CodeViewDebug.cpp revision 360784
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
95using namespace llvm;
96using namespace llvm::codeview;
97
98namespace {
99class CVMCAdapter : public CodeViewRecordStreamer {
100public:
101  CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable)
102      : OS(&OS), TypeTable(TypeTable) {}
103
104  void EmitBytes(StringRef Data) { OS->EmitBytes(Data); }
105
106  void EmitIntValue(uint64_t Value, unsigned Size) {
107    OS->EmitIntValueInHex(Value, Size);
108  }
109
110  void EmitBinaryData(StringRef Data) { OS->EmitBinaryData(Data); }
111
112  void AddComment(const Twine &T) { OS->AddComment(T); }
113
114  void AddRawComment(const Twine &T) { OS->emitRawComment(T); }
115
116  bool isVerboseAsm() { return OS->isVerboseAsm(); }
117
118  std::string getTypeName(TypeIndex TI) {
119    std::string TypeName;
120    if (!TI.isNoneType()) {
121      if (TI.isSimple())
122        TypeName = TypeIndex::simpleTypeName(TI);
123      else
124        TypeName = TypeTable.getTypeName(TI);
125    }
126    return TypeName;
127  }
128
129private:
130  MCStreamer *OS = nullptr;
131  TypeCollection &TypeTable;
132};
133} // namespace
134
135static CPUType mapArchToCVCPUType(Triple::ArchType Type) {
136  switch (Type) {
137  case Triple::ArchType::x86:
138    return CPUType::Pentium3;
139  case Triple::ArchType::x86_64:
140    return CPUType::X64;
141  case Triple::ArchType::thumb:
142    return CPUType::Thumb;
143  case Triple::ArchType::aarch64:
144    return CPUType::ARM64;
145  default:
146    report_fatal_error("target architecture doesn't map to a CodeView CPUType");
147  }
148}
149
150CodeViewDebug::CodeViewDebug(AsmPrinter *AP)
151    : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {
152  // If module doesn't have named metadata anchors or COFF debug section
153  // is not available, skip any debug info related stuff.
154  if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") ||
155      !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) {
156    Asm = nullptr;
157    MMI->setDebugInfoAvailability(false);
158    return;
159  }
160  // Tell MMI that we have debug info.
161  MMI->setDebugInfoAvailability(true);
162
163  TheCPU =
164      mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch());
165
166  collectGlobalVariableInfo();
167
168  // Check if we should emit type record hashes.
169  ConstantInt *GH = mdconst::extract_or_null<ConstantInt>(
170      MMI->getModule()->getModuleFlag("CodeViewGHash"));
171  EmitDebugGlobalHashes = GH && !GH->isZero();
172}
173
174StringRef CodeViewDebug::getFullFilepath(const DIFile *File) {
175  std::string &Filepath = FileToFilepathMap[File];
176  if (!Filepath.empty())
177    return Filepath;
178
179  StringRef Dir = File->getDirectory(), Filename = File->getFilename();
180
181  // If this is a Unix-style path, just use it as is. Don't try to canonicalize
182  // it textually because one of the path components could be a symlink.
183  if (Dir.startswith("/") || Filename.startswith("/")) {
184    if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix))
185      return Filename;
186    Filepath = Dir;
187    if (Dir.back() != '/')
188      Filepath += '/';
189    Filepath += Filename;
190    return Filepath;
191  }
192
193  // Clang emits directory and relative filename info into the IR, but CodeView
194  // operates on full paths.  We could change Clang to emit full paths too, but
195  // that would increase the IR size and probably not needed for other users.
196  // For now, just concatenate and canonicalize the path here.
197  if (Filename.find(':') == 1)
198    Filepath = Filename;
199  else
200    Filepath = (Dir + "\\" + Filename).str();
201
202  // Canonicalize the path.  We have to do it textually because we may no longer
203  // have access the file in the filesystem.
204  // First, replace all slashes with backslashes.
205  std::replace(Filepath.begin(), Filepath.end(), '/', '\\');
206
207  // Remove all "\.\" with "\".
208  size_t Cursor = 0;
209  while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos)
210    Filepath.erase(Cursor, 2);
211
212  // Replace all "\XXX\..\" with "\".  Don't try too hard though as the original
213  // path should be well-formatted, e.g. start with a drive letter, etc.
214  Cursor = 0;
215  while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) {
216    // Something's wrong if the path starts with "\..\", abort.
217    if (Cursor == 0)
218      break;
219
220    size_t PrevSlash = Filepath.rfind('\\', Cursor - 1);
221    if (PrevSlash == std::string::npos)
222      // Something's wrong, abort.
223      break;
224
225    Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash);
226    // The next ".." might be following the one we've just erased.
227    Cursor = PrevSlash;
228  }
229
230  // Remove all duplicate backslashes.
231  Cursor = 0;
232  while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos)
233    Filepath.erase(Cursor, 1);
234
235  return Filepath;
236}
237
238unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) {
239  StringRef FullPath = getFullFilepath(F);
240  unsigned NextId = FileIdMap.size() + 1;
241  auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId));
242  if (Insertion.second) {
243    // We have to compute the full filepath and emit a .cv_file directive.
244    ArrayRef<uint8_t> ChecksumAsBytes;
245    FileChecksumKind CSKind = FileChecksumKind::None;
246    if (F->getChecksum()) {
247      std::string Checksum = fromHex(F->getChecksum()->Value);
248      void *CKMem = OS.getContext().allocate(Checksum.size(), 1);
249      memcpy(CKMem, Checksum.data(), Checksum.size());
250      ChecksumAsBytes = ArrayRef<uint8_t>(
251          reinterpret_cast<const uint8_t *>(CKMem), Checksum.size());
252      switch (F->getChecksum()->Kind) {
253      case DIFile::CSK_MD5:  CSKind = FileChecksumKind::MD5; break;
254      case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break;
255      }
256    }
257    bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes,
258                                          static_cast<unsigned>(CSKind));
259    (void)Success;
260    assert(Success && ".cv_file directive failed");
261  }
262  return Insertion.first->second;
263}
264
265CodeViewDebug::InlineSite &
266CodeViewDebug::getInlineSite(const DILocation *InlinedAt,
267                             const DISubprogram *Inlinee) {
268  auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()});
269  InlineSite *Site = &SiteInsertion.first->second;
270  if (SiteInsertion.second) {
271    unsigned ParentFuncId = CurFn->FuncId;
272    if (const DILocation *OuterIA = InlinedAt->getInlinedAt())
273      ParentFuncId =
274          getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram())
275              .SiteFuncId;
276
277    Site->SiteFuncId = NextFuncId++;
278    OS.EmitCVInlineSiteIdDirective(
279        Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()),
280        InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc());
281    Site->Inlinee = Inlinee;
282    InlinedSubprograms.insert(Inlinee);
283    getFuncIdForSubprogram(Inlinee);
284  }
285  return *Site;
286}
287
288static StringRef getPrettyScopeName(const DIScope *Scope) {
289  StringRef ScopeName = Scope->getName();
290  if (!ScopeName.empty())
291    return ScopeName;
292
293  switch (Scope->getTag()) {
294  case dwarf::DW_TAG_enumeration_type:
295  case dwarf::DW_TAG_class_type:
296  case dwarf::DW_TAG_structure_type:
297  case dwarf::DW_TAG_union_type:
298    return "<unnamed-tag>";
299  case dwarf::DW_TAG_namespace:
300    return "`anonymous namespace'";
301  }
302
303  return StringRef();
304}
305
306static const DISubprogram *getQualifiedNameComponents(
307    const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) {
308  const DISubprogram *ClosestSubprogram = nullptr;
309  while (Scope != nullptr) {
310    if (ClosestSubprogram == nullptr)
311      ClosestSubprogram = dyn_cast<DISubprogram>(Scope);
312    StringRef ScopeName = getPrettyScopeName(Scope);
313    if (!ScopeName.empty())
314      QualifiedNameComponents.push_back(ScopeName);
315    Scope = Scope->getScope();
316  }
317  return ClosestSubprogram;
318}
319
320static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents,
321                                    StringRef TypeName) {
322  std::string FullyQualifiedName;
323  for (StringRef QualifiedNameComponent :
324       llvm::reverse(QualifiedNameComponents)) {
325    FullyQualifiedName.append(QualifiedNameComponent);
326    FullyQualifiedName.append("::");
327  }
328  FullyQualifiedName.append(TypeName);
329  return FullyQualifiedName;
330}
331
332static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) {
333  SmallVector<StringRef, 5> QualifiedNameComponents;
334  getQualifiedNameComponents(Scope, QualifiedNameComponents);
335  return getQualifiedName(QualifiedNameComponents, Name);
336}
337
338struct CodeViewDebug::TypeLoweringScope {
339  TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; }
340  ~TypeLoweringScope() {
341    // Don't decrement TypeEmissionLevel until after emitting deferred types, so
342    // inner TypeLoweringScopes don't attempt to emit deferred types.
343    if (CVD.TypeEmissionLevel == 1)
344      CVD.emitDeferredCompleteTypes();
345    --CVD.TypeEmissionLevel;
346  }
347  CodeViewDebug &CVD;
348};
349
350static std::string getFullyQualifiedName(const DIScope *Ty) {
351  const DIScope *Scope = Ty->getScope();
352  return getFullyQualifiedName(Scope, getPrettyScopeName(Ty));
353}
354
355TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) {
356  // No scope means global scope and that uses the zero index.
357  if (!Scope || isa<DIFile>(Scope))
358    return TypeIndex();
359
360  assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type");
361
362  // Check if we've already translated this scope.
363  auto I = TypeIndices.find({Scope, nullptr});
364  if (I != TypeIndices.end())
365    return I->second;
366
367  // Build the fully qualified name of the scope.
368  std::string ScopeName = getFullyQualifiedName(Scope);
369  StringIdRecord SID(TypeIndex(), ScopeName);
370  auto TI = TypeTable.writeLeafType(SID);
371  return recordTypeIndexForDINode(Scope, TI);
372}
373
374TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) {
375  assert(SP);
376
377  // Check if we've already translated this subprogram.
378  auto I = TypeIndices.find({SP, nullptr});
379  if (I != TypeIndices.end())
380    return I->second;
381
382  // The display name includes function template arguments. Drop them to match
383  // MSVC.
384  StringRef DisplayName = SP->getName().split('<').first;
385
386  const DIScope *Scope = SP->getScope();
387  TypeIndex TI;
388  if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) {
389    // If the scope is a DICompositeType, then this must be a method. Member
390    // function types take some special handling, and require access to the
391    // subprogram.
392    TypeIndex ClassType = getTypeIndex(Class);
393    MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class),
394                               DisplayName);
395    TI = TypeTable.writeLeafType(MFuncId);
396  } else {
397    // Otherwise, this must be a free function.
398    TypeIndex ParentScope = getScopeIndex(Scope);
399    FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName);
400    TI = TypeTable.writeLeafType(FuncId);
401  }
402
403  return recordTypeIndexForDINode(SP, TI);
404}
405
406static bool isNonTrivial(const DICompositeType *DCTy) {
407  return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial);
408}
409
410static FunctionOptions
411getFunctionOptions(const DISubroutineType *Ty,
412                   const DICompositeType *ClassTy = nullptr,
413                   StringRef SPName = StringRef("")) {
414  FunctionOptions FO = FunctionOptions::None;
415  const DIType *ReturnTy = nullptr;
416  if (auto TypeArray = Ty->getTypeArray()) {
417    if (TypeArray.size())
418      ReturnTy = TypeArray[0];
419  }
420
421  if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) {
422    if (isNonTrivial(ReturnDCTy))
423      FO |= FunctionOptions::CxxReturnUdt;
424  }
425
426  // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison.
427  if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) {
428    FO |= FunctionOptions::Constructor;
429
430  // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag.
431
432  }
433  return FO;
434}
435
436TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP,
437                                               const DICompositeType *Class) {
438  // Always use the method declaration as the key for the function type. The
439  // method declaration contains the this adjustment.
440  if (SP->getDeclaration())
441    SP = SP->getDeclaration();
442  assert(!SP->getDeclaration() && "should use declaration as key");
443
444  // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide
445  // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}.
446  auto I = TypeIndices.find({SP, Class});
447  if (I != TypeIndices.end())
448    return I->second;
449
450  // Make sure complete type info for the class is emitted *after* the member
451  // function type, as the complete class type is likely to reference this
452  // member function type.
453  TypeLoweringScope S(*this);
454  const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0;
455
456  FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName());
457  TypeIndex TI = lowerTypeMemberFunction(
458      SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO);
459  return recordTypeIndexForDINode(SP, TI, Class);
460}
461
462TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node,
463                                                  TypeIndex TI,
464                                                  const DIType *ClassTy) {
465  auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI});
466  (void)InsertResult;
467  assert(InsertResult.second && "DINode was already assigned a type index");
468  return TI;
469}
470
471unsigned CodeViewDebug::getPointerSizeInBytes() {
472  return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8;
473}
474
475void CodeViewDebug::recordLocalVariable(LocalVariable &&Var,
476                                        const LexicalScope *LS) {
477  if (const DILocation *InlinedAt = LS->getInlinedAt()) {
478    // This variable was inlined. Associate it with the InlineSite.
479    const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram();
480    InlineSite &Site = getInlineSite(InlinedAt, Inlinee);
481    Site.InlinedLocals.emplace_back(Var);
482  } else {
483    // This variable goes into the corresponding lexical scope.
484    ScopeVariables[LS].emplace_back(Var);
485  }
486}
487
488static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs,
489                               const DILocation *Loc) {
490  auto B = Locs.begin(), E = Locs.end();
491  if (std::find(B, E, Loc) == E)
492    Locs.push_back(Loc);
493}
494
495void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL,
496                                        const MachineFunction *MF) {
497  // Skip this instruction if it has the same location as the previous one.
498  if (!DL || DL == PrevInstLoc)
499    return;
500
501  const DIScope *Scope = DL.get()->getScope();
502  if (!Scope)
503    return;
504
505  // Skip this line if it is longer than the maximum we can record.
506  LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true);
507  if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() ||
508      LI.isNeverStepInto())
509    return;
510
511  ColumnInfo CI(DL.getCol(), /*EndColumn=*/0);
512  if (CI.getStartColumn() != DL.getCol())
513    return;
514
515  if (!CurFn->HaveLineInfo)
516    CurFn->HaveLineInfo = true;
517  unsigned FileId = 0;
518  if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile())
519    FileId = CurFn->LastFileId;
520  else
521    FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile());
522  PrevInstLoc = DL;
523
524  unsigned FuncId = CurFn->FuncId;
525  if (const DILocation *SiteLoc = DL->getInlinedAt()) {
526    const DILocation *Loc = DL.get();
527
528    // If this location was actually inlined from somewhere else, give it the ID
529    // of the inline call site.
530    FuncId =
531        getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId;
532
533    // Ensure we have links in the tree of inline call sites.
534    bool FirstLoc = true;
535    while ((SiteLoc = Loc->getInlinedAt())) {
536      InlineSite &Site =
537          getInlineSite(SiteLoc, Loc->getScope()->getSubprogram());
538      if (!FirstLoc)
539        addLocIfNotPresent(Site.ChildSites, Loc);
540      FirstLoc = false;
541      Loc = SiteLoc;
542    }
543    addLocIfNotPresent(CurFn->ChildSites, Loc);
544  }
545
546  OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(),
547                        /*PrologueEnd=*/false, /*IsStmt=*/false,
548                        DL->getFilename(), SMLoc());
549}
550
551void CodeViewDebug::emitCodeViewMagicVersion() {
552  OS.EmitValueToAlignment(4);
553  OS.AddComment("Debug section magic");
554  OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4);
555}
556
557void CodeViewDebug::endModule() {
558  if (!Asm || !MMI->hasDebugInfo())
559    return;
560
561  assert(Asm != nullptr);
562
563  // The COFF .debug$S section consists of several subsections, each starting
564  // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length
565  // of the payload followed by the payload itself.  The subsections are 4-byte
566  // aligned.
567
568  // Use the generic .debug$S section, and make a subsection for all the inlined
569  // subprograms.
570  switchToDebugSectionForSymbol(nullptr);
571
572  MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols);
573  emitCompilerInformation();
574  endCVSubsection(CompilerInfo);
575
576  emitInlineeLinesSubsection();
577
578  // Emit per-function debug information.
579  for (auto &P : FnDebugInfo)
580    if (!P.first->isDeclarationForLinker())
581      emitDebugInfoForFunction(P.first, *P.second);
582
583  // Emit global variable debug information.
584  setCurrentSubprogram(nullptr);
585  emitDebugInfoForGlobals();
586
587  // Emit retained types.
588  emitDebugInfoForRetainedTypes();
589
590  // Switch back to the generic .debug$S section after potentially processing
591  // comdat symbol sections.
592  switchToDebugSectionForSymbol(nullptr);
593
594  // Emit UDT records for any types used by global variables.
595  if (!GlobalUDTs.empty()) {
596    MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
597    emitDebugInfoForUDTs(GlobalUDTs);
598    endCVSubsection(SymbolsEnd);
599  }
600
601  // This subsection holds a file index to offset in string table table.
602  OS.AddComment("File index to string table offset subsection");
603  OS.EmitCVFileChecksumsDirective();
604
605  // This subsection holds the string table.
606  OS.AddComment("String table");
607  OS.EmitCVStringTableDirective();
608
609  // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol
610  // subsection in the generic .debug$S section at the end. There is no
611  // particular reason for this ordering other than to match MSVC.
612  emitBuildInfo();
613
614  // Emit type information and hashes last, so that any types we translate while
615  // emitting function info are included.
616  emitTypeInformation();
617
618  if (EmitDebugGlobalHashes)
619    emitTypeGlobalHashes();
620
621  clear();
622}
623
624static void
625emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S,
626                             unsigned MaxFixedRecordLength = 0xF00) {
627  // The maximum CV record length is 0xFF00. Most of the strings we emit appear
628  // after a fixed length portion of the record. The fixed length portion should
629  // always be less than 0xF00 (3840) bytes, so truncate the string so that the
630  // overall record size is less than the maximum allowed.
631  SmallString<32> NullTerminatedString(
632      S.take_front(MaxRecordLength - MaxFixedRecordLength - 1));
633  NullTerminatedString.push_back('\0');
634  OS.EmitBytes(NullTerminatedString);
635}
636
637void CodeViewDebug::emitTypeInformation() {
638  if (TypeTable.empty())
639    return;
640
641  // Start the .debug$T or .debug$P section with 0x4.
642  OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection());
643  emitCodeViewMagicVersion();
644
645  TypeTableCollection Table(TypeTable.records());
646  TypeVisitorCallbackPipeline Pipeline;
647
648  // To emit type record using Codeview MCStreamer adapter
649  CVMCAdapter CVMCOS(OS, Table);
650  TypeRecordMapping typeMapping(CVMCOS);
651  Pipeline.addCallbackToPipeline(typeMapping);
652
653  Optional<TypeIndex> B = Table.getFirst();
654  while (B) {
655    // This will fail if the record data is invalid.
656    CVType Record = Table.getType(*B);
657
658    Error E = codeview::visitTypeRecord(Record, *B, Pipeline);
659
660    if (E) {
661      logAllUnhandledErrors(std::move(E), errs(), "error: ");
662      llvm_unreachable("produced malformed type record");
663    }
664
665    B = Table.getNext(*B);
666  }
667}
668
669void CodeViewDebug::emitTypeGlobalHashes() {
670  if (TypeTable.empty())
671    return;
672
673  // Start the .debug$H section with the version and hash algorithm, currently
674  // hardcoded to version 0, SHA1.
675  OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection());
676
677  OS.EmitValueToAlignment(4);
678  OS.AddComment("Magic");
679  OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4);
680  OS.AddComment("Section Version");
681  OS.EmitIntValue(0, 2);
682  OS.AddComment("Hash Algorithm");
683  OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2);
684
685  TypeIndex TI(TypeIndex::FirstNonSimpleIndex);
686  for (const auto &GHR : TypeTable.hashes()) {
687    if (OS.isVerboseAsm()) {
688      // Emit an EOL-comment describing which TypeIndex this hash corresponds
689      // to, as well as the stringified SHA1 hash.
690      SmallString<32> Comment;
691      raw_svector_ostream CommentOS(Comment);
692      CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR);
693      OS.AddComment(Comment);
694      ++TI;
695    }
696    assert(GHR.Hash.size() == 8);
697    StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()),
698                GHR.Hash.size());
699    OS.EmitBinaryData(S);
700  }
701}
702
703static SourceLanguage MapDWLangToCVLang(unsigned DWLang) {
704  switch (DWLang) {
705  case dwarf::DW_LANG_C:
706  case dwarf::DW_LANG_C89:
707  case dwarf::DW_LANG_C99:
708  case dwarf::DW_LANG_C11:
709  case dwarf::DW_LANG_ObjC:
710    return SourceLanguage::C;
711  case dwarf::DW_LANG_C_plus_plus:
712  case dwarf::DW_LANG_C_plus_plus_03:
713  case dwarf::DW_LANG_C_plus_plus_11:
714  case dwarf::DW_LANG_C_plus_plus_14:
715    return SourceLanguage::Cpp;
716  case dwarf::DW_LANG_Fortran77:
717  case dwarf::DW_LANG_Fortran90:
718  case dwarf::DW_LANG_Fortran03:
719  case dwarf::DW_LANG_Fortran08:
720    return SourceLanguage::Fortran;
721  case dwarf::DW_LANG_Pascal83:
722    return SourceLanguage::Pascal;
723  case dwarf::DW_LANG_Cobol74:
724  case dwarf::DW_LANG_Cobol85:
725    return SourceLanguage::Cobol;
726  case dwarf::DW_LANG_Java:
727    return SourceLanguage::Java;
728  case dwarf::DW_LANG_D:
729    return SourceLanguage::D;
730  case dwarf::DW_LANG_Swift:
731    return SourceLanguage::Swift;
732  default:
733    // There's no CodeView representation for this language, and CV doesn't
734    // have an "unknown" option for the language field, so we'll use MASM,
735    // as it's very low level.
736    return SourceLanguage::Masm;
737  }
738}
739
740namespace {
741struct Version {
742  int Part[4];
743};
744} // end anonymous namespace
745
746// Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out
747// the version number.
748static Version parseVersion(StringRef Name) {
749  Version V = {{0}};
750  int N = 0;
751  for (const char C : Name) {
752    if (isdigit(C)) {
753      V.Part[N] *= 10;
754      V.Part[N] += C - '0';
755    } else if (C == '.') {
756      ++N;
757      if (N >= 4)
758        return V;
759    } else if (N > 0)
760      return V;
761  }
762  return V;
763}
764
765void CodeViewDebug::emitCompilerInformation() {
766  MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3);
767  uint32_t Flags = 0;
768
769  NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
770  const MDNode *Node = *CUs->operands().begin();
771  const auto *CU = cast<DICompileUnit>(Node);
772
773  // The low byte of the flags indicates the source language.
774  Flags = MapDWLangToCVLang(CU->getSourceLanguage());
775  // TODO:  Figure out which other flags need to be set.
776
777  OS.AddComment("Flags and language");
778  OS.EmitIntValue(Flags, 4);
779
780  OS.AddComment("CPUType");
781  OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2);
782
783  StringRef CompilerVersion = CU->getProducer();
784  Version FrontVer = parseVersion(CompilerVersion);
785  OS.AddComment("Frontend version");
786  for (int N = 0; N < 4; ++N)
787    OS.EmitIntValue(FrontVer.Part[N], 2);
788
789  // Some Microsoft tools, like Binscope, expect a backend version number of at
790  // least 8.something, so we'll coerce the LLVM version into a form that
791  // guarantees it'll be big enough without really lying about the version.
792  int Major = 1000 * LLVM_VERSION_MAJOR +
793              10 * LLVM_VERSION_MINOR +
794              LLVM_VERSION_PATCH;
795  // Clamp it for builds that use unusually large version numbers.
796  Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max());
797  Version BackVer = {{ Major, 0, 0, 0 }};
798  OS.AddComment("Backend version");
799  for (int N = 0; N < 4; ++N)
800    OS.EmitIntValue(BackVer.Part[N], 2);
801
802  OS.AddComment("Null-terminated compiler version string");
803  emitNullTerminatedSymbolName(OS, CompilerVersion);
804
805  endSymbolRecord(CompilerEnd);
806}
807
808static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable,
809                                    StringRef S) {
810  StringIdRecord SIR(TypeIndex(0x0), S);
811  return TypeTable.writeLeafType(SIR);
812}
813
814void CodeViewDebug::emitBuildInfo() {
815  // First, make LF_BUILDINFO. It's a sequence of strings with various bits of
816  // build info. The known prefix is:
817  // - Absolute path of current directory
818  // - Compiler path
819  // - Main source file path, relative to CWD or absolute
820  // - Type server PDB file
821  // - Canonical compiler command line
822  // If frontend and backend compilation are separated (think llc or LTO), it's
823  // not clear if the compiler path should refer to the executable for the
824  // frontend or the backend. Leave it blank for now.
825  TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {};
826  NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
827  const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs.
828  const auto *CU = cast<DICompileUnit>(Node);
829  const DIFile *MainSourceFile = CU->getFile();
830  BuildInfoArgs[BuildInfoRecord::CurrentDirectory] =
831      getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory());
832  BuildInfoArgs[BuildInfoRecord::SourceFile] =
833      getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename());
834  // FIXME: Path to compiler and command line. PDB is intentionally blank unless
835  // we implement /Zi type servers.
836  BuildInfoRecord BIR(BuildInfoArgs);
837  TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR);
838
839  // Make a new .debug$S subsection for the S_BUILDINFO record, which points
840  // from the module symbols into the type stream.
841  MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
842  MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO);
843  OS.AddComment("LF_BUILDINFO index");
844  OS.EmitIntValue(BuildInfoIndex.getIndex(), 4);
845  endSymbolRecord(BIEnd);
846  endCVSubsection(BISubsecEnd);
847}
848
849void CodeViewDebug::emitInlineeLinesSubsection() {
850  if (InlinedSubprograms.empty())
851    return;
852
853  OS.AddComment("Inlinee lines subsection");
854  MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines);
855
856  // We emit the checksum info for files.  This is used by debuggers to
857  // determine if a pdb matches the source before loading it.  Visual Studio,
858  // for instance, will display a warning that the breakpoints are not valid if
859  // the pdb does not match the source.
860  OS.AddComment("Inlinee lines signature");
861  OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4);
862
863  for (const DISubprogram *SP : InlinedSubprograms) {
864    assert(TypeIndices.count({SP, nullptr}));
865    TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}];
866
867    OS.AddBlankLine();
868    unsigned FileId = maybeRecordFile(SP->getFile());
869    OS.AddComment("Inlined function " + SP->getName() + " starts at " +
870                  SP->getFilename() + Twine(':') + Twine(SP->getLine()));
871    OS.AddBlankLine();
872    OS.AddComment("Type index of inlined function");
873    OS.EmitIntValue(InlineeIdx.getIndex(), 4);
874    OS.AddComment("Offset into filechecksum table");
875    OS.EmitCVFileChecksumOffsetDirective(FileId);
876    OS.AddComment("Starting line number");
877    OS.EmitIntValue(SP->getLine(), 4);
878  }
879
880  endCVSubsection(InlineEnd);
881}
882
883void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI,
884                                        const DILocation *InlinedAt,
885                                        const InlineSite &Site) {
886  assert(TypeIndices.count({Site.Inlinee, nullptr}));
887  TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}];
888
889  // SymbolRecord
890  MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE);
891
892  OS.AddComment("PtrParent");
893  OS.EmitIntValue(0, 4);
894  OS.AddComment("PtrEnd");
895  OS.EmitIntValue(0, 4);
896  OS.AddComment("Inlinee type index");
897  OS.EmitIntValue(InlineeIdx.getIndex(), 4);
898
899  unsigned FileId = maybeRecordFile(Site.Inlinee->getFile());
900  unsigned StartLineNum = Site.Inlinee->getLine();
901
902  OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum,
903                                    FI.Begin, FI.End);
904
905  endSymbolRecord(InlineEnd);
906
907  emitLocalVariableList(FI, Site.InlinedLocals);
908
909  // Recurse on child inlined call sites before closing the scope.
910  for (const DILocation *ChildSite : Site.ChildSites) {
911    auto I = FI.InlineSites.find(ChildSite);
912    assert(I != FI.InlineSites.end() &&
913           "child site not in function inline site map");
914    emitInlinedCallSite(FI, ChildSite, I->second);
915  }
916
917  // Close the scope.
918  emitEndSymbolRecord(SymbolKind::S_INLINESITE_END);
919}
920
921void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) {
922  // If we have a symbol, it may be in a section that is COMDAT. If so, find the
923  // comdat key. A section may be comdat because of -ffunction-sections or
924  // because it is comdat in the IR.
925  MCSectionCOFF *GVSec =
926      GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr;
927  const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr;
928
929  MCSectionCOFF *DebugSec = cast<MCSectionCOFF>(
930      Asm->getObjFileLowering().getCOFFDebugSymbolsSection());
931  DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym);
932
933  OS.SwitchSection(DebugSec);
934
935  // Emit the magic version number if this is the first time we've switched to
936  // this section.
937  if (ComdatDebugSections.insert(DebugSec).second)
938    emitCodeViewMagicVersion();
939}
940
941// Emit an S_THUNK32/S_END symbol pair for a thunk routine.
942// The only supported thunk ordinal is currently the standard type.
943void CodeViewDebug::emitDebugInfoForThunk(const Function *GV,
944                                          FunctionInfo &FI,
945                                          const MCSymbol *Fn) {
946  std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName());
947  const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind.
948
949  OS.AddComment("Symbol subsection for " + Twine(FuncName));
950  MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
951
952  // Emit S_THUNK32
953  MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32);
954  OS.AddComment("PtrParent");
955  OS.EmitIntValue(0, 4);
956  OS.AddComment("PtrEnd");
957  OS.EmitIntValue(0, 4);
958  OS.AddComment("PtrNext");
959  OS.EmitIntValue(0, 4);
960  OS.AddComment("Thunk section relative address");
961  OS.EmitCOFFSecRel32(Fn, /*Offset=*/0);
962  OS.AddComment("Thunk section index");
963  OS.EmitCOFFSectionIndex(Fn);
964  OS.AddComment("Code size");
965  OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2);
966  OS.AddComment("Ordinal");
967  OS.EmitIntValue(unsigned(ordinal), 1);
968  OS.AddComment("Function name");
969  emitNullTerminatedSymbolName(OS, FuncName);
970  // Additional fields specific to the thunk ordinal would go here.
971  endSymbolRecord(ThunkRecordEnd);
972
973  // Local variables/inlined routines are purposely omitted here.  The point of
974  // marking this as a thunk is so Visual Studio will NOT stop in this routine.
975
976  // Emit S_PROC_ID_END
977  emitEndSymbolRecord(SymbolKind::S_PROC_ID_END);
978
979  endCVSubsection(SymbolsEnd);
980}
981
982void CodeViewDebug::emitDebugInfoForFunction(const Function *GV,
983                                             FunctionInfo &FI) {
984  // For each function there is a separate subsection which holds the PC to
985  // file:line table.
986  const MCSymbol *Fn = Asm->getSymbol(GV);
987  assert(Fn);
988
989  // Switch to the to a comdat section, if appropriate.
990  switchToDebugSectionForSymbol(Fn);
991
992  std::string FuncName;
993  auto *SP = GV->getSubprogram();
994  assert(SP);
995  setCurrentSubprogram(SP);
996
997  if (SP->isThunk()) {
998    emitDebugInfoForThunk(GV, FI, Fn);
999    return;
1000  }
1001
1002  // If we have a display name, build the fully qualified name by walking the
1003  // chain of scopes.
1004  if (!SP->getName().empty())
1005    FuncName = getFullyQualifiedName(SP->getScope(), SP->getName());
1006
1007  // If our DISubprogram name is empty, use the mangled name.
1008  if (FuncName.empty())
1009    FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName());
1010
1011  // Emit FPO data, but only on 32-bit x86. No other platforms use it.
1012  if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86)
1013    OS.EmitCVFPOData(Fn);
1014
1015  // Emit a symbol subsection, required by VS2012+ to find function boundaries.
1016  OS.AddComment("Symbol subsection for " + Twine(FuncName));
1017  MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
1018  {
1019    SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID
1020                                                : SymbolKind::S_GPROC32_ID;
1021    MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind);
1022
1023    // These fields are filled in by tools like CVPACK which run after the fact.
1024    OS.AddComment("PtrParent");
1025    OS.EmitIntValue(0, 4);
1026    OS.AddComment("PtrEnd");
1027    OS.EmitIntValue(0, 4);
1028    OS.AddComment("PtrNext");
1029    OS.EmitIntValue(0, 4);
1030    // This is the important bit that tells the debugger where the function
1031    // code is located and what's its size:
1032    OS.AddComment("Code size");
1033    OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4);
1034    OS.AddComment("Offset after prologue");
1035    OS.EmitIntValue(0, 4);
1036    OS.AddComment("Offset before epilogue");
1037    OS.EmitIntValue(0, 4);
1038    OS.AddComment("Function type index");
1039    OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4);
1040    OS.AddComment("Function section relative address");
1041    OS.EmitCOFFSecRel32(Fn, /*Offset=*/0);
1042    OS.AddComment("Function section index");
1043    OS.EmitCOFFSectionIndex(Fn);
1044    OS.AddComment("Flags");
1045    OS.EmitIntValue(0, 1);
1046    // Emit the function display name as a null-terminated string.
1047    OS.AddComment("Function name");
1048    // Truncate the name so we won't overflow the record length field.
1049    emitNullTerminatedSymbolName(OS, FuncName);
1050    endSymbolRecord(ProcRecordEnd);
1051
1052    MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC);
1053    // Subtract out the CSR size since MSVC excludes that and we include it.
1054    OS.AddComment("FrameSize");
1055    OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4);
1056    OS.AddComment("Padding");
1057    OS.EmitIntValue(0, 4);
1058    OS.AddComment("Offset of padding");
1059    OS.EmitIntValue(0, 4);
1060    OS.AddComment("Bytes of callee saved registers");
1061    OS.EmitIntValue(FI.CSRSize, 4);
1062    OS.AddComment("Exception handler offset");
1063    OS.EmitIntValue(0, 4);
1064    OS.AddComment("Exception handler section");
1065    OS.EmitIntValue(0, 2);
1066    OS.AddComment("Flags (defines frame register)");
1067    OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4);
1068    endSymbolRecord(FrameProcEnd);
1069
1070    emitLocalVariableList(FI, FI.Locals);
1071    emitGlobalVariableList(FI.Globals);
1072    emitLexicalBlockList(FI.ChildBlocks, FI);
1073
1074    // Emit inlined call site information. Only emit functions inlined directly
1075    // into the parent function. We'll emit the other sites recursively as part
1076    // of their parent inline site.
1077    for (const DILocation *InlinedAt : FI.ChildSites) {
1078      auto I = FI.InlineSites.find(InlinedAt);
1079      assert(I != FI.InlineSites.end() &&
1080             "child site not in function inline site map");
1081      emitInlinedCallSite(FI, InlinedAt, I->second);
1082    }
1083
1084    for (auto Annot : FI.Annotations) {
1085      MCSymbol *Label = Annot.first;
1086      MDTuple *Strs = cast<MDTuple>(Annot.second);
1087      MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION);
1088      OS.EmitCOFFSecRel32(Label, /*Offset=*/0);
1089      // FIXME: Make sure we don't overflow the max record size.
1090      OS.EmitCOFFSectionIndex(Label);
1091      OS.EmitIntValue(Strs->getNumOperands(), 2);
1092      for (Metadata *MD : Strs->operands()) {
1093        // MDStrings are null terminated, so we can do EmitBytes and get the
1094        // nice .asciz directive.
1095        StringRef Str = cast<MDString>(MD)->getString();
1096        assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString");
1097        OS.EmitBytes(StringRef(Str.data(), Str.size() + 1));
1098      }
1099      endSymbolRecord(AnnotEnd);
1100    }
1101
1102    for (auto HeapAllocSite : FI.HeapAllocSites) {
1103      const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite);
1104      const MCSymbol *EndLabel = std::get<1>(HeapAllocSite);
1105      const DIType *DITy = std::get<2>(HeapAllocSite);
1106      MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE);
1107      OS.AddComment("Call site offset");
1108      OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0);
1109      OS.AddComment("Call site section index");
1110      OS.EmitCOFFSectionIndex(BeginLabel);
1111      OS.AddComment("Call instruction length");
1112      OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2);
1113      OS.AddComment("Type index");
1114      OS.EmitIntValue(getCompleteTypeIndex(DITy).getIndex(), 4);
1115      endSymbolRecord(HeapAllocEnd);
1116    }
1117
1118    if (SP != nullptr)
1119      emitDebugInfoForUDTs(LocalUDTs);
1120
1121    // We're done with this function.
1122    emitEndSymbolRecord(SymbolKind::S_PROC_ID_END);
1123  }
1124  endCVSubsection(SymbolsEnd);
1125
1126  // We have an assembler directive that takes care of the whole line table.
1127  OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End);
1128}
1129
1130CodeViewDebug::LocalVarDefRange
1131CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) {
1132  LocalVarDefRange DR;
1133  DR.InMemory = -1;
1134  DR.DataOffset = Offset;
1135  assert(DR.DataOffset == Offset && "truncation");
1136  DR.IsSubfield = 0;
1137  DR.StructOffset = 0;
1138  DR.CVRegister = CVRegister;
1139  return DR;
1140}
1141
1142void CodeViewDebug::collectVariableInfoFromMFTable(
1143    DenseSet<InlinedEntity> &Processed) {
1144  const MachineFunction &MF = *Asm->MF;
1145  const TargetSubtargetInfo &TSI = MF.getSubtarget();
1146  const TargetFrameLowering *TFI = TSI.getFrameLowering();
1147  const TargetRegisterInfo *TRI = TSI.getRegisterInfo();
1148
1149  for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) {
1150    if (!VI.Var)
1151      continue;
1152    assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) &&
1153           "Expected inlined-at fields to agree");
1154
1155    Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt()));
1156    LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc);
1157
1158    // If variable scope is not found then skip this variable.
1159    if (!Scope)
1160      continue;
1161
1162    // If the variable has an attached offset expression, extract it.
1163    // FIXME: Try to handle DW_OP_deref as well.
1164    int64_t ExprOffset = 0;
1165    bool Deref = false;
1166    if (VI.Expr) {
1167      // If there is one DW_OP_deref element, use offset of 0 and keep going.
1168      if (VI.Expr->getNumElements() == 1 &&
1169          VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref)
1170        Deref = true;
1171      else if (!VI.Expr->extractIfOffset(ExprOffset))
1172        continue;
1173    }
1174
1175    // Get the frame register used and the offset.
1176    unsigned FrameReg = 0;
1177    int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg);
1178    uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg);
1179
1180    // Calculate the label ranges.
1181    LocalVarDefRange DefRange =
1182        createDefRangeMem(CVReg, FrameOffset + ExprOffset);
1183
1184    for (const InsnRange &Range : Scope->getRanges()) {
1185      const MCSymbol *Begin = getLabelBeforeInsn(Range.first);
1186      const MCSymbol *End = getLabelAfterInsn(Range.second);
1187      End = End ? End : Asm->getFunctionEnd();
1188      DefRange.Ranges.emplace_back(Begin, End);
1189    }
1190
1191    LocalVariable Var;
1192    Var.DIVar = VI.Var;
1193    Var.DefRanges.emplace_back(std::move(DefRange));
1194    if (Deref)
1195      Var.UseReferenceType = true;
1196
1197    recordLocalVariable(std::move(Var), Scope);
1198  }
1199}
1200
1201static bool canUseReferenceType(const DbgVariableLocation &Loc) {
1202  return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0;
1203}
1204
1205static bool needsReferenceType(const DbgVariableLocation &Loc) {
1206  return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0;
1207}
1208
1209void CodeViewDebug::calculateRanges(
1210    LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) {
1211  const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo();
1212
1213  // Calculate the definition ranges.
1214  for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) {
1215    const auto &Entry = *I;
1216    if (!Entry.isDbgValue())
1217      continue;
1218    const MachineInstr *DVInst = Entry.getInstr();
1219    assert(DVInst->isDebugValue() && "Invalid History entry");
1220    // FIXME: Find a way to represent constant variables, since they are
1221    // relatively common.
1222    Optional<DbgVariableLocation> Location =
1223        DbgVariableLocation::extractFromMachineInstruction(*DVInst);
1224    if (!Location)
1225      continue;
1226
1227    // CodeView can only express variables in register and variables in memory
1228    // at a constant offset from a register. However, for variables passed
1229    // indirectly by pointer, it is common for that pointer to be spilled to a
1230    // stack location. For the special case of one offseted load followed by a
1231    // zero offset load (a pointer spilled to the stack), we change the type of
1232    // the local variable from a value type to a reference type. This tricks the
1233    // debugger into doing the load for us.
1234    if (Var.UseReferenceType) {
1235      // We're using a reference type. Drop the last zero offset load.
1236      if (canUseReferenceType(*Location))
1237        Location->LoadChain.pop_back();
1238      else
1239        continue;
1240    } else if (needsReferenceType(*Location)) {
1241      // This location can't be expressed without switching to a reference type.
1242      // Start over using that.
1243      Var.UseReferenceType = true;
1244      Var.DefRanges.clear();
1245      calculateRanges(Var, Entries);
1246      return;
1247    }
1248
1249    // We can only handle a register or an offseted load of a register.
1250    if (Location->Register == 0 || Location->LoadChain.size() > 1)
1251      continue;
1252    {
1253      LocalVarDefRange DR;
1254      DR.CVRegister = TRI->getCodeViewRegNum(Location->Register);
1255      DR.InMemory = !Location->LoadChain.empty();
1256      DR.DataOffset =
1257          !Location->LoadChain.empty() ? Location->LoadChain.back() : 0;
1258      if (Location->FragmentInfo) {
1259        DR.IsSubfield = true;
1260        DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8;
1261      } else {
1262        DR.IsSubfield = false;
1263        DR.StructOffset = 0;
1264      }
1265
1266      if (Var.DefRanges.empty() ||
1267          Var.DefRanges.back().isDifferentLocation(DR)) {
1268        Var.DefRanges.emplace_back(std::move(DR));
1269      }
1270    }
1271
1272    // Compute the label range.
1273    const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr());
1274    const MCSymbol *End;
1275    if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) {
1276      auto &EndingEntry = Entries[Entry.getEndIndex()];
1277      End = EndingEntry.isDbgValue()
1278                ? getLabelBeforeInsn(EndingEntry.getInstr())
1279                : getLabelAfterInsn(EndingEntry.getInstr());
1280    } else
1281      End = Asm->getFunctionEnd();
1282
1283    // If the last range end is our begin, just extend the last range.
1284    // Otherwise make a new range.
1285    SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R =
1286        Var.DefRanges.back().Ranges;
1287    if (!R.empty() && R.back().second == Begin)
1288      R.back().second = End;
1289    else
1290      R.emplace_back(Begin, End);
1291
1292    // FIXME: Do more range combining.
1293  }
1294}
1295
1296void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) {
1297  DenseSet<InlinedEntity> Processed;
1298  // Grab the variable info that was squirreled away in the MMI side-table.
1299  collectVariableInfoFromMFTable(Processed);
1300
1301  for (const auto &I : DbgValues) {
1302    InlinedEntity IV = I.first;
1303    if (Processed.count(IV))
1304      continue;
1305    const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first);
1306    const DILocation *InlinedAt = IV.second;
1307
1308    // Instruction ranges, specifying where IV is accessible.
1309    const auto &Entries = I.second;
1310
1311    LexicalScope *Scope = nullptr;
1312    if (InlinedAt)
1313      Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt);
1314    else
1315      Scope = LScopes.findLexicalScope(DIVar->getScope());
1316    // If variable scope is not found then skip this variable.
1317    if (!Scope)
1318      continue;
1319
1320    LocalVariable Var;
1321    Var.DIVar = DIVar;
1322
1323    calculateRanges(Var, Entries);
1324    recordLocalVariable(std::move(Var), Scope);
1325  }
1326}
1327
1328void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) {
1329  const TargetSubtargetInfo &TSI = MF->getSubtarget();
1330  const TargetRegisterInfo *TRI = TSI.getRegisterInfo();
1331  const MachineFrameInfo &MFI = MF->getFrameInfo();
1332  const Function &GV = MF->getFunction();
1333  auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()});
1334  assert(Insertion.second && "function already has info");
1335  CurFn = Insertion.first->second.get();
1336  CurFn->FuncId = NextFuncId++;
1337  CurFn->Begin = Asm->getFunctionBegin();
1338
1339  // The S_FRAMEPROC record reports the stack size, and how many bytes of
1340  // callee-saved registers were used. For targets that don't use a PUSH
1341  // instruction (AArch64), this will be zero.
1342  CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters();
1343  CurFn->FrameSize = MFI.getStackSize();
1344  CurFn->OffsetAdjustment = MFI.getOffsetAdjustment();
1345  CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF);
1346
1347  // For this function S_FRAMEPROC record, figure out which codeview register
1348  // will be the frame pointer.
1349  CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None.
1350  CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None.
1351  if (CurFn->FrameSize > 0) {
1352    if (!TSI.getFrameLowering()->hasFP(*MF)) {
1353      CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr;
1354      CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr;
1355    } else {
1356      // If there is an FP, parameters are always relative to it.
1357      CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr;
1358      if (CurFn->HasStackRealignment) {
1359        // If the stack needs realignment, locals are relative to SP or VFRAME.
1360        CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr;
1361      } else {
1362        // Otherwise, locals are relative to EBP, and we probably have VLAs or
1363        // other stack adjustments.
1364        CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr;
1365      }
1366    }
1367  }
1368
1369  // Compute other frame procedure options.
1370  FrameProcedureOptions FPO = FrameProcedureOptions::None;
1371  if (MFI.hasVarSizedObjects())
1372    FPO |= FrameProcedureOptions::HasAlloca;
1373  if (MF->exposesReturnsTwice())
1374    FPO |= FrameProcedureOptions::HasSetJmp;
1375  // FIXME: Set HasLongJmp if we ever track that info.
1376  if (MF->hasInlineAsm())
1377    FPO |= FrameProcedureOptions::HasInlineAssembly;
1378  if (GV.hasPersonalityFn()) {
1379    if (isAsynchronousEHPersonality(
1380            classifyEHPersonality(GV.getPersonalityFn())))
1381      FPO |= FrameProcedureOptions::HasStructuredExceptionHandling;
1382    else
1383      FPO |= FrameProcedureOptions::HasExceptionHandling;
1384  }
1385  if (GV.hasFnAttribute(Attribute::InlineHint))
1386    FPO |= FrameProcedureOptions::MarkedInline;
1387  if (GV.hasFnAttribute(Attribute::Naked))
1388    FPO |= FrameProcedureOptions::Naked;
1389  if (MFI.hasStackProtectorIndex())
1390    FPO |= FrameProcedureOptions::SecurityChecks;
1391  FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U);
1392  FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U);
1393  if (Asm->TM.getOptLevel() != CodeGenOpt::None &&
1394      !GV.hasOptSize() && !GV.hasOptNone())
1395    FPO |= FrameProcedureOptions::OptimizedForSpeed;
1396  // FIXME: Set GuardCfg when it is implemented.
1397  CurFn->FrameProcOpts = FPO;
1398
1399  OS.EmitCVFuncIdDirective(CurFn->FuncId);
1400
1401  // Find the end of the function prolog.  First known non-DBG_VALUE and
1402  // non-frame setup location marks the beginning of the function body.
1403  // FIXME: is there a simpler a way to do this? Can we just search
1404  // for the first instruction of the function, not the last of the prolog?
1405  DebugLoc PrologEndLoc;
1406  bool EmptyPrologue = true;
1407  for (const auto &MBB : *MF) {
1408    for (const auto &MI : MBB) {
1409      if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) &&
1410          MI.getDebugLoc()) {
1411        PrologEndLoc = MI.getDebugLoc();
1412        break;
1413      } else if (!MI.isMetaInstruction()) {
1414        EmptyPrologue = false;
1415      }
1416    }
1417  }
1418
1419  // Record beginning of function if we have a non-empty prologue.
1420  if (PrologEndLoc && !EmptyPrologue) {
1421    DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc();
1422    maybeRecordLocation(FnStartDL, MF);
1423  }
1424
1425  // Find heap alloc sites and emit labels around them.
1426  for (const auto &MBB : *MF) {
1427    for (const auto &MI : MBB) {
1428      if (MI.getHeapAllocMarker()) {
1429        requestLabelBeforeInsn(&MI);
1430        requestLabelAfterInsn(&MI);
1431      }
1432    }
1433  }
1434}
1435
1436static bool shouldEmitUdt(const DIType *T) {
1437  if (!T)
1438    return false;
1439
1440  // MSVC does not emit UDTs for typedefs that are scoped to classes.
1441  if (T->getTag() == dwarf::DW_TAG_typedef) {
1442    if (DIScope *Scope = T->getScope()) {
1443      switch (Scope->getTag()) {
1444      case dwarf::DW_TAG_structure_type:
1445      case dwarf::DW_TAG_class_type:
1446      case dwarf::DW_TAG_union_type:
1447        return false;
1448      }
1449    }
1450  }
1451
1452  while (true) {
1453    if (!T || T->isForwardDecl())
1454      return false;
1455
1456    const DIDerivedType *DT = dyn_cast<DIDerivedType>(T);
1457    if (!DT)
1458      return true;
1459    T = DT->getBaseType();
1460  }
1461  return true;
1462}
1463
1464void CodeViewDebug::addToUDTs(const DIType *Ty) {
1465  // Don't record empty UDTs.
1466  if (Ty->getName().empty())
1467    return;
1468  if (!shouldEmitUdt(Ty))
1469    return;
1470
1471  SmallVector<StringRef, 5> QualifiedNameComponents;
1472  const DISubprogram *ClosestSubprogram =
1473      getQualifiedNameComponents(Ty->getScope(), QualifiedNameComponents);
1474
1475  std::string FullyQualifiedName =
1476      getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty));
1477
1478  if (ClosestSubprogram == nullptr) {
1479    GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty);
1480  } else if (ClosestSubprogram == CurrentSubprogram) {
1481    LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty);
1482  }
1483
1484  // TODO: What if the ClosestSubprogram is neither null or the current
1485  // subprogram?  Currently, the UDT just gets dropped on the floor.
1486  //
1487  // The current behavior is not desirable.  To get maximal fidelity, we would
1488  // need to perform all type translation before beginning emission of .debug$S
1489  // and then make LocalUDTs a member of FunctionInfo
1490}
1491
1492TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) {
1493  // Generic dispatch for lowering an unknown type.
1494  switch (Ty->getTag()) {
1495  case dwarf::DW_TAG_array_type:
1496    return lowerTypeArray(cast<DICompositeType>(Ty));
1497  case dwarf::DW_TAG_typedef:
1498    return lowerTypeAlias(cast<DIDerivedType>(Ty));
1499  case dwarf::DW_TAG_base_type:
1500    return lowerTypeBasic(cast<DIBasicType>(Ty));
1501  case dwarf::DW_TAG_pointer_type:
1502    if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type")
1503      return lowerTypeVFTableShape(cast<DIDerivedType>(Ty));
1504    LLVM_FALLTHROUGH;
1505  case dwarf::DW_TAG_reference_type:
1506  case dwarf::DW_TAG_rvalue_reference_type:
1507    return lowerTypePointer(cast<DIDerivedType>(Ty));
1508  case dwarf::DW_TAG_ptr_to_member_type:
1509    return lowerTypeMemberPointer(cast<DIDerivedType>(Ty));
1510  case dwarf::DW_TAG_restrict_type:
1511  case dwarf::DW_TAG_const_type:
1512  case dwarf::DW_TAG_volatile_type:
1513  // TODO: add support for DW_TAG_atomic_type here
1514    return lowerTypeModifier(cast<DIDerivedType>(Ty));
1515  case dwarf::DW_TAG_subroutine_type:
1516    if (ClassTy) {
1517      // The member function type of a member function pointer has no
1518      // ThisAdjustment.
1519      return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy,
1520                                     /*ThisAdjustment=*/0,
1521                                     /*IsStaticMethod=*/false);
1522    }
1523    return lowerTypeFunction(cast<DISubroutineType>(Ty));
1524  case dwarf::DW_TAG_enumeration_type:
1525    return lowerTypeEnum(cast<DICompositeType>(Ty));
1526  case dwarf::DW_TAG_class_type:
1527  case dwarf::DW_TAG_structure_type:
1528    return lowerTypeClass(cast<DICompositeType>(Ty));
1529  case dwarf::DW_TAG_union_type:
1530    return lowerTypeUnion(cast<DICompositeType>(Ty));
1531  case dwarf::DW_TAG_unspecified_type:
1532    if (Ty->getName() == "decltype(nullptr)")
1533      return TypeIndex::NullptrT();
1534    return TypeIndex::None();
1535  default:
1536    // Use the null type index.
1537    return TypeIndex();
1538  }
1539}
1540
1541TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) {
1542  TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType());
1543  StringRef TypeName = Ty->getName();
1544
1545  addToUDTs(Ty);
1546
1547  if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) &&
1548      TypeName == "HRESULT")
1549    return TypeIndex(SimpleTypeKind::HResult);
1550  if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) &&
1551      TypeName == "wchar_t")
1552    return TypeIndex(SimpleTypeKind::WideCharacter);
1553
1554  return UnderlyingTypeIndex;
1555}
1556
1557TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) {
1558  const DIType *ElementType = Ty->getBaseType();
1559  TypeIndex ElementTypeIndex = getTypeIndex(ElementType);
1560  // IndexType is size_t, which depends on the bitness of the target.
1561  TypeIndex IndexType = getPointerSizeInBytes() == 8
1562                            ? TypeIndex(SimpleTypeKind::UInt64Quad)
1563                            : TypeIndex(SimpleTypeKind::UInt32Long);
1564
1565  uint64_t ElementSize = getBaseTypeSize(ElementType) / 8;
1566
1567  // Add subranges to array type.
1568  DINodeArray Elements = Ty->getElements();
1569  for (int i = Elements.size() - 1; i >= 0; --i) {
1570    const DINode *Element = Elements[i];
1571    assert(Element->getTag() == dwarf::DW_TAG_subrange_type);
1572
1573    const DISubrange *Subrange = cast<DISubrange>(Element);
1574    assert(Subrange->getLowerBound() == 0 &&
1575           "codeview doesn't support subranges with lower bounds");
1576    int64_t Count = -1;
1577    if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>())
1578      Count = CI->getSExtValue();
1579
1580    // Forward declarations of arrays without a size and VLAs use a count of -1.
1581    // Emit a count of zero in these cases to match what MSVC does for arrays
1582    // without a size. MSVC doesn't support VLAs, so it's not clear what we
1583    // should do for them even if we could distinguish them.
1584    if (Count == -1)
1585      Count = 0;
1586
1587    // Update the element size and element type index for subsequent subranges.
1588    ElementSize *= Count;
1589
1590    // If this is the outermost array, use the size from the array. It will be
1591    // more accurate if we had a VLA or an incomplete element type size.
1592    uint64_t ArraySize =
1593        (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize;
1594
1595    StringRef Name = (i == 0) ? Ty->getName() : "";
1596    ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name);
1597    ElementTypeIndex = TypeTable.writeLeafType(AR);
1598  }
1599
1600  return ElementTypeIndex;
1601}
1602
1603TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) {
1604  TypeIndex Index;
1605  dwarf::TypeKind Kind;
1606  uint32_t ByteSize;
1607
1608  Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding());
1609  ByteSize = Ty->getSizeInBits() / 8;
1610
1611  SimpleTypeKind STK = SimpleTypeKind::None;
1612  switch (Kind) {
1613  case dwarf::DW_ATE_address:
1614    // FIXME: Translate
1615    break;
1616  case dwarf::DW_ATE_boolean:
1617    switch (ByteSize) {
1618    case 1:  STK = SimpleTypeKind::Boolean8;   break;
1619    case 2:  STK = SimpleTypeKind::Boolean16;  break;
1620    case 4:  STK = SimpleTypeKind::Boolean32;  break;
1621    case 8:  STK = SimpleTypeKind::Boolean64;  break;
1622    case 16: STK = SimpleTypeKind::Boolean128; break;
1623    }
1624    break;
1625  case dwarf::DW_ATE_complex_float:
1626    switch (ByteSize) {
1627    case 2:  STK = SimpleTypeKind::Complex16;  break;
1628    case 4:  STK = SimpleTypeKind::Complex32;  break;
1629    case 8:  STK = SimpleTypeKind::Complex64;  break;
1630    case 10: STK = SimpleTypeKind::Complex80;  break;
1631    case 16: STK = SimpleTypeKind::Complex128; break;
1632    }
1633    break;
1634  case dwarf::DW_ATE_float:
1635    switch (ByteSize) {
1636    case 2:  STK = SimpleTypeKind::Float16;  break;
1637    case 4:  STK = SimpleTypeKind::Float32;  break;
1638    case 6:  STK = SimpleTypeKind::Float48;  break;
1639    case 8:  STK = SimpleTypeKind::Float64;  break;
1640    case 10: STK = SimpleTypeKind::Float80;  break;
1641    case 16: STK = SimpleTypeKind::Float128; break;
1642    }
1643    break;
1644  case dwarf::DW_ATE_signed:
1645    switch (ByteSize) {
1646    case 1:  STK = SimpleTypeKind::SignedCharacter; break;
1647    case 2:  STK = SimpleTypeKind::Int16Short;      break;
1648    case 4:  STK = SimpleTypeKind::Int32;           break;
1649    case 8:  STK = SimpleTypeKind::Int64Quad;       break;
1650    case 16: STK = SimpleTypeKind::Int128Oct;       break;
1651    }
1652    break;
1653  case dwarf::DW_ATE_unsigned:
1654    switch (ByteSize) {
1655    case 1:  STK = SimpleTypeKind::UnsignedCharacter; break;
1656    case 2:  STK = SimpleTypeKind::UInt16Short;       break;
1657    case 4:  STK = SimpleTypeKind::UInt32;            break;
1658    case 8:  STK = SimpleTypeKind::UInt64Quad;        break;
1659    case 16: STK = SimpleTypeKind::UInt128Oct;        break;
1660    }
1661    break;
1662  case dwarf::DW_ATE_UTF:
1663    switch (ByteSize) {
1664    case 2: STK = SimpleTypeKind::Character16; break;
1665    case 4: STK = SimpleTypeKind::Character32; break;
1666    }
1667    break;
1668  case dwarf::DW_ATE_signed_char:
1669    if (ByteSize == 1)
1670      STK = SimpleTypeKind::SignedCharacter;
1671    break;
1672  case dwarf::DW_ATE_unsigned_char:
1673    if (ByteSize == 1)
1674      STK = SimpleTypeKind::UnsignedCharacter;
1675    break;
1676  default:
1677    break;
1678  }
1679
1680  // Apply some fixups based on the source-level type name.
1681  if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int")
1682    STK = SimpleTypeKind::Int32Long;
1683  if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int")
1684    STK = SimpleTypeKind::UInt32Long;
1685  if (STK == SimpleTypeKind::UInt16Short &&
1686      (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t"))
1687    STK = SimpleTypeKind::WideCharacter;
1688  if ((STK == SimpleTypeKind::SignedCharacter ||
1689       STK == SimpleTypeKind::UnsignedCharacter) &&
1690      Ty->getName() == "char")
1691    STK = SimpleTypeKind::NarrowCharacter;
1692
1693  return TypeIndex(STK);
1694}
1695
1696TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty,
1697                                          PointerOptions PO) {
1698  TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType());
1699
1700  // Pointers to simple types without any options can use SimpleTypeMode, rather
1701  // than having a dedicated pointer type record.
1702  if (PointeeTI.isSimple() && PO == PointerOptions::None &&
1703      PointeeTI.getSimpleMode() == SimpleTypeMode::Direct &&
1704      Ty->getTag() == dwarf::DW_TAG_pointer_type) {
1705    SimpleTypeMode Mode = Ty->getSizeInBits() == 64
1706                              ? SimpleTypeMode::NearPointer64
1707                              : SimpleTypeMode::NearPointer32;
1708    return TypeIndex(PointeeTI.getSimpleKind(), Mode);
1709  }
1710
1711  PointerKind PK =
1712      Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32;
1713  PointerMode PM = PointerMode::Pointer;
1714  switch (Ty->getTag()) {
1715  default: llvm_unreachable("not a pointer tag type");
1716  case dwarf::DW_TAG_pointer_type:
1717    PM = PointerMode::Pointer;
1718    break;
1719  case dwarf::DW_TAG_reference_type:
1720    PM = PointerMode::LValueReference;
1721    break;
1722  case dwarf::DW_TAG_rvalue_reference_type:
1723    PM = PointerMode::RValueReference;
1724    break;
1725  }
1726
1727  if (Ty->isObjectPointer())
1728    PO |= PointerOptions::Const;
1729
1730  PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8);
1731  return TypeTable.writeLeafType(PR);
1732}
1733
1734static PointerToMemberRepresentation
1735translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) {
1736  // SizeInBytes being zero generally implies that the member pointer type was
1737  // incomplete, which can happen if it is part of a function prototype. In this
1738  // case, use the unknown model instead of the general model.
1739  if (IsPMF) {
1740    switch (Flags & DINode::FlagPtrToMemberRep) {
1741    case 0:
1742      return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown
1743                              : PointerToMemberRepresentation::GeneralFunction;
1744    case DINode::FlagSingleInheritance:
1745      return PointerToMemberRepresentation::SingleInheritanceFunction;
1746    case DINode::FlagMultipleInheritance:
1747      return PointerToMemberRepresentation::MultipleInheritanceFunction;
1748    case DINode::FlagVirtualInheritance:
1749      return PointerToMemberRepresentation::VirtualInheritanceFunction;
1750    }
1751  } else {
1752    switch (Flags & DINode::FlagPtrToMemberRep) {
1753    case 0:
1754      return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown
1755                              : PointerToMemberRepresentation::GeneralData;
1756    case DINode::FlagSingleInheritance:
1757      return PointerToMemberRepresentation::SingleInheritanceData;
1758    case DINode::FlagMultipleInheritance:
1759      return PointerToMemberRepresentation::MultipleInheritanceData;
1760    case DINode::FlagVirtualInheritance:
1761      return PointerToMemberRepresentation::VirtualInheritanceData;
1762    }
1763  }
1764  llvm_unreachable("invalid ptr to member representation");
1765}
1766
1767TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty,
1768                                                PointerOptions PO) {
1769  assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type);
1770  TypeIndex ClassTI = getTypeIndex(Ty->getClassType());
1771  TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType());
1772  PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64
1773                                                : PointerKind::Near32;
1774  bool IsPMF = isa<DISubroutineType>(Ty->getBaseType());
1775  PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction
1776                         : PointerMode::PointerToDataMember;
1777
1778  assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big");
1779  uint8_t SizeInBytes = Ty->getSizeInBits() / 8;
1780  MemberPointerInfo MPI(
1781      ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags()));
1782  PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI);
1783  return TypeTable.writeLeafType(PR);
1784}
1785
1786/// Given a DWARF calling convention, get the CodeView equivalent. If we don't
1787/// have a translation, use the NearC convention.
1788static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) {
1789  switch (DwarfCC) {
1790  case dwarf::DW_CC_normal:             return CallingConvention::NearC;
1791  case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast;
1792  case dwarf::DW_CC_BORLAND_thiscall:   return CallingConvention::ThisCall;
1793  case dwarf::DW_CC_BORLAND_stdcall:    return CallingConvention::NearStdCall;
1794  case dwarf::DW_CC_BORLAND_pascal:     return CallingConvention::NearPascal;
1795  case dwarf::DW_CC_LLVM_vectorcall:    return CallingConvention::NearVector;
1796  }
1797  return CallingConvention::NearC;
1798}
1799
1800TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) {
1801  ModifierOptions Mods = ModifierOptions::None;
1802  PointerOptions PO = PointerOptions::None;
1803  bool IsModifier = true;
1804  const DIType *BaseTy = Ty;
1805  while (IsModifier && BaseTy) {
1806    // FIXME: Need to add DWARF tags for __unaligned and _Atomic
1807    switch (BaseTy->getTag()) {
1808    case dwarf::DW_TAG_const_type:
1809      Mods |= ModifierOptions::Const;
1810      PO |= PointerOptions::Const;
1811      break;
1812    case dwarf::DW_TAG_volatile_type:
1813      Mods |= ModifierOptions::Volatile;
1814      PO |= PointerOptions::Volatile;
1815      break;
1816    case dwarf::DW_TAG_restrict_type:
1817      // Only pointer types be marked with __restrict. There is no known flag
1818      // for __restrict in LF_MODIFIER records.
1819      PO |= PointerOptions::Restrict;
1820      break;
1821    default:
1822      IsModifier = false;
1823      break;
1824    }
1825    if (IsModifier)
1826      BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType();
1827  }
1828
1829  // Check if the inner type will use an LF_POINTER record. If so, the
1830  // qualifiers will go in the LF_POINTER record. This comes up for types like
1831  // 'int *const' and 'int *__restrict', not the more common cases like 'const
1832  // char *'.
1833  if (BaseTy) {
1834    switch (BaseTy->getTag()) {
1835    case dwarf::DW_TAG_pointer_type:
1836    case dwarf::DW_TAG_reference_type:
1837    case dwarf::DW_TAG_rvalue_reference_type:
1838      return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO);
1839    case dwarf::DW_TAG_ptr_to_member_type:
1840      return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO);
1841    default:
1842      break;
1843    }
1844  }
1845
1846  TypeIndex ModifiedTI = getTypeIndex(BaseTy);
1847
1848  // Return the base type index if there aren't any modifiers. For example, the
1849  // metadata could contain restrict wrappers around non-pointer types.
1850  if (Mods == ModifierOptions::None)
1851    return ModifiedTI;
1852
1853  ModifierRecord MR(ModifiedTI, Mods);
1854  return TypeTable.writeLeafType(MR);
1855}
1856
1857TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) {
1858  SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices;
1859  for (const DIType *ArgType : Ty->getTypeArray())
1860    ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType));
1861
1862  // MSVC uses type none for variadic argument.
1863  if (ReturnAndArgTypeIndices.size() > 1 &&
1864      ReturnAndArgTypeIndices.back() == TypeIndex::Void()) {
1865    ReturnAndArgTypeIndices.back() = TypeIndex::None();
1866  }
1867  TypeIndex ReturnTypeIndex = TypeIndex::Void();
1868  ArrayRef<TypeIndex> ArgTypeIndices = None;
1869  if (!ReturnAndArgTypeIndices.empty()) {
1870    auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices);
1871    ReturnTypeIndex = ReturnAndArgTypesRef.front();
1872    ArgTypeIndices = ReturnAndArgTypesRef.drop_front();
1873  }
1874
1875  ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices);
1876  TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec);
1877
1878  CallingConvention CC = dwarfCCToCodeView(Ty->getCC());
1879
1880  FunctionOptions FO = getFunctionOptions(Ty);
1881  ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(),
1882                            ArgListIndex);
1883  return TypeTable.writeLeafType(Procedure);
1884}
1885
1886TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty,
1887                                                 const DIType *ClassTy,
1888                                                 int ThisAdjustment,
1889                                                 bool IsStaticMethod,
1890                                                 FunctionOptions FO) {
1891  // Lower the containing class type.
1892  TypeIndex ClassType = getTypeIndex(ClassTy);
1893
1894  DITypeRefArray ReturnAndArgs = Ty->getTypeArray();
1895
1896  unsigned Index = 0;
1897  SmallVector<TypeIndex, 8> ArgTypeIndices;
1898  TypeIndex ReturnTypeIndex = TypeIndex::Void();
1899  if (ReturnAndArgs.size() > Index) {
1900    ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]);
1901  }
1902
1903  // If the first argument is a pointer type and this isn't a static method,
1904  // treat it as the special 'this' parameter, which is encoded separately from
1905  // the arguments.
1906  TypeIndex ThisTypeIndex;
1907  if (!IsStaticMethod && ReturnAndArgs.size() > Index) {
1908    if (const DIDerivedType *PtrTy =
1909            dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) {
1910      if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) {
1911        ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty);
1912        Index++;
1913      }
1914    }
1915  }
1916
1917  while (Index < ReturnAndArgs.size())
1918    ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++]));
1919
1920  // MSVC uses type none for variadic argument.
1921  if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void())
1922    ArgTypeIndices.back() = TypeIndex::None();
1923
1924  ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices);
1925  TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec);
1926
1927  CallingConvention CC = dwarfCCToCodeView(Ty->getCC());
1928
1929  MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO,
1930                           ArgTypeIndices.size(), ArgListIndex, ThisAdjustment);
1931  return TypeTable.writeLeafType(MFR);
1932}
1933
1934TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) {
1935  unsigned VSlotCount =
1936      Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize());
1937  SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near);
1938
1939  VFTableShapeRecord VFTSR(Slots);
1940  return TypeTable.writeLeafType(VFTSR);
1941}
1942
1943static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) {
1944  switch (Flags & DINode::FlagAccessibility) {
1945  case DINode::FlagPrivate:   return MemberAccess::Private;
1946  case DINode::FlagPublic:    return MemberAccess::Public;
1947  case DINode::FlagProtected: return MemberAccess::Protected;
1948  case 0:
1949    // If there was no explicit access control, provide the default for the tag.
1950    return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private
1951                                                 : MemberAccess::Public;
1952  }
1953  llvm_unreachable("access flags are exclusive");
1954}
1955
1956static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) {
1957  if (SP->isArtificial())
1958    return MethodOptions::CompilerGenerated;
1959
1960  // FIXME: Handle other MethodOptions.
1961
1962  return MethodOptions::None;
1963}
1964
1965static MethodKind translateMethodKindFlags(const DISubprogram *SP,
1966                                           bool Introduced) {
1967  if (SP->getFlags() & DINode::FlagStaticMember)
1968    return MethodKind::Static;
1969
1970  switch (SP->getVirtuality()) {
1971  case dwarf::DW_VIRTUALITY_none:
1972    break;
1973  case dwarf::DW_VIRTUALITY_virtual:
1974    return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual;
1975  case dwarf::DW_VIRTUALITY_pure_virtual:
1976    return Introduced ? MethodKind::PureIntroducingVirtual
1977                      : MethodKind::PureVirtual;
1978  default:
1979    llvm_unreachable("unhandled virtuality case");
1980  }
1981
1982  return MethodKind::Vanilla;
1983}
1984
1985static TypeRecordKind getRecordKind(const DICompositeType *Ty) {
1986  switch (Ty->getTag()) {
1987  case dwarf::DW_TAG_class_type:     return TypeRecordKind::Class;
1988  case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct;
1989  }
1990  llvm_unreachable("unexpected tag");
1991}
1992
1993/// Return ClassOptions that should be present on both the forward declaration
1994/// and the defintion of a tag type.
1995static ClassOptions getCommonClassOptions(const DICompositeType *Ty) {
1996  ClassOptions CO = ClassOptions::None;
1997
1998  // MSVC always sets this flag, even for local types. Clang doesn't always
1999  // appear to give every type a linkage name, which may be problematic for us.
2000  // FIXME: Investigate the consequences of not following them here.
2001  if (!Ty->getIdentifier().empty())
2002    CO |= ClassOptions::HasUniqueName;
2003
2004  // Put the Nested flag on a type if it appears immediately inside a tag type.
2005  // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass
2006  // here. That flag is only set on definitions, and not forward declarations.
2007  const DIScope *ImmediateScope = Ty->getScope();
2008  if (ImmediateScope && isa<DICompositeType>(ImmediateScope))
2009    CO |= ClassOptions::Nested;
2010
2011  // Put the Scoped flag on function-local types. MSVC puts this flag for enum
2012  // type only when it has an immediate function scope. Clang never puts enums
2013  // inside DILexicalBlock scopes. Enum types, as generated by clang, are
2014  // always in function, class, or file scopes.
2015  if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) {
2016    if (ImmediateScope && isa<DISubprogram>(ImmediateScope))
2017      CO |= ClassOptions::Scoped;
2018  } else {
2019    for (const DIScope *Scope = ImmediateScope; Scope != nullptr;
2020         Scope = Scope->getScope()) {
2021      if (isa<DISubprogram>(Scope)) {
2022        CO |= ClassOptions::Scoped;
2023        break;
2024      }
2025    }
2026  }
2027
2028  return CO;
2029}
2030
2031void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) {
2032  switch (Ty->getTag()) {
2033  case dwarf::DW_TAG_class_type:
2034  case dwarf::DW_TAG_structure_type:
2035  case dwarf::DW_TAG_union_type:
2036  case dwarf::DW_TAG_enumeration_type:
2037    break;
2038  default:
2039    return;
2040  }
2041
2042  if (const auto *File = Ty->getFile()) {
2043    StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File));
2044    TypeIndex SIDI = TypeTable.writeLeafType(SIDR);
2045
2046    UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine());
2047    TypeTable.writeLeafType(USLR);
2048  }
2049}
2050
2051TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) {
2052  ClassOptions CO = getCommonClassOptions(Ty);
2053  TypeIndex FTI;
2054  unsigned EnumeratorCount = 0;
2055
2056  if (Ty->isForwardDecl()) {
2057    CO |= ClassOptions::ForwardReference;
2058  } else {
2059    ContinuationRecordBuilder ContinuationBuilder;
2060    ContinuationBuilder.begin(ContinuationRecordKind::FieldList);
2061    for (const DINode *Element : Ty->getElements()) {
2062      // We assume that the frontend provides all members in source declaration
2063      // order, which is what MSVC does.
2064      if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) {
2065        EnumeratorRecord ER(MemberAccess::Public,
2066                            APSInt::getUnsigned(Enumerator->getValue()),
2067                            Enumerator->getName());
2068        ContinuationBuilder.writeMemberType(ER);
2069        EnumeratorCount++;
2070      }
2071    }
2072    FTI = TypeTable.insertRecord(ContinuationBuilder);
2073  }
2074
2075  std::string FullName = getFullyQualifiedName(Ty);
2076
2077  EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(),
2078                getTypeIndex(Ty->getBaseType()));
2079  TypeIndex EnumTI = TypeTable.writeLeafType(ER);
2080
2081  addUDTSrcLine(Ty, EnumTI);
2082
2083  return EnumTI;
2084}
2085
2086//===----------------------------------------------------------------------===//
2087// ClassInfo
2088//===----------------------------------------------------------------------===//
2089
2090struct llvm::ClassInfo {
2091  struct MemberInfo {
2092    const DIDerivedType *MemberTypeNode;
2093    uint64_t BaseOffset;
2094  };
2095  // [MemberInfo]
2096  using MemberList = std::vector<MemberInfo>;
2097
2098  using MethodsList = TinyPtrVector<const DISubprogram *>;
2099  // MethodName -> MethodsList
2100  using MethodsMap = MapVector<MDString *, MethodsList>;
2101
2102  /// Base classes.
2103  std::vector<const DIDerivedType *> Inheritance;
2104
2105  /// Direct members.
2106  MemberList Members;
2107  // Direct overloaded methods gathered by name.
2108  MethodsMap Methods;
2109
2110  TypeIndex VShapeTI;
2111
2112  std::vector<const DIType *> NestedTypes;
2113};
2114
2115void CodeViewDebug::clear() {
2116  assert(CurFn == nullptr);
2117  FileIdMap.clear();
2118  FnDebugInfo.clear();
2119  FileToFilepathMap.clear();
2120  LocalUDTs.clear();
2121  GlobalUDTs.clear();
2122  TypeIndices.clear();
2123  CompleteTypeIndices.clear();
2124  ScopeGlobals.clear();
2125}
2126
2127void CodeViewDebug::collectMemberInfo(ClassInfo &Info,
2128                                      const DIDerivedType *DDTy) {
2129  if (!DDTy->getName().empty()) {
2130    Info.Members.push_back({DDTy, 0});
2131    return;
2132  }
2133
2134  // An unnamed member may represent a nested struct or union. Attempt to
2135  // interpret the unnamed member as a DICompositeType possibly wrapped in
2136  // qualifier types. Add all the indirect fields to the current record if that
2137  // succeeds, and drop the member if that fails.
2138  assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!");
2139  uint64_t Offset = DDTy->getOffsetInBits();
2140  const DIType *Ty = DDTy->getBaseType();
2141  bool FullyResolved = false;
2142  while (!FullyResolved) {
2143    switch (Ty->getTag()) {
2144    case dwarf::DW_TAG_const_type:
2145    case dwarf::DW_TAG_volatile_type:
2146      // FIXME: we should apply the qualifier types to the indirect fields
2147      // rather than dropping them.
2148      Ty = cast<DIDerivedType>(Ty)->getBaseType();
2149      break;
2150    default:
2151      FullyResolved = true;
2152      break;
2153    }
2154  }
2155
2156  const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty);
2157  if (!DCTy)
2158    return;
2159
2160  ClassInfo NestedInfo = collectClassInfo(DCTy);
2161  for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members)
2162    Info.Members.push_back(
2163        {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset});
2164}
2165
2166ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) {
2167  ClassInfo Info;
2168  // Add elements to structure type.
2169  DINodeArray Elements = Ty->getElements();
2170  for (auto *Element : Elements) {
2171    // We assume that the frontend provides all members in source declaration
2172    // order, which is what MSVC does.
2173    if (!Element)
2174      continue;
2175    if (auto *SP = dyn_cast<DISubprogram>(Element)) {
2176      Info.Methods[SP->getRawName()].push_back(SP);
2177    } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) {
2178      if (DDTy->getTag() == dwarf::DW_TAG_member) {
2179        collectMemberInfo(Info, DDTy);
2180      } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) {
2181        Info.Inheritance.push_back(DDTy);
2182      } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type &&
2183                 DDTy->getName() == "__vtbl_ptr_type") {
2184        Info.VShapeTI = getTypeIndex(DDTy);
2185      } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) {
2186        Info.NestedTypes.push_back(DDTy);
2187      } else if (DDTy->getTag() == dwarf::DW_TAG_friend) {
2188        // Ignore friend members. It appears that MSVC emitted info about
2189        // friends in the past, but modern versions do not.
2190      }
2191    } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) {
2192      Info.NestedTypes.push_back(Composite);
2193    }
2194    // Skip other unrecognized kinds of elements.
2195  }
2196  return Info;
2197}
2198
2199static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) {
2200  // This routine is used by lowerTypeClass and lowerTypeUnion to determine
2201  // if a complete type should be emitted instead of a forward reference.
2202  return Ty->getName().empty() && Ty->getIdentifier().empty() &&
2203      !Ty->isForwardDecl();
2204}
2205
2206TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) {
2207  // Emit the complete type for unnamed structs.  C++ classes with methods
2208  // which have a circular reference back to the class type are expected to
2209  // be named by the front-end and should not be "unnamed".  C unnamed
2210  // structs should not have circular references.
2211  if (shouldAlwaysEmitCompleteClassType(Ty)) {
2212    // If this unnamed complete type is already in the process of being defined
2213    // then the description of the type is malformed and cannot be emitted
2214    // into CodeView correctly so report a fatal error.
2215    auto I = CompleteTypeIndices.find(Ty);
2216    if (I != CompleteTypeIndices.end() && I->second == TypeIndex())
2217      report_fatal_error("cannot debug circular reference to unnamed type");
2218    return getCompleteTypeIndex(Ty);
2219  }
2220
2221  // First, construct the forward decl.  Don't look into Ty to compute the
2222  // forward decl options, since it might not be available in all TUs.
2223  TypeRecordKind Kind = getRecordKind(Ty);
2224  ClassOptions CO =
2225      ClassOptions::ForwardReference | getCommonClassOptions(Ty);
2226  std::string FullName = getFullyQualifiedName(Ty);
2227  ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0,
2228                 FullName, Ty->getIdentifier());
2229  TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR);
2230  if (!Ty->isForwardDecl())
2231    DeferredCompleteTypes.push_back(Ty);
2232  return FwdDeclTI;
2233}
2234
2235TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) {
2236  // Construct the field list and complete type record.
2237  TypeRecordKind Kind = getRecordKind(Ty);
2238  ClassOptions CO = getCommonClassOptions(Ty);
2239  TypeIndex FieldTI;
2240  TypeIndex VShapeTI;
2241  unsigned FieldCount;
2242  bool ContainsNestedClass;
2243  std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) =
2244      lowerRecordFieldList(Ty);
2245
2246  if (ContainsNestedClass)
2247    CO |= ClassOptions::ContainsNestedClass;
2248
2249  // MSVC appears to set this flag by searching any destructor or method with
2250  // FunctionOptions::Constructor among the emitted members. Clang AST has all
2251  // the members, however special member functions are not yet emitted into
2252  // debug information. For now checking a class's non-triviality seems enough.
2253  // FIXME: not true for a nested unnamed struct.
2254  if (isNonTrivial(Ty))
2255    CO |= ClassOptions::HasConstructorOrDestructor;
2256
2257  std::string FullName = getFullyQualifiedName(Ty);
2258
2259  uint64_t SizeInBytes = Ty->getSizeInBits() / 8;
2260
2261  ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI,
2262                 SizeInBytes, FullName, Ty->getIdentifier());
2263  TypeIndex ClassTI = TypeTable.writeLeafType(CR);
2264
2265  addUDTSrcLine(Ty, ClassTI);
2266
2267  addToUDTs(Ty);
2268
2269  return ClassTI;
2270}
2271
2272TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) {
2273  // Emit the complete type for unnamed unions.
2274  if (shouldAlwaysEmitCompleteClassType(Ty))
2275    return getCompleteTypeIndex(Ty);
2276
2277  ClassOptions CO =
2278      ClassOptions::ForwardReference | getCommonClassOptions(Ty);
2279  std::string FullName = getFullyQualifiedName(Ty);
2280  UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier());
2281  TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR);
2282  if (!Ty->isForwardDecl())
2283    DeferredCompleteTypes.push_back(Ty);
2284  return FwdDeclTI;
2285}
2286
2287TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) {
2288  ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty);
2289  TypeIndex FieldTI;
2290  unsigned FieldCount;
2291  bool ContainsNestedClass;
2292  std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) =
2293      lowerRecordFieldList(Ty);
2294
2295  if (ContainsNestedClass)
2296    CO |= ClassOptions::ContainsNestedClass;
2297
2298  uint64_t SizeInBytes = Ty->getSizeInBits() / 8;
2299  std::string FullName = getFullyQualifiedName(Ty);
2300
2301  UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName,
2302                 Ty->getIdentifier());
2303  TypeIndex UnionTI = TypeTable.writeLeafType(UR);
2304
2305  addUDTSrcLine(Ty, UnionTI);
2306
2307  addToUDTs(Ty);
2308
2309  return UnionTI;
2310}
2311
2312std::tuple<TypeIndex, TypeIndex, unsigned, bool>
2313CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) {
2314  // Manually count members. MSVC appears to count everything that generates a
2315  // field list record. Each individual overload in a method overload group
2316  // contributes to this count, even though the overload group is a single field
2317  // list record.
2318  unsigned MemberCount = 0;
2319  ClassInfo Info = collectClassInfo(Ty);
2320  ContinuationRecordBuilder ContinuationBuilder;
2321  ContinuationBuilder.begin(ContinuationRecordKind::FieldList);
2322
2323  // Create base classes.
2324  for (const DIDerivedType *I : Info.Inheritance) {
2325    if (I->getFlags() & DINode::FlagVirtual) {
2326      // Virtual base.
2327      unsigned VBPtrOffset = I->getVBPtrOffset();
2328      // FIXME: Despite the accessor name, the offset is really in bytes.
2329      unsigned VBTableIndex = I->getOffsetInBits() / 4;
2330      auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase
2331                            ? TypeRecordKind::IndirectVirtualBaseClass
2332                            : TypeRecordKind::VirtualBaseClass;
2333      VirtualBaseClassRecord VBCR(
2334          RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()),
2335          getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset,
2336          VBTableIndex);
2337
2338      ContinuationBuilder.writeMemberType(VBCR);
2339      MemberCount++;
2340    } else {
2341      assert(I->getOffsetInBits() % 8 == 0 &&
2342             "bases must be on byte boundaries");
2343      BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()),
2344                          getTypeIndex(I->getBaseType()),
2345                          I->getOffsetInBits() / 8);
2346      ContinuationBuilder.writeMemberType(BCR);
2347      MemberCount++;
2348    }
2349  }
2350
2351  // Create members.
2352  for (ClassInfo::MemberInfo &MemberInfo : Info.Members) {
2353    const DIDerivedType *Member = MemberInfo.MemberTypeNode;
2354    TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType());
2355    StringRef MemberName = Member->getName();
2356    MemberAccess Access =
2357        translateAccessFlags(Ty->getTag(), Member->getFlags());
2358
2359    if (Member->isStaticMember()) {
2360      StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName);
2361      ContinuationBuilder.writeMemberType(SDMR);
2362      MemberCount++;
2363      continue;
2364    }
2365
2366    // Virtual function pointer member.
2367    if ((Member->getFlags() & DINode::FlagArtificial) &&
2368        Member->getName().startswith("_vptr$")) {
2369      VFPtrRecord VFPR(getTypeIndex(Member->getBaseType()));
2370      ContinuationBuilder.writeMemberType(VFPR);
2371      MemberCount++;
2372      continue;
2373    }
2374
2375    // Data member.
2376    uint64_t MemberOffsetInBits =
2377        Member->getOffsetInBits() + MemberInfo.BaseOffset;
2378    if (Member->isBitField()) {
2379      uint64_t StartBitOffset = MemberOffsetInBits;
2380      if (const auto *CI =
2381              dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) {
2382        MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset;
2383      }
2384      StartBitOffset -= MemberOffsetInBits;
2385      BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(),
2386                         StartBitOffset);
2387      MemberBaseType = TypeTable.writeLeafType(BFR);
2388    }
2389    uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8;
2390    DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes,
2391                         MemberName);
2392    ContinuationBuilder.writeMemberType(DMR);
2393    MemberCount++;
2394  }
2395
2396  // Create methods
2397  for (auto &MethodItr : Info.Methods) {
2398    StringRef Name = MethodItr.first->getString();
2399
2400    std::vector<OneMethodRecord> Methods;
2401    for (const DISubprogram *SP : MethodItr.second) {
2402      TypeIndex MethodType = getMemberFunctionType(SP, Ty);
2403      bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual;
2404
2405      unsigned VFTableOffset = -1;
2406      if (Introduced)
2407        VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes();
2408
2409      Methods.push_back(OneMethodRecord(
2410          MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()),
2411          translateMethodKindFlags(SP, Introduced),
2412          translateMethodOptionFlags(SP), VFTableOffset, Name));
2413      MemberCount++;
2414    }
2415    assert(!Methods.empty() && "Empty methods map entry");
2416    if (Methods.size() == 1)
2417      ContinuationBuilder.writeMemberType(Methods[0]);
2418    else {
2419      // FIXME: Make this use its own ContinuationBuilder so that
2420      // MethodOverloadList can be split correctly.
2421      MethodOverloadListRecord MOLR(Methods);
2422      TypeIndex MethodList = TypeTable.writeLeafType(MOLR);
2423
2424      OverloadedMethodRecord OMR(Methods.size(), MethodList, Name);
2425      ContinuationBuilder.writeMemberType(OMR);
2426    }
2427  }
2428
2429  // Create nested classes.
2430  for (const DIType *Nested : Info.NestedTypes) {
2431    NestedTypeRecord R(getTypeIndex(Nested), Nested->getName());
2432    ContinuationBuilder.writeMemberType(R);
2433    MemberCount++;
2434  }
2435
2436  TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder);
2437  return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount,
2438                         !Info.NestedTypes.empty());
2439}
2440
2441TypeIndex CodeViewDebug::getVBPTypeIndex() {
2442  if (!VBPType.getIndex()) {
2443    // Make a 'const int *' type.
2444    ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const);
2445    TypeIndex ModifiedTI = TypeTable.writeLeafType(MR);
2446
2447    PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64
2448                                                  : PointerKind::Near32;
2449    PointerMode PM = PointerMode::Pointer;
2450    PointerOptions PO = PointerOptions::None;
2451    PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes());
2452    VBPType = TypeTable.writeLeafType(PR);
2453  }
2454
2455  return VBPType;
2456}
2457
2458TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) {
2459  // The null DIType is the void type. Don't try to hash it.
2460  if (!Ty)
2461    return TypeIndex::Void();
2462
2463  // Check if we've already translated this type. Don't try to do a
2464  // get-or-create style insertion that caches the hash lookup across the
2465  // lowerType call. It will update the TypeIndices map.
2466  auto I = TypeIndices.find({Ty, ClassTy});
2467  if (I != TypeIndices.end())
2468    return I->second;
2469
2470  TypeLoweringScope S(*this);
2471  TypeIndex TI = lowerType(Ty, ClassTy);
2472  return recordTypeIndexForDINode(Ty, TI, ClassTy);
2473}
2474
2475codeview::TypeIndex
2476CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy,
2477                                      const DISubroutineType *SubroutineTy) {
2478  assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type &&
2479         "this type must be a pointer type");
2480
2481  PointerOptions Options = PointerOptions::None;
2482  if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference)
2483    Options = PointerOptions::LValueRefThisPointer;
2484  else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference)
2485    Options = PointerOptions::RValueRefThisPointer;
2486
2487  // Check if we've already translated this type.  If there is no ref qualifier
2488  // on the function then we look up this pointer type with no associated class
2489  // so that the TypeIndex for the this pointer can be shared with the type
2490  // index for other pointers to this class type.  If there is a ref qualifier
2491  // then we lookup the pointer using the subroutine as the parent type.
2492  auto I = TypeIndices.find({PtrTy, SubroutineTy});
2493  if (I != TypeIndices.end())
2494    return I->second;
2495
2496  TypeLoweringScope S(*this);
2497  TypeIndex TI = lowerTypePointer(PtrTy, Options);
2498  return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy);
2499}
2500
2501TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) {
2502  PointerRecord PR(getTypeIndex(Ty),
2503                   getPointerSizeInBytes() == 8 ? PointerKind::Near64
2504                                                : PointerKind::Near32,
2505                   PointerMode::LValueReference, PointerOptions::None,
2506                   Ty->getSizeInBits() / 8);
2507  return TypeTable.writeLeafType(PR);
2508}
2509
2510TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) {
2511  // The null DIType is the void type. Don't try to hash it.
2512  if (!Ty)
2513    return TypeIndex::Void();
2514
2515  // Look through typedefs when getting the complete type index. Call
2516  // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are
2517  // emitted only once.
2518  if (Ty->getTag() == dwarf::DW_TAG_typedef)
2519    (void)getTypeIndex(Ty);
2520  while (Ty->getTag() == dwarf::DW_TAG_typedef)
2521    Ty = cast<DIDerivedType>(Ty)->getBaseType();
2522
2523  // If this is a non-record type, the complete type index is the same as the
2524  // normal type index. Just call getTypeIndex.
2525  switch (Ty->getTag()) {
2526  case dwarf::DW_TAG_class_type:
2527  case dwarf::DW_TAG_structure_type:
2528  case dwarf::DW_TAG_union_type:
2529    break;
2530  default:
2531    return getTypeIndex(Ty);
2532  }
2533
2534  const auto *CTy = cast<DICompositeType>(Ty);
2535
2536  TypeLoweringScope S(*this);
2537
2538  // Make sure the forward declaration is emitted first. It's unclear if this
2539  // is necessary, but MSVC does it, and we should follow suit until we can show
2540  // otherwise.
2541  // We only emit a forward declaration for named types.
2542  if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) {
2543    TypeIndex FwdDeclTI = getTypeIndex(CTy);
2544
2545    // Just use the forward decl if we don't have complete type info. This
2546    // might happen if the frontend is using modules and expects the complete
2547    // definition to be emitted elsewhere.
2548    if (CTy->isForwardDecl())
2549      return FwdDeclTI;
2550  }
2551
2552  // Check if we've already translated the complete record type.
2553  // Insert the type with a null TypeIndex to signify that the type is currently
2554  // being lowered.
2555  auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()});
2556  if (!InsertResult.second)
2557    return InsertResult.first->second;
2558
2559  TypeIndex TI;
2560  switch (CTy->getTag()) {
2561  case dwarf::DW_TAG_class_type:
2562  case dwarf::DW_TAG_structure_type:
2563    TI = lowerCompleteTypeClass(CTy);
2564    break;
2565  case dwarf::DW_TAG_union_type:
2566    TI = lowerCompleteTypeUnion(CTy);
2567    break;
2568  default:
2569    llvm_unreachable("not a record");
2570  }
2571
2572  // Update the type index associated with this CompositeType.  This cannot
2573  // use the 'InsertResult' iterator above because it is potentially
2574  // invalidated by map insertions which can occur while lowering the class
2575  // type above.
2576  CompleteTypeIndices[CTy] = TI;
2577  return TI;
2578}
2579
2580/// Emit all the deferred complete record types. Try to do this in FIFO order,
2581/// and do this until fixpoint, as each complete record type typically
2582/// references
2583/// many other record types.
2584void CodeViewDebug::emitDeferredCompleteTypes() {
2585  SmallVector<const DICompositeType *, 4> TypesToEmit;
2586  while (!DeferredCompleteTypes.empty()) {
2587    std::swap(DeferredCompleteTypes, TypesToEmit);
2588    for (const DICompositeType *RecordTy : TypesToEmit)
2589      getCompleteTypeIndex(RecordTy);
2590    TypesToEmit.clear();
2591  }
2592}
2593
2594void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI,
2595                                          ArrayRef<LocalVariable> Locals) {
2596  // Get the sorted list of parameters and emit them first.
2597  SmallVector<const LocalVariable *, 6> Params;
2598  for (const LocalVariable &L : Locals)
2599    if (L.DIVar->isParameter())
2600      Params.push_back(&L);
2601  llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) {
2602    return L->DIVar->getArg() < R->DIVar->getArg();
2603  });
2604  for (const LocalVariable *L : Params)
2605    emitLocalVariable(FI, *L);
2606
2607  // Next emit all non-parameters in the order that we found them.
2608  for (const LocalVariable &L : Locals)
2609    if (!L.DIVar->isParameter())
2610      emitLocalVariable(FI, L);
2611}
2612
2613void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI,
2614                                      const LocalVariable &Var) {
2615  // LocalSym record, see SymbolRecord.h for more info.
2616  MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL);
2617
2618  LocalSymFlags Flags = LocalSymFlags::None;
2619  if (Var.DIVar->isParameter())
2620    Flags |= LocalSymFlags::IsParameter;
2621  if (Var.DefRanges.empty())
2622    Flags |= LocalSymFlags::IsOptimizedOut;
2623
2624  OS.AddComment("TypeIndex");
2625  TypeIndex TI = Var.UseReferenceType
2626                     ? getTypeIndexForReferenceTo(Var.DIVar->getType())
2627                     : getCompleteTypeIndex(Var.DIVar->getType());
2628  OS.EmitIntValue(TI.getIndex(), 4);
2629  OS.AddComment("Flags");
2630  OS.EmitIntValue(static_cast<uint16_t>(Flags), 2);
2631  // Truncate the name so we won't overflow the record length field.
2632  emitNullTerminatedSymbolName(OS, Var.DIVar->getName());
2633  endSymbolRecord(LocalEnd);
2634
2635  // Calculate the on disk prefix of the appropriate def range record. The
2636  // records and on disk formats are described in SymbolRecords.h. BytePrefix
2637  // should be big enough to hold all forms without memory allocation.
2638  SmallString<20> BytePrefix;
2639  for (const LocalVarDefRange &DefRange : Var.DefRanges) {
2640    BytePrefix.clear();
2641    if (DefRange.InMemory) {
2642      int Offset = DefRange.DataOffset;
2643      unsigned Reg = DefRange.CVRegister;
2644
2645      // 32-bit x86 call sequences often use PUSH instructions, which disrupt
2646      // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0,
2647      // instead. In frames without stack realignment, $T0 will be the CFA.
2648      if (RegisterId(Reg) == RegisterId::ESP) {
2649        Reg = unsigned(RegisterId::VFRAME);
2650        Offset += FI.OffsetAdjustment;
2651      }
2652
2653      // If we can use the chosen frame pointer for the frame and this isn't a
2654      // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record.
2655      // Otherwise, use S_DEFRANGE_REGISTER_REL.
2656      EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU);
2657      if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None &&
2658          (bool(Flags & LocalSymFlags::IsParameter)
2659               ? (EncFP == FI.EncodedParamFramePtrReg)
2660               : (EncFP == FI.EncodedLocalFramePtrReg))) {
2661        DefRangeFramePointerRelHeader DRHdr;
2662        DRHdr.Offset = Offset;
2663        OS.EmitCVDefRangeDirective(DefRange.Ranges, DRHdr);
2664      } else {
2665        uint16_t RegRelFlags = 0;
2666        if (DefRange.IsSubfield) {
2667          RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag |
2668                        (DefRange.StructOffset
2669                         << DefRangeRegisterRelSym::OffsetInParentShift);
2670        }
2671        DefRangeRegisterRelHeader DRHdr;
2672        DRHdr.Register = Reg;
2673        DRHdr.Flags = RegRelFlags;
2674        DRHdr.BasePointerOffset = Offset;
2675        OS.EmitCVDefRangeDirective(DefRange.Ranges, DRHdr);
2676      }
2677    } else {
2678      assert(DefRange.DataOffset == 0 && "unexpected offset into register");
2679      if (DefRange.IsSubfield) {
2680        DefRangeSubfieldRegisterHeader DRHdr;
2681        DRHdr.Register = DefRange.CVRegister;
2682        DRHdr.MayHaveNoName = 0;
2683        DRHdr.OffsetInParent = DefRange.StructOffset;
2684        OS.EmitCVDefRangeDirective(DefRange.Ranges, DRHdr);
2685      } else {
2686        DefRangeRegisterHeader DRHdr;
2687        DRHdr.Register = DefRange.CVRegister;
2688        DRHdr.MayHaveNoName = 0;
2689        OS.EmitCVDefRangeDirective(DefRange.Ranges, DRHdr);
2690      }
2691    }
2692  }
2693}
2694
2695void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks,
2696                                         const FunctionInfo& FI) {
2697  for (LexicalBlock *Block : Blocks)
2698    emitLexicalBlock(*Block, FI);
2699}
2700
2701/// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a
2702/// lexical block scope.
2703void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block,
2704                                     const FunctionInfo& FI) {
2705  MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32);
2706  OS.AddComment("PtrParent");
2707  OS.EmitIntValue(0, 4);                                  // PtrParent
2708  OS.AddComment("PtrEnd");
2709  OS.EmitIntValue(0, 4);                                  // PtrEnd
2710  OS.AddComment("Code size");
2711  OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4);   // Code Size
2712  OS.AddComment("Function section relative address");
2713  OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0);         // Func Offset
2714  OS.AddComment("Function section index");
2715  OS.EmitCOFFSectionIndex(FI.Begin);                      // Func Symbol
2716  OS.AddComment("Lexical block name");
2717  emitNullTerminatedSymbolName(OS, Block.Name);           // Name
2718  endSymbolRecord(RecordEnd);
2719
2720  // Emit variables local to this lexical block.
2721  emitLocalVariableList(FI, Block.Locals);
2722  emitGlobalVariableList(Block.Globals);
2723
2724  // Emit lexical blocks contained within this block.
2725  emitLexicalBlockList(Block.Children, FI);
2726
2727  // Close the lexical block scope.
2728  emitEndSymbolRecord(SymbolKind::S_END);
2729}
2730
2731/// Convenience routine for collecting lexical block information for a list
2732/// of lexical scopes.
2733void CodeViewDebug::collectLexicalBlockInfo(
2734        SmallVectorImpl<LexicalScope *> &Scopes,
2735        SmallVectorImpl<LexicalBlock *> &Blocks,
2736        SmallVectorImpl<LocalVariable> &Locals,
2737        SmallVectorImpl<CVGlobalVariable> &Globals) {
2738  for (LexicalScope *Scope : Scopes)
2739    collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals);
2740}
2741
2742/// Populate the lexical blocks and local variable lists of the parent with
2743/// information about the specified lexical scope.
2744void CodeViewDebug::collectLexicalBlockInfo(
2745    LexicalScope &Scope,
2746    SmallVectorImpl<LexicalBlock *> &ParentBlocks,
2747    SmallVectorImpl<LocalVariable> &ParentLocals,
2748    SmallVectorImpl<CVGlobalVariable> &ParentGlobals) {
2749  if (Scope.isAbstractScope())
2750    return;
2751
2752  // Gather information about the lexical scope including local variables,
2753  // global variables, and address ranges.
2754  bool IgnoreScope = false;
2755  auto LI = ScopeVariables.find(&Scope);
2756  SmallVectorImpl<LocalVariable> *Locals =
2757      LI != ScopeVariables.end() ? &LI->second : nullptr;
2758  auto GI = ScopeGlobals.find(Scope.getScopeNode());
2759  SmallVectorImpl<CVGlobalVariable> *Globals =
2760      GI != ScopeGlobals.end() ? GI->second.get() : nullptr;
2761  const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode());
2762  const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges();
2763
2764  // Ignore lexical scopes which do not contain variables.
2765  if (!Locals && !Globals)
2766    IgnoreScope = true;
2767
2768  // Ignore lexical scopes which are not lexical blocks.
2769  if (!DILB)
2770    IgnoreScope = true;
2771
2772  // Ignore scopes which have too many address ranges to represent in the
2773  // current CodeView format or do not have a valid address range.
2774  //
2775  // For lexical scopes with multiple address ranges you may be tempted to
2776  // construct a single range covering every instruction where the block is
2777  // live and everything in between.  Unfortunately, Visual Studio only
2778  // displays variables from the first matching lexical block scope.  If the
2779  // first lexical block contains exception handling code or cold code which
2780  // is moved to the bottom of the routine creating a single range covering
2781  // nearly the entire routine, then it will hide all other lexical blocks
2782  // and the variables they contain.
2783  if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second))
2784    IgnoreScope = true;
2785
2786  if (IgnoreScope) {
2787    // This scope can be safely ignored and eliminating it will reduce the
2788    // size of the debug information. Be sure to collect any variable and scope
2789    // information from the this scope or any of its children and collapse them
2790    // into the parent scope.
2791    if (Locals)
2792      ParentLocals.append(Locals->begin(), Locals->end());
2793    if (Globals)
2794      ParentGlobals.append(Globals->begin(), Globals->end());
2795    collectLexicalBlockInfo(Scope.getChildren(),
2796                            ParentBlocks,
2797                            ParentLocals,
2798                            ParentGlobals);
2799    return;
2800  }
2801
2802  // Create a new CodeView lexical block for this lexical scope.  If we've
2803  // seen this DILexicalBlock before then the scope tree is malformed and
2804  // we can handle this gracefully by not processing it a second time.
2805  auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()});
2806  if (!BlockInsertion.second)
2807    return;
2808
2809  // Create a lexical block containing the variables and collect the the
2810  // lexical block information for the children.
2811  const InsnRange &Range = Ranges.front();
2812  assert(Range.first && Range.second);
2813  LexicalBlock &Block = BlockInsertion.first->second;
2814  Block.Begin = getLabelBeforeInsn(Range.first);
2815  Block.End = getLabelAfterInsn(Range.second);
2816  assert(Block.Begin && "missing label for scope begin");
2817  assert(Block.End && "missing label for scope end");
2818  Block.Name = DILB->getName();
2819  if (Locals)
2820    Block.Locals = std::move(*Locals);
2821  if (Globals)
2822    Block.Globals = std::move(*Globals);
2823  ParentBlocks.push_back(&Block);
2824  collectLexicalBlockInfo(Scope.getChildren(),
2825                          Block.Children,
2826                          Block.Locals,
2827                          Block.Globals);
2828}
2829
2830void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) {
2831  const Function &GV = MF->getFunction();
2832  assert(FnDebugInfo.count(&GV));
2833  assert(CurFn == FnDebugInfo[&GV].get());
2834
2835  collectVariableInfo(GV.getSubprogram());
2836
2837  // Build the lexical block structure to emit for this routine.
2838  if (LexicalScope *CFS = LScopes.getCurrentFunctionScope())
2839    collectLexicalBlockInfo(*CFS,
2840                            CurFn->ChildBlocks,
2841                            CurFn->Locals,
2842                            CurFn->Globals);
2843
2844  // Clear the scope and variable information from the map which will not be
2845  // valid after we have finished processing this routine.  This also prepares
2846  // the map for the subsequent routine.
2847  ScopeVariables.clear();
2848
2849  // Don't emit anything if we don't have any line tables.
2850  // Thunks are compiler-generated and probably won't have source correlation.
2851  if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) {
2852    FnDebugInfo.erase(&GV);
2853    CurFn = nullptr;
2854    return;
2855  }
2856
2857  // Find heap alloc sites and add to list.
2858  for (const auto &MBB : *MF) {
2859    for (const auto &MI : MBB) {
2860      if (MDNode *MD = MI.getHeapAllocMarker()) {
2861        CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI),
2862                                                        getLabelAfterInsn(&MI),
2863                                                        dyn_cast<DIType>(MD)));
2864      }
2865    }
2866  }
2867
2868  CurFn->Annotations = MF->getCodeViewAnnotations();
2869
2870  CurFn->End = Asm->getFunctionEnd();
2871
2872  CurFn = nullptr;
2873}
2874
2875// Usable locations are valid with non-zero line numbers. A line number of zero
2876// corresponds to optimized code that doesn't have a distinct source location.
2877// In this case, we try to use the previous or next source location depending on
2878// the context.
2879static bool isUsableDebugLoc(DebugLoc DL) {
2880  return DL && DL.getLine() != 0;
2881}
2882
2883void CodeViewDebug::beginInstruction(const MachineInstr *MI) {
2884  DebugHandlerBase::beginInstruction(MI);
2885
2886  // Ignore DBG_VALUE and DBG_LABEL locations and function prologue.
2887  if (!Asm || !CurFn || MI->isDebugInstr() ||
2888      MI->getFlag(MachineInstr::FrameSetup))
2889    return;
2890
2891  // If the first instruction of a new MBB has no location, find the first
2892  // instruction with a location and use that.
2893  DebugLoc DL = MI->getDebugLoc();
2894  if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) {
2895    for (const auto &NextMI : *MI->getParent()) {
2896      if (NextMI.isDebugInstr())
2897        continue;
2898      DL = NextMI.getDebugLoc();
2899      if (isUsableDebugLoc(DL))
2900        break;
2901    }
2902    // FIXME: Handle the case where the BB has no valid locations. This would
2903    // probably require doing a real dataflow analysis.
2904  }
2905  PrevInstBB = MI->getParent();
2906
2907  // If we still don't have a debug location, don't record a location.
2908  if (!isUsableDebugLoc(DL))
2909    return;
2910
2911  maybeRecordLocation(DL, Asm->MF);
2912}
2913
2914MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) {
2915  MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(),
2916           *EndLabel = MMI->getContext().createTempSymbol();
2917  OS.EmitIntValue(unsigned(Kind), 4);
2918  OS.AddComment("Subsection size");
2919  OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4);
2920  OS.EmitLabel(BeginLabel);
2921  return EndLabel;
2922}
2923
2924void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) {
2925  OS.EmitLabel(EndLabel);
2926  // Every subsection must be aligned to a 4-byte boundary.
2927  OS.EmitValueToAlignment(4);
2928}
2929
2930static StringRef getSymbolName(SymbolKind SymKind) {
2931  for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames())
2932    if (EE.Value == SymKind)
2933      return EE.Name;
2934  return "";
2935}
2936
2937MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) {
2938  MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(),
2939           *EndLabel = MMI->getContext().createTempSymbol();
2940  OS.AddComment("Record length");
2941  OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2);
2942  OS.EmitLabel(BeginLabel);
2943  if (OS.isVerboseAsm())
2944    OS.AddComment("Record kind: " + getSymbolName(SymKind));
2945  OS.EmitIntValue(unsigned(SymKind), 2);
2946  return EndLabel;
2947}
2948
2949void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) {
2950  // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid
2951  // an extra copy of every symbol record in LLD. This increases object file
2952  // size by less than 1% in the clang build, and is compatible with the Visual
2953  // C++ linker.
2954  OS.EmitValueToAlignment(4);
2955  OS.EmitLabel(SymEnd);
2956}
2957
2958void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) {
2959  OS.AddComment("Record length");
2960  OS.EmitIntValue(2, 2);
2961  if (OS.isVerboseAsm())
2962    OS.AddComment("Record kind: " + getSymbolName(EndKind));
2963  OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind
2964}
2965
2966void CodeViewDebug::emitDebugInfoForUDTs(
2967    ArrayRef<std::pair<std::string, const DIType *>> UDTs) {
2968  for (const auto &UDT : UDTs) {
2969    const DIType *T = UDT.second;
2970    assert(shouldEmitUdt(T));
2971
2972    MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT);
2973    OS.AddComment("Type");
2974    OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4);
2975    emitNullTerminatedSymbolName(OS, UDT.first);
2976    endSymbolRecord(UDTRecordEnd);
2977  }
2978}
2979
2980void CodeViewDebug::collectGlobalVariableInfo() {
2981  DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *>
2982      GlobalMap;
2983  for (const GlobalVariable &GV : MMI->getModule()->globals()) {
2984    SmallVector<DIGlobalVariableExpression *, 1> GVEs;
2985    GV.getDebugInfo(GVEs);
2986    for (const auto *GVE : GVEs)
2987      GlobalMap[GVE] = &GV;
2988  }
2989
2990  NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
2991  for (const MDNode *Node : CUs->operands()) {
2992    const auto *CU = cast<DICompileUnit>(Node);
2993    for (const auto *GVE : CU->getGlobalVariables()) {
2994      const DIGlobalVariable *DIGV = GVE->getVariable();
2995      const DIExpression *DIE = GVE->getExpression();
2996
2997      // Emit constant global variables in a global symbol section.
2998      if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) {
2999        CVGlobalVariable CVGV = {DIGV, DIE};
3000        GlobalVariables.emplace_back(std::move(CVGV));
3001      }
3002
3003      const auto *GV = GlobalMap.lookup(GVE);
3004      if (!GV || GV->isDeclarationForLinker())
3005        continue;
3006
3007      DIScope *Scope = DIGV->getScope();
3008      SmallVector<CVGlobalVariable, 1> *VariableList;
3009      if (Scope && isa<DILocalScope>(Scope)) {
3010        // Locate a global variable list for this scope, creating one if
3011        // necessary.
3012        auto Insertion = ScopeGlobals.insert(
3013            {Scope, std::unique_ptr<GlobalVariableList>()});
3014        if (Insertion.second)
3015          Insertion.first->second = std::make_unique<GlobalVariableList>();
3016        VariableList = Insertion.first->second.get();
3017      } else if (GV->hasComdat())
3018        // Emit this global variable into a COMDAT section.
3019        VariableList = &ComdatVariables;
3020      else
3021        // Emit this global variable in a single global symbol section.
3022        VariableList = &GlobalVariables;
3023      CVGlobalVariable CVGV = {DIGV, GV};
3024      VariableList->emplace_back(std::move(CVGV));
3025    }
3026  }
3027}
3028
3029void CodeViewDebug::emitDebugInfoForGlobals() {
3030  // First, emit all globals that are not in a comdat in a single symbol
3031  // substream. MSVC doesn't like it if the substream is empty, so only open
3032  // it if we have at least one global to emit.
3033  switchToDebugSectionForSymbol(nullptr);
3034  if (!GlobalVariables.empty()) {
3035    OS.AddComment("Symbol subsection for globals");
3036    MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols);
3037    emitGlobalVariableList(GlobalVariables);
3038    endCVSubsection(EndLabel);
3039  }
3040
3041  // Second, emit each global that is in a comdat into its own .debug$S
3042  // section along with its own symbol substream.
3043  for (const CVGlobalVariable &CVGV : ComdatVariables) {
3044    const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>();
3045    MCSymbol *GVSym = Asm->getSymbol(GV);
3046    OS.AddComment("Symbol subsection for " +
3047                  Twine(GlobalValue::dropLLVMManglingEscape(GV->getName())));
3048    switchToDebugSectionForSymbol(GVSym);
3049    MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols);
3050    // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions.
3051    emitDebugInfoForGlobal(CVGV);
3052    endCVSubsection(EndLabel);
3053  }
3054}
3055
3056void CodeViewDebug::emitDebugInfoForRetainedTypes() {
3057  NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
3058  for (const MDNode *Node : CUs->operands()) {
3059    for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) {
3060      if (DIType *RT = dyn_cast<DIType>(Ty)) {
3061        getTypeIndex(RT);
3062        // FIXME: Add to global/local DTU list.
3063      }
3064    }
3065  }
3066}
3067
3068// Emit each global variable in the specified array.
3069void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) {
3070  for (const CVGlobalVariable &CVGV : Globals) {
3071    // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions.
3072    emitDebugInfoForGlobal(CVGV);
3073  }
3074}
3075
3076void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) {
3077  const DIGlobalVariable *DIGV = CVGV.DIGV;
3078  if (const GlobalVariable *GV =
3079          CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) {
3080    // DataSym record, see SymbolRecord.h for more info. Thread local data
3081    // happens to have the same format as global data.
3082    MCSymbol *GVSym = Asm->getSymbol(GV);
3083    SymbolKind DataSym = GV->isThreadLocal()
3084                             ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32
3085                                                      : SymbolKind::S_GTHREAD32)
3086                             : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32
3087                                                      : SymbolKind::S_GDATA32);
3088    MCSymbol *DataEnd = beginSymbolRecord(DataSym);
3089    OS.AddComment("Type");
3090    OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4);
3091    OS.AddComment("DataOffset");
3092    OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0);
3093    OS.AddComment("Segment");
3094    OS.EmitCOFFSectionIndex(GVSym);
3095    OS.AddComment("Name");
3096    const unsigned LengthOfDataRecord = 12;
3097    emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord);
3098    endSymbolRecord(DataEnd);
3099  } else {
3100    // FIXME: Currently this only emits the global variables in the IR metadata.
3101    // This should also emit enums and static data members.
3102    const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>();
3103    assert(DIE->isConstant() &&
3104           "Global constant variables must contain a constant expression.");
3105    uint64_t Val = DIE->getElement(1);
3106
3107    MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT);
3108    OS.AddComment("Type");
3109    OS.EmitIntValue(getTypeIndex(DIGV->getType()).getIndex(), 4);
3110    OS.AddComment("Value");
3111
3112    // Encoded integers shouldn't need more than 10 bytes.
3113    uint8_t data[10];
3114    BinaryStreamWriter Writer(data, llvm::support::endianness::little);
3115    CodeViewRecordIO IO(Writer);
3116    cantFail(IO.mapEncodedInteger(Val));
3117    StringRef SRef((char *)data, Writer.getOffset());
3118    OS.EmitBinaryData(SRef);
3119
3120    OS.AddComment("Name");
3121    const DIScope *Scope = DIGV->getScope();
3122    // For static data members, get the scope from the declaration.
3123    if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>(
3124            DIGV->getRawStaticDataMemberDeclaration()))
3125      Scope = MemberDecl->getScope();
3126    emitNullTerminatedSymbolName(OS,
3127                                 getFullyQualifiedName(Scope, DIGV->getName()));
3128    endSymbolRecord(SConstantEnd);
3129  }
3130}
3131