1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "DwarfExpression.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/ArrayRef.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/DenseSet.h" 19 #include "llvm/ADT/MapVector.h" 20 #include "llvm/ADT/None.h" 21 #include "llvm/ADT/Optional.h" 22 #include "llvm/ADT/STLExtras.h" 23 #include "llvm/ADT/SmallString.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/StringRef.h" 26 #include "llvm/ADT/TinyPtrVector.h" 27 #include "llvm/ADT/Triple.h" 28 #include "llvm/ADT/Twine.h" 29 #include "llvm/BinaryFormat/COFF.h" 30 #include "llvm/BinaryFormat/Dwarf.h" 31 #include "llvm/CodeGen/AsmPrinter.h" 32 #include "llvm/CodeGen/LexicalScopes.h" 33 #include "llvm/CodeGen/MachineFrameInfo.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineInstr.h" 36 #include "llvm/CodeGen/MachineModuleInfo.h" 37 #include "llvm/CodeGen/MachineOperand.h" 38 #include "llvm/CodeGen/TargetFrameLowering.h" 39 #include "llvm/CodeGen/TargetRegisterInfo.h" 40 #include "llvm/CodeGen/TargetSubtargetInfo.h" 41 #include "llvm/Config/llvm-config.h" 42 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 43 #include "llvm/DebugInfo/CodeView/CodeView.h" 44 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 45 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 46 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 47 #include "llvm/DebugInfo/CodeView/EnumTables.h" 48 #include "llvm/DebugInfo/CodeView/Line.h" 49 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 50 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 51 #include "llvm/DebugInfo/CodeView/TypeIndex.h" 52 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 53 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 54 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 55 #include "llvm/IR/Constants.h" 56 #include "llvm/IR/DataLayout.h" 57 #include "llvm/IR/DebugInfoMetadata.h" 58 #include "llvm/IR/DebugLoc.h" 59 #include "llvm/IR/Function.h" 60 #include "llvm/IR/GlobalValue.h" 61 #include "llvm/IR/GlobalVariable.h" 62 #include "llvm/IR/Metadata.h" 63 #include "llvm/IR/Module.h" 64 #include "llvm/MC/MCAsmInfo.h" 65 #include "llvm/MC/MCContext.h" 66 #include "llvm/MC/MCSectionCOFF.h" 67 #include "llvm/MC/MCStreamer.h" 68 #include "llvm/MC/MCSymbol.h" 69 #include "llvm/Support/BinaryByteStream.h" 70 #include "llvm/Support/BinaryStreamReader.h" 71 #include "llvm/Support/BinaryStreamWriter.h" 72 #include "llvm/Support/Casting.h" 73 #include "llvm/Support/CommandLine.h" 74 #include "llvm/Support/Compiler.h" 75 #include "llvm/Support/Endian.h" 76 #include "llvm/Support/Error.h" 77 #include "llvm/Support/ErrorHandling.h" 78 #include "llvm/Support/FormatVariadic.h" 79 #include "llvm/Support/Path.h" 80 #include "llvm/Support/SMLoc.h" 81 #include "llvm/Support/ScopedPrinter.h" 82 #include "llvm/Target/TargetLoweringObjectFile.h" 83 #include "llvm/Target/TargetMachine.h" 84 #include <algorithm> 85 #include <cassert> 86 #include <cctype> 87 #include <cstddef> 88 #include <cstdint> 89 #include <iterator> 90 #include <limits> 91 #include <string> 92 #include <utility> 93 #include <vector> 94 95 using namespace llvm; 96 using namespace llvm::codeview; 97 98 namespace { 99 class CVMCAdapter : public CodeViewRecordStreamer { 100 public: 101 CVMCAdapter(MCStreamer &OS) : OS(&OS) {} 102 103 void EmitBytes(StringRef Data) { OS->EmitBytes(Data); } 104 105 void EmitIntValue(uint64_t Value, unsigned Size) { 106 OS->EmitIntValue(Value, Size); 107 } 108 109 void EmitBinaryData(StringRef Data) { OS->EmitBinaryData(Data); } 110 111 private: 112 MCStreamer *OS = nullptr; 113 }; 114 } // namespace 115 116 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 117 switch (Type) { 118 case Triple::ArchType::x86: 119 return CPUType::Pentium3; 120 case Triple::ArchType::x86_64: 121 return CPUType::X64; 122 case Triple::ArchType::thumb: 123 return CPUType::Thumb; 124 case Triple::ArchType::aarch64: 125 return CPUType::ARM64; 126 default: 127 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 128 } 129 } 130 131 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 132 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 133 // If module doesn't have named metadata anchors or COFF debug section 134 // is not available, skip any debug info related stuff. 135 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 136 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 137 Asm = nullptr; 138 MMI->setDebugInfoAvailability(false); 139 return; 140 } 141 // Tell MMI that we have debug info. 142 MMI->setDebugInfoAvailability(true); 143 144 TheCPU = 145 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 146 147 collectGlobalVariableInfo(); 148 149 // Check if we should emit type record hashes. 150 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 151 MMI->getModule()->getModuleFlag("CodeViewGHash")); 152 EmitDebugGlobalHashes = GH && !GH->isZero(); 153 } 154 155 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 156 std::string &Filepath = FileToFilepathMap[File]; 157 if (!Filepath.empty()) 158 return Filepath; 159 160 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 161 162 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 163 // it textually because one of the path components could be a symlink. 164 if (Dir.startswith("/") || Filename.startswith("/")) { 165 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 166 return Filename; 167 Filepath = Dir; 168 if (Dir.back() != '/') 169 Filepath += '/'; 170 Filepath += Filename; 171 return Filepath; 172 } 173 174 // Clang emits directory and relative filename info into the IR, but CodeView 175 // operates on full paths. We could change Clang to emit full paths too, but 176 // that would increase the IR size and probably not needed for other users. 177 // For now, just concatenate and canonicalize the path here. 178 if (Filename.find(':') == 1) 179 Filepath = Filename; 180 else 181 Filepath = (Dir + "\\" + Filename).str(); 182 183 // Canonicalize the path. We have to do it textually because we may no longer 184 // have access the file in the filesystem. 185 // First, replace all slashes with backslashes. 186 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 187 188 // Remove all "\.\" with "\". 189 size_t Cursor = 0; 190 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 191 Filepath.erase(Cursor, 2); 192 193 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 194 // path should be well-formatted, e.g. start with a drive letter, etc. 195 Cursor = 0; 196 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 197 // Something's wrong if the path starts with "\..\", abort. 198 if (Cursor == 0) 199 break; 200 201 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 202 if (PrevSlash == std::string::npos) 203 // Something's wrong, abort. 204 break; 205 206 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 207 // The next ".." might be following the one we've just erased. 208 Cursor = PrevSlash; 209 } 210 211 // Remove all duplicate backslashes. 212 Cursor = 0; 213 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 214 Filepath.erase(Cursor, 1); 215 216 return Filepath; 217 } 218 219 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 220 StringRef FullPath = getFullFilepath(F); 221 unsigned NextId = FileIdMap.size() + 1; 222 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 223 if (Insertion.second) { 224 // We have to compute the full filepath and emit a .cv_file directive. 225 ArrayRef<uint8_t> ChecksumAsBytes; 226 FileChecksumKind CSKind = FileChecksumKind::None; 227 if (F->getChecksum()) { 228 std::string Checksum = fromHex(F->getChecksum()->Value); 229 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 230 memcpy(CKMem, Checksum.data(), Checksum.size()); 231 ChecksumAsBytes = ArrayRef<uint8_t>( 232 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 233 switch (F->getChecksum()->Kind) { 234 case DIFile::CSK_MD5: CSKind = FileChecksumKind::MD5; break; 235 case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break; 236 } 237 } 238 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 239 static_cast<unsigned>(CSKind)); 240 (void)Success; 241 assert(Success && ".cv_file directive failed"); 242 } 243 return Insertion.first->second; 244 } 245 246 CodeViewDebug::InlineSite & 247 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 248 const DISubprogram *Inlinee) { 249 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 250 InlineSite *Site = &SiteInsertion.first->second; 251 if (SiteInsertion.second) { 252 unsigned ParentFuncId = CurFn->FuncId; 253 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 254 ParentFuncId = 255 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 256 .SiteFuncId; 257 258 Site->SiteFuncId = NextFuncId++; 259 OS.EmitCVInlineSiteIdDirective( 260 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 261 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 262 Site->Inlinee = Inlinee; 263 InlinedSubprograms.insert(Inlinee); 264 getFuncIdForSubprogram(Inlinee); 265 } 266 return *Site; 267 } 268 269 static StringRef getPrettyScopeName(const DIScope *Scope) { 270 StringRef ScopeName = Scope->getName(); 271 if (!ScopeName.empty()) 272 return ScopeName; 273 274 switch (Scope->getTag()) { 275 case dwarf::DW_TAG_enumeration_type: 276 case dwarf::DW_TAG_class_type: 277 case dwarf::DW_TAG_structure_type: 278 case dwarf::DW_TAG_union_type: 279 return "<unnamed-tag>"; 280 case dwarf::DW_TAG_namespace: 281 return "`anonymous namespace'"; 282 } 283 284 return StringRef(); 285 } 286 287 static const DISubprogram *getQualifiedNameComponents( 288 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 289 const DISubprogram *ClosestSubprogram = nullptr; 290 while (Scope != nullptr) { 291 if (ClosestSubprogram == nullptr) 292 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 293 StringRef ScopeName = getPrettyScopeName(Scope); 294 if (!ScopeName.empty()) 295 QualifiedNameComponents.push_back(ScopeName); 296 Scope = Scope->getScope(); 297 } 298 return ClosestSubprogram; 299 } 300 301 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents, 302 StringRef TypeName) { 303 std::string FullyQualifiedName; 304 for (StringRef QualifiedNameComponent : 305 llvm::reverse(QualifiedNameComponents)) { 306 FullyQualifiedName.append(QualifiedNameComponent); 307 FullyQualifiedName.append("::"); 308 } 309 FullyQualifiedName.append(TypeName); 310 return FullyQualifiedName; 311 } 312 313 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) { 314 SmallVector<StringRef, 5> QualifiedNameComponents; 315 getQualifiedNameComponents(Scope, QualifiedNameComponents); 316 return getQualifiedName(QualifiedNameComponents, Name); 317 } 318 319 struct CodeViewDebug::TypeLoweringScope { 320 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 321 ~TypeLoweringScope() { 322 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 323 // inner TypeLoweringScopes don't attempt to emit deferred types. 324 if (CVD.TypeEmissionLevel == 1) 325 CVD.emitDeferredCompleteTypes(); 326 --CVD.TypeEmissionLevel; 327 } 328 CodeViewDebug &CVD; 329 }; 330 331 static std::string getFullyQualifiedName(const DIScope *Ty) { 332 const DIScope *Scope = Ty->getScope(); 333 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 334 } 335 336 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 337 // No scope means global scope and that uses the zero index. 338 if (!Scope || isa<DIFile>(Scope)) 339 return TypeIndex(); 340 341 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 342 343 // Check if we've already translated this scope. 344 auto I = TypeIndices.find({Scope, nullptr}); 345 if (I != TypeIndices.end()) 346 return I->second; 347 348 // Build the fully qualified name of the scope. 349 std::string ScopeName = getFullyQualifiedName(Scope); 350 StringIdRecord SID(TypeIndex(), ScopeName); 351 auto TI = TypeTable.writeLeafType(SID); 352 return recordTypeIndexForDINode(Scope, TI); 353 } 354 355 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 356 assert(SP); 357 358 // Check if we've already translated this subprogram. 359 auto I = TypeIndices.find({SP, nullptr}); 360 if (I != TypeIndices.end()) 361 return I->second; 362 363 // The display name includes function template arguments. Drop them to match 364 // MSVC. 365 StringRef DisplayName = SP->getName().split('<').first; 366 367 const DIScope *Scope = SP->getScope(); 368 TypeIndex TI; 369 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 370 // If the scope is a DICompositeType, then this must be a method. Member 371 // function types take some special handling, and require access to the 372 // subprogram. 373 TypeIndex ClassType = getTypeIndex(Class); 374 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 375 DisplayName); 376 TI = TypeTable.writeLeafType(MFuncId); 377 } else { 378 // Otherwise, this must be a free function. 379 TypeIndex ParentScope = getScopeIndex(Scope); 380 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 381 TI = TypeTable.writeLeafType(FuncId); 382 } 383 384 return recordTypeIndexForDINode(SP, TI); 385 } 386 387 static bool isNonTrivial(const DICompositeType *DCTy) { 388 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 389 } 390 391 static FunctionOptions 392 getFunctionOptions(const DISubroutineType *Ty, 393 const DICompositeType *ClassTy = nullptr, 394 StringRef SPName = StringRef("")) { 395 FunctionOptions FO = FunctionOptions::None; 396 const DIType *ReturnTy = nullptr; 397 if (auto TypeArray = Ty->getTypeArray()) { 398 if (TypeArray.size()) 399 ReturnTy = TypeArray[0]; 400 } 401 402 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) { 403 if (isNonTrivial(ReturnDCTy)) 404 FO |= FunctionOptions::CxxReturnUdt; 405 } 406 407 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 408 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 409 FO |= FunctionOptions::Constructor; 410 411 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 412 413 } 414 return FO; 415 } 416 417 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 418 const DICompositeType *Class) { 419 // Always use the method declaration as the key for the function type. The 420 // method declaration contains the this adjustment. 421 if (SP->getDeclaration()) 422 SP = SP->getDeclaration(); 423 assert(!SP->getDeclaration() && "should use declaration as key"); 424 425 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 426 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 427 auto I = TypeIndices.find({SP, Class}); 428 if (I != TypeIndices.end()) 429 return I->second; 430 431 // Make sure complete type info for the class is emitted *after* the member 432 // function type, as the complete class type is likely to reference this 433 // member function type. 434 TypeLoweringScope S(*this); 435 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 436 437 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 438 TypeIndex TI = lowerTypeMemberFunction( 439 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 440 return recordTypeIndexForDINode(SP, TI, Class); 441 } 442 443 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 444 TypeIndex TI, 445 const DIType *ClassTy) { 446 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 447 (void)InsertResult; 448 assert(InsertResult.second && "DINode was already assigned a type index"); 449 return TI; 450 } 451 452 unsigned CodeViewDebug::getPointerSizeInBytes() { 453 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 454 } 455 456 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 457 const LexicalScope *LS) { 458 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 459 // This variable was inlined. Associate it with the InlineSite. 460 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 461 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 462 Site.InlinedLocals.emplace_back(Var); 463 } else { 464 // This variable goes into the corresponding lexical scope. 465 ScopeVariables[LS].emplace_back(Var); 466 } 467 } 468 469 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 470 const DILocation *Loc) { 471 auto B = Locs.begin(), E = Locs.end(); 472 if (std::find(B, E, Loc) == E) 473 Locs.push_back(Loc); 474 } 475 476 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 477 const MachineFunction *MF) { 478 // Skip this instruction if it has the same location as the previous one. 479 if (!DL || DL == PrevInstLoc) 480 return; 481 482 const DIScope *Scope = DL.get()->getScope(); 483 if (!Scope) 484 return; 485 486 // Skip this line if it is longer than the maximum we can record. 487 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 488 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 489 LI.isNeverStepInto()) 490 return; 491 492 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 493 if (CI.getStartColumn() != DL.getCol()) 494 return; 495 496 if (!CurFn->HaveLineInfo) 497 CurFn->HaveLineInfo = true; 498 unsigned FileId = 0; 499 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 500 FileId = CurFn->LastFileId; 501 else 502 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 503 PrevInstLoc = DL; 504 505 unsigned FuncId = CurFn->FuncId; 506 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 507 const DILocation *Loc = DL.get(); 508 509 // If this location was actually inlined from somewhere else, give it the ID 510 // of the inline call site. 511 FuncId = 512 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 513 514 // Ensure we have links in the tree of inline call sites. 515 bool FirstLoc = true; 516 while ((SiteLoc = Loc->getInlinedAt())) { 517 InlineSite &Site = 518 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 519 if (!FirstLoc) 520 addLocIfNotPresent(Site.ChildSites, Loc); 521 FirstLoc = false; 522 Loc = SiteLoc; 523 } 524 addLocIfNotPresent(CurFn->ChildSites, Loc); 525 } 526 527 OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 528 /*PrologueEnd=*/false, /*IsStmt=*/false, 529 DL->getFilename(), SMLoc()); 530 } 531 532 void CodeViewDebug::emitCodeViewMagicVersion() { 533 OS.EmitValueToAlignment(4); 534 OS.AddComment("Debug section magic"); 535 OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4); 536 } 537 538 void CodeViewDebug::endModule() { 539 if (!Asm || !MMI->hasDebugInfo()) 540 return; 541 542 assert(Asm != nullptr); 543 544 // The COFF .debug$S section consists of several subsections, each starting 545 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 546 // of the payload followed by the payload itself. The subsections are 4-byte 547 // aligned. 548 549 // Use the generic .debug$S section, and make a subsection for all the inlined 550 // subprograms. 551 switchToDebugSectionForSymbol(nullptr); 552 553 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 554 emitCompilerInformation(); 555 endCVSubsection(CompilerInfo); 556 557 emitInlineeLinesSubsection(); 558 559 // Emit per-function debug information. 560 for (auto &P : FnDebugInfo) 561 if (!P.first->isDeclarationForLinker()) 562 emitDebugInfoForFunction(P.first, *P.second); 563 564 // Emit global variable debug information. 565 setCurrentSubprogram(nullptr); 566 emitDebugInfoForGlobals(); 567 568 // Emit retained types. 569 emitDebugInfoForRetainedTypes(); 570 571 // Switch back to the generic .debug$S section after potentially processing 572 // comdat symbol sections. 573 switchToDebugSectionForSymbol(nullptr); 574 575 // Emit UDT records for any types used by global variables. 576 if (!GlobalUDTs.empty()) { 577 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 578 emitDebugInfoForUDTs(GlobalUDTs); 579 endCVSubsection(SymbolsEnd); 580 } 581 582 // This subsection holds a file index to offset in string table table. 583 OS.AddComment("File index to string table offset subsection"); 584 OS.EmitCVFileChecksumsDirective(); 585 586 // This subsection holds the string table. 587 OS.AddComment("String table"); 588 OS.EmitCVStringTableDirective(); 589 590 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 591 // subsection in the generic .debug$S section at the end. There is no 592 // particular reason for this ordering other than to match MSVC. 593 emitBuildInfo(); 594 595 // Emit type information and hashes last, so that any types we translate while 596 // emitting function info are included. 597 emitTypeInformation(); 598 599 if (EmitDebugGlobalHashes) 600 emitTypeGlobalHashes(); 601 602 clear(); 603 } 604 605 static void 606 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 607 unsigned MaxFixedRecordLength = 0xF00) { 608 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 609 // after a fixed length portion of the record. The fixed length portion should 610 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 611 // overall record size is less than the maximum allowed. 612 SmallString<32> NullTerminatedString( 613 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 614 NullTerminatedString.push_back('\0'); 615 OS.EmitBytes(NullTerminatedString); 616 } 617 618 void CodeViewDebug::emitTypeInformation() { 619 if (TypeTable.empty()) 620 return; 621 622 // Start the .debug$T or .debug$P section with 0x4. 623 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 624 emitCodeViewMagicVersion(); 625 626 SmallString<8> CommentPrefix; 627 if (OS.isVerboseAsm()) { 628 CommentPrefix += '\t'; 629 CommentPrefix += Asm->MAI->getCommentString(); 630 CommentPrefix += ' '; 631 } 632 633 TypeTableCollection Table(TypeTable.records()); 634 Optional<TypeIndex> B = Table.getFirst(); 635 while (B) { 636 // This will fail if the record data is invalid. 637 CVType Record = Table.getType(*B); 638 639 TypeVisitorCallbackPipeline Pipeline; 640 CVMCAdapter CVMCOS(OS); 641 TypeRecordMapping typeMapping(CVMCOS); 642 SmallString<512> CommentBlock; 643 raw_svector_ostream CommentOS(CommentBlock); 644 645 if (OS.isVerboseAsm()) { 646 // Emit a block comment describing the type record for readability. 647 ScopedPrinter SP(CommentOS); 648 SP.setPrefix(CommentPrefix); 649 TypeDumpVisitor TDV(Table, &SP, false); 650 Pipeline.addCallbackToPipeline(TDV); 651 } 652 Pipeline.addCallbackToPipeline(typeMapping); 653 654 auto RecordLen = Record.length(); 655 auto RecordKind = Record.kind(); 656 OS.EmitIntValue(RecordLen - 2, 2); 657 OS.EmitIntValue(RecordKind, sizeof(RecordKind)); 658 659 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 660 661 if (E) { 662 logAllUnhandledErrors(std::move(E), errs(), "error: "); 663 llvm_unreachable("produced malformed type record"); 664 } 665 666 if (OS.isVerboseAsm()) { 667 // emitRawComment will insert its own tab and comment string before 668 // the first line, so strip off our first one. It also prints its own 669 // newline. 670 OS.emitRawComment( 671 CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim()); 672 } 673 B = Table.getNext(*B); 674 } 675 } 676 677 void CodeViewDebug::emitTypeGlobalHashes() { 678 if (TypeTable.empty()) 679 return; 680 681 // Start the .debug$H section with the version and hash algorithm, currently 682 // hardcoded to version 0, SHA1. 683 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 684 685 OS.EmitValueToAlignment(4); 686 OS.AddComment("Magic"); 687 OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4); 688 OS.AddComment("Section Version"); 689 OS.EmitIntValue(0, 2); 690 OS.AddComment("Hash Algorithm"); 691 OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2); 692 693 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 694 for (const auto &GHR : TypeTable.hashes()) { 695 if (OS.isVerboseAsm()) { 696 // Emit an EOL-comment describing which TypeIndex this hash corresponds 697 // to, as well as the stringified SHA1 hash. 698 SmallString<32> Comment; 699 raw_svector_ostream CommentOS(Comment); 700 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 701 OS.AddComment(Comment); 702 ++TI; 703 } 704 assert(GHR.Hash.size() == 8); 705 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 706 GHR.Hash.size()); 707 OS.EmitBinaryData(S); 708 } 709 } 710 711 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 712 switch (DWLang) { 713 case dwarf::DW_LANG_C: 714 case dwarf::DW_LANG_C89: 715 case dwarf::DW_LANG_C99: 716 case dwarf::DW_LANG_C11: 717 case dwarf::DW_LANG_ObjC: 718 return SourceLanguage::C; 719 case dwarf::DW_LANG_C_plus_plus: 720 case dwarf::DW_LANG_C_plus_plus_03: 721 case dwarf::DW_LANG_C_plus_plus_11: 722 case dwarf::DW_LANG_C_plus_plus_14: 723 return SourceLanguage::Cpp; 724 case dwarf::DW_LANG_Fortran77: 725 case dwarf::DW_LANG_Fortran90: 726 case dwarf::DW_LANG_Fortran03: 727 case dwarf::DW_LANG_Fortran08: 728 return SourceLanguage::Fortran; 729 case dwarf::DW_LANG_Pascal83: 730 return SourceLanguage::Pascal; 731 case dwarf::DW_LANG_Cobol74: 732 case dwarf::DW_LANG_Cobol85: 733 return SourceLanguage::Cobol; 734 case dwarf::DW_LANG_Java: 735 return SourceLanguage::Java; 736 case dwarf::DW_LANG_D: 737 return SourceLanguage::D; 738 case dwarf::DW_LANG_Swift: 739 return SourceLanguage::Swift; 740 default: 741 // There's no CodeView representation for this language, and CV doesn't 742 // have an "unknown" option for the language field, so we'll use MASM, 743 // as it's very low level. 744 return SourceLanguage::Masm; 745 } 746 } 747 748 namespace { 749 struct Version { 750 int Part[4]; 751 }; 752 } // end anonymous namespace 753 754 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 755 // the version number. 756 static Version parseVersion(StringRef Name) { 757 Version V = {{0}}; 758 int N = 0; 759 for (const char C : Name) { 760 if (isdigit(C)) { 761 V.Part[N] *= 10; 762 V.Part[N] += C - '0'; 763 } else if (C == '.') { 764 ++N; 765 if (N >= 4) 766 return V; 767 } else if (N > 0) 768 return V; 769 } 770 return V; 771 } 772 773 void CodeViewDebug::emitCompilerInformation() { 774 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 775 uint32_t Flags = 0; 776 777 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 778 const MDNode *Node = *CUs->operands().begin(); 779 const auto *CU = cast<DICompileUnit>(Node); 780 781 // The low byte of the flags indicates the source language. 782 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 783 // TODO: Figure out which other flags need to be set. 784 785 OS.AddComment("Flags and language"); 786 OS.EmitIntValue(Flags, 4); 787 788 OS.AddComment("CPUType"); 789 OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2); 790 791 StringRef CompilerVersion = CU->getProducer(); 792 Version FrontVer = parseVersion(CompilerVersion); 793 OS.AddComment("Frontend version"); 794 for (int N = 0; N < 4; ++N) 795 OS.EmitIntValue(FrontVer.Part[N], 2); 796 797 // Some Microsoft tools, like Binscope, expect a backend version number of at 798 // least 8.something, so we'll coerce the LLVM version into a form that 799 // guarantees it'll be big enough without really lying about the version. 800 int Major = 1000 * LLVM_VERSION_MAJOR + 801 10 * LLVM_VERSION_MINOR + 802 LLVM_VERSION_PATCH; 803 // Clamp it for builds that use unusually large version numbers. 804 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 805 Version BackVer = {{ Major, 0, 0, 0 }}; 806 OS.AddComment("Backend version"); 807 for (int N = 0; N < 4; ++N) 808 OS.EmitIntValue(BackVer.Part[N], 2); 809 810 OS.AddComment("Null-terminated compiler version string"); 811 emitNullTerminatedSymbolName(OS, CompilerVersion); 812 813 endSymbolRecord(CompilerEnd); 814 } 815 816 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 817 StringRef S) { 818 StringIdRecord SIR(TypeIndex(0x0), S); 819 return TypeTable.writeLeafType(SIR); 820 } 821 822 void CodeViewDebug::emitBuildInfo() { 823 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 824 // build info. The known prefix is: 825 // - Absolute path of current directory 826 // - Compiler path 827 // - Main source file path, relative to CWD or absolute 828 // - Type server PDB file 829 // - Canonical compiler command line 830 // If frontend and backend compilation are separated (think llc or LTO), it's 831 // not clear if the compiler path should refer to the executable for the 832 // frontend or the backend. Leave it blank for now. 833 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 834 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 835 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 836 const auto *CU = cast<DICompileUnit>(Node); 837 const DIFile *MainSourceFile = CU->getFile(); 838 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 839 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 840 BuildInfoArgs[BuildInfoRecord::SourceFile] = 841 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 842 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 843 // we implement /Zi type servers. 844 BuildInfoRecord BIR(BuildInfoArgs); 845 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 846 847 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 848 // from the module symbols into the type stream. 849 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 850 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 851 OS.AddComment("LF_BUILDINFO index"); 852 OS.EmitIntValue(BuildInfoIndex.getIndex(), 4); 853 endSymbolRecord(BIEnd); 854 endCVSubsection(BISubsecEnd); 855 } 856 857 void CodeViewDebug::emitInlineeLinesSubsection() { 858 if (InlinedSubprograms.empty()) 859 return; 860 861 OS.AddComment("Inlinee lines subsection"); 862 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 863 864 // We emit the checksum info for files. This is used by debuggers to 865 // determine if a pdb matches the source before loading it. Visual Studio, 866 // for instance, will display a warning that the breakpoints are not valid if 867 // the pdb does not match the source. 868 OS.AddComment("Inlinee lines signature"); 869 OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4); 870 871 for (const DISubprogram *SP : InlinedSubprograms) { 872 assert(TypeIndices.count({SP, nullptr})); 873 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 874 875 OS.AddBlankLine(); 876 unsigned FileId = maybeRecordFile(SP->getFile()); 877 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 878 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 879 OS.AddBlankLine(); 880 OS.AddComment("Type index of inlined function"); 881 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 882 OS.AddComment("Offset into filechecksum table"); 883 OS.EmitCVFileChecksumOffsetDirective(FileId); 884 OS.AddComment("Starting line number"); 885 OS.EmitIntValue(SP->getLine(), 4); 886 } 887 888 endCVSubsection(InlineEnd); 889 } 890 891 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 892 const DILocation *InlinedAt, 893 const InlineSite &Site) { 894 assert(TypeIndices.count({Site.Inlinee, nullptr})); 895 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 896 897 // SymbolRecord 898 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 899 900 OS.AddComment("PtrParent"); 901 OS.EmitIntValue(0, 4); 902 OS.AddComment("PtrEnd"); 903 OS.EmitIntValue(0, 4); 904 OS.AddComment("Inlinee type index"); 905 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 906 907 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 908 unsigned StartLineNum = Site.Inlinee->getLine(); 909 910 OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 911 FI.Begin, FI.End); 912 913 endSymbolRecord(InlineEnd); 914 915 emitLocalVariableList(FI, Site.InlinedLocals); 916 917 // Recurse on child inlined call sites before closing the scope. 918 for (const DILocation *ChildSite : Site.ChildSites) { 919 auto I = FI.InlineSites.find(ChildSite); 920 assert(I != FI.InlineSites.end() && 921 "child site not in function inline site map"); 922 emitInlinedCallSite(FI, ChildSite, I->second); 923 } 924 925 // Close the scope. 926 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 927 } 928 929 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 930 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 931 // comdat key. A section may be comdat because of -ffunction-sections or 932 // because it is comdat in the IR. 933 MCSectionCOFF *GVSec = 934 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 935 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 936 937 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 938 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 939 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 940 941 OS.SwitchSection(DebugSec); 942 943 // Emit the magic version number if this is the first time we've switched to 944 // this section. 945 if (ComdatDebugSections.insert(DebugSec).second) 946 emitCodeViewMagicVersion(); 947 } 948 949 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 950 // The only supported thunk ordinal is currently the standard type. 951 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 952 FunctionInfo &FI, 953 const MCSymbol *Fn) { 954 std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 955 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 956 957 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 958 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 959 960 // Emit S_THUNK32 961 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 962 OS.AddComment("PtrParent"); 963 OS.EmitIntValue(0, 4); 964 OS.AddComment("PtrEnd"); 965 OS.EmitIntValue(0, 4); 966 OS.AddComment("PtrNext"); 967 OS.EmitIntValue(0, 4); 968 OS.AddComment("Thunk section relative address"); 969 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 970 OS.AddComment("Thunk section index"); 971 OS.EmitCOFFSectionIndex(Fn); 972 OS.AddComment("Code size"); 973 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 974 OS.AddComment("Ordinal"); 975 OS.EmitIntValue(unsigned(ordinal), 1); 976 OS.AddComment("Function name"); 977 emitNullTerminatedSymbolName(OS, FuncName); 978 // Additional fields specific to the thunk ordinal would go here. 979 endSymbolRecord(ThunkRecordEnd); 980 981 // Local variables/inlined routines are purposely omitted here. The point of 982 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 983 984 // Emit S_PROC_ID_END 985 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 986 987 endCVSubsection(SymbolsEnd); 988 } 989 990 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 991 FunctionInfo &FI) { 992 // For each function there is a separate subsection which holds the PC to 993 // file:line table. 994 const MCSymbol *Fn = Asm->getSymbol(GV); 995 assert(Fn); 996 997 // Switch to the to a comdat section, if appropriate. 998 switchToDebugSectionForSymbol(Fn); 999 1000 std::string FuncName; 1001 auto *SP = GV->getSubprogram(); 1002 assert(SP); 1003 setCurrentSubprogram(SP); 1004 1005 if (SP->isThunk()) { 1006 emitDebugInfoForThunk(GV, FI, Fn); 1007 return; 1008 } 1009 1010 // If we have a display name, build the fully qualified name by walking the 1011 // chain of scopes. 1012 if (!SP->getName().empty()) 1013 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1014 1015 // If our DISubprogram name is empty, use the mangled name. 1016 if (FuncName.empty()) 1017 FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 1018 1019 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1020 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1021 OS.EmitCVFPOData(Fn); 1022 1023 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1024 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1025 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1026 { 1027 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1028 : SymbolKind::S_GPROC32_ID; 1029 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 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 endSymbolRecord(ProcRecordEnd); 1059 1060 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1061 // Subtract out the CSR size since MSVC excludes that and we include it. 1062 OS.AddComment("FrameSize"); 1063 OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4); 1064 OS.AddComment("Padding"); 1065 OS.EmitIntValue(0, 4); 1066 OS.AddComment("Offset of padding"); 1067 OS.EmitIntValue(0, 4); 1068 OS.AddComment("Bytes of callee saved registers"); 1069 OS.EmitIntValue(FI.CSRSize, 4); 1070 OS.AddComment("Exception handler offset"); 1071 OS.EmitIntValue(0, 4); 1072 OS.AddComment("Exception handler section"); 1073 OS.EmitIntValue(0, 2); 1074 OS.AddComment("Flags (defines frame register)"); 1075 OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4); 1076 endSymbolRecord(FrameProcEnd); 1077 1078 emitLocalVariableList(FI, FI.Locals); 1079 emitGlobalVariableList(FI.Globals); 1080 emitLexicalBlockList(FI.ChildBlocks, FI); 1081 1082 // Emit inlined call site information. Only emit functions inlined directly 1083 // into the parent function. We'll emit the other sites recursively as part 1084 // of their parent inline site. 1085 for (const DILocation *InlinedAt : FI.ChildSites) { 1086 auto I = FI.InlineSites.find(InlinedAt); 1087 assert(I != FI.InlineSites.end() && 1088 "child site not in function inline site map"); 1089 emitInlinedCallSite(FI, InlinedAt, I->second); 1090 } 1091 1092 for (auto Annot : FI.Annotations) { 1093 MCSymbol *Label = Annot.first; 1094 MDTuple *Strs = cast<MDTuple>(Annot.second); 1095 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1096 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1097 // FIXME: Make sure we don't overflow the max record size. 1098 OS.EmitCOFFSectionIndex(Label); 1099 OS.EmitIntValue(Strs->getNumOperands(), 2); 1100 for (Metadata *MD : Strs->operands()) { 1101 // MDStrings are null terminated, so we can do EmitBytes and get the 1102 // nice .asciz directive. 1103 StringRef Str = cast<MDString>(MD)->getString(); 1104 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1105 OS.EmitBytes(StringRef(Str.data(), Str.size() + 1)); 1106 } 1107 endSymbolRecord(AnnotEnd); 1108 } 1109 1110 for (auto HeapAllocSite : FI.HeapAllocSites) { 1111 MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1112 MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1113 1114 // The labels might not be defined if the instruction was replaced 1115 // somewhere in the codegen pipeline. 1116 if (!BeginLabel->isDefined() || !EndLabel->isDefined()) 1117 continue; 1118 1119 DIType *DITy = std::get<2>(HeapAllocSite); 1120 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1121 OS.AddComment("Call site offset"); 1122 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1123 OS.AddComment("Call site section index"); 1124 OS.EmitCOFFSectionIndex(BeginLabel); 1125 OS.AddComment("Call instruction length"); 1126 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1127 OS.AddComment("Type index"); 1128 OS.EmitIntValue(getCompleteTypeIndex(DITy).getIndex(), 4); 1129 endSymbolRecord(HeapAllocEnd); 1130 } 1131 1132 if (SP != nullptr) 1133 emitDebugInfoForUDTs(LocalUDTs); 1134 1135 // We're done with this function. 1136 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1137 } 1138 endCVSubsection(SymbolsEnd); 1139 1140 // We have an assembler directive that takes care of the whole line table. 1141 OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1142 } 1143 1144 CodeViewDebug::LocalVarDefRange 1145 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1146 LocalVarDefRange DR; 1147 DR.InMemory = -1; 1148 DR.DataOffset = Offset; 1149 assert(DR.DataOffset == Offset && "truncation"); 1150 DR.IsSubfield = 0; 1151 DR.StructOffset = 0; 1152 DR.CVRegister = CVRegister; 1153 return DR; 1154 } 1155 1156 void CodeViewDebug::collectVariableInfoFromMFTable( 1157 DenseSet<InlinedEntity> &Processed) { 1158 const MachineFunction &MF = *Asm->MF; 1159 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1160 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1161 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1162 1163 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1164 if (!VI.Var) 1165 continue; 1166 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1167 "Expected inlined-at fields to agree"); 1168 1169 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1170 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1171 1172 // If variable scope is not found then skip this variable. 1173 if (!Scope) 1174 continue; 1175 1176 // If the variable has an attached offset expression, extract it. 1177 // FIXME: Try to handle DW_OP_deref as well. 1178 int64_t ExprOffset = 0; 1179 bool Deref = false; 1180 if (VI.Expr) { 1181 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1182 if (VI.Expr->getNumElements() == 1 && 1183 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1184 Deref = true; 1185 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1186 continue; 1187 } 1188 1189 // Get the frame register used and the offset. 1190 unsigned FrameReg = 0; 1191 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1192 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1193 1194 // Calculate the label ranges. 1195 LocalVarDefRange DefRange = 1196 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1197 1198 for (const InsnRange &Range : Scope->getRanges()) { 1199 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1200 const MCSymbol *End = getLabelAfterInsn(Range.second); 1201 End = End ? End : Asm->getFunctionEnd(); 1202 DefRange.Ranges.emplace_back(Begin, End); 1203 } 1204 1205 LocalVariable Var; 1206 Var.DIVar = VI.Var; 1207 Var.DefRanges.emplace_back(std::move(DefRange)); 1208 if (Deref) 1209 Var.UseReferenceType = true; 1210 1211 recordLocalVariable(std::move(Var), Scope); 1212 } 1213 } 1214 1215 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1216 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1217 } 1218 1219 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1220 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1221 } 1222 1223 void CodeViewDebug::calculateRanges( 1224 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1225 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1226 1227 // Calculate the definition ranges. 1228 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1229 const auto &Entry = *I; 1230 if (!Entry.isDbgValue()) 1231 continue; 1232 const MachineInstr *DVInst = Entry.getInstr(); 1233 assert(DVInst->isDebugValue() && "Invalid History entry"); 1234 // FIXME: Find a way to represent constant variables, since they are 1235 // relatively common. 1236 Optional<DbgVariableLocation> Location = 1237 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1238 if (!Location) 1239 continue; 1240 1241 // CodeView can only express variables in register and variables in memory 1242 // at a constant offset from a register. However, for variables passed 1243 // indirectly by pointer, it is common for that pointer to be spilled to a 1244 // stack location. For the special case of one offseted load followed by a 1245 // zero offset load (a pointer spilled to the stack), we change the type of 1246 // the local variable from a value type to a reference type. This tricks the 1247 // debugger into doing the load for us. 1248 if (Var.UseReferenceType) { 1249 // We're using a reference type. Drop the last zero offset load. 1250 if (canUseReferenceType(*Location)) 1251 Location->LoadChain.pop_back(); 1252 else 1253 continue; 1254 } else if (needsReferenceType(*Location)) { 1255 // This location can't be expressed without switching to a reference type. 1256 // Start over using that. 1257 Var.UseReferenceType = true; 1258 Var.DefRanges.clear(); 1259 calculateRanges(Var, Entries); 1260 return; 1261 } 1262 1263 // We can only handle a register or an offseted load of a register. 1264 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1265 continue; 1266 { 1267 LocalVarDefRange DR; 1268 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1269 DR.InMemory = !Location->LoadChain.empty(); 1270 DR.DataOffset = 1271 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1272 if (Location->FragmentInfo) { 1273 DR.IsSubfield = true; 1274 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1275 } else { 1276 DR.IsSubfield = false; 1277 DR.StructOffset = 0; 1278 } 1279 1280 if (Var.DefRanges.empty() || 1281 Var.DefRanges.back().isDifferentLocation(DR)) { 1282 Var.DefRanges.emplace_back(std::move(DR)); 1283 } 1284 } 1285 1286 // Compute the label range. 1287 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1288 const MCSymbol *End; 1289 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1290 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1291 End = EndingEntry.isDbgValue() 1292 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1293 : getLabelAfterInsn(EndingEntry.getInstr()); 1294 } else 1295 End = Asm->getFunctionEnd(); 1296 1297 // If the last range end is our begin, just extend the last range. 1298 // Otherwise make a new range. 1299 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1300 Var.DefRanges.back().Ranges; 1301 if (!R.empty() && R.back().second == Begin) 1302 R.back().second = End; 1303 else 1304 R.emplace_back(Begin, End); 1305 1306 // FIXME: Do more range combining. 1307 } 1308 } 1309 1310 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1311 DenseSet<InlinedEntity> Processed; 1312 // Grab the variable info that was squirreled away in the MMI side-table. 1313 collectVariableInfoFromMFTable(Processed); 1314 1315 for (const auto &I : DbgValues) { 1316 InlinedEntity IV = I.first; 1317 if (Processed.count(IV)) 1318 continue; 1319 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1320 const DILocation *InlinedAt = IV.second; 1321 1322 // Instruction ranges, specifying where IV is accessible. 1323 const auto &Entries = I.second; 1324 1325 LexicalScope *Scope = nullptr; 1326 if (InlinedAt) 1327 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1328 else 1329 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1330 // If variable scope is not found then skip this variable. 1331 if (!Scope) 1332 continue; 1333 1334 LocalVariable Var; 1335 Var.DIVar = DIVar; 1336 1337 calculateRanges(Var, Entries); 1338 recordLocalVariable(std::move(Var), Scope); 1339 } 1340 } 1341 1342 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1343 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1344 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1345 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1346 const Function &GV = MF->getFunction(); 1347 auto Insertion = FnDebugInfo.insert({&GV, llvm::make_unique<FunctionInfo>()}); 1348 assert(Insertion.second && "function already has info"); 1349 CurFn = Insertion.first->second.get(); 1350 CurFn->FuncId = NextFuncId++; 1351 CurFn->Begin = Asm->getFunctionBegin(); 1352 1353 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1354 // callee-saved registers were used. For targets that don't use a PUSH 1355 // instruction (AArch64), this will be zero. 1356 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1357 CurFn->FrameSize = MFI.getStackSize(); 1358 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1359 CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF); 1360 1361 // For this function S_FRAMEPROC record, figure out which codeview register 1362 // will be the frame pointer. 1363 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1364 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1365 if (CurFn->FrameSize > 0) { 1366 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1367 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1368 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1369 } else { 1370 // If there is an FP, parameters are always relative to it. 1371 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1372 if (CurFn->HasStackRealignment) { 1373 // If the stack needs realignment, locals are relative to SP or VFRAME. 1374 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1375 } else { 1376 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1377 // other stack adjustments. 1378 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1379 } 1380 } 1381 } 1382 1383 // Compute other frame procedure options. 1384 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1385 if (MFI.hasVarSizedObjects()) 1386 FPO |= FrameProcedureOptions::HasAlloca; 1387 if (MF->exposesReturnsTwice()) 1388 FPO |= FrameProcedureOptions::HasSetJmp; 1389 // FIXME: Set HasLongJmp if we ever track that info. 1390 if (MF->hasInlineAsm()) 1391 FPO |= FrameProcedureOptions::HasInlineAssembly; 1392 if (GV.hasPersonalityFn()) { 1393 if (isAsynchronousEHPersonality( 1394 classifyEHPersonality(GV.getPersonalityFn()))) 1395 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1396 else 1397 FPO |= FrameProcedureOptions::HasExceptionHandling; 1398 } 1399 if (GV.hasFnAttribute(Attribute::InlineHint)) 1400 FPO |= FrameProcedureOptions::MarkedInline; 1401 if (GV.hasFnAttribute(Attribute::Naked)) 1402 FPO |= FrameProcedureOptions::Naked; 1403 if (MFI.hasStackProtectorIndex()) 1404 FPO |= FrameProcedureOptions::SecurityChecks; 1405 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1406 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1407 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1408 !GV.hasOptSize() && !GV.hasOptNone()) 1409 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1410 // FIXME: Set GuardCfg when it is implemented. 1411 CurFn->FrameProcOpts = FPO; 1412 1413 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1414 1415 // Find the end of the function prolog. First known non-DBG_VALUE and 1416 // non-frame setup location marks the beginning of the function body. 1417 // FIXME: is there a simpler a way to do this? Can we just search 1418 // for the first instruction of the function, not the last of the prolog? 1419 DebugLoc PrologEndLoc; 1420 bool EmptyPrologue = true; 1421 for (const auto &MBB : *MF) { 1422 for (const auto &MI : MBB) { 1423 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1424 MI.getDebugLoc()) { 1425 PrologEndLoc = MI.getDebugLoc(); 1426 break; 1427 } else if (!MI.isMetaInstruction()) { 1428 EmptyPrologue = false; 1429 } 1430 } 1431 } 1432 1433 // Record beginning of function if we have a non-empty prologue. 1434 if (PrologEndLoc && !EmptyPrologue) { 1435 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1436 maybeRecordLocation(FnStartDL, MF); 1437 } 1438 } 1439 1440 static bool shouldEmitUdt(const DIType *T) { 1441 if (!T) 1442 return false; 1443 1444 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1445 if (T->getTag() == dwarf::DW_TAG_typedef) { 1446 if (DIScope *Scope = T->getScope()) { 1447 switch (Scope->getTag()) { 1448 case dwarf::DW_TAG_structure_type: 1449 case dwarf::DW_TAG_class_type: 1450 case dwarf::DW_TAG_union_type: 1451 return false; 1452 } 1453 } 1454 } 1455 1456 while (true) { 1457 if (!T || T->isForwardDecl()) 1458 return false; 1459 1460 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1461 if (!DT) 1462 return true; 1463 T = DT->getBaseType(); 1464 } 1465 return true; 1466 } 1467 1468 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1469 // Don't record empty UDTs. 1470 if (Ty->getName().empty()) 1471 return; 1472 if (!shouldEmitUdt(Ty)) 1473 return; 1474 1475 SmallVector<StringRef, 5> QualifiedNameComponents; 1476 const DISubprogram *ClosestSubprogram = 1477 getQualifiedNameComponents(Ty->getScope(), QualifiedNameComponents); 1478 1479 std::string FullyQualifiedName = 1480 getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty)); 1481 1482 if (ClosestSubprogram == nullptr) { 1483 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1484 } else if (ClosestSubprogram == CurrentSubprogram) { 1485 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1486 } 1487 1488 // TODO: What if the ClosestSubprogram is neither null or the current 1489 // subprogram? Currently, the UDT just gets dropped on the floor. 1490 // 1491 // The current behavior is not desirable. To get maximal fidelity, we would 1492 // need to perform all type translation before beginning emission of .debug$S 1493 // and then make LocalUDTs a member of FunctionInfo 1494 } 1495 1496 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1497 // Generic dispatch for lowering an unknown type. 1498 switch (Ty->getTag()) { 1499 case dwarf::DW_TAG_array_type: 1500 return lowerTypeArray(cast<DICompositeType>(Ty)); 1501 case dwarf::DW_TAG_typedef: 1502 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1503 case dwarf::DW_TAG_base_type: 1504 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1505 case dwarf::DW_TAG_pointer_type: 1506 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1507 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1508 LLVM_FALLTHROUGH; 1509 case dwarf::DW_TAG_reference_type: 1510 case dwarf::DW_TAG_rvalue_reference_type: 1511 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1512 case dwarf::DW_TAG_ptr_to_member_type: 1513 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1514 case dwarf::DW_TAG_restrict_type: 1515 case dwarf::DW_TAG_const_type: 1516 case dwarf::DW_TAG_volatile_type: 1517 // TODO: add support for DW_TAG_atomic_type here 1518 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1519 case dwarf::DW_TAG_subroutine_type: 1520 if (ClassTy) { 1521 // The member function type of a member function pointer has no 1522 // ThisAdjustment. 1523 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1524 /*ThisAdjustment=*/0, 1525 /*IsStaticMethod=*/false); 1526 } 1527 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1528 case dwarf::DW_TAG_enumeration_type: 1529 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1530 case dwarf::DW_TAG_class_type: 1531 case dwarf::DW_TAG_structure_type: 1532 return lowerTypeClass(cast<DICompositeType>(Ty)); 1533 case dwarf::DW_TAG_union_type: 1534 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1535 case dwarf::DW_TAG_unspecified_type: 1536 if (Ty->getName() == "decltype(nullptr)") 1537 return TypeIndex::NullptrT(); 1538 return TypeIndex::None(); 1539 default: 1540 // Use the null type index. 1541 return TypeIndex(); 1542 } 1543 } 1544 1545 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1546 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1547 StringRef TypeName = Ty->getName(); 1548 1549 addToUDTs(Ty); 1550 1551 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1552 TypeName == "HRESULT") 1553 return TypeIndex(SimpleTypeKind::HResult); 1554 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1555 TypeName == "wchar_t") 1556 return TypeIndex(SimpleTypeKind::WideCharacter); 1557 1558 return UnderlyingTypeIndex; 1559 } 1560 1561 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1562 const DIType *ElementType = Ty->getBaseType(); 1563 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1564 // IndexType is size_t, which depends on the bitness of the target. 1565 TypeIndex IndexType = getPointerSizeInBytes() == 8 1566 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1567 : TypeIndex(SimpleTypeKind::UInt32Long); 1568 1569 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1570 1571 // Add subranges to array type. 1572 DINodeArray Elements = Ty->getElements(); 1573 for (int i = Elements.size() - 1; i >= 0; --i) { 1574 const DINode *Element = Elements[i]; 1575 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1576 1577 const DISubrange *Subrange = cast<DISubrange>(Element); 1578 assert(Subrange->getLowerBound() == 0 && 1579 "codeview doesn't support subranges with lower bounds"); 1580 int64_t Count = -1; 1581 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>()) 1582 Count = CI->getSExtValue(); 1583 1584 // Forward declarations of arrays without a size and VLAs use a count of -1. 1585 // Emit a count of zero in these cases to match what MSVC does for arrays 1586 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1587 // should do for them even if we could distinguish them. 1588 if (Count == -1) 1589 Count = 0; 1590 1591 // Update the element size and element type index for subsequent subranges. 1592 ElementSize *= Count; 1593 1594 // If this is the outermost array, use the size from the array. It will be 1595 // more accurate if we had a VLA or an incomplete element type size. 1596 uint64_t ArraySize = 1597 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1598 1599 StringRef Name = (i == 0) ? Ty->getName() : ""; 1600 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1601 ElementTypeIndex = TypeTable.writeLeafType(AR); 1602 } 1603 1604 return ElementTypeIndex; 1605 } 1606 1607 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1608 TypeIndex Index; 1609 dwarf::TypeKind Kind; 1610 uint32_t ByteSize; 1611 1612 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1613 ByteSize = Ty->getSizeInBits() / 8; 1614 1615 SimpleTypeKind STK = SimpleTypeKind::None; 1616 switch (Kind) { 1617 case dwarf::DW_ATE_address: 1618 // FIXME: Translate 1619 break; 1620 case dwarf::DW_ATE_boolean: 1621 switch (ByteSize) { 1622 case 1: STK = SimpleTypeKind::Boolean8; break; 1623 case 2: STK = SimpleTypeKind::Boolean16; break; 1624 case 4: STK = SimpleTypeKind::Boolean32; break; 1625 case 8: STK = SimpleTypeKind::Boolean64; break; 1626 case 16: STK = SimpleTypeKind::Boolean128; break; 1627 } 1628 break; 1629 case dwarf::DW_ATE_complex_float: 1630 switch (ByteSize) { 1631 case 2: STK = SimpleTypeKind::Complex16; break; 1632 case 4: STK = SimpleTypeKind::Complex32; break; 1633 case 8: STK = SimpleTypeKind::Complex64; break; 1634 case 10: STK = SimpleTypeKind::Complex80; break; 1635 case 16: STK = SimpleTypeKind::Complex128; break; 1636 } 1637 break; 1638 case dwarf::DW_ATE_float: 1639 switch (ByteSize) { 1640 case 2: STK = SimpleTypeKind::Float16; break; 1641 case 4: STK = SimpleTypeKind::Float32; break; 1642 case 6: STK = SimpleTypeKind::Float48; break; 1643 case 8: STK = SimpleTypeKind::Float64; break; 1644 case 10: STK = SimpleTypeKind::Float80; break; 1645 case 16: STK = SimpleTypeKind::Float128; break; 1646 } 1647 break; 1648 case dwarf::DW_ATE_signed: 1649 switch (ByteSize) { 1650 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1651 case 2: STK = SimpleTypeKind::Int16Short; break; 1652 case 4: STK = SimpleTypeKind::Int32; break; 1653 case 8: STK = SimpleTypeKind::Int64Quad; break; 1654 case 16: STK = SimpleTypeKind::Int128Oct; break; 1655 } 1656 break; 1657 case dwarf::DW_ATE_unsigned: 1658 switch (ByteSize) { 1659 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1660 case 2: STK = SimpleTypeKind::UInt16Short; break; 1661 case 4: STK = SimpleTypeKind::UInt32; break; 1662 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1663 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1664 } 1665 break; 1666 case dwarf::DW_ATE_UTF: 1667 switch (ByteSize) { 1668 case 2: STK = SimpleTypeKind::Character16; break; 1669 case 4: STK = SimpleTypeKind::Character32; break; 1670 } 1671 break; 1672 case dwarf::DW_ATE_signed_char: 1673 if (ByteSize == 1) 1674 STK = SimpleTypeKind::SignedCharacter; 1675 break; 1676 case dwarf::DW_ATE_unsigned_char: 1677 if (ByteSize == 1) 1678 STK = SimpleTypeKind::UnsignedCharacter; 1679 break; 1680 default: 1681 break; 1682 } 1683 1684 // Apply some fixups based on the source-level type name. 1685 if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int") 1686 STK = SimpleTypeKind::Int32Long; 1687 if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int") 1688 STK = SimpleTypeKind::UInt32Long; 1689 if (STK == SimpleTypeKind::UInt16Short && 1690 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1691 STK = SimpleTypeKind::WideCharacter; 1692 if ((STK == SimpleTypeKind::SignedCharacter || 1693 STK == SimpleTypeKind::UnsignedCharacter) && 1694 Ty->getName() == "char") 1695 STK = SimpleTypeKind::NarrowCharacter; 1696 1697 return TypeIndex(STK); 1698 } 1699 1700 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1701 PointerOptions PO) { 1702 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1703 1704 // Pointers to simple types without any options can use SimpleTypeMode, rather 1705 // than having a dedicated pointer type record. 1706 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1707 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1708 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1709 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1710 ? SimpleTypeMode::NearPointer64 1711 : SimpleTypeMode::NearPointer32; 1712 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1713 } 1714 1715 PointerKind PK = 1716 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1717 PointerMode PM = PointerMode::Pointer; 1718 switch (Ty->getTag()) { 1719 default: llvm_unreachable("not a pointer tag type"); 1720 case dwarf::DW_TAG_pointer_type: 1721 PM = PointerMode::Pointer; 1722 break; 1723 case dwarf::DW_TAG_reference_type: 1724 PM = PointerMode::LValueReference; 1725 break; 1726 case dwarf::DW_TAG_rvalue_reference_type: 1727 PM = PointerMode::RValueReference; 1728 break; 1729 } 1730 1731 if (Ty->isObjectPointer()) 1732 PO |= PointerOptions::Const; 1733 1734 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1735 return TypeTable.writeLeafType(PR); 1736 } 1737 1738 static PointerToMemberRepresentation 1739 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1740 // SizeInBytes being zero generally implies that the member pointer type was 1741 // incomplete, which can happen if it is part of a function prototype. In this 1742 // case, use the unknown model instead of the general model. 1743 if (IsPMF) { 1744 switch (Flags & DINode::FlagPtrToMemberRep) { 1745 case 0: 1746 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1747 : PointerToMemberRepresentation::GeneralFunction; 1748 case DINode::FlagSingleInheritance: 1749 return PointerToMemberRepresentation::SingleInheritanceFunction; 1750 case DINode::FlagMultipleInheritance: 1751 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1752 case DINode::FlagVirtualInheritance: 1753 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1754 } 1755 } else { 1756 switch (Flags & DINode::FlagPtrToMemberRep) { 1757 case 0: 1758 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1759 : PointerToMemberRepresentation::GeneralData; 1760 case DINode::FlagSingleInheritance: 1761 return PointerToMemberRepresentation::SingleInheritanceData; 1762 case DINode::FlagMultipleInheritance: 1763 return PointerToMemberRepresentation::MultipleInheritanceData; 1764 case DINode::FlagVirtualInheritance: 1765 return PointerToMemberRepresentation::VirtualInheritanceData; 1766 } 1767 } 1768 llvm_unreachable("invalid ptr to member representation"); 1769 } 1770 1771 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1772 PointerOptions PO) { 1773 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1774 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1775 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType()); 1776 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1777 : PointerKind::Near32; 1778 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1779 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1780 : PointerMode::PointerToDataMember; 1781 1782 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1783 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1784 MemberPointerInfo MPI( 1785 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1786 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1787 return TypeTable.writeLeafType(PR); 1788 } 1789 1790 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1791 /// have a translation, use the NearC convention. 1792 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1793 switch (DwarfCC) { 1794 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1795 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1796 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1797 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1798 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1799 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1800 } 1801 return CallingConvention::NearC; 1802 } 1803 1804 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1805 ModifierOptions Mods = ModifierOptions::None; 1806 PointerOptions PO = PointerOptions::None; 1807 bool IsModifier = true; 1808 const DIType *BaseTy = Ty; 1809 while (IsModifier && BaseTy) { 1810 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1811 switch (BaseTy->getTag()) { 1812 case dwarf::DW_TAG_const_type: 1813 Mods |= ModifierOptions::Const; 1814 PO |= PointerOptions::Const; 1815 break; 1816 case dwarf::DW_TAG_volatile_type: 1817 Mods |= ModifierOptions::Volatile; 1818 PO |= PointerOptions::Volatile; 1819 break; 1820 case dwarf::DW_TAG_restrict_type: 1821 // Only pointer types be marked with __restrict. There is no known flag 1822 // for __restrict in LF_MODIFIER records. 1823 PO |= PointerOptions::Restrict; 1824 break; 1825 default: 1826 IsModifier = false; 1827 break; 1828 } 1829 if (IsModifier) 1830 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1831 } 1832 1833 // Check if the inner type will use an LF_POINTER record. If so, the 1834 // qualifiers will go in the LF_POINTER record. This comes up for types like 1835 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1836 // char *'. 1837 if (BaseTy) { 1838 switch (BaseTy->getTag()) { 1839 case dwarf::DW_TAG_pointer_type: 1840 case dwarf::DW_TAG_reference_type: 1841 case dwarf::DW_TAG_rvalue_reference_type: 1842 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1843 case dwarf::DW_TAG_ptr_to_member_type: 1844 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1845 default: 1846 break; 1847 } 1848 } 1849 1850 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1851 1852 // Return the base type index if there aren't any modifiers. For example, the 1853 // metadata could contain restrict wrappers around non-pointer types. 1854 if (Mods == ModifierOptions::None) 1855 return ModifiedTI; 1856 1857 ModifierRecord MR(ModifiedTI, Mods); 1858 return TypeTable.writeLeafType(MR); 1859 } 1860 1861 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1862 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1863 for (const DIType *ArgType : Ty->getTypeArray()) 1864 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1865 1866 // MSVC uses type none for variadic argument. 1867 if (ReturnAndArgTypeIndices.size() > 1 && 1868 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1869 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1870 } 1871 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1872 ArrayRef<TypeIndex> ArgTypeIndices = None; 1873 if (!ReturnAndArgTypeIndices.empty()) { 1874 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1875 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1876 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1877 } 1878 1879 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1880 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1881 1882 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1883 1884 FunctionOptions FO = getFunctionOptions(Ty); 1885 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 1886 ArgListIndex); 1887 return TypeTable.writeLeafType(Procedure); 1888 } 1889 1890 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 1891 const DIType *ClassTy, 1892 int ThisAdjustment, 1893 bool IsStaticMethod, 1894 FunctionOptions FO) { 1895 // Lower the containing class type. 1896 TypeIndex ClassType = getTypeIndex(ClassTy); 1897 1898 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 1899 1900 unsigned Index = 0; 1901 SmallVector<TypeIndex, 8> ArgTypeIndices; 1902 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1903 if (ReturnAndArgs.size() > Index) { 1904 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 1905 } 1906 1907 // If the first argument is a pointer type and this isn't a static method, 1908 // treat it as the special 'this' parameter, which is encoded separately from 1909 // the arguments. 1910 TypeIndex ThisTypeIndex; 1911 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 1912 if (const DIDerivedType *PtrTy = 1913 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 1914 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 1915 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 1916 Index++; 1917 } 1918 } 1919 } 1920 1921 while (Index < ReturnAndArgs.size()) 1922 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 1923 1924 // MSVC uses type none for variadic argument. 1925 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 1926 ArgTypeIndices.back() = TypeIndex::None(); 1927 1928 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1929 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1930 1931 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1932 1933 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1934 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1935 return TypeTable.writeLeafType(MFR); 1936 } 1937 1938 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1939 unsigned VSlotCount = 1940 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1941 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1942 1943 VFTableShapeRecord VFTSR(Slots); 1944 return TypeTable.writeLeafType(VFTSR); 1945 } 1946 1947 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1948 switch (Flags & DINode::FlagAccessibility) { 1949 case DINode::FlagPrivate: return MemberAccess::Private; 1950 case DINode::FlagPublic: return MemberAccess::Public; 1951 case DINode::FlagProtected: return MemberAccess::Protected; 1952 case 0: 1953 // If there was no explicit access control, provide the default for the tag. 1954 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1955 : MemberAccess::Public; 1956 } 1957 llvm_unreachable("access flags are exclusive"); 1958 } 1959 1960 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1961 if (SP->isArtificial()) 1962 return MethodOptions::CompilerGenerated; 1963 1964 // FIXME: Handle other MethodOptions. 1965 1966 return MethodOptions::None; 1967 } 1968 1969 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 1970 bool Introduced) { 1971 if (SP->getFlags() & DINode::FlagStaticMember) 1972 return MethodKind::Static; 1973 1974 switch (SP->getVirtuality()) { 1975 case dwarf::DW_VIRTUALITY_none: 1976 break; 1977 case dwarf::DW_VIRTUALITY_virtual: 1978 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 1979 case dwarf::DW_VIRTUALITY_pure_virtual: 1980 return Introduced ? MethodKind::PureIntroducingVirtual 1981 : MethodKind::PureVirtual; 1982 default: 1983 llvm_unreachable("unhandled virtuality case"); 1984 } 1985 1986 return MethodKind::Vanilla; 1987 } 1988 1989 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 1990 switch (Ty->getTag()) { 1991 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 1992 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 1993 } 1994 llvm_unreachable("unexpected tag"); 1995 } 1996 1997 /// Return ClassOptions that should be present on both the forward declaration 1998 /// and the defintion of a tag type. 1999 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2000 ClassOptions CO = ClassOptions::None; 2001 2002 // MSVC always sets this flag, even for local types. Clang doesn't always 2003 // appear to give every type a linkage name, which may be problematic for us. 2004 // FIXME: Investigate the consequences of not following them here. 2005 if (!Ty->getIdentifier().empty()) 2006 CO |= ClassOptions::HasUniqueName; 2007 2008 // Put the Nested flag on a type if it appears immediately inside a tag type. 2009 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2010 // here. That flag is only set on definitions, and not forward declarations. 2011 const DIScope *ImmediateScope = Ty->getScope(); 2012 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2013 CO |= ClassOptions::Nested; 2014 2015 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2016 // type only when it has an immediate function scope. Clang never puts enums 2017 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2018 // always in function, class, or file scopes. 2019 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2020 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2021 CO |= ClassOptions::Scoped; 2022 } else { 2023 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2024 Scope = Scope->getScope()) { 2025 if (isa<DISubprogram>(Scope)) { 2026 CO |= ClassOptions::Scoped; 2027 break; 2028 } 2029 } 2030 } 2031 2032 return CO; 2033 } 2034 2035 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2036 switch (Ty->getTag()) { 2037 case dwarf::DW_TAG_class_type: 2038 case dwarf::DW_TAG_structure_type: 2039 case dwarf::DW_TAG_union_type: 2040 case dwarf::DW_TAG_enumeration_type: 2041 break; 2042 default: 2043 return; 2044 } 2045 2046 if (const auto *File = Ty->getFile()) { 2047 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2048 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2049 2050 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2051 TypeTable.writeLeafType(USLR); 2052 } 2053 } 2054 2055 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2056 ClassOptions CO = getCommonClassOptions(Ty); 2057 TypeIndex FTI; 2058 unsigned EnumeratorCount = 0; 2059 2060 if (Ty->isForwardDecl()) { 2061 CO |= ClassOptions::ForwardReference; 2062 } else { 2063 ContinuationRecordBuilder ContinuationBuilder; 2064 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2065 for (const DINode *Element : Ty->getElements()) { 2066 // We assume that the frontend provides all members in source declaration 2067 // order, which is what MSVC does. 2068 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2069 EnumeratorRecord ER(MemberAccess::Public, 2070 APSInt::getUnsigned(Enumerator->getValue()), 2071 Enumerator->getName()); 2072 ContinuationBuilder.writeMemberType(ER); 2073 EnumeratorCount++; 2074 } 2075 } 2076 FTI = TypeTable.insertRecord(ContinuationBuilder); 2077 } 2078 2079 std::string FullName = getFullyQualifiedName(Ty); 2080 2081 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2082 getTypeIndex(Ty->getBaseType())); 2083 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2084 2085 addUDTSrcLine(Ty, EnumTI); 2086 2087 return EnumTI; 2088 } 2089 2090 //===----------------------------------------------------------------------===// 2091 // ClassInfo 2092 //===----------------------------------------------------------------------===// 2093 2094 struct llvm::ClassInfo { 2095 struct MemberInfo { 2096 const DIDerivedType *MemberTypeNode; 2097 uint64_t BaseOffset; 2098 }; 2099 // [MemberInfo] 2100 using MemberList = std::vector<MemberInfo>; 2101 2102 using MethodsList = TinyPtrVector<const DISubprogram *>; 2103 // MethodName -> MethodsList 2104 using MethodsMap = MapVector<MDString *, MethodsList>; 2105 2106 /// Base classes. 2107 std::vector<const DIDerivedType *> Inheritance; 2108 2109 /// Direct members. 2110 MemberList Members; 2111 // Direct overloaded methods gathered by name. 2112 MethodsMap Methods; 2113 2114 TypeIndex VShapeTI; 2115 2116 std::vector<const DIType *> NestedTypes; 2117 }; 2118 2119 void CodeViewDebug::clear() { 2120 assert(CurFn == nullptr); 2121 FileIdMap.clear(); 2122 FnDebugInfo.clear(); 2123 FileToFilepathMap.clear(); 2124 LocalUDTs.clear(); 2125 GlobalUDTs.clear(); 2126 TypeIndices.clear(); 2127 CompleteTypeIndices.clear(); 2128 ScopeGlobals.clear(); 2129 } 2130 2131 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2132 const DIDerivedType *DDTy) { 2133 if (!DDTy->getName().empty()) { 2134 Info.Members.push_back({DDTy, 0}); 2135 return; 2136 } 2137 2138 // An unnamed member may represent a nested struct or union. Attempt to 2139 // interpret the unnamed member as a DICompositeType possibly wrapped in 2140 // qualifier types. Add all the indirect fields to the current record if that 2141 // succeeds, and drop the member if that fails. 2142 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2143 uint64_t Offset = DDTy->getOffsetInBits(); 2144 const DIType *Ty = DDTy->getBaseType(); 2145 bool FullyResolved = false; 2146 while (!FullyResolved) { 2147 switch (Ty->getTag()) { 2148 case dwarf::DW_TAG_const_type: 2149 case dwarf::DW_TAG_volatile_type: 2150 // FIXME: we should apply the qualifier types to the indirect fields 2151 // rather than dropping them. 2152 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2153 break; 2154 default: 2155 FullyResolved = true; 2156 break; 2157 } 2158 } 2159 2160 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2161 if (!DCTy) 2162 return; 2163 2164 ClassInfo NestedInfo = collectClassInfo(DCTy); 2165 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2166 Info.Members.push_back( 2167 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2168 } 2169 2170 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2171 ClassInfo Info; 2172 // Add elements to structure type. 2173 DINodeArray Elements = Ty->getElements(); 2174 for (auto *Element : Elements) { 2175 // We assume that the frontend provides all members in source declaration 2176 // order, which is what MSVC does. 2177 if (!Element) 2178 continue; 2179 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2180 Info.Methods[SP->getRawName()].push_back(SP); 2181 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2182 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2183 collectMemberInfo(Info, DDTy); 2184 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2185 Info.Inheritance.push_back(DDTy); 2186 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2187 DDTy->getName() == "__vtbl_ptr_type") { 2188 Info.VShapeTI = getTypeIndex(DDTy); 2189 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2190 Info.NestedTypes.push_back(DDTy); 2191 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2192 // Ignore friend members. It appears that MSVC emitted info about 2193 // friends in the past, but modern versions do not. 2194 } 2195 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2196 Info.NestedTypes.push_back(Composite); 2197 } 2198 // Skip other unrecognized kinds of elements. 2199 } 2200 return Info; 2201 } 2202 2203 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2204 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2205 // if a complete type should be emitted instead of a forward reference. 2206 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2207 !Ty->isForwardDecl(); 2208 } 2209 2210 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2211 // Emit the complete type for unnamed structs. C++ classes with methods 2212 // which have a circular reference back to the class type are expected to 2213 // be named by the front-end and should not be "unnamed". C unnamed 2214 // structs should not have circular references. 2215 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2216 // If this unnamed complete type is already in the process of being defined 2217 // then the description of the type is malformed and cannot be emitted 2218 // into CodeView correctly so report a fatal error. 2219 auto I = CompleteTypeIndices.find(Ty); 2220 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2221 report_fatal_error("cannot debug circular reference to unnamed type"); 2222 return getCompleteTypeIndex(Ty); 2223 } 2224 2225 // First, construct the forward decl. Don't look into Ty to compute the 2226 // forward decl options, since it might not be available in all TUs. 2227 TypeRecordKind Kind = getRecordKind(Ty); 2228 ClassOptions CO = 2229 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2230 std::string FullName = getFullyQualifiedName(Ty); 2231 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2232 FullName, Ty->getIdentifier()); 2233 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2234 if (!Ty->isForwardDecl()) 2235 DeferredCompleteTypes.push_back(Ty); 2236 return FwdDeclTI; 2237 } 2238 2239 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2240 // Construct the field list and complete type record. 2241 TypeRecordKind Kind = getRecordKind(Ty); 2242 ClassOptions CO = getCommonClassOptions(Ty); 2243 TypeIndex FieldTI; 2244 TypeIndex VShapeTI; 2245 unsigned FieldCount; 2246 bool ContainsNestedClass; 2247 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2248 lowerRecordFieldList(Ty); 2249 2250 if (ContainsNestedClass) 2251 CO |= ClassOptions::ContainsNestedClass; 2252 2253 // MSVC appears to set this flag by searching any destructor or method with 2254 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2255 // the members, however special member functions are not yet emitted into 2256 // debug information. For now checking a class's non-triviality seems enough. 2257 // FIXME: not true for a nested unnamed struct. 2258 if (isNonTrivial(Ty)) 2259 CO |= ClassOptions::HasConstructorOrDestructor; 2260 2261 std::string FullName = getFullyQualifiedName(Ty); 2262 2263 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2264 2265 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2266 SizeInBytes, FullName, Ty->getIdentifier()); 2267 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2268 2269 addUDTSrcLine(Ty, ClassTI); 2270 2271 addToUDTs(Ty); 2272 2273 return ClassTI; 2274 } 2275 2276 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2277 // Emit the complete type for unnamed unions. 2278 if (shouldAlwaysEmitCompleteClassType(Ty)) 2279 return getCompleteTypeIndex(Ty); 2280 2281 ClassOptions CO = 2282 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2283 std::string FullName = getFullyQualifiedName(Ty); 2284 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2285 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2286 if (!Ty->isForwardDecl()) 2287 DeferredCompleteTypes.push_back(Ty); 2288 return FwdDeclTI; 2289 } 2290 2291 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2292 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2293 TypeIndex FieldTI; 2294 unsigned FieldCount; 2295 bool ContainsNestedClass; 2296 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2297 lowerRecordFieldList(Ty); 2298 2299 if (ContainsNestedClass) 2300 CO |= ClassOptions::ContainsNestedClass; 2301 2302 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2303 std::string FullName = getFullyQualifiedName(Ty); 2304 2305 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2306 Ty->getIdentifier()); 2307 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2308 2309 addUDTSrcLine(Ty, UnionTI); 2310 2311 addToUDTs(Ty); 2312 2313 return UnionTI; 2314 } 2315 2316 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2317 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2318 // Manually count members. MSVC appears to count everything that generates a 2319 // field list record. Each individual overload in a method overload group 2320 // contributes to this count, even though the overload group is a single field 2321 // list record. 2322 unsigned MemberCount = 0; 2323 ClassInfo Info = collectClassInfo(Ty); 2324 ContinuationRecordBuilder ContinuationBuilder; 2325 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2326 2327 // Create base classes. 2328 for (const DIDerivedType *I : Info.Inheritance) { 2329 if (I->getFlags() & DINode::FlagVirtual) { 2330 // Virtual base. 2331 unsigned VBPtrOffset = I->getVBPtrOffset(); 2332 // FIXME: Despite the accessor name, the offset is really in bytes. 2333 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2334 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2335 ? TypeRecordKind::IndirectVirtualBaseClass 2336 : TypeRecordKind::VirtualBaseClass; 2337 VirtualBaseClassRecord VBCR( 2338 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2339 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2340 VBTableIndex); 2341 2342 ContinuationBuilder.writeMemberType(VBCR); 2343 MemberCount++; 2344 } else { 2345 assert(I->getOffsetInBits() % 8 == 0 && 2346 "bases must be on byte boundaries"); 2347 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2348 getTypeIndex(I->getBaseType()), 2349 I->getOffsetInBits() / 8); 2350 ContinuationBuilder.writeMemberType(BCR); 2351 MemberCount++; 2352 } 2353 } 2354 2355 // Create members. 2356 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2357 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2358 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2359 StringRef MemberName = Member->getName(); 2360 MemberAccess Access = 2361 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2362 2363 if (Member->isStaticMember()) { 2364 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2365 ContinuationBuilder.writeMemberType(SDMR); 2366 MemberCount++; 2367 continue; 2368 } 2369 2370 // Virtual function pointer member. 2371 if ((Member->getFlags() & DINode::FlagArtificial) && 2372 Member->getName().startswith("_vptr$")) { 2373 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2374 ContinuationBuilder.writeMemberType(VFPR); 2375 MemberCount++; 2376 continue; 2377 } 2378 2379 // Data member. 2380 uint64_t MemberOffsetInBits = 2381 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2382 if (Member->isBitField()) { 2383 uint64_t StartBitOffset = MemberOffsetInBits; 2384 if (const auto *CI = 2385 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2386 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2387 } 2388 StartBitOffset -= MemberOffsetInBits; 2389 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2390 StartBitOffset); 2391 MemberBaseType = TypeTable.writeLeafType(BFR); 2392 } 2393 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2394 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2395 MemberName); 2396 ContinuationBuilder.writeMemberType(DMR); 2397 MemberCount++; 2398 } 2399 2400 // Create methods 2401 for (auto &MethodItr : Info.Methods) { 2402 StringRef Name = MethodItr.first->getString(); 2403 2404 std::vector<OneMethodRecord> Methods; 2405 for (const DISubprogram *SP : MethodItr.second) { 2406 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2407 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2408 2409 unsigned VFTableOffset = -1; 2410 if (Introduced) 2411 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2412 2413 Methods.push_back(OneMethodRecord( 2414 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2415 translateMethodKindFlags(SP, Introduced), 2416 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2417 MemberCount++; 2418 } 2419 assert(!Methods.empty() && "Empty methods map entry"); 2420 if (Methods.size() == 1) 2421 ContinuationBuilder.writeMemberType(Methods[0]); 2422 else { 2423 // FIXME: Make this use its own ContinuationBuilder so that 2424 // MethodOverloadList can be split correctly. 2425 MethodOverloadListRecord MOLR(Methods); 2426 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2427 2428 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2429 ContinuationBuilder.writeMemberType(OMR); 2430 } 2431 } 2432 2433 // Create nested classes. 2434 for (const DIType *Nested : Info.NestedTypes) { 2435 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2436 ContinuationBuilder.writeMemberType(R); 2437 MemberCount++; 2438 } 2439 2440 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2441 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2442 !Info.NestedTypes.empty()); 2443 } 2444 2445 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2446 if (!VBPType.getIndex()) { 2447 // Make a 'const int *' type. 2448 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2449 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2450 2451 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2452 : PointerKind::Near32; 2453 PointerMode PM = PointerMode::Pointer; 2454 PointerOptions PO = PointerOptions::None; 2455 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2456 VBPType = TypeTable.writeLeafType(PR); 2457 } 2458 2459 return VBPType; 2460 } 2461 2462 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2463 // The null DIType is the void type. Don't try to hash it. 2464 if (!Ty) 2465 return TypeIndex::Void(); 2466 2467 // Check if we've already translated this type. Don't try to do a 2468 // get-or-create style insertion that caches the hash lookup across the 2469 // lowerType call. It will update the TypeIndices map. 2470 auto I = TypeIndices.find({Ty, ClassTy}); 2471 if (I != TypeIndices.end()) 2472 return I->second; 2473 2474 TypeLoweringScope S(*this); 2475 TypeIndex TI = lowerType(Ty, ClassTy); 2476 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2477 } 2478 2479 codeview::TypeIndex 2480 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2481 const DISubroutineType *SubroutineTy) { 2482 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2483 "this type must be a pointer type"); 2484 2485 PointerOptions Options = PointerOptions::None; 2486 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2487 Options = PointerOptions::LValueRefThisPointer; 2488 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2489 Options = PointerOptions::RValueRefThisPointer; 2490 2491 // Check if we've already translated this type. If there is no ref qualifier 2492 // on the function then we look up this pointer type with no associated class 2493 // so that the TypeIndex for the this pointer can be shared with the type 2494 // index for other pointers to this class type. If there is a ref qualifier 2495 // then we lookup the pointer using the subroutine as the parent type. 2496 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2497 if (I != TypeIndices.end()) 2498 return I->second; 2499 2500 TypeLoweringScope S(*this); 2501 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2502 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2503 } 2504 2505 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2506 PointerRecord PR(getTypeIndex(Ty), 2507 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2508 : PointerKind::Near32, 2509 PointerMode::LValueReference, PointerOptions::None, 2510 Ty->getSizeInBits() / 8); 2511 return TypeTable.writeLeafType(PR); 2512 } 2513 2514 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2515 // The null DIType is the void type. Don't try to hash it. 2516 if (!Ty) 2517 return TypeIndex::Void(); 2518 2519 // Look through typedefs when getting the complete type index. Call 2520 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2521 // emitted only once. 2522 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2523 (void)getTypeIndex(Ty); 2524 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2525 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2526 2527 // If this is a non-record type, the complete type index is the same as the 2528 // normal type index. Just call getTypeIndex. 2529 switch (Ty->getTag()) { 2530 case dwarf::DW_TAG_class_type: 2531 case dwarf::DW_TAG_structure_type: 2532 case dwarf::DW_TAG_union_type: 2533 break; 2534 default: 2535 return getTypeIndex(Ty); 2536 } 2537 2538 const auto *CTy = cast<DICompositeType>(Ty); 2539 2540 TypeLoweringScope S(*this); 2541 2542 // Make sure the forward declaration is emitted first. It's unclear if this 2543 // is necessary, but MSVC does it, and we should follow suit until we can show 2544 // otherwise. 2545 // We only emit a forward declaration for named types. 2546 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2547 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2548 2549 // Just use the forward decl if we don't have complete type info. This 2550 // might happen if the frontend is using modules and expects the complete 2551 // definition to be emitted elsewhere. 2552 if (CTy->isForwardDecl()) 2553 return FwdDeclTI; 2554 } 2555 2556 // Check if we've already translated the complete record type. 2557 // Insert the type with a null TypeIndex to signify that the type is currently 2558 // being lowered. 2559 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2560 if (!InsertResult.second) 2561 return InsertResult.first->second; 2562 2563 TypeIndex TI; 2564 switch (CTy->getTag()) { 2565 case dwarf::DW_TAG_class_type: 2566 case dwarf::DW_TAG_structure_type: 2567 TI = lowerCompleteTypeClass(CTy); 2568 break; 2569 case dwarf::DW_TAG_union_type: 2570 TI = lowerCompleteTypeUnion(CTy); 2571 break; 2572 default: 2573 llvm_unreachable("not a record"); 2574 } 2575 2576 // Update the type index associated with this CompositeType. This cannot 2577 // use the 'InsertResult' iterator above because it is potentially 2578 // invalidated by map insertions which can occur while lowering the class 2579 // type above. 2580 CompleteTypeIndices[CTy] = TI; 2581 return TI; 2582 } 2583 2584 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2585 /// and do this until fixpoint, as each complete record type typically 2586 /// references 2587 /// many other record types. 2588 void CodeViewDebug::emitDeferredCompleteTypes() { 2589 SmallVector<const DICompositeType *, 4> TypesToEmit; 2590 while (!DeferredCompleteTypes.empty()) { 2591 std::swap(DeferredCompleteTypes, TypesToEmit); 2592 for (const DICompositeType *RecordTy : TypesToEmit) 2593 getCompleteTypeIndex(RecordTy); 2594 TypesToEmit.clear(); 2595 } 2596 } 2597 2598 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2599 ArrayRef<LocalVariable> Locals) { 2600 // Get the sorted list of parameters and emit them first. 2601 SmallVector<const LocalVariable *, 6> Params; 2602 for (const LocalVariable &L : Locals) 2603 if (L.DIVar->isParameter()) 2604 Params.push_back(&L); 2605 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2606 return L->DIVar->getArg() < R->DIVar->getArg(); 2607 }); 2608 for (const LocalVariable *L : Params) 2609 emitLocalVariable(FI, *L); 2610 2611 // Next emit all non-parameters in the order that we found them. 2612 for (const LocalVariable &L : Locals) 2613 if (!L.DIVar->isParameter()) 2614 emitLocalVariable(FI, L); 2615 } 2616 2617 /// Only call this on endian-specific types like ulittle16_t and little32_t, or 2618 /// structs composed of them. 2619 template <typename T> 2620 static void copyBytesForDefRange(SmallString<20> &BytePrefix, 2621 SymbolKind SymKind, const T &DefRangeHeader) { 2622 BytePrefix.resize(2 + sizeof(T)); 2623 ulittle16_t SymKindLE = ulittle16_t(SymKind); 2624 memcpy(&BytePrefix[0], &SymKindLE, 2); 2625 memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T)); 2626 } 2627 2628 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2629 const LocalVariable &Var) { 2630 // LocalSym record, see SymbolRecord.h for more info. 2631 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2632 2633 LocalSymFlags Flags = LocalSymFlags::None; 2634 if (Var.DIVar->isParameter()) 2635 Flags |= LocalSymFlags::IsParameter; 2636 if (Var.DefRanges.empty()) 2637 Flags |= LocalSymFlags::IsOptimizedOut; 2638 2639 OS.AddComment("TypeIndex"); 2640 TypeIndex TI = Var.UseReferenceType 2641 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2642 : getCompleteTypeIndex(Var.DIVar->getType()); 2643 OS.EmitIntValue(TI.getIndex(), 4); 2644 OS.AddComment("Flags"); 2645 OS.EmitIntValue(static_cast<uint16_t>(Flags), 2); 2646 // Truncate the name so we won't overflow the record length field. 2647 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2648 endSymbolRecord(LocalEnd); 2649 2650 // Calculate the on disk prefix of the appropriate def range record. The 2651 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2652 // should be big enough to hold all forms without memory allocation. 2653 SmallString<20> BytePrefix; 2654 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2655 BytePrefix.clear(); 2656 if (DefRange.InMemory) { 2657 int Offset = DefRange.DataOffset; 2658 unsigned Reg = DefRange.CVRegister; 2659 2660 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2661 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2662 // instead. In frames without stack realignment, $T0 will be the CFA. 2663 if (RegisterId(Reg) == RegisterId::ESP) { 2664 Reg = unsigned(RegisterId::VFRAME); 2665 Offset += FI.OffsetAdjustment; 2666 } 2667 2668 // If we can use the chosen frame pointer for the frame and this isn't a 2669 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2670 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2671 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2672 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2673 (bool(Flags & LocalSymFlags::IsParameter) 2674 ? (EncFP == FI.EncodedParamFramePtrReg) 2675 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2676 little32_t FPOffset = little32_t(Offset); 2677 copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset); 2678 } else { 2679 uint16_t RegRelFlags = 0; 2680 if (DefRange.IsSubfield) { 2681 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2682 (DefRange.StructOffset 2683 << DefRangeRegisterRelSym::OffsetInParentShift); 2684 } 2685 DefRangeRegisterRelSym::Header DRHdr; 2686 DRHdr.Register = Reg; 2687 DRHdr.Flags = RegRelFlags; 2688 DRHdr.BasePointerOffset = Offset; 2689 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr); 2690 } 2691 } else { 2692 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2693 if (DefRange.IsSubfield) { 2694 DefRangeSubfieldRegisterSym::Header DRHdr; 2695 DRHdr.Register = DefRange.CVRegister; 2696 DRHdr.MayHaveNoName = 0; 2697 DRHdr.OffsetInParent = DefRange.StructOffset; 2698 copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr); 2699 } else { 2700 DefRangeRegisterSym::Header DRHdr; 2701 DRHdr.Register = DefRange.CVRegister; 2702 DRHdr.MayHaveNoName = 0; 2703 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr); 2704 } 2705 } 2706 OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix); 2707 } 2708 } 2709 2710 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2711 const FunctionInfo& FI) { 2712 for (LexicalBlock *Block : Blocks) 2713 emitLexicalBlock(*Block, FI); 2714 } 2715 2716 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2717 /// lexical block scope. 2718 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2719 const FunctionInfo& FI) { 2720 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2721 OS.AddComment("PtrParent"); 2722 OS.EmitIntValue(0, 4); // PtrParent 2723 OS.AddComment("PtrEnd"); 2724 OS.EmitIntValue(0, 4); // PtrEnd 2725 OS.AddComment("Code size"); 2726 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2727 OS.AddComment("Function section relative address"); 2728 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2729 OS.AddComment("Function section index"); 2730 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2731 OS.AddComment("Lexical block name"); 2732 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2733 endSymbolRecord(RecordEnd); 2734 2735 // Emit variables local to this lexical block. 2736 emitLocalVariableList(FI, Block.Locals); 2737 emitGlobalVariableList(Block.Globals); 2738 2739 // Emit lexical blocks contained within this block. 2740 emitLexicalBlockList(Block.Children, FI); 2741 2742 // Close the lexical block scope. 2743 emitEndSymbolRecord(SymbolKind::S_END); 2744 } 2745 2746 /// Convenience routine for collecting lexical block information for a list 2747 /// of lexical scopes. 2748 void CodeViewDebug::collectLexicalBlockInfo( 2749 SmallVectorImpl<LexicalScope *> &Scopes, 2750 SmallVectorImpl<LexicalBlock *> &Blocks, 2751 SmallVectorImpl<LocalVariable> &Locals, 2752 SmallVectorImpl<CVGlobalVariable> &Globals) { 2753 for (LexicalScope *Scope : Scopes) 2754 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2755 } 2756 2757 /// Populate the lexical blocks and local variable lists of the parent with 2758 /// information about the specified lexical scope. 2759 void CodeViewDebug::collectLexicalBlockInfo( 2760 LexicalScope &Scope, 2761 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2762 SmallVectorImpl<LocalVariable> &ParentLocals, 2763 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2764 if (Scope.isAbstractScope()) 2765 return; 2766 2767 // Gather information about the lexical scope including local variables, 2768 // global variables, and address ranges. 2769 bool IgnoreScope = false; 2770 auto LI = ScopeVariables.find(&Scope); 2771 SmallVectorImpl<LocalVariable> *Locals = 2772 LI != ScopeVariables.end() ? &LI->second : nullptr; 2773 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2774 SmallVectorImpl<CVGlobalVariable> *Globals = 2775 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2776 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2777 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2778 2779 // Ignore lexical scopes which do not contain variables. 2780 if (!Locals && !Globals) 2781 IgnoreScope = true; 2782 2783 // Ignore lexical scopes which are not lexical blocks. 2784 if (!DILB) 2785 IgnoreScope = true; 2786 2787 // Ignore scopes which have too many address ranges to represent in the 2788 // current CodeView format or do not have a valid address range. 2789 // 2790 // For lexical scopes with multiple address ranges you may be tempted to 2791 // construct a single range covering every instruction where the block is 2792 // live and everything in between. Unfortunately, Visual Studio only 2793 // displays variables from the first matching lexical block scope. If the 2794 // first lexical block contains exception handling code or cold code which 2795 // is moved to the bottom of the routine creating a single range covering 2796 // nearly the entire routine, then it will hide all other lexical blocks 2797 // and the variables they contain. 2798 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2799 IgnoreScope = true; 2800 2801 if (IgnoreScope) { 2802 // This scope can be safely ignored and eliminating it will reduce the 2803 // size of the debug information. Be sure to collect any variable and scope 2804 // information from the this scope or any of its children and collapse them 2805 // into the parent scope. 2806 if (Locals) 2807 ParentLocals.append(Locals->begin(), Locals->end()); 2808 if (Globals) 2809 ParentGlobals.append(Globals->begin(), Globals->end()); 2810 collectLexicalBlockInfo(Scope.getChildren(), 2811 ParentBlocks, 2812 ParentLocals, 2813 ParentGlobals); 2814 return; 2815 } 2816 2817 // Create a new CodeView lexical block for this lexical scope. If we've 2818 // seen this DILexicalBlock before then the scope tree is malformed and 2819 // we can handle this gracefully by not processing it a second time. 2820 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2821 if (!BlockInsertion.second) 2822 return; 2823 2824 // Create a lexical block containing the variables and collect the the 2825 // lexical block information for the children. 2826 const InsnRange &Range = Ranges.front(); 2827 assert(Range.first && Range.second); 2828 LexicalBlock &Block = BlockInsertion.first->second; 2829 Block.Begin = getLabelBeforeInsn(Range.first); 2830 Block.End = getLabelAfterInsn(Range.second); 2831 assert(Block.Begin && "missing label for scope begin"); 2832 assert(Block.End && "missing label for scope end"); 2833 Block.Name = DILB->getName(); 2834 if (Locals) 2835 Block.Locals = std::move(*Locals); 2836 if (Globals) 2837 Block.Globals = std::move(*Globals); 2838 ParentBlocks.push_back(&Block); 2839 collectLexicalBlockInfo(Scope.getChildren(), 2840 Block.Children, 2841 Block.Locals, 2842 Block.Globals); 2843 } 2844 2845 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2846 const Function &GV = MF->getFunction(); 2847 assert(FnDebugInfo.count(&GV)); 2848 assert(CurFn == FnDebugInfo[&GV].get()); 2849 2850 collectVariableInfo(GV.getSubprogram()); 2851 2852 // Build the lexical block structure to emit for this routine. 2853 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2854 collectLexicalBlockInfo(*CFS, 2855 CurFn->ChildBlocks, 2856 CurFn->Locals, 2857 CurFn->Globals); 2858 2859 // Clear the scope and variable information from the map which will not be 2860 // valid after we have finished processing this routine. This also prepares 2861 // the map for the subsequent routine. 2862 ScopeVariables.clear(); 2863 2864 // Don't emit anything if we don't have any line tables. 2865 // Thunks are compiler-generated and probably won't have source correlation. 2866 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2867 FnDebugInfo.erase(&GV); 2868 CurFn = nullptr; 2869 return; 2870 } 2871 2872 CurFn->Annotations = MF->getCodeViewAnnotations(); 2873 CurFn->HeapAllocSites = MF->getCodeViewHeapAllocSites(); 2874 2875 CurFn->End = Asm->getFunctionEnd(); 2876 2877 CurFn = nullptr; 2878 } 2879 2880 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2881 DebugHandlerBase::beginInstruction(MI); 2882 2883 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2884 if (!Asm || !CurFn || MI->isDebugInstr() || 2885 MI->getFlag(MachineInstr::FrameSetup)) 2886 return; 2887 2888 // If the first instruction of a new MBB has no location, find the first 2889 // instruction with a location and use that. 2890 DebugLoc DL = MI->getDebugLoc(); 2891 if (!DL && MI->getParent() != PrevInstBB) { 2892 for (const auto &NextMI : *MI->getParent()) { 2893 if (NextMI.isDebugInstr()) 2894 continue; 2895 DL = NextMI.getDebugLoc(); 2896 if (DL) 2897 break; 2898 } 2899 } 2900 PrevInstBB = MI->getParent(); 2901 2902 // If we still don't have a debug location, don't record a location. 2903 if (!DL) 2904 return; 2905 2906 maybeRecordLocation(DL, Asm->MF); 2907 } 2908 2909 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2910 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2911 *EndLabel = MMI->getContext().createTempSymbol(); 2912 OS.EmitIntValue(unsigned(Kind), 4); 2913 OS.AddComment("Subsection size"); 2914 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2915 OS.EmitLabel(BeginLabel); 2916 return EndLabel; 2917 } 2918 2919 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2920 OS.EmitLabel(EndLabel); 2921 // Every subsection must be aligned to a 4-byte boundary. 2922 OS.EmitValueToAlignment(4); 2923 } 2924 2925 static StringRef getSymbolName(SymbolKind SymKind) { 2926 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 2927 if (EE.Value == SymKind) 2928 return EE.Name; 2929 return ""; 2930 } 2931 2932 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 2933 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2934 *EndLabel = MMI->getContext().createTempSymbol(); 2935 OS.AddComment("Record length"); 2936 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 2937 OS.EmitLabel(BeginLabel); 2938 if (OS.isVerboseAsm()) 2939 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 2940 OS.EmitIntValue(unsigned(SymKind), 2); 2941 return EndLabel; 2942 } 2943 2944 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 2945 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 2946 // an extra copy of every symbol record in LLD. This increases object file 2947 // size by less than 1% in the clang build, and is compatible with the Visual 2948 // C++ linker. 2949 OS.EmitValueToAlignment(4); 2950 OS.EmitLabel(SymEnd); 2951 } 2952 2953 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 2954 OS.AddComment("Record length"); 2955 OS.EmitIntValue(2, 2); 2956 if (OS.isVerboseAsm()) 2957 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 2958 OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind 2959 } 2960 2961 void CodeViewDebug::emitDebugInfoForUDTs( 2962 ArrayRef<std::pair<std::string, const DIType *>> UDTs) { 2963 for (const auto &UDT : UDTs) { 2964 const DIType *T = UDT.second; 2965 assert(shouldEmitUdt(T)); 2966 2967 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 2968 OS.AddComment("Type"); 2969 OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4); 2970 emitNullTerminatedSymbolName(OS, UDT.first); 2971 endSymbolRecord(UDTRecordEnd); 2972 } 2973 } 2974 2975 void CodeViewDebug::collectGlobalVariableInfo() { 2976 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 2977 GlobalMap; 2978 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 2979 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 2980 GV.getDebugInfo(GVEs); 2981 for (const auto *GVE : GVEs) 2982 GlobalMap[GVE] = &GV; 2983 } 2984 2985 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2986 for (const MDNode *Node : CUs->operands()) { 2987 const auto *CU = cast<DICompileUnit>(Node); 2988 for (const auto *GVE : CU->getGlobalVariables()) { 2989 const DIGlobalVariable *DIGV = GVE->getVariable(); 2990 const DIExpression *DIE = GVE->getExpression(); 2991 2992 // Emit constant global variables in a global symbol section. 2993 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 2994 CVGlobalVariable CVGV = {DIGV, DIE}; 2995 GlobalVariables.emplace_back(std::move(CVGV)); 2996 } 2997 2998 const auto *GV = GlobalMap.lookup(GVE); 2999 if (!GV || GV->isDeclarationForLinker()) 3000 continue; 3001 3002 DIScope *Scope = DIGV->getScope(); 3003 SmallVector<CVGlobalVariable, 1> *VariableList; 3004 if (Scope && isa<DILocalScope>(Scope)) { 3005 // Locate a global variable list for this scope, creating one if 3006 // necessary. 3007 auto Insertion = ScopeGlobals.insert( 3008 {Scope, std::unique_ptr<GlobalVariableList>()}); 3009 if (Insertion.second) 3010 Insertion.first->second = llvm::make_unique<GlobalVariableList>(); 3011 VariableList = Insertion.first->second.get(); 3012 } else if (GV->hasComdat()) 3013 // Emit this global variable into a COMDAT section. 3014 VariableList = &ComdatVariables; 3015 else 3016 // Emit this global variable in a single global symbol section. 3017 VariableList = &GlobalVariables; 3018 CVGlobalVariable CVGV = {DIGV, GV}; 3019 VariableList->emplace_back(std::move(CVGV)); 3020 } 3021 } 3022 } 3023 3024 void CodeViewDebug::emitDebugInfoForGlobals() { 3025 // First, emit all globals that are not in a comdat in a single symbol 3026 // substream. MSVC doesn't like it if the substream is empty, so only open 3027 // it if we have at least one global to emit. 3028 switchToDebugSectionForSymbol(nullptr); 3029 if (!GlobalVariables.empty()) { 3030 OS.AddComment("Symbol subsection for globals"); 3031 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3032 emitGlobalVariableList(GlobalVariables); 3033 endCVSubsection(EndLabel); 3034 } 3035 3036 // Second, emit each global that is in a comdat into its own .debug$S 3037 // section along with its own symbol substream. 3038 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3039 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3040 MCSymbol *GVSym = Asm->getSymbol(GV); 3041 OS.AddComment("Symbol subsection for " + 3042 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3043 switchToDebugSectionForSymbol(GVSym); 3044 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3045 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3046 emitDebugInfoForGlobal(CVGV); 3047 endCVSubsection(EndLabel); 3048 } 3049 } 3050 3051 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3052 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3053 for (const MDNode *Node : CUs->operands()) { 3054 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3055 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3056 getTypeIndex(RT); 3057 // FIXME: Add to global/local DTU list. 3058 } 3059 } 3060 } 3061 } 3062 3063 // Emit each global variable in the specified array. 3064 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3065 for (const CVGlobalVariable &CVGV : Globals) { 3066 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3067 emitDebugInfoForGlobal(CVGV); 3068 } 3069 } 3070 3071 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3072 const DIGlobalVariable *DIGV = CVGV.DIGV; 3073 if (const GlobalVariable *GV = 3074 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3075 // DataSym record, see SymbolRecord.h for more info. Thread local data 3076 // happens to have the same format as global data. 3077 MCSymbol *GVSym = Asm->getSymbol(GV); 3078 SymbolKind DataSym = GV->isThreadLocal() 3079 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3080 : SymbolKind::S_GTHREAD32) 3081 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3082 : SymbolKind::S_GDATA32); 3083 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3084 OS.AddComment("Type"); 3085 OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4); 3086 OS.AddComment("DataOffset"); 3087 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 3088 OS.AddComment("Segment"); 3089 OS.EmitCOFFSectionIndex(GVSym); 3090 OS.AddComment("Name"); 3091 const unsigned LengthOfDataRecord = 12; 3092 emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord); 3093 endSymbolRecord(DataEnd); 3094 } else { 3095 // FIXME: Currently this only emits the global variables in the IR metadata. 3096 // This should also emit enums and static data members. 3097 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3098 assert(DIE->isConstant() && 3099 "Global constant variables must contain a constant expression."); 3100 uint64_t Val = DIE->getElement(1); 3101 3102 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3103 OS.AddComment("Type"); 3104 OS.EmitIntValue(getTypeIndex(DIGV->getType()).getIndex(), 4); 3105 OS.AddComment("Value"); 3106 3107 // Encoded integers shouldn't need more than 10 bytes. 3108 uint8_t data[10]; 3109 BinaryStreamWriter Writer(data, llvm::support::endianness::little); 3110 CodeViewRecordIO IO(Writer); 3111 cantFail(IO.mapEncodedInteger(Val)); 3112 StringRef SRef((char *)data, Writer.getOffset()); 3113 OS.EmitBinaryData(SRef); 3114 3115 OS.AddComment("Name"); 3116 const DIScope *Scope = DIGV->getScope(); 3117 // For static data members, get the scope from the declaration. 3118 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3119 DIGV->getRawStaticDataMemberDeclaration())) 3120 Scope = MemberDecl->getScope(); 3121 emitNullTerminatedSymbolName(OS, 3122 getFullyQualifiedName(Scope, DIGV->getName())); 3123 endSymbolRecord(SConstantEnd); 3124 } 3125 } 3126