), 2455 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2456 : PointerKind::Near32, 2457 PointerMode::LValueReference, PointerOptions::None, 2458 Ty->getSizeInBits() / 8); 2459 return TypeTable.writeLeafType(PR); 2460 } 2461 2462 TypeIndex CodeViewDebug::getCompleteTypeIndex(DITypeRef TypeRef) { 2463 const DIType *Ty = TypeRef.resolve(); 2464 2465 // The null DIType is the void type. Don't try to hash it. 2466 if (!Ty) 2467 return TypeIndex::Void(); 2468 2469 // If this is a non-record type, the complete type index is the same as the 2470 // normal type index. Just call getTypeIndex. 2471 switch (Ty->getTag()) { 2472 case dwarf::DW_TAG_class_type: 2473 case dwarf::DW_TAG_structure_type: 2474 case dwarf::DW_TAG_union_type: 2475 break; 2476 default: 2477 return getTypeIndex(Ty); 2478 } 2479 2480 // Check if we've already translated the complete record type. 2481 const auto *CTy = cast<DICompositeType>(Ty); 2482 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2483 if (!InsertResult.second) 2484 return InsertResult.first->second; 2485 2486 TypeLoweringScope S(*this); 2487 2488 // Make sure the forward declaration is emitted first. It's unclear if this 2489 // is necessary, but MSVC does it, and we should follow suit until we can show 2490 // otherwise. 2491 // We only emit a forward declaration for named types. 2492 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2493 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2494 2495 // Just use the forward decl if we don't have complete type info. This 2496 // might happen if the frontend is using modules and expects the complete 2497 // definition to be emitted elsewhere. 2498 if (CTy->isForwardDecl()) 2499 return FwdDeclTI; 2500 } 2501 2502 TypeIndex TI; 2503 switch (CTy->getTag()) { 2504 case dwarf::DW_TAG_class_type: 2505 case dwarf::DW_TAG_structure_type: 2506 TI = lowerCompleteTypeClass(CTy); 2507 break; 2508 case dwarf::DW_TAG_union_type: 2509 TI = lowerCompleteTypeUnion(CTy); 2510 break; 2511 default: 2512 llvm_unreachable("not a record"); 2513 } 2514 2515 // Update the type index associated with this CompositeType. This cannot 2516 // use the 'InsertResult' iterator above because it is potentially 2517 // invalidated by map insertions which can occur while lowering the class 2518 // type above. 2519 CompleteTypeIndices[CTy] = TI; 2520 return TI; 2521 } 2522 2523 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2524 /// and do this until fixpoint, as each complete record type typically 2525 /// references 2526 /// many other record types. 2527 void CodeViewDebug::emitDeferredCompleteTypes() { 2528 SmallVector<const DICompositeType *, 4> TypesToEmit; 2529 while (!DeferredCompleteTypes.empty()) { 2530 std::swap(DeferredCompleteTypes, TypesToEmit); 2531 for (const DICompositeType *RecordTy : TypesToEmit) 2532 getCompleteTypeIndex(RecordTy); 2533 TypesToEmit.clear(); 2534 } 2535 } 2536 2537 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2538 ArrayRef<LocalVariable> Locals) { 2539 // Get the sorted list of parameters and emit them first. 2540 SmallVector<const LocalVariable *, 6> Params; 2541 for (const LocalVariable &L : Locals) 2542 if (L.DIVar->isParameter()) 2543 Params.push_back(&L); 2544 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2545 return L->DIVar->getArg() < R->DIVar->getArg(); 2546 }); 2547 for (const LocalVariable *L : Params) 2548 emitLocalVariable(FI, *L); 2549 2550 // Next emit all non-parameters in the order that we found them. 2551 for (const LocalVariable &L : Locals) 2552 if (!L.DIVar->isParameter()) 2553 emitLocalVariable(FI, L); 2554 } 2555 2556 /// Only call this on endian-specific types like ulittle16_t and little32_t, or 2557 /// structs composed of them. 2558 template <typename T> 2559 static void copyBytesForDefRange(SmallString<20> &BytePrefix, 2560 SymbolKind SymKind, const T &DefRangeHeader) { 2561 BytePrefix.resize(2 + sizeof(T)); 2562 ulittle16_t SymKindLE = ulittle16_t(SymKind); 2563 memcpy(&BytePrefix[0], &SymKindLE, 2); 2564 memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T)); 2565 } 2566 2567 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2568 const LocalVariable &Var) { 2569 // LocalSym record, see SymbolRecord.h for more info. 2570 MCSymbol *LocalBegin = MMI->getContext().createTempSymbol(), 2571 *LocalEnd = MMI->getContext().createTempSymbol(); 2572 OS.AddComment("Record length"); 2573 OS.emitAbsoluteSymbolDiff(LocalEnd, LocalBegin, 2); 2574 OS.EmitLabel(LocalBegin); 2575 2576 OS.AddComment("Record kind: S_LOCAL"); 2577 OS.EmitIntValue(unsigned(SymbolKind::S_LOCAL), 2); 2578 2579 LocalSymFlags Flags = LocalSymFlags::None; 2580 if (Var.DIVar->isParameter()) 2581 Flags |= LocalSymFlags::IsParameter; 2582 if (Var.DefRanges.empty()) 2583 Flags |= LocalSymFlags::IsOptimizedOut; 2584 2585 OS.AddComment("TypeIndex"); 2586 TypeIndex TI = Var.UseReferenceType 2587 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2588 : getCompleteTypeIndex(Var.DIVar->getType()); 2589 OS.EmitIntValue(TI.getIndex(), 4); 2590 OS.AddComment("Flags"); 2591 OS.EmitIntValue(static_cast<uint16_t>(Flags), 2); 2592 // Truncate the name so we won't overflow the record length field. 2593 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2594 OS.EmitLabel(LocalEnd); 2595 2596 // Calculate the on disk prefix of the appropriate def range record. The 2597 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2598 // should be big enough to hold all forms without memory allocation. 2599 SmallString<20> BytePrefix; 2600 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2601 BytePrefix.clear(); 2602 if (DefRange.InMemory) { 2603 int Offset = DefRange.DataOffset; 2604 unsigned Reg = DefRange.CVRegister; 2605 2606 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2607 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2608 // instead. In frames without stack realignment, $T0 will be the CFA. 2609 if (RegisterId(Reg) == RegisterId::ESP) { 2610 Reg = unsigned(RegisterId::VFRAME); 2611 Offset += FI.OffsetAdjustment; 2612 } 2613 2614 // If we can use the chosen frame pointer for the frame and this isn't a 2615 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2616 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2617 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2618 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2619 (bool(Flags & LocalSymFlags::IsParameter) 2620 ? (EncFP == FI.EncodedParamFramePtrReg) 2621 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2622 little32_t FPOffset = little32_t(Offset); 2623 copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset); 2624 } else { 2625 uint16_t RegRelFlags = 0; 2626 if (DefRange.IsSubfield) { 2627 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2628 (DefRange.StructOffset 2629 << DefRangeRegisterRelSym::OffsetInParentShift); 2630 } 2631 DefRangeRegisterRelSym::Header DRHdr; 2632 DRHdr.Register = Reg; 2633 DRHdr.Flags = RegRelFlags; 2634 DRHdr.BasePointerOffset = Offset; 2635 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr); 2636 } 2637 } else { 2638 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2639 if (DefRange.IsSubfield) { 2640 DefRangeSubfieldRegisterSym::Header DRHdr; 2641 DRHdr.Register = DefRange.CVRegister; 2642 DRHdr.MayHaveNoName = 0; 2643 DRHdr.OffsetInParent = DefRange.StructOffset; 2644 copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr); 2645 } else { 2646 DefRangeRegisterSym::Header DRHdr; 2647 DRHdr.Register = DefRange.CVRegister; 2648 DRHdr.MayHaveNoName = 0; 2649 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr); 2650 } 2651 } 2652 OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix); 2653 } 2654 } 2655 2656 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2657 const FunctionInfo& FI) { 2658 for (LexicalBlock *Block : Blocks) 2659 emitLexicalBlock(*Block, FI); 2660 } 2661 2662 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2663 /// lexical block scope. 2664 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2665 const FunctionInfo& FI) { 2666 MCSymbol *RecordBegin = MMI->getContext().createTempSymbol(), 2667 *RecordEnd = MMI->getContext().createTempSymbol(); 2668 2669 // Lexical block symbol record. 2670 OS.AddComment("Record length"); 2671 OS.emitAbsoluteSymbolDiff(RecordEnd, RecordBegin, 2); // Record Length 2672 OS.EmitLabel(RecordBegin); 2673 OS.AddComment("Record kind: S_BLOCK32"); 2674 OS.EmitIntValue(SymbolKind::S_BLOCK32, 2); // Record Kind 2675 OS.AddComment("PtrParent"); 2676 OS.EmitIntValue(0, 4); // PtrParent 2677 OS.AddComment("PtrEnd"); 2678 OS.EmitIntValue(0, 4); // PtrEnd 2679 OS.AddComment("Code size"); 2680 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2681 OS.AddComment("Function section relative address"); 2682 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2683 OS.AddComment("Function section index"); 2684 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2685 OS.AddComment("Lexical block name"); 2686 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2687 OS.EmitLabel(RecordEnd); 2688 2689 // Emit variables local to this lexical block. 2690 emitLocalVariableList(FI, Block.Locals); 2691 2692 // Emit lexical blocks contained within this block. 2693 emitLexicalBlockList(Block.Children, FI); 2694 2695 // Close the lexical block scope. 2696 OS.AddComment("Record length"); 2697 OS.EmitIntValue(2, 2); // Record Length 2698 OS.AddComment("Record kind: S_END"); 2699 OS.EmitIntValue(SymbolKind::S_END, 2); // Record Kind 2700 } 2701 2702 /// Convenience routine for collecting lexical block information for a list 2703 /// of lexical scopes. 2704 void CodeViewDebug::collectLexicalBlockInfo( 2705 SmallVectorImpl<LexicalScope *> &Scopes, 2706 SmallVectorImpl<LexicalBlock *> &Blocks, 2707 SmallVectorImpl<LocalVariable> &Locals) { 2708 for (LexicalScope *Scope : Scopes) 2709 collectLexicalBlockInfo(*Scope, Blocks, Locals); 2710 } 2711 2712 /// Populate the lexical blocks and local variable lists of the parent with 2713 /// information about the specified lexical scope. 2714 void CodeViewDebug::collectLexicalBlockInfo( 2715 LexicalScope &Scope, 2716 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2717 SmallVectorImpl<LocalVariable> &ParentLocals) { 2718 if (Scope.isAbstractScope()) 2719 return; 2720 2721 auto LocalsIter = ScopeVariables.find(&Scope); 2722 if (LocalsIter == ScopeVariables.end()) { 2723 // This scope does not contain variables and can be eliminated. 2724 collectLexicalBlockInfo(Scope.getChildren(), ParentBlocks, ParentLocals); 2725 return; 2726 } 2727 SmallVectorImpl<LocalVariable> &Locals = LocalsIter->second; 2728 2729 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2730 if (!DILB) { 2731 // This scope is not a lexical block and can be eliminated, but keep any 2732 // local variables it contains. 2733 ParentLocals.append(Locals.begin(), Locals.end()); 2734 collectLexicalBlockInfo(Scope.getChildren(), ParentBlocks, ParentLocals); 2735 return; 2736 } 2737 2738 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2739 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) { 2740 // This lexical block scope has too many address ranges to represent in the 2741 // current CodeView format or does not have a valid address range. 2742 // Eliminate this lexical scope and promote any locals it contains to the 2743 // parent scope. 2744 // 2745 // For lexical scopes with multiple address ranges you may be tempted to 2746 // construct a single range covering every instruction where the block is 2747 // live and everything in between. Unfortunately, Visual Studio only 2748 // displays variables from the first matching lexical block scope. If the 2749 // first lexical block contains exception handling code or cold code which 2750 // is moved to the bottom of the routine creating a single range covering 2751 // nearly the entire routine, then it will hide all other lexical blocks 2752 // and the variables they contain. 2753 // 2754 ParentLocals.append(Locals.begin(), Locals.end()); 2755 collectLexicalBlockInfo(Scope.getChildren(), ParentBlocks, ParentLocals); 2756 return; 2757 } 2758 2759 // Create a new CodeView lexical block for this lexical scope. If we've 2760 // seen this DILexicalBlock before then the scope tree is malformed and 2761 // we can handle this gracefully by not processing it a second time. 2762 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2763 if (!BlockInsertion.second) 2764 return; 2765 2766 // Create a lexical block containing the local variables and collect the 2767 // the lexical block information for the children. 2768 const InsnRange &Range = Ranges.front(); 2769 assert(Range.first && Range.second); 2770 LexicalBlock &Block = BlockInsertion.first->second; 2771 Block.Begin = getLabelBeforeInsn(Range.first); 2772 Block.End = getLabelAfterInsn(Range.second); 2773 assert(Block.Begin && "missing label for scope begin"); 2774 assert(Block.End && "missing label for scope end"); 2775 Block.Name = DILB->getName(); 2776 Block.Locals = std::move(Locals); 2777 ParentBlocks.push_back(&Block); 2778 collectLexicalBlockInfo(Scope.getChildren(), Block.Children, Block.Locals); 2779 } 2780 2781 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2782 const Function &GV = MF->getFunction(); 2783 assert(FnDebugInfo.count(&GV)); 2784 assert(CurFn == FnDebugInfo[&GV].get()); 2785 2786 collectVariableInfo(GV.getSubprogram()); 2787 2788 // Build the lexical block structure to emit for this routine. 2789 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2790 collectLexicalBlockInfo(*CFS, CurFn->ChildBlocks, CurFn->Locals); 2791 2792 // Clear the scope and variable information from the map which will not be 2793 // valid after we have finished processing this routine. This also prepares 2794 // the map for the subsequent routine. 2795 ScopeVariables.clear(); 2796 2797 // Don't emit anything if we don't have any line tables. 2798 // Thunks are compiler-generated and probably won't have source correlation. 2799 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2800 FnDebugInfo.erase(&GV); 2801 CurFn = nullptr; 2802 return; 2803 } 2804 2805 CurFn->Annotations = MF->getCodeViewAnnotations(); 2806 2807 CurFn->End = Asm->getFunctionEnd(); 2808 2809 CurFn = nullptr; 2810 } 2811 2812 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2813 DebugHandlerBase::beginInstruction(MI); 2814 2815 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2816 if (!Asm || !CurFn || MI->isDebugInstr() || 2817 MI->getFlag(MachineInstr::FrameSetup)) 2818 return; 2819 2820 // If the first instruction of a new MBB has no location, find the first 2821 // instruction with a location and use that. 2822 DebugLoc DL = MI->getDebugLoc(); 2823 if (!DL && MI->getParent() != PrevInstBB) { 2824 for (const auto &NextMI : *MI->getParent()) { 2825 if (NextMI.isDebugInstr()) 2826 continue; 2827 DL = NextMI.getDebugLoc(); 2828 if (DL) 2829 break; 2830 } 2831 } 2832 PrevInstBB = MI->getParent(); 2833 2834 // If we still don't have a debug location, don't record a location. 2835 if (!DL) 2836 return; 2837 2838 maybeRecordLocation(DL, Asm->MF); 2839 } 2840 2841 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2842 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2843 *EndLabel = MMI->getContext().createTempSymbol(); 2844 OS.EmitIntValue(unsigned(Kind), 4); 2845 OS.AddComment("Subsection size"); 2846 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2847 OS.EmitLabel(BeginLabel); 2848 return EndLabel; 2849 } 2850 2851 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2852 OS.EmitLabel(EndLabel); 2853 // Every subsection must be aligned to a 4-byte boundary. 2854 OS.EmitValueToAlignment(4); 2855 } 2856 2857 void CodeViewDebug::emitDebugInfoForUDTs( 2858 ArrayRef<std::pair<std::string, const DIType *>> UDTs) { 2859 for (const auto &UDT : UDTs) { 2860 const DIType *T = UDT.second; 2861 assert(shouldEmitUdt(T)); 2862 2863 MCSymbol *UDTRecordBegin = MMI->getContext().createTempSymbol(), 2864 *UDTRecordEnd = MMI->getContext().createTempSymbol(); 2865 OS.AddComment("Record length"); 2866 OS.emitAbsoluteSymbolDiff(UDTRecordEnd, UDTRecordBegin, 2); 2867 OS.EmitLabel(UDTRecordBegin); 2868 2869 OS.AddComment("Record kind: S_UDT"); 2870 OS.EmitIntValue(unsigned(SymbolKind::S_UDT), 2); 2871 2872 OS.AddComment("Type"); 2873 OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4); 2874 2875 emitNullTerminatedSymbolName(OS, UDT.first); 2876 OS.EmitLabel(UDTRecordEnd); 2877 } 2878 } 2879 2880 void CodeViewDebug::emitDebugInfoForGlobals() { 2881 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 2882 GlobalMap; 2883 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 2884 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 2885 GV.getDebugInfo(GVEs); 2886 for (const auto *GVE : GVEs) 2887 GlobalMap[GVE] = &GV; 2888 } 2889 2890 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2891 for (const MDNode *Node : CUs->operands()) { 2892 const auto *CU = cast<DICompileUnit>(Node); 2893 2894 // First, emit all globals that are not in a comdat in a single symbol 2895 // substream. MSVC doesn't like it if the substream is empty, so only open 2896 // it if we have at least one global to emit. 2897 switchToDebugSectionForSymbol(nullptr); 2898 MCSymbol *EndLabel = nullptr; 2899 for (const auto *GVE : CU->getGlobalVariables()) { 2900 if (const auto *GV = GlobalMap.lookup(GVE)) 2901 if (!GV->hasComdat() && !GV->isDeclarationForLinker()) { 2902 if (!EndLabel) { 2903 OS.AddComment("Symbol subsection for globals"); 2904 EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 2905 } 2906 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 2907 emitDebugInfoForGlobal(GVE->getVariable(), GV, Asm->getSymbol(GV)); 2908 } 2909 } 2910 if (EndLabel) 2911 endCVSubsection(EndLabel); 2912 2913 // Second, emit each global that is in a comdat into its own .debug$S 2914 // section along with its own symbol substream. 2915 for (const auto *GVE : CU->getGlobalVariables()) { 2916 if (const auto *GV = GlobalMap.lookup(GVE)) { 2917 if (GV->hasComdat()) { 2918 MCSymbol *GVSym = Asm->getSymbol(GV); 2919 OS.AddComment("Symbol subsection for " + 2920 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 2921 switchToDebugSectionForSymbol(GVSym); 2922 EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 2923 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 2924 emitDebugInfoForGlobal(GVE->getVariable(), GV, GVSym); 2925 endCVSubsection(EndLabel); 2926 } 2927 } 2928 } 2929 } 2930 } 2931 2932 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 2933 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2934 for (const MDNode *Node : CUs->operands()) { 2935 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 2936 if (DIType *RT = dyn_cast<DIType>(Ty)) { 2937 getTypeIndex(RT); 2938 // FIXME: Add to global/local DTU list. 2939 } 2940 } 2941 } 2942 } 2943 2944 void CodeViewDebug::emitDebugInfoForGlobal(const DIGlobalVariable *DIGV, 2945 const GlobalVariable *GV, 2946 MCSymbol *GVSym) { 2947 // DataSym record, see SymbolRecord.h for more info. 2948 // FIXME: Thread local data, etc 2949 MCSymbol *DataBegin = MMI->getContext().createTempSymbol(), 2950 *DataEnd = MMI->getContext().createTempSymbol(); 2951 const unsigned FixedLengthOfThisRecord = 12; 2952 OS.AddComment("Record length"); 2953 OS.emitAbsoluteSymbolDiff(DataEnd, DataBegin, 2); 2954 OS.EmitLabel(DataBegin); 2955 if (DIGV->isLocalToUnit()) { 2956 if (GV->isThreadLocal()) { 2957 OS.AddComment("Record kind: S_LTHREAD32"); 2958 OS.EmitIntValue(unsigned(SymbolKind::S_LTHREAD32), 2); 2959 } else { 2960 OS.AddComment("Record kind: S_LDATA32"); 2961 OS.EmitIntValue(unsigned(SymbolKind::S_LDATA32), 2); 2962 } 2963 } else { 2964 if (GV->isThreadLocal()) { 2965 OS.AddComment("Record kind: S_GTHREAD32"); 2966 OS.EmitIntValue(unsigned(SymbolKind::S_GTHREAD32), 2); 2967 } else { 2968 OS.AddComment("Record kind: S_GDATA32"); 2969 OS.EmitIntValue(unsigned(SymbolKind::S_GDATA32), 2); 2970 } 2971 } 2972 OS.AddComment("Type"); 2973 OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4); 2974 OS.AddComment("DataOffset"); 2975 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 2976 OS.AddComment("Segment"); 2977 OS.EmitCOFFSectionIndex(GVSym); 2978 OS.AddComment("Name"); 2979 emitNullTerminatedSymbolName(OS, DIGV->getName(), FixedLengthOfThisRecord); 2980 OS.EmitLabel(DataEnd); 2981 } 2982