xref: /llvm-project/llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp (revision 3739979c2031ebcf0ed40d30d549cf82bb1a122c)
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/ContinuationRecordBuilder.h"
45 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h"
46 #include "llvm/DebugInfo/CodeView/EnumTables.h"
47 #include "llvm/DebugInfo/CodeView/Line.h"
48 #include "llvm/DebugInfo/CodeView/SymbolRecord.h"
49 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h"
50 #include "llvm/DebugInfo/CodeView/TypeIndex.h"
51 #include "llvm/DebugInfo/CodeView/TypeRecord.h"
52 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h"
53 #include "llvm/IR/Constants.h"
54 #include "llvm/IR/DataLayout.h"
55 #include "llvm/IR/DebugInfoMetadata.h"
56 #include "llvm/IR/DebugLoc.h"
57 #include "llvm/IR/Function.h"
58 #include "llvm/IR/GlobalValue.h"
59 #include "llvm/IR/GlobalVariable.h"
60 #include "llvm/IR/Metadata.h"
61 #include "llvm/IR/Module.h"
62 #include "llvm/MC/MCAsmInfo.h"
63 #include "llvm/MC/MCContext.h"
64 #include "llvm/MC/MCSectionCOFF.h"
65 #include "llvm/MC/MCStreamer.h"
66 #include "llvm/MC/MCSymbol.h"
67 #include "llvm/Support/BinaryByteStream.h"
68 #include "llvm/Support/BinaryStreamReader.h"
69 #include "llvm/Support/Casting.h"
70 #include "llvm/Support/CommandLine.h"
71 #include "llvm/Support/Compiler.h"
72 #include "llvm/Support/Endian.h"
73 #include "llvm/Support/Error.h"
74 #include "llvm/Support/ErrorHandling.h"
75 #include "llvm/Support/FormatVariadic.h"
76 #include "llvm/Support/Path.h"
77 #include "llvm/Support/SMLoc.h"
78 #include "llvm/Support/ScopedPrinter.h"
79 #include "llvm/Target/TargetLoweringObjectFile.h"
80 #include "llvm/Target/TargetMachine.h"
81 #include <algorithm>
82 #include <cassert>
83 #include <cctype>
84 #include <cstddef>
85 #include <cstdint>
86 #include <iterator>
87 #include <limits>
88 #include <string>
89 #include <utility>
90 #include <vector>
91 
92 using namespace llvm;
93 using namespace llvm::codeview;
94 
95 static CPUType mapArchToCVCPUType(Triple::ArchType Type) {
96   switch (Type) {
97   case Triple::ArchType::x86:
98     return CPUType::Pentium3;
99   case Triple::ArchType::x86_64:
100     return CPUType::X64;
101   case Triple::ArchType::thumb:
102     return CPUType::Thumb;
103   case Triple::ArchType::aarch64:
104     return CPUType::ARM64;
105   default:
106     report_fatal_error("target architecture doesn't map to a CodeView CPUType");
107   }
108 }
109 
110 CodeViewDebug::CodeViewDebug(AsmPrinter *AP)
111     : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {
112   // If module doesn't have named metadata anchors or COFF debug section
113   // is not available, skip any debug info related stuff.
114   if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") ||
115       !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) {
116     Asm = nullptr;
117     MMI->setDebugInfoAvailability(false);
118     return;
119   }
120   // Tell MMI that we have debug info.
121   MMI->setDebugInfoAvailability(true);
122 
123   TheCPU =
124       mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch());
125 
126   collectGlobalVariableInfo();
127 
128   // Check if we should emit type record hashes.
129   ConstantInt *GH = mdconst::extract_or_null<ConstantInt>(
130       MMI->getModule()->getModuleFlag("CodeViewGHash"));
131   EmitDebugGlobalHashes = GH && !GH->isZero();
132 }
133 
134 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) {
135   std::string &Filepath = FileToFilepathMap[File];
136   if (!Filepath.empty())
137     return Filepath;
138 
139   StringRef Dir = File->getDirectory(), Filename = File->getFilename();
140 
141   // If this is a Unix-style path, just use it as is. Don't try to canonicalize
142   // it textually because one of the path components could be a symlink.
143   if (Dir.startswith("/") || Filename.startswith("/")) {
144     if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix))
145       return Filename;
146     Filepath = Dir;
147     if (Dir.back() != '/')
148       Filepath += '/';
149     Filepath += Filename;
150     return Filepath;
151   }
152 
153   // Clang emits directory and relative filename info into the IR, but CodeView
154   // operates on full paths.  We could change Clang to emit full paths too, but
155   // that would increase the IR size and probably not needed for other users.
156   // For now, just concatenate and canonicalize the path here.
157   if (Filename.find(':') == 1)
158     Filepath = Filename;
159   else
160     Filepath = (Dir + "\\" + Filename).str();
161 
162   // Canonicalize the path.  We have to do it textually because we may no longer
163   // have access the file in the filesystem.
164   // First, replace all slashes with backslashes.
165   std::replace(Filepath.begin(), Filepath.end(), '/', '\\');
166 
167   // Remove all "\.\" with "\".
168   size_t Cursor = 0;
169   while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos)
170     Filepath.erase(Cursor, 2);
171 
172   // Replace all "\XXX\..\" with "\".  Don't try too hard though as the original
173   // path should be well-formatted, e.g. start with a drive letter, etc.
174   Cursor = 0;
175   while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) {
176     // Something's wrong if the path starts with "\..\", abort.
177     if (Cursor == 0)
178       break;
179 
180     size_t PrevSlash = Filepath.rfind('\\', Cursor - 1);
181     if (PrevSlash == std::string::npos)
182       // Something's wrong, abort.
183       break;
184 
185     Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash);
186     // The next ".." might be following the one we've just erased.
187     Cursor = PrevSlash;
188   }
189 
190   // Remove all duplicate backslashes.
191   Cursor = 0;
192   while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos)
193     Filepath.erase(Cursor, 1);
194 
195   return Filepath;
196 }
197 
198 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) {
199   StringRef FullPath = getFullFilepath(F);
200   unsigned NextId = FileIdMap.size() + 1;
201   auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId));
202   if (Insertion.second) {
203     // We have to compute the full filepath and emit a .cv_file directive.
204     ArrayRef<uint8_t> ChecksumAsBytes;
205     FileChecksumKind CSKind = FileChecksumKind::None;
206     if (F->getChecksum()) {
207       std::string Checksum = fromHex(F->getChecksum()->Value);
208       void *CKMem = OS.getContext().allocate(Checksum.size(), 1);
209       memcpy(CKMem, Checksum.data(), Checksum.size());
210       ChecksumAsBytes = ArrayRef<uint8_t>(
211           reinterpret_cast<const uint8_t *>(CKMem), Checksum.size());
212       switch (F->getChecksum()->Kind) {
213       case DIFile::CSK_MD5:  CSKind = FileChecksumKind::MD5; break;
214       case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break;
215       }
216     }
217     bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes,
218                                           static_cast<unsigned>(CSKind));
219     (void)Success;
220     assert(Success && ".cv_file directive failed");
221   }
222   return Insertion.first->second;
223 }
224 
225 CodeViewDebug::InlineSite &
226 CodeViewDebug::getInlineSite(const DILocation *InlinedAt,
227                              const DISubprogram *Inlinee) {
228   auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()});
229   InlineSite *Site = &SiteInsertion.first->second;
230   if (SiteInsertion.second) {
231     unsigned ParentFuncId = CurFn->FuncId;
232     if (const DILocation *OuterIA = InlinedAt->getInlinedAt())
233       ParentFuncId =
234           getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram())
235               .SiteFuncId;
236 
237     Site->SiteFuncId = NextFuncId++;
238     OS.EmitCVInlineSiteIdDirective(
239         Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()),
240         InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc());
241     Site->Inlinee = Inlinee;
242     InlinedSubprograms.insert(Inlinee);
243     getFuncIdForSubprogram(Inlinee);
244   }
245   return *Site;
246 }
247 
248 static StringRef getPrettyScopeName(const DIScope *Scope) {
249   StringRef ScopeName = Scope->getName();
250   if (!ScopeName.empty())
251     return ScopeName;
252 
253   switch (Scope->getTag()) {
254   case dwarf::DW_TAG_enumeration_type:
255   case dwarf::DW_TAG_class_type:
256   case dwarf::DW_TAG_structure_type:
257   case dwarf::DW_TAG_union_type:
258     return "<unnamed-tag>";
259   case dwarf::DW_TAG_namespace:
260     return "`anonymous namespace'";
261   }
262 
263   return StringRef();
264 }
265 
266 static const DISubprogram *getQualifiedNameComponents(
267     const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) {
268   const DISubprogram *ClosestSubprogram = nullptr;
269   while (Scope != nullptr) {
270     if (ClosestSubprogram == nullptr)
271       ClosestSubprogram = dyn_cast<DISubprogram>(Scope);
272     StringRef ScopeName = getPrettyScopeName(Scope);
273     if (!ScopeName.empty())
274       QualifiedNameComponents.push_back(ScopeName);
275     Scope = Scope->getScope().resolve();
276   }
277   return ClosestSubprogram;
278 }
279 
280 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents,
281                                     StringRef TypeName) {
282   std::string FullyQualifiedName;
283   for (StringRef QualifiedNameComponent :
284        llvm::reverse(QualifiedNameComponents)) {
285     FullyQualifiedName.append(QualifiedNameComponent);
286     FullyQualifiedName.append("::");
287   }
288   FullyQualifiedName.append(TypeName);
289   return FullyQualifiedName;
290 }
291 
292 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) {
293   SmallVector<StringRef, 5> QualifiedNameComponents;
294   getQualifiedNameComponents(Scope, QualifiedNameComponents);
295   return getQualifiedName(QualifiedNameComponents, Name);
296 }
297 
298 struct CodeViewDebug::TypeLoweringScope {
299   TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; }
300   ~TypeLoweringScope() {
301     // Don't decrement TypeEmissionLevel until after emitting deferred types, so
302     // inner TypeLoweringScopes don't attempt to emit deferred types.
303     if (CVD.TypeEmissionLevel == 1)
304       CVD.emitDeferredCompleteTypes();
305     --CVD.TypeEmissionLevel;
306   }
307   CodeViewDebug &CVD;
308 };
309 
310 static std::string getFullyQualifiedName(const DIScope *Ty) {
311   const DIScope *Scope = Ty->getScope().resolve();
312   return getFullyQualifiedName(Scope, getPrettyScopeName(Ty));
313 }
314 
315 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) {
316   // No scope means global scope and that uses the zero index.
317   if (!Scope || isa<DIFile>(Scope))
318     return TypeIndex();
319 
320   assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type");
321 
322   // Check if we've already translated this scope.
323   auto I = TypeIndices.find({Scope, nullptr});
324   if (I != TypeIndices.end())
325     return I->second;
326 
327   // Build the fully qualified name of the scope.
328   std::string ScopeName = getFullyQualifiedName(Scope);
329   StringIdRecord SID(TypeIndex(), ScopeName);
330   auto TI = TypeTable.writeLeafType(SID);
331   return recordTypeIndexForDINode(Scope, TI);
332 }
333 
334 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) {
335   assert(SP);
336 
337   // Check if we've already translated this subprogram.
338   auto I = TypeIndices.find({SP, nullptr});
339   if (I != TypeIndices.end())
340     return I->second;
341 
342   // The display name includes function template arguments. Drop them to match
343   // MSVC.
344   StringRef DisplayName = SP->getName().split('<').first;
345 
346   const DIScope *Scope = SP->getScope().resolve();
347   TypeIndex TI;
348   if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) {
349     // If the scope is a DICompositeType, then this must be a method. Member
350     // function types take some special handling, and require access to the
351     // subprogram.
352     TypeIndex ClassType = getTypeIndex(Class);
353     MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class),
354                                DisplayName);
355     TI = TypeTable.writeLeafType(MFuncId);
356   } else {
357     // Otherwise, this must be a free function.
358     TypeIndex ParentScope = getScopeIndex(Scope);
359     FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName);
360     TI = TypeTable.writeLeafType(FuncId);
361   }
362 
363   return recordTypeIndexForDINode(SP, TI);
364 }
365 
366 static bool isNonTrivial(const DICompositeType *DCTy) {
367   return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial);
368 }
369 
370 static FunctionOptions
371 getFunctionOptions(const DISubroutineType *Ty,
372                    const DICompositeType *ClassTy = nullptr,
373                    StringRef SPName = StringRef("")) {
374   FunctionOptions FO = FunctionOptions::None;
375   const DIType *ReturnTy = nullptr;
376   if (auto TypeArray = Ty->getTypeArray()) {
377     if (TypeArray.size())
378       ReturnTy = TypeArray[0].resolve();
379   }
380 
381   if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) {
382     if (isNonTrivial(ReturnDCTy))
383       FO |= FunctionOptions::CxxReturnUdt;
384   }
385 
386   // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison.
387   if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) {
388     FO |= FunctionOptions::Constructor;
389 
390   // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag.
391 
392   }
393   return FO;
394 }
395 
396 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP,
397                                                const DICompositeType *Class) {
398   // Always use the method declaration as the key for the function type. The
399   // method declaration contains the this adjustment.
400   if (SP->getDeclaration())
401     SP = SP->getDeclaration();
402   assert(!SP->getDeclaration() && "should use declaration as key");
403 
404   // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide
405   // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}.
406   auto I = TypeIndices.find({SP, Class});
407   if (I != TypeIndices.end())
408     return I->second;
409 
410   // Make sure complete type info for the class is emitted *after* the member
411   // function type, as the complete class type is likely to reference this
412   // member function type.
413   TypeLoweringScope S(*this);
414   const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0;
415 
416   FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName());
417   TypeIndex TI = lowerTypeMemberFunction(
418       SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO);
419   return recordTypeIndexForDINode(SP, TI, Class);
420 }
421 
422 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node,
423                                                   TypeIndex TI,
424                                                   const DIType *ClassTy) {
425   auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI});
426   (void)InsertResult;
427   assert(InsertResult.second && "DINode was already assigned a type index");
428   return TI;
429 }
430 
431 unsigned CodeViewDebug::getPointerSizeInBytes() {
432   return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8;
433 }
434 
435 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var,
436                                         const LexicalScope *LS) {
437   if (const DILocation *InlinedAt = LS->getInlinedAt()) {
438     // This variable was inlined. Associate it with the InlineSite.
439     const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram();
440     InlineSite &Site = getInlineSite(InlinedAt, Inlinee);
441     Site.InlinedLocals.emplace_back(Var);
442   } else {
443     // This variable goes into the corresponding lexical scope.
444     ScopeVariables[LS].emplace_back(Var);
445   }
446 }
447 
448 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs,
449                                const DILocation *Loc) {
450   auto B = Locs.begin(), E = Locs.end();
451   if (std::find(B, E, Loc) == E)
452     Locs.push_back(Loc);
453 }
454 
455 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL,
456                                         const MachineFunction *MF) {
457   // Skip this instruction if it has the same location as the previous one.
458   if (!DL || DL == PrevInstLoc)
459     return;
460 
461   const DIScope *Scope = DL.get()->getScope();
462   if (!Scope)
463     return;
464 
465   // Skip this line if it is longer than the maximum we can record.
466   LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true);
467   if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() ||
468       LI.isNeverStepInto())
469     return;
470 
471   ColumnInfo CI(DL.getCol(), /*EndColumn=*/0);
472   if (CI.getStartColumn() != DL.getCol())
473     return;
474 
475   if (!CurFn->HaveLineInfo)
476     CurFn->HaveLineInfo = true;
477   unsigned FileId = 0;
478   if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile())
479     FileId = CurFn->LastFileId;
480   else
481     FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile());
482   PrevInstLoc = DL;
483 
484   unsigned FuncId = CurFn->FuncId;
485   if (const DILocation *SiteLoc = DL->getInlinedAt()) {
486     const DILocation *Loc = DL.get();
487 
488     // If this location was actually inlined from somewhere else, give it the ID
489     // of the inline call site.
490     FuncId =
491         getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId;
492 
493     // Ensure we have links in the tree of inline call sites.
494     bool FirstLoc = true;
495     while ((SiteLoc = Loc->getInlinedAt())) {
496       InlineSite &Site =
497           getInlineSite(SiteLoc, Loc->getScope()->getSubprogram());
498       if (!FirstLoc)
499         addLocIfNotPresent(Site.ChildSites, Loc);
500       FirstLoc = false;
501       Loc = SiteLoc;
502     }
503     addLocIfNotPresent(CurFn->ChildSites, Loc);
504   }
505 
506   OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(),
507                         /*PrologueEnd=*/false, /*IsStmt=*/false,
508                         DL->getFilename(), SMLoc());
509 }
510 
511 void CodeViewDebug::emitCodeViewMagicVersion() {
512   OS.EmitValueToAlignment(4);
513   OS.AddComment("Debug section magic");
514   OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4);
515 }
516 
517 void CodeViewDebug::endModule() {
518   if (!Asm || !MMI->hasDebugInfo())
519     return;
520 
521   assert(Asm != nullptr);
522 
523   // The COFF .debug$S section consists of several subsections, each starting
524   // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length
525   // of the payload followed by the payload itself.  The subsections are 4-byte
526   // aligned.
527 
528   // Use the generic .debug$S section, and make a subsection for all the inlined
529   // subprograms.
530   switchToDebugSectionForSymbol(nullptr);
531 
532   MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols);
533   emitCompilerInformation();
534   endCVSubsection(CompilerInfo);
535 
536   emitInlineeLinesSubsection();
537 
538   // Emit per-function debug information.
539   for (auto &P : FnDebugInfo)
540     if (!P.first->isDeclarationForLinker())
541       emitDebugInfoForFunction(P.first, *P.second);
542 
543   // Emit global variable debug information.
544   setCurrentSubprogram(nullptr);
545   emitDebugInfoForGlobals();
546 
547   // Emit retained types.
548   emitDebugInfoForRetainedTypes();
549 
550   // Switch back to the generic .debug$S section after potentially processing
551   // comdat symbol sections.
552   switchToDebugSectionForSymbol(nullptr);
553 
554   // Emit UDT records for any types used by global variables.
555   if (!GlobalUDTs.empty()) {
556     MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
557     emitDebugInfoForUDTs(GlobalUDTs);
558     endCVSubsection(SymbolsEnd);
559   }
560 
561   // This subsection holds a file index to offset in string table table.
562   OS.AddComment("File index to string table offset subsection");
563   OS.EmitCVFileChecksumsDirective();
564 
565   // This subsection holds the string table.
566   OS.AddComment("String table");
567   OS.EmitCVStringTableDirective();
568 
569   // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol
570   // subsection in the generic .debug$S section at the end. There is no
571   // particular reason for this ordering other than to match MSVC.
572   emitBuildInfo();
573 
574   // Emit type information and hashes last, so that any types we translate while
575   // emitting function info are included.
576   emitTypeInformation();
577 
578   if (EmitDebugGlobalHashes)
579     emitTypeGlobalHashes();
580 
581   clear();
582 }
583 
584 static void emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S,
585     unsigned MaxFixedRecordLength = 0xF00) {
586   // The maximum CV record length is 0xFF00. Most of the strings we emit appear
587   // after a fixed length portion of the record. The fixed length portion should
588   // always be less than 0xF00 (3840) bytes, so truncate the string so that the
589   // overall record size is less than the maximum allowed.
590   SmallString<32> NullTerminatedString(
591       S.take_front(MaxRecordLength - MaxFixedRecordLength - 1));
592   NullTerminatedString.push_back('\0');
593   OS.EmitBytes(NullTerminatedString);
594 }
595 
596 void CodeViewDebug::emitTypeInformation() {
597   if (TypeTable.empty())
598     return;
599 
600   // Start the .debug$T or .debug$P section with 0x4.
601   OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection());
602   emitCodeViewMagicVersion();
603 
604   SmallString<8> CommentPrefix;
605   if (OS.isVerboseAsm()) {
606     CommentPrefix += '\t';
607     CommentPrefix += Asm->MAI->getCommentString();
608     CommentPrefix += ' ';
609   }
610 
611   TypeTableCollection Table(TypeTable.records());
612   Optional<TypeIndex> B = Table.getFirst();
613   while (B) {
614     // This will fail if the record data is invalid.
615     CVType Record = Table.getType(*B);
616 
617     if (OS.isVerboseAsm()) {
618       // Emit a block comment describing the type record for readability.
619       SmallString<512> CommentBlock;
620       raw_svector_ostream CommentOS(CommentBlock);
621       ScopedPrinter SP(CommentOS);
622       SP.setPrefix(CommentPrefix);
623       TypeDumpVisitor TDV(Table, &SP, false);
624 
625       Error E = codeview::visitTypeRecord(Record, *B, TDV);
626       if (E) {
627         logAllUnhandledErrors(std::move(E), errs(), "error: ");
628         llvm_unreachable("produced malformed type record");
629       }
630       // emitRawComment will insert its own tab and comment string before
631       // the first line, so strip off our first one. It also prints its own
632       // newline.
633       OS.emitRawComment(
634           CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim());
635     }
636     OS.EmitBinaryData(Record.str_data());
637     B = Table.getNext(*B);
638   }
639 }
640 
641 void CodeViewDebug::emitTypeGlobalHashes() {
642   if (TypeTable.empty())
643     return;
644 
645   // Start the .debug$H section with the version and hash algorithm, currently
646   // hardcoded to version 0, SHA1.
647   OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection());
648 
649   OS.EmitValueToAlignment(4);
650   OS.AddComment("Magic");
651   OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4);
652   OS.AddComment("Section Version");
653   OS.EmitIntValue(0, 2);
654   OS.AddComment("Hash Algorithm");
655   OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2);
656 
657   TypeIndex TI(TypeIndex::FirstNonSimpleIndex);
658   for (const auto &GHR : TypeTable.hashes()) {
659     if (OS.isVerboseAsm()) {
660       // Emit an EOL-comment describing which TypeIndex this hash corresponds
661       // to, as well as the stringified SHA1 hash.
662       SmallString<32> Comment;
663       raw_svector_ostream CommentOS(Comment);
664       CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR);
665       OS.AddComment(Comment);
666       ++TI;
667     }
668     assert(GHR.Hash.size() == 8);
669     StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()),
670                 GHR.Hash.size());
671     OS.EmitBinaryData(S);
672   }
673 }
674 
675 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) {
676   switch (DWLang) {
677   case dwarf::DW_LANG_C:
678   case dwarf::DW_LANG_C89:
679   case dwarf::DW_LANG_C99:
680   case dwarf::DW_LANG_C11:
681   case dwarf::DW_LANG_ObjC:
682     return SourceLanguage::C;
683   case dwarf::DW_LANG_C_plus_plus:
684   case dwarf::DW_LANG_C_plus_plus_03:
685   case dwarf::DW_LANG_C_plus_plus_11:
686   case dwarf::DW_LANG_C_plus_plus_14:
687     return SourceLanguage::Cpp;
688   case dwarf::DW_LANG_Fortran77:
689   case dwarf::DW_LANG_Fortran90:
690   case dwarf::DW_LANG_Fortran03:
691   case dwarf::DW_LANG_Fortran08:
692     return SourceLanguage::Fortran;
693   case dwarf::DW_LANG_Pascal83:
694     return SourceLanguage::Pascal;
695   case dwarf::DW_LANG_Cobol74:
696   case dwarf::DW_LANG_Cobol85:
697     return SourceLanguage::Cobol;
698   case dwarf::DW_LANG_Java:
699     return SourceLanguage::Java;
700   case dwarf::DW_LANG_D:
701     return SourceLanguage::D;
702   case dwarf::DW_LANG_Swift:
703     return SourceLanguage::Swift;
704   default:
705     // There's no CodeView representation for this language, and CV doesn't
706     // have an "unknown" option for the language field, so we'll use MASM,
707     // as it's very low level.
708     return SourceLanguage::Masm;
709   }
710 }
711 
712 namespace {
713 struct Version {
714   int Part[4];
715 };
716 } // end anonymous namespace
717 
718 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out
719 // the version number.
720 static Version parseVersion(StringRef Name) {
721   Version V = {{0}};
722   int N = 0;
723   for (const char C : Name) {
724     if (isdigit(C)) {
725       V.Part[N] *= 10;
726       V.Part[N] += C - '0';
727     } else if (C == '.') {
728       ++N;
729       if (N >= 4)
730         return V;
731     } else if (N > 0)
732       return V;
733   }
734   return V;
735 }
736 
737 void CodeViewDebug::emitCompilerInformation() {
738   MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3);
739   uint32_t Flags = 0;
740 
741   NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
742   const MDNode *Node = *CUs->operands().begin();
743   const auto *CU = cast<DICompileUnit>(Node);
744 
745   // The low byte of the flags indicates the source language.
746   Flags = MapDWLangToCVLang(CU->getSourceLanguage());
747   // TODO:  Figure out which other flags need to be set.
748 
749   OS.AddComment("Flags and language");
750   OS.EmitIntValue(Flags, 4);
751 
752   OS.AddComment("CPUType");
753   OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2);
754 
755   StringRef CompilerVersion = CU->getProducer();
756   Version FrontVer = parseVersion(CompilerVersion);
757   OS.AddComment("Frontend version");
758   for (int N = 0; N < 4; ++N)
759     OS.EmitIntValue(FrontVer.Part[N], 2);
760 
761   // Some Microsoft tools, like Binscope, expect a backend version number of at
762   // least 8.something, so we'll coerce the LLVM version into a form that
763   // guarantees it'll be big enough without really lying about the version.
764   int Major = 1000 * LLVM_VERSION_MAJOR +
765               10 * LLVM_VERSION_MINOR +
766               LLVM_VERSION_PATCH;
767   // Clamp it for builds that use unusually large version numbers.
768   Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max());
769   Version BackVer = {{ Major, 0, 0, 0 }};
770   OS.AddComment("Backend version");
771   for (int N = 0; N < 4; ++N)
772     OS.EmitIntValue(BackVer.Part[N], 2);
773 
774   OS.AddComment("Null-terminated compiler version string");
775   emitNullTerminatedSymbolName(OS, CompilerVersion);
776 
777   endSymbolRecord(CompilerEnd);
778 }
779 
780 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable,
781                                     StringRef S) {
782   StringIdRecord SIR(TypeIndex(0x0), S);
783   return TypeTable.writeLeafType(SIR);
784 }
785 
786 void CodeViewDebug::emitBuildInfo() {
787   // First, make LF_BUILDINFO. It's a sequence of strings with various bits of
788   // build info. The known prefix is:
789   // - Absolute path of current directory
790   // - Compiler path
791   // - Main source file path, relative to CWD or absolute
792   // - Type server PDB file
793   // - Canonical compiler command line
794   // If frontend and backend compilation are separated (think llc or LTO), it's
795   // not clear if the compiler path should refer to the executable for the
796   // frontend or the backend. Leave it blank for now.
797   TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {};
798   NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
799   const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs.
800   const auto *CU = cast<DICompileUnit>(Node);
801   const DIFile *MainSourceFile = CU->getFile();
802   BuildInfoArgs[BuildInfoRecord::CurrentDirectory] =
803       getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory());
804   BuildInfoArgs[BuildInfoRecord::SourceFile] =
805       getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename());
806   // FIXME: Path to compiler and command line. PDB is intentionally blank unless
807   // we implement /Zi type servers.
808   BuildInfoRecord BIR(BuildInfoArgs);
809   TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR);
810 
811   // Make a new .debug$S subsection for the S_BUILDINFO record, which points
812   // from the module symbols into the type stream.
813   MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
814   MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO);
815   OS.AddComment("LF_BUILDINFO index");
816   OS.EmitIntValue(BuildInfoIndex.getIndex(), 4);
817   endSymbolRecord(BIEnd);
818   endCVSubsection(BISubsecEnd);
819 }
820 
821 void CodeViewDebug::emitInlineeLinesSubsection() {
822   if (InlinedSubprograms.empty())
823     return;
824 
825   OS.AddComment("Inlinee lines subsection");
826   MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines);
827 
828   // We emit the checksum info for files.  This is used by debuggers to
829   // determine if a pdb matches the source before loading it.  Visual Studio,
830   // for instance, will display a warning that the breakpoints are not valid if
831   // the pdb does not match the source.
832   OS.AddComment("Inlinee lines signature");
833   OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4);
834 
835   for (const DISubprogram *SP : InlinedSubprograms) {
836     assert(TypeIndices.count({SP, nullptr}));
837     TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}];
838 
839     OS.AddBlankLine();
840     unsigned FileId = maybeRecordFile(SP->getFile());
841     OS.AddComment("Inlined function " + SP->getName() + " starts at " +
842                   SP->getFilename() + Twine(':') + Twine(SP->getLine()));
843     OS.AddBlankLine();
844     OS.AddComment("Type index of inlined function");
845     OS.EmitIntValue(InlineeIdx.getIndex(), 4);
846     OS.AddComment("Offset into filechecksum table");
847     OS.EmitCVFileChecksumOffsetDirective(FileId);
848     OS.AddComment("Starting line number");
849     OS.EmitIntValue(SP->getLine(), 4);
850   }
851 
852   endCVSubsection(InlineEnd);
853 }
854 
855 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI,
856                                         const DILocation *InlinedAt,
857                                         const InlineSite &Site) {
858   assert(TypeIndices.count({Site.Inlinee, nullptr}));
859   TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}];
860 
861   // SymbolRecord
862   MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE);
863 
864   OS.AddComment("PtrParent");
865   OS.EmitIntValue(0, 4);
866   OS.AddComment("PtrEnd");
867   OS.EmitIntValue(0, 4);
868   OS.AddComment("Inlinee type index");
869   OS.EmitIntValue(InlineeIdx.getIndex(), 4);
870 
871   unsigned FileId = maybeRecordFile(Site.Inlinee->getFile());
872   unsigned StartLineNum = Site.Inlinee->getLine();
873 
874   OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum,
875                                     FI.Begin, FI.End);
876 
877   endSymbolRecord(InlineEnd);
878 
879   emitLocalVariableList(FI, Site.InlinedLocals);
880 
881   // Recurse on child inlined call sites before closing the scope.
882   for (const DILocation *ChildSite : Site.ChildSites) {
883     auto I = FI.InlineSites.find(ChildSite);
884     assert(I != FI.InlineSites.end() &&
885            "child site not in function inline site map");
886     emitInlinedCallSite(FI, ChildSite, I->second);
887   }
888 
889   // Close the scope.
890   emitEndSymbolRecord(SymbolKind::S_INLINESITE_END);
891 }
892 
893 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) {
894   // If we have a symbol, it may be in a section that is COMDAT. If so, find the
895   // comdat key. A section may be comdat because of -ffunction-sections or
896   // because it is comdat in the IR.
897   MCSectionCOFF *GVSec =
898       GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr;
899   const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr;
900 
901   MCSectionCOFF *DebugSec = cast<MCSectionCOFF>(
902       Asm->getObjFileLowering().getCOFFDebugSymbolsSection());
903   DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym);
904 
905   OS.SwitchSection(DebugSec);
906 
907   // Emit the magic version number if this is the first time we've switched to
908   // this section.
909   if (ComdatDebugSections.insert(DebugSec).second)
910     emitCodeViewMagicVersion();
911 }
912 
913 // Emit an S_THUNK32/S_END symbol pair for a thunk routine.
914 // The only supported thunk ordinal is currently the standard type.
915 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV,
916                                           FunctionInfo &FI,
917                                           const MCSymbol *Fn) {
918   std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName());
919   const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind.
920 
921   OS.AddComment("Symbol subsection for " + Twine(FuncName));
922   MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
923 
924   // Emit S_THUNK32
925   MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32);
926   OS.AddComment("PtrParent");
927   OS.EmitIntValue(0, 4);
928   OS.AddComment("PtrEnd");
929   OS.EmitIntValue(0, 4);
930   OS.AddComment("PtrNext");
931   OS.EmitIntValue(0, 4);
932   OS.AddComment("Thunk section relative address");
933   OS.EmitCOFFSecRel32(Fn, /*Offset=*/0);
934   OS.AddComment("Thunk section index");
935   OS.EmitCOFFSectionIndex(Fn);
936   OS.AddComment("Code size");
937   OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2);
938   OS.AddComment("Ordinal");
939   OS.EmitIntValue(unsigned(ordinal), 1);
940   OS.AddComment("Function name");
941   emitNullTerminatedSymbolName(OS, FuncName);
942   // Additional fields specific to the thunk ordinal would go here.
943   endSymbolRecord(ThunkRecordEnd);
944 
945   // Local variables/inlined routines are purposely omitted here.  The point of
946   // marking this as a thunk is so Visual Studio will NOT stop in this routine.
947 
948   // Emit S_PROC_ID_END
949   emitEndSymbolRecord(SymbolKind::S_PROC_ID_END);
950 
951   endCVSubsection(SymbolsEnd);
952 }
953 
954 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV,
955                                              FunctionInfo &FI) {
956   // For each function there is a separate subsection which holds the PC to
957   // file:line table.
958   const MCSymbol *Fn = Asm->getSymbol(GV);
959   assert(Fn);
960 
961   // Switch to the to a comdat section, if appropriate.
962   switchToDebugSectionForSymbol(Fn);
963 
964   std::string FuncName;
965   auto *SP = GV->getSubprogram();
966   assert(SP);
967   setCurrentSubprogram(SP);
968 
969   if (SP->isThunk()) {
970     emitDebugInfoForThunk(GV, FI, Fn);
971     return;
972   }
973 
974   // If we have a display name, build the fully qualified name by walking the
975   // chain of scopes.
976   if (!SP->getName().empty())
977     FuncName =
978         getFullyQualifiedName(SP->getScope().resolve(), SP->getName());
979 
980   // If our DISubprogram name is empty, use the mangled name.
981   if (FuncName.empty())
982     FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName());
983 
984   // Emit FPO data, but only on 32-bit x86. No other platforms use it.
985   if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86)
986     OS.EmitCVFPOData(Fn);
987 
988   // Emit a symbol subsection, required by VS2012+ to find function boundaries.
989   OS.AddComment("Symbol subsection for " + Twine(FuncName));
990   MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
991   {
992     SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID
993                                                 : SymbolKind::S_GPROC32_ID;
994     MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind);
995 
996     // These fields are filled in by tools like CVPACK which run after the fact.
997     OS.AddComment("PtrParent");
998     OS.EmitIntValue(0, 4);
999     OS.AddComment("PtrEnd");
1000     OS.EmitIntValue(0, 4);
1001     OS.AddComment("PtrNext");
1002     OS.EmitIntValue(0, 4);
1003     // This is the important bit that tells the debugger where the function
1004     // code is located and what's its size:
1005     OS.AddComment("Code size");
1006     OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4);
1007     OS.AddComment("Offset after prologue");
1008     OS.EmitIntValue(0, 4);
1009     OS.AddComment("Offset before epilogue");
1010     OS.EmitIntValue(0, 4);
1011     OS.AddComment("Function type index");
1012     OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4);
1013     OS.AddComment("Function section relative address");
1014     OS.EmitCOFFSecRel32(Fn, /*Offset=*/0);
1015     OS.AddComment("Function section index");
1016     OS.EmitCOFFSectionIndex(Fn);
1017     OS.AddComment("Flags");
1018     OS.EmitIntValue(0, 1);
1019     // Emit the function display name as a null-terminated string.
1020     OS.AddComment("Function name");
1021     // Truncate the name so we won't overflow the record length field.
1022     emitNullTerminatedSymbolName(OS, FuncName);
1023     endSymbolRecord(ProcRecordEnd);
1024 
1025     MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC);
1026     // Subtract out the CSR size since MSVC excludes that and we include it.
1027     OS.AddComment("FrameSize");
1028     OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4);
1029     OS.AddComment("Padding");
1030     OS.EmitIntValue(0, 4);
1031     OS.AddComment("Offset of padding");
1032     OS.EmitIntValue(0, 4);
1033     OS.AddComment("Bytes of callee saved registers");
1034     OS.EmitIntValue(FI.CSRSize, 4);
1035     OS.AddComment("Exception handler offset");
1036     OS.EmitIntValue(0, 4);
1037     OS.AddComment("Exception handler section");
1038     OS.EmitIntValue(0, 2);
1039     OS.AddComment("Flags (defines frame register)");
1040     OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4);
1041     endSymbolRecord(FrameProcEnd);
1042 
1043     emitLocalVariableList(FI, FI.Locals);
1044     emitGlobalVariableList(FI.Globals);
1045     emitLexicalBlockList(FI.ChildBlocks, FI);
1046 
1047     // Emit inlined call site information. Only emit functions inlined directly
1048     // into the parent function. We'll emit the other sites recursively as part
1049     // of their parent inline site.
1050     for (const DILocation *InlinedAt : FI.ChildSites) {
1051       auto I = FI.InlineSites.find(InlinedAt);
1052       assert(I != FI.InlineSites.end() &&
1053              "child site not in function inline site map");
1054       emitInlinedCallSite(FI, InlinedAt, I->second);
1055     }
1056 
1057     for (auto Annot : FI.Annotations) {
1058       MCSymbol *Label = Annot.first;
1059       MDTuple *Strs = cast<MDTuple>(Annot.second);
1060       MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION);
1061       OS.EmitCOFFSecRel32(Label, /*Offset=*/0);
1062       // FIXME: Make sure we don't overflow the max record size.
1063       OS.EmitCOFFSectionIndex(Label);
1064       OS.EmitIntValue(Strs->getNumOperands(), 2);
1065       for (Metadata *MD : Strs->operands()) {
1066         // MDStrings are null terminated, so we can do EmitBytes and get the
1067         // nice .asciz directive.
1068         StringRef Str = cast<MDString>(MD)->getString();
1069         assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString");
1070         OS.EmitBytes(StringRef(Str.data(), Str.size() + 1));
1071       }
1072       endSymbolRecord(AnnotEnd);
1073     }
1074 
1075     if (SP != nullptr)
1076       emitDebugInfoForUDTs(LocalUDTs);
1077 
1078     // We're done with this function.
1079     emitEndSymbolRecord(SymbolKind::S_PROC_ID_END);
1080   }
1081   endCVSubsection(SymbolsEnd);
1082 
1083   // We have an assembler directive that takes care of the whole line table.
1084   OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End);
1085 }
1086 
1087 CodeViewDebug::LocalVarDefRange
1088 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) {
1089   LocalVarDefRange DR;
1090   DR.InMemory = -1;
1091   DR.DataOffset = Offset;
1092   assert(DR.DataOffset == Offset && "truncation");
1093   DR.IsSubfield = 0;
1094   DR.StructOffset = 0;
1095   DR.CVRegister = CVRegister;
1096   return DR;
1097 }
1098 
1099 void CodeViewDebug::collectVariableInfoFromMFTable(
1100     DenseSet<InlinedEntity> &Processed) {
1101   const MachineFunction &MF = *Asm->MF;
1102   const TargetSubtargetInfo &TSI = MF.getSubtarget();
1103   const TargetFrameLowering *TFI = TSI.getFrameLowering();
1104   const TargetRegisterInfo *TRI = TSI.getRegisterInfo();
1105 
1106   for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) {
1107     if (!VI.Var)
1108       continue;
1109     assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) &&
1110            "Expected inlined-at fields to agree");
1111 
1112     Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt()));
1113     LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc);
1114 
1115     // If variable scope is not found then skip this variable.
1116     if (!Scope)
1117       continue;
1118 
1119     // If the variable has an attached offset expression, extract it.
1120     // FIXME: Try to handle DW_OP_deref as well.
1121     int64_t ExprOffset = 0;
1122     if (VI.Expr)
1123       if (!VI.Expr->extractIfOffset(ExprOffset))
1124         continue;
1125 
1126     // Get the frame register used and the offset.
1127     unsigned FrameReg = 0;
1128     int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg);
1129     uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg);
1130 
1131     // Calculate the label ranges.
1132     LocalVarDefRange DefRange =
1133         createDefRangeMem(CVReg, FrameOffset + ExprOffset);
1134     for (const InsnRange &Range : Scope->getRanges()) {
1135       const MCSymbol *Begin = getLabelBeforeInsn(Range.first);
1136       const MCSymbol *End = getLabelAfterInsn(Range.second);
1137       End = End ? End : Asm->getFunctionEnd();
1138       DefRange.Ranges.emplace_back(Begin, End);
1139     }
1140 
1141     LocalVariable Var;
1142     Var.DIVar = VI.Var;
1143     Var.DefRanges.emplace_back(std::move(DefRange));
1144     recordLocalVariable(std::move(Var), Scope);
1145   }
1146 }
1147 
1148 static bool canUseReferenceType(const DbgVariableLocation &Loc) {
1149   return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0;
1150 }
1151 
1152 static bool needsReferenceType(const DbgVariableLocation &Loc) {
1153   return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0;
1154 }
1155 
1156 void CodeViewDebug::calculateRanges(
1157     LocalVariable &Var, const DbgValueHistoryMap::InstrRanges &Ranges) {
1158   const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo();
1159 
1160   // Calculate the definition ranges.
1161   for (auto I = Ranges.begin(), E = Ranges.end(); I != E; ++I) {
1162     const auto &Range = *I;
1163     const MachineInstr *DVInst = Range.getBegin();
1164     assert(DVInst->isDebugValue() && "Invalid History entry");
1165     // FIXME: Find a way to represent constant variables, since they are
1166     // relatively common.
1167     Optional<DbgVariableLocation> Location =
1168         DbgVariableLocation::extractFromMachineInstruction(*DVInst);
1169     if (!Location)
1170       continue;
1171 
1172     // CodeView can only express variables in register and variables in memory
1173     // at a constant offset from a register. However, for variables passed
1174     // indirectly by pointer, it is common for that pointer to be spilled to a
1175     // stack location. For the special case of one offseted load followed by a
1176     // zero offset load (a pointer spilled to the stack), we change the type of
1177     // the local variable from a value type to a reference type. This tricks the
1178     // debugger into doing the load for us.
1179     if (Var.UseReferenceType) {
1180       // We're using a reference type. Drop the last zero offset load.
1181       if (canUseReferenceType(*Location))
1182         Location->LoadChain.pop_back();
1183       else
1184         continue;
1185     } else if (needsReferenceType(*Location)) {
1186       // This location can't be expressed without switching to a reference type.
1187       // Start over using that.
1188       Var.UseReferenceType = true;
1189       Var.DefRanges.clear();
1190       calculateRanges(Var, Ranges);
1191       return;
1192     }
1193 
1194     // We can only handle a register or an offseted load of a register.
1195     if (Location->Register == 0 || Location->LoadChain.size() > 1)
1196       continue;
1197     {
1198       LocalVarDefRange DR;
1199       DR.CVRegister = TRI->getCodeViewRegNum(Location->Register);
1200       DR.InMemory = !Location->LoadChain.empty();
1201       DR.DataOffset =
1202           !Location->LoadChain.empty() ? Location->LoadChain.back() : 0;
1203       if (Location->FragmentInfo) {
1204         DR.IsSubfield = true;
1205         DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8;
1206       } else {
1207         DR.IsSubfield = false;
1208         DR.StructOffset = 0;
1209       }
1210 
1211       if (Var.DefRanges.empty() ||
1212           Var.DefRanges.back().isDifferentLocation(DR)) {
1213         Var.DefRanges.emplace_back(std::move(DR));
1214       }
1215     }
1216 
1217     // Compute the label range.
1218     const MCSymbol *Begin = getLabelBeforeInsn(Range.getBegin());
1219     const MCSymbol *End = getLabelAfterInsn(Range.getEnd());
1220     if (!End) {
1221       // This range is valid until the next overlapping bitpiece. In the
1222       // common case, ranges will not be bitpieces, so they will overlap.
1223       auto J = std::next(I);
1224       const DIExpression *DIExpr = DVInst->getDebugExpression();
1225       while (J != E &&
1226              !DIExpr->fragmentsOverlap(J->getBegin()->getDebugExpression()))
1227         ++J;
1228       if (J != E)
1229         End = getLabelBeforeInsn(J->getBegin());
1230       else
1231         End = Asm->getFunctionEnd();
1232     }
1233 
1234     // If the last range end is our begin, just extend the last range.
1235     // Otherwise make a new range.
1236     SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R =
1237         Var.DefRanges.back().Ranges;
1238     if (!R.empty() && R.back().second == Begin)
1239       R.back().second = End;
1240     else
1241       R.emplace_back(Begin, End);
1242 
1243     // FIXME: Do more range combining.
1244   }
1245 }
1246 
1247 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) {
1248   DenseSet<InlinedEntity> Processed;
1249   // Grab the variable info that was squirreled away in the MMI side-table.
1250   collectVariableInfoFromMFTable(Processed);
1251 
1252   for (const auto &I : DbgValues) {
1253     InlinedEntity IV = I.first;
1254     if (Processed.count(IV))
1255       continue;
1256     const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first);
1257     const DILocation *InlinedAt = IV.second;
1258 
1259     // Instruction ranges, specifying where IV is accessible.
1260     const auto &Ranges = I.second;
1261 
1262     LexicalScope *Scope = nullptr;
1263     if (InlinedAt)
1264       Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt);
1265     else
1266       Scope = LScopes.findLexicalScope(DIVar->getScope());
1267     // If variable scope is not found then skip this variable.
1268     if (!Scope)
1269       continue;
1270 
1271     LocalVariable Var;
1272     Var.DIVar = DIVar;
1273 
1274     calculateRanges(Var, Ranges);
1275     recordLocalVariable(std::move(Var), Scope);
1276   }
1277 }
1278 
1279 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) {
1280   const TargetSubtargetInfo &TSI = MF->getSubtarget();
1281   const TargetRegisterInfo *TRI = TSI.getRegisterInfo();
1282   const MachineFrameInfo &MFI = MF->getFrameInfo();
1283   const Function &GV = MF->getFunction();
1284   auto Insertion = FnDebugInfo.insert({&GV, llvm::make_unique<FunctionInfo>()});
1285   assert(Insertion.second && "function already has info");
1286   CurFn = Insertion.first->second.get();
1287   CurFn->FuncId = NextFuncId++;
1288   CurFn->Begin = Asm->getFunctionBegin();
1289 
1290   // The S_FRAMEPROC record reports the stack size, and how many bytes of
1291   // callee-saved registers were used. For targets that don't use a PUSH
1292   // instruction (AArch64), this will be zero.
1293   CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters();
1294   CurFn->FrameSize = MFI.getStackSize();
1295   CurFn->OffsetAdjustment = MFI.getOffsetAdjustment();
1296   CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF);
1297 
1298   // For this function S_FRAMEPROC record, figure out which codeview register
1299   // will be the frame pointer.
1300   CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None.
1301   CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None.
1302   if (CurFn->FrameSize > 0) {
1303     if (!TSI.getFrameLowering()->hasFP(*MF)) {
1304       CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr;
1305       CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr;
1306     } else {
1307       // If there is an FP, parameters are always relative to it.
1308       CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr;
1309       if (CurFn->HasStackRealignment) {
1310         // If the stack needs realignment, locals are relative to SP or VFRAME.
1311         CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr;
1312       } else {
1313         // Otherwise, locals are relative to EBP, and we probably have VLAs or
1314         // other stack adjustments.
1315         CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr;
1316       }
1317     }
1318   }
1319 
1320   // Compute other frame procedure options.
1321   FrameProcedureOptions FPO = FrameProcedureOptions::None;
1322   if (MFI.hasVarSizedObjects())
1323     FPO |= FrameProcedureOptions::HasAlloca;
1324   if (MF->exposesReturnsTwice())
1325     FPO |= FrameProcedureOptions::HasSetJmp;
1326   // FIXME: Set HasLongJmp if we ever track that info.
1327   if (MF->hasInlineAsm())
1328     FPO |= FrameProcedureOptions::HasInlineAssembly;
1329   if (GV.hasPersonalityFn()) {
1330     if (isAsynchronousEHPersonality(
1331             classifyEHPersonality(GV.getPersonalityFn())))
1332       FPO |= FrameProcedureOptions::HasStructuredExceptionHandling;
1333     else
1334       FPO |= FrameProcedureOptions::HasExceptionHandling;
1335   }
1336   if (GV.hasFnAttribute(Attribute::InlineHint))
1337     FPO |= FrameProcedureOptions::MarkedInline;
1338   if (GV.hasFnAttribute(Attribute::Naked))
1339     FPO |= FrameProcedureOptions::Naked;
1340   if (MFI.hasStackProtectorIndex())
1341     FPO |= FrameProcedureOptions::SecurityChecks;
1342   FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U);
1343   FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U);
1344   if (Asm->TM.getOptLevel() != CodeGenOpt::None &&
1345       !GV.hasOptSize() && !GV.hasOptNone())
1346     FPO |= FrameProcedureOptions::OptimizedForSpeed;
1347   // FIXME: Set GuardCfg when it is implemented.
1348   CurFn->FrameProcOpts = FPO;
1349 
1350   OS.EmitCVFuncIdDirective(CurFn->FuncId);
1351 
1352   // Find the end of the function prolog.  First known non-DBG_VALUE and
1353   // non-frame setup location marks the beginning of the function body.
1354   // FIXME: is there a simpler a way to do this? Can we just search
1355   // for the first instruction of the function, not the last of the prolog?
1356   DebugLoc PrologEndLoc;
1357   bool EmptyPrologue = true;
1358   for (const auto &MBB : *MF) {
1359     for (const auto &MI : MBB) {
1360       if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) &&
1361           MI.getDebugLoc()) {
1362         PrologEndLoc = MI.getDebugLoc();
1363         break;
1364       } else if (!MI.isMetaInstruction()) {
1365         EmptyPrologue = false;
1366       }
1367     }
1368   }
1369 
1370   // Record beginning of function if we have a non-empty prologue.
1371   if (PrologEndLoc && !EmptyPrologue) {
1372     DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc();
1373     maybeRecordLocation(FnStartDL, MF);
1374   }
1375 }
1376 
1377 static bool shouldEmitUdt(const DIType *T) {
1378   if (!T)
1379     return false;
1380 
1381   // MSVC does not emit UDTs for typedefs that are scoped to classes.
1382   if (T->getTag() == dwarf::DW_TAG_typedef) {
1383     if (DIScope *Scope = T->getScope().resolve()) {
1384       switch (Scope->getTag()) {
1385       case dwarf::DW_TAG_structure_type:
1386       case dwarf::DW_TAG_class_type:
1387       case dwarf::DW_TAG_union_type:
1388         return false;
1389       }
1390     }
1391   }
1392 
1393   while (true) {
1394     if (!T || T->isForwardDecl())
1395       return false;
1396 
1397     const DIDerivedType *DT = dyn_cast<DIDerivedType>(T);
1398     if (!DT)
1399       return true;
1400     T = DT->getBaseType().resolve();
1401   }
1402   return true;
1403 }
1404 
1405 void CodeViewDebug::addToUDTs(const DIType *Ty) {
1406   // Don't record empty UDTs.
1407   if (Ty->getName().empty())
1408     return;
1409   if (!shouldEmitUdt(Ty))
1410     return;
1411 
1412   SmallVector<StringRef, 5> QualifiedNameComponents;
1413   const DISubprogram *ClosestSubprogram = getQualifiedNameComponents(
1414       Ty->getScope().resolve(), QualifiedNameComponents);
1415 
1416   std::string FullyQualifiedName =
1417       getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty));
1418 
1419   if (ClosestSubprogram == nullptr) {
1420     GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty);
1421   } else if (ClosestSubprogram == CurrentSubprogram) {
1422     LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty);
1423   }
1424 
1425   // TODO: What if the ClosestSubprogram is neither null or the current
1426   // subprogram?  Currently, the UDT just gets dropped on the floor.
1427   //
1428   // The current behavior is not desirable.  To get maximal fidelity, we would
1429   // need to perform all type translation before beginning emission of .debug$S
1430   // and then make LocalUDTs a member of FunctionInfo
1431 }
1432 
1433 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) {
1434   // Generic dispatch for lowering an unknown type.
1435   switch (Ty->getTag()) {
1436   case dwarf::DW_TAG_array_type:
1437     return lowerTypeArray(cast<DICompositeType>(Ty));
1438   case dwarf::DW_TAG_typedef:
1439     return lowerTypeAlias(cast<DIDerivedType>(Ty));
1440   case dwarf::DW_TAG_base_type:
1441     return lowerTypeBasic(cast<DIBasicType>(Ty));
1442   case dwarf::DW_TAG_pointer_type:
1443     if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type")
1444       return lowerTypeVFTableShape(cast<DIDerivedType>(Ty));
1445     LLVM_FALLTHROUGH;
1446   case dwarf::DW_TAG_reference_type:
1447   case dwarf::DW_TAG_rvalue_reference_type:
1448     return lowerTypePointer(cast<DIDerivedType>(Ty));
1449   case dwarf::DW_TAG_ptr_to_member_type:
1450     return lowerTypeMemberPointer(cast<DIDerivedType>(Ty));
1451   case dwarf::DW_TAG_restrict_type:
1452   case dwarf::DW_TAG_const_type:
1453   case dwarf::DW_TAG_volatile_type:
1454   // TODO: add support for DW_TAG_atomic_type here
1455     return lowerTypeModifier(cast<DIDerivedType>(Ty));
1456   case dwarf::DW_TAG_subroutine_type:
1457     if (ClassTy) {
1458       // The member function type of a member function pointer has no
1459       // ThisAdjustment.
1460       return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy,
1461                                      /*ThisAdjustment=*/0,
1462                                      /*IsStaticMethod=*/false);
1463     }
1464     return lowerTypeFunction(cast<DISubroutineType>(Ty));
1465   case dwarf::DW_TAG_enumeration_type:
1466     return lowerTypeEnum(cast<DICompositeType>(Ty));
1467   case dwarf::DW_TAG_class_type:
1468   case dwarf::DW_TAG_structure_type:
1469     return lowerTypeClass(cast<DICompositeType>(Ty));
1470   case dwarf::DW_TAG_union_type:
1471     return lowerTypeUnion(cast<DICompositeType>(Ty));
1472   case dwarf::DW_TAG_unspecified_type:
1473     if (Ty->getName() == "decltype(nullptr)")
1474       return TypeIndex::NullptrT();
1475     return TypeIndex::None();
1476   default:
1477     // Use the null type index.
1478     return TypeIndex();
1479   }
1480 }
1481 
1482 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) {
1483   DITypeRef UnderlyingTypeRef = Ty->getBaseType();
1484   TypeIndex UnderlyingTypeIndex = getTypeIndex(UnderlyingTypeRef);
1485   StringRef TypeName = Ty->getName();
1486 
1487   addToUDTs(Ty);
1488 
1489   if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) &&
1490       TypeName == "HRESULT")
1491     return TypeIndex(SimpleTypeKind::HResult);
1492   if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) &&
1493       TypeName == "wchar_t")
1494     return TypeIndex(SimpleTypeKind::WideCharacter);
1495 
1496   return UnderlyingTypeIndex;
1497 }
1498 
1499 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) {
1500   DITypeRef ElementTypeRef = Ty->getBaseType();
1501   TypeIndex ElementTypeIndex = getTypeIndex(ElementTypeRef);
1502   // IndexType is size_t, which depends on the bitness of the target.
1503   TypeIndex IndexType = getPointerSizeInBytes() == 8
1504                             ? TypeIndex(SimpleTypeKind::UInt64Quad)
1505                             : TypeIndex(SimpleTypeKind::UInt32Long);
1506 
1507   uint64_t ElementSize = getBaseTypeSize(ElementTypeRef) / 8;
1508 
1509   // Add subranges to array type.
1510   DINodeArray Elements = Ty->getElements();
1511   for (int i = Elements.size() - 1; i >= 0; --i) {
1512     const DINode *Element = Elements[i];
1513     assert(Element->getTag() == dwarf::DW_TAG_subrange_type);
1514 
1515     const DISubrange *Subrange = cast<DISubrange>(Element);
1516     assert(Subrange->getLowerBound() == 0 &&
1517            "codeview doesn't support subranges with lower bounds");
1518     int64_t Count = -1;
1519     if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>())
1520       Count = CI->getSExtValue();
1521 
1522     // Forward declarations of arrays without a size and VLAs use a count of -1.
1523     // Emit a count of zero in these cases to match what MSVC does for arrays
1524     // without a size. MSVC doesn't support VLAs, so it's not clear what we
1525     // should do for them even if we could distinguish them.
1526     if (Count == -1)
1527       Count = 0;
1528 
1529     // Update the element size and element type index for subsequent subranges.
1530     ElementSize *= Count;
1531 
1532     // If this is the outermost array, use the size from the array. It will be
1533     // more accurate if we had a VLA or an incomplete element type size.
1534     uint64_t ArraySize =
1535         (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize;
1536 
1537     StringRef Name = (i == 0) ? Ty->getName() : "";
1538     ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name);
1539     ElementTypeIndex = TypeTable.writeLeafType(AR);
1540   }
1541 
1542   return ElementTypeIndex;
1543 }
1544 
1545 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) {
1546   TypeIndex Index;
1547   dwarf::TypeKind Kind;
1548   uint32_t ByteSize;
1549 
1550   Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding());
1551   ByteSize = Ty->getSizeInBits() / 8;
1552 
1553   SimpleTypeKind STK = SimpleTypeKind::None;
1554   switch (Kind) {
1555   case dwarf::DW_ATE_address:
1556     // FIXME: Translate
1557     break;
1558   case dwarf::DW_ATE_boolean:
1559     switch (ByteSize) {
1560     case 1:  STK = SimpleTypeKind::Boolean8;   break;
1561     case 2:  STK = SimpleTypeKind::Boolean16;  break;
1562     case 4:  STK = SimpleTypeKind::Boolean32;  break;
1563     case 8:  STK = SimpleTypeKind::Boolean64;  break;
1564     case 16: STK = SimpleTypeKind::Boolean128; break;
1565     }
1566     break;
1567   case dwarf::DW_ATE_complex_float:
1568     switch (ByteSize) {
1569     case 2:  STK = SimpleTypeKind::Complex16;  break;
1570     case 4:  STK = SimpleTypeKind::Complex32;  break;
1571     case 8:  STK = SimpleTypeKind::Complex64;  break;
1572     case 10: STK = SimpleTypeKind::Complex80;  break;
1573     case 16: STK = SimpleTypeKind::Complex128; break;
1574     }
1575     break;
1576   case dwarf::DW_ATE_float:
1577     switch (ByteSize) {
1578     case 2:  STK = SimpleTypeKind::Float16;  break;
1579     case 4:  STK = SimpleTypeKind::Float32;  break;
1580     case 6:  STK = SimpleTypeKind::Float48;  break;
1581     case 8:  STK = SimpleTypeKind::Float64;  break;
1582     case 10: STK = SimpleTypeKind::Float80;  break;
1583     case 16: STK = SimpleTypeKind::Float128; break;
1584     }
1585     break;
1586   case dwarf::DW_ATE_signed:
1587     switch (ByteSize) {
1588     case 1:  STK = SimpleTypeKind::SignedCharacter; break;
1589     case 2:  STK = SimpleTypeKind::Int16Short;      break;
1590     case 4:  STK = SimpleTypeKind::Int32;           break;
1591     case 8:  STK = SimpleTypeKind::Int64Quad;       break;
1592     case 16: STK = SimpleTypeKind::Int128Oct;       break;
1593     }
1594     break;
1595   case dwarf::DW_ATE_unsigned:
1596     switch (ByteSize) {
1597     case 1:  STK = SimpleTypeKind::UnsignedCharacter; break;
1598     case 2:  STK = SimpleTypeKind::UInt16Short;       break;
1599     case 4:  STK = SimpleTypeKind::UInt32;            break;
1600     case 8:  STK = SimpleTypeKind::UInt64Quad;        break;
1601     case 16: STK = SimpleTypeKind::UInt128Oct;        break;
1602     }
1603     break;
1604   case dwarf::DW_ATE_UTF:
1605     switch (ByteSize) {
1606     case 2: STK = SimpleTypeKind::Character16; break;
1607     case 4: STK = SimpleTypeKind::Character32; break;
1608     }
1609     break;
1610   case dwarf::DW_ATE_signed_char:
1611     if (ByteSize == 1)
1612       STK = SimpleTypeKind::SignedCharacter;
1613     break;
1614   case dwarf::DW_ATE_unsigned_char:
1615     if (ByteSize == 1)
1616       STK = SimpleTypeKind::UnsignedCharacter;
1617     break;
1618   default:
1619     break;
1620   }
1621 
1622   // Apply some fixups based on the source-level type name.
1623   if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int")
1624     STK = SimpleTypeKind::Int32Long;
1625   if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int")
1626     STK = SimpleTypeKind::UInt32Long;
1627   if (STK == SimpleTypeKind::UInt16Short &&
1628       (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t"))
1629     STK = SimpleTypeKind::WideCharacter;
1630   if ((STK == SimpleTypeKind::SignedCharacter ||
1631        STK == SimpleTypeKind::UnsignedCharacter) &&
1632       Ty->getName() == "char")
1633     STK = SimpleTypeKind::NarrowCharacter;
1634 
1635   return TypeIndex(STK);
1636 }
1637 
1638 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty,
1639                                           PointerOptions PO) {
1640   TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType());
1641 
1642   // Pointers to simple types without any options can use SimpleTypeMode, rather
1643   // than having a dedicated pointer type record.
1644   if (PointeeTI.isSimple() && PO == PointerOptions::None &&
1645       PointeeTI.getSimpleMode() == SimpleTypeMode::Direct &&
1646       Ty->getTag() == dwarf::DW_TAG_pointer_type) {
1647     SimpleTypeMode Mode = Ty->getSizeInBits() == 64
1648                               ? SimpleTypeMode::NearPointer64
1649                               : SimpleTypeMode::NearPointer32;
1650     return TypeIndex(PointeeTI.getSimpleKind(), Mode);
1651   }
1652 
1653   PointerKind PK =
1654       Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32;
1655   PointerMode PM = PointerMode::Pointer;
1656   switch (Ty->getTag()) {
1657   default: llvm_unreachable("not a pointer tag type");
1658   case dwarf::DW_TAG_pointer_type:
1659     PM = PointerMode::Pointer;
1660     break;
1661   case dwarf::DW_TAG_reference_type:
1662     PM = PointerMode::LValueReference;
1663     break;
1664   case dwarf::DW_TAG_rvalue_reference_type:
1665     PM = PointerMode::RValueReference;
1666     break;
1667   }
1668 
1669   if (Ty->isObjectPointer())
1670     PO |= PointerOptions::Const;
1671 
1672   PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8);
1673   return TypeTable.writeLeafType(PR);
1674 }
1675 
1676 static PointerToMemberRepresentation
1677 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) {
1678   // SizeInBytes being zero generally implies that the member pointer type was
1679   // incomplete, which can happen if it is part of a function prototype. In this
1680   // case, use the unknown model instead of the general model.
1681   if (IsPMF) {
1682     switch (Flags & DINode::FlagPtrToMemberRep) {
1683     case 0:
1684       return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown
1685                               : PointerToMemberRepresentation::GeneralFunction;
1686     case DINode::FlagSingleInheritance:
1687       return PointerToMemberRepresentation::SingleInheritanceFunction;
1688     case DINode::FlagMultipleInheritance:
1689       return PointerToMemberRepresentation::MultipleInheritanceFunction;
1690     case DINode::FlagVirtualInheritance:
1691       return PointerToMemberRepresentation::VirtualInheritanceFunction;
1692     }
1693   } else {
1694     switch (Flags & DINode::FlagPtrToMemberRep) {
1695     case 0:
1696       return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown
1697                               : PointerToMemberRepresentation::GeneralData;
1698     case DINode::FlagSingleInheritance:
1699       return PointerToMemberRepresentation::SingleInheritanceData;
1700     case DINode::FlagMultipleInheritance:
1701       return PointerToMemberRepresentation::MultipleInheritanceData;
1702     case DINode::FlagVirtualInheritance:
1703       return PointerToMemberRepresentation::VirtualInheritanceData;
1704     }
1705   }
1706   llvm_unreachable("invalid ptr to member representation");
1707 }
1708 
1709 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty,
1710                                                 PointerOptions PO) {
1711   assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type);
1712   TypeIndex ClassTI = getTypeIndex(Ty->getClassType());
1713   TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType());
1714   PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64
1715                                                 : PointerKind::Near32;
1716   bool IsPMF = isa<DISubroutineType>(Ty->getBaseType());
1717   PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction
1718                          : PointerMode::PointerToDataMember;
1719 
1720   assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big");
1721   uint8_t SizeInBytes = Ty->getSizeInBits() / 8;
1722   MemberPointerInfo MPI(
1723       ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags()));
1724   PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI);
1725   return TypeTable.writeLeafType(PR);
1726 }
1727 
1728 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't
1729 /// have a translation, use the NearC convention.
1730 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) {
1731   switch (DwarfCC) {
1732   case dwarf::DW_CC_normal:             return CallingConvention::NearC;
1733   case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast;
1734   case dwarf::DW_CC_BORLAND_thiscall:   return CallingConvention::ThisCall;
1735   case dwarf::DW_CC_BORLAND_stdcall:    return CallingConvention::NearStdCall;
1736   case dwarf::DW_CC_BORLAND_pascal:     return CallingConvention::NearPascal;
1737   case dwarf::DW_CC_LLVM_vectorcall:    return CallingConvention::NearVector;
1738   }
1739   return CallingConvention::NearC;
1740 }
1741 
1742 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) {
1743   ModifierOptions Mods = ModifierOptions::None;
1744   PointerOptions PO = PointerOptions::None;
1745   bool IsModifier = true;
1746   const DIType *BaseTy = Ty;
1747   while (IsModifier && BaseTy) {
1748     // FIXME: Need to add DWARF tags for __unaligned and _Atomic
1749     switch (BaseTy->getTag()) {
1750     case dwarf::DW_TAG_const_type:
1751       Mods |= ModifierOptions::Const;
1752       PO |= PointerOptions::Const;
1753       break;
1754     case dwarf::DW_TAG_volatile_type:
1755       Mods |= ModifierOptions::Volatile;
1756       PO |= PointerOptions::Volatile;
1757       break;
1758     case dwarf::DW_TAG_restrict_type:
1759       // Only pointer types be marked with __restrict. There is no known flag
1760       // for __restrict in LF_MODIFIER records.
1761       PO |= PointerOptions::Restrict;
1762       break;
1763     default:
1764       IsModifier = false;
1765       break;
1766     }
1767     if (IsModifier)
1768       BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType().resolve();
1769   }
1770 
1771   // Check if the inner type will use an LF_POINTER record. If so, the
1772   // qualifiers will go in the LF_POINTER record. This comes up for types like
1773   // 'int *const' and 'int *__restrict', not the more common cases like 'const
1774   // char *'.
1775   if (BaseTy) {
1776     switch (BaseTy->getTag()) {
1777     case dwarf::DW_TAG_pointer_type:
1778     case dwarf::DW_TAG_reference_type:
1779     case dwarf::DW_TAG_rvalue_reference_type:
1780       return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO);
1781     case dwarf::DW_TAG_ptr_to_member_type:
1782       return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO);
1783     default:
1784       break;
1785     }
1786   }
1787 
1788   TypeIndex ModifiedTI = getTypeIndex(BaseTy);
1789 
1790   // Return the base type index if there aren't any modifiers. For example, the
1791   // metadata could contain restrict wrappers around non-pointer types.
1792   if (Mods == ModifierOptions::None)
1793     return ModifiedTI;
1794 
1795   ModifierRecord MR(ModifiedTI, Mods);
1796   return TypeTable.writeLeafType(MR);
1797 }
1798 
1799 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) {
1800   SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices;
1801   for (DITypeRef ArgTypeRef : Ty->getTypeArray())
1802     ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgTypeRef));
1803 
1804   // MSVC uses type none for variadic argument.
1805   if (ReturnAndArgTypeIndices.size() > 1 &&
1806       ReturnAndArgTypeIndices.back() == TypeIndex::Void()) {
1807     ReturnAndArgTypeIndices.back() = TypeIndex::None();
1808   }
1809   TypeIndex ReturnTypeIndex = TypeIndex::Void();
1810   ArrayRef<TypeIndex> ArgTypeIndices = None;
1811   if (!ReturnAndArgTypeIndices.empty()) {
1812     auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices);
1813     ReturnTypeIndex = ReturnAndArgTypesRef.front();
1814     ArgTypeIndices = ReturnAndArgTypesRef.drop_front();
1815   }
1816 
1817   ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices);
1818   TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec);
1819 
1820   CallingConvention CC = dwarfCCToCodeView(Ty->getCC());
1821 
1822   FunctionOptions FO = getFunctionOptions(Ty);
1823   ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(),
1824                             ArgListIndex);
1825   return TypeTable.writeLeafType(Procedure);
1826 }
1827 
1828 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty,
1829                                                  const DIType *ClassTy,
1830                                                  int ThisAdjustment,
1831                                                  bool IsStaticMethod,
1832                                                  FunctionOptions FO) {
1833   // Lower the containing class type.
1834   TypeIndex ClassType = getTypeIndex(ClassTy);
1835 
1836   DITypeRefArray ReturnAndArgs = Ty->getTypeArray();
1837 
1838   unsigned Index = 0;
1839   SmallVector<TypeIndex, 8> ArgTypeIndices;
1840   TypeIndex ReturnTypeIndex = TypeIndex::Void();
1841   if (ReturnAndArgs.size() > Index) {
1842     ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]);
1843   }
1844 
1845   // If the first argument is a pointer type and this isn't a static method,
1846   // treat it as the special 'this' parameter, which is encoded separately from
1847   // the arguments.
1848   TypeIndex ThisTypeIndex;
1849   if (!IsStaticMethod && ReturnAndArgs.size() > Index) {
1850     if (const DIDerivedType *PtrTy =
1851             dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index].resolve())) {
1852       if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) {
1853         ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty);
1854         Index++;
1855       }
1856     }
1857   }
1858 
1859   while (Index < ReturnAndArgs.size())
1860     ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++]));
1861 
1862   // MSVC uses type none for variadic argument.
1863   if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void())
1864     ArgTypeIndices.back() = TypeIndex::None();
1865 
1866   ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices);
1867   TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec);
1868 
1869   CallingConvention CC = dwarfCCToCodeView(Ty->getCC());
1870 
1871   MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO,
1872                            ArgTypeIndices.size(), ArgListIndex, ThisAdjustment);
1873   return TypeTable.writeLeafType(MFR);
1874 }
1875 
1876 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) {
1877   unsigned VSlotCount =
1878       Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize());
1879   SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near);
1880 
1881   VFTableShapeRecord VFTSR(Slots);
1882   return TypeTable.writeLeafType(VFTSR);
1883 }
1884 
1885 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) {
1886   switch (Flags & DINode::FlagAccessibility) {
1887   case DINode::FlagPrivate:   return MemberAccess::Private;
1888   case DINode::FlagPublic:    return MemberAccess::Public;
1889   case DINode::FlagProtected: return MemberAccess::Protected;
1890   case 0:
1891     // If there was no explicit access control, provide the default for the tag.
1892     return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private
1893                                                  : MemberAccess::Public;
1894   }
1895   llvm_unreachable("access flags are exclusive");
1896 }
1897 
1898 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) {
1899   if (SP->isArtificial())
1900     return MethodOptions::CompilerGenerated;
1901 
1902   // FIXME: Handle other MethodOptions.
1903 
1904   return MethodOptions::None;
1905 }
1906 
1907 static MethodKind translateMethodKindFlags(const DISubprogram *SP,
1908                                            bool Introduced) {
1909   if (SP->getFlags() & DINode::FlagStaticMember)
1910     return MethodKind::Static;
1911 
1912   switch (SP->getVirtuality()) {
1913   case dwarf::DW_VIRTUALITY_none:
1914     break;
1915   case dwarf::DW_VIRTUALITY_virtual:
1916     return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual;
1917   case dwarf::DW_VIRTUALITY_pure_virtual:
1918     return Introduced ? MethodKind::PureIntroducingVirtual
1919                       : MethodKind::PureVirtual;
1920   default:
1921     llvm_unreachable("unhandled virtuality case");
1922   }
1923 
1924   return MethodKind::Vanilla;
1925 }
1926 
1927 static TypeRecordKind getRecordKind(const DICompositeType *Ty) {
1928   switch (Ty->getTag()) {
1929   case dwarf::DW_TAG_class_type:     return TypeRecordKind::Class;
1930   case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct;
1931   }
1932   llvm_unreachable("unexpected tag");
1933 }
1934 
1935 /// Return ClassOptions that should be present on both the forward declaration
1936 /// and the defintion of a tag type.
1937 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) {
1938   ClassOptions CO = ClassOptions::None;
1939 
1940   // MSVC always sets this flag, even for local types. Clang doesn't always
1941   // appear to give every type a linkage name, which may be problematic for us.
1942   // FIXME: Investigate the consequences of not following them here.
1943   if (!Ty->getIdentifier().empty())
1944     CO |= ClassOptions::HasUniqueName;
1945 
1946   // Put the Nested flag on a type if it appears immediately inside a tag type.
1947   // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass
1948   // here. That flag is only set on definitions, and not forward declarations.
1949   const DIScope *ImmediateScope = Ty->getScope().resolve();
1950   if (ImmediateScope && isa<DICompositeType>(ImmediateScope))
1951     CO |= ClassOptions::Nested;
1952 
1953   // Put the Scoped flag on function-local types. MSVC puts this flag for enum
1954   // type only when it has an immediate function scope. Clang never puts enums
1955   // inside DILexicalBlock scopes. Enum types, as generated by clang, are
1956   // always in function, class, or file scopes.
1957   if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) {
1958     if (ImmediateScope && isa<DISubprogram>(ImmediateScope))
1959       CO |= ClassOptions::Scoped;
1960   } else {
1961     for (const DIScope *Scope = ImmediateScope; Scope != nullptr;
1962          Scope = Scope->getScope().resolve()) {
1963       if (isa<DISubprogram>(Scope)) {
1964         CO |= ClassOptions::Scoped;
1965         break;
1966       }
1967     }
1968   }
1969 
1970   return CO;
1971 }
1972 
1973 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) {
1974   switch (Ty->getTag()) {
1975   case dwarf::DW_TAG_class_type:
1976   case dwarf::DW_TAG_structure_type:
1977   case dwarf::DW_TAG_union_type:
1978   case dwarf::DW_TAG_enumeration_type:
1979     break;
1980   default:
1981     return;
1982   }
1983 
1984   if (const auto *File = Ty->getFile()) {
1985     StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File));
1986     TypeIndex SIDI = TypeTable.writeLeafType(SIDR);
1987 
1988     UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine());
1989     TypeTable.writeLeafType(USLR);
1990   }
1991 }
1992 
1993 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) {
1994   ClassOptions CO = getCommonClassOptions(Ty);
1995   TypeIndex FTI;
1996   unsigned EnumeratorCount = 0;
1997 
1998   if (Ty->isForwardDecl()) {
1999     CO |= ClassOptions::ForwardReference;
2000   } else {
2001     ContinuationRecordBuilder ContinuationBuilder;
2002     ContinuationBuilder.begin(ContinuationRecordKind::FieldList);
2003     for (const DINode *Element : Ty->getElements()) {
2004       // We assume that the frontend provides all members in source declaration
2005       // order, which is what MSVC does.
2006       if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) {
2007         EnumeratorRecord ER(MemberAccess::Public,
2008                             APSInt::getUnsigned(Enumerator->getValue()),
2009                             Enumerator->getName());
2010         ContinuationBuilder.writeMemberType(ER);
2011         EnumeratorCount++;
2012       }
2013     }
2014     FTI = TypeTable.insertRecord(ContinuationBuilder);
2015   }
2016 
2017   std::string FullName = getFullyQualifiedName(Ty);
2018 
2019   EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(),
2020                 getTypeIndex(Ty->getBaseType()));
2021   TypeIndex EnumTI = TypeTable.writeLeafType(ER);
2022 
2023   addUDTSrcLine(Ty, EnumTI);
2024 
2025   return EnumTI;
2026 }
2027 
2028 //===----------------------------------------------------------------------===//
2029 // ClassInfo
2030 //===----------------------------------------------------------------------===//
2031 
2032 struct llvm::ClassInfo {
2033   struct MemberInfo {
2034     const DIDerivedType *MemberTypeNode;
2035     uint64_t BaseOffset;
2036   };
2037   // [MemberInfo]
2038   using MemberList = std::vector<MemberInfo>;
2039 
2040   using MethodsList = TinyPtrVector<const DISubprogram *>;
2041   // MethodName -> MethodsList
2042   using MethodsMap = MapVector<MDString *, MethodsList>;
2043 
2044   /// Base classes.
2045   std::vector<const DIDerivedType *> Inheritance;
2046 
2047   /// Direct members.
2048   MemberList Members;
2049   // Direct overloaded methods gathered by name.
2050   MethodsMap Methods;
2051 
2052   TypeIndex VShapeTI;
2053 
2054   std::vector<const DIType *> NestedTypes;
2055 };
2056 
2057 void CodeViewDebug::clear() {
2058   assert(CurFn == nullptr);
2059   FileIdMap.clear();
2060   FnDebugInfo.clear();
2061   FileToFilepathMap.clear();
2062   LocalUDTs.clear();
2063   GlobalUDTs.clear();
2064   TypeIndices.clear();
2065   CompleteTypeIndices.clear();
2066   ScopeGlobals.clear();
2067 }
2068 
2069 void CodeViewDebug::collectMemberInfo(ClassInfo &Info,
2070                                       const DIDerivedType *DDTy) {
2071   if (!DDTy->getName().empty()) {
2072     Info.Members.push_back({DDTy, 0});
2073     return;
2074   }
2075 
2076   // An unnamed member may represent a nested struct or union. Attempt to
2077   // interpret the unnamed member as a DICompositeType possibly wrapped in
2078   // qualifier types. Add all the indirect fields to the current record if that
2079   // succeeds, and drop the member if that fails.
2080   assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!");
2081   uint64_t Offset = DDTy->getOffsetInBits();
2082   const DIType *Ty = DDTy->getBaseType().resolve();
2083   bool FullyResolved = false;
2084   while (!FullyResolved) {
2085     switch (Ty->getTag()) {
2086     case dwarf::DW_TAG_const_type:
2087     case dwarf::DW_TAG_volatile_type:
2088       // FIXME: we should apply the qualifier types to the indirect fields
2089       // rather than dropping them.
2090       Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve();
2091       break;
2092     default:
2093       FullyResolved = true;
2094       break;
2095     }
2096   }
2097 
2098   const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty);
2099   if (!DCTy)
2100     return;
2101 
2102   ClassInfo NestedInfo = collectClassInfo(DCTy);
2103   for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members)
2104     Info.Members.push_back(
2105         {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset});
2106 }
2107 
2108 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) {
2109   ClassInfo Info;
2110   // Add elements to structure type.
2111   DINodeArray Elements = Ty->getElements();
2112   for (auto *Element : Elements) {
2113     // We assume that the frontend provides all members in source declaration
2114     // order, which is what MSVC does.
2115     if (!Element)
2116       continue;
2117     if (auto *SP = dyn_cast<DISubprogram>(Element)) {
2118       Info.Methods[SP->getRawName()].push_back(SP);
2119     } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) {
2120       if (DDTy->getTag() == dwarf::DW_TAG_member) {
2121         collectMemberInfo(Info, DDTy);
2122       } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) {
2123         Info.Inheritance.push_back(DDTy);
2124       } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type &&
2125                  DDTy->getName() == "__vtbl_ptr_type") {
2126         Info.VShapeTI = getTypeIndex(DDTy);
2127       } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) {
2128         Info.NestedTypes.push_back(DDTy);
2129       } else if (DDTy->getTag() == dwarf::DW_TAG_friend) {
2130         // Ignore friend members. It appears that MSVC emitted info about
2131         // friends in the past, but modern versions do not.
2132       }
2133     } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) {
2134       Info.NestedTypes.push_back(Composite);
2135     }
2136     // Skip other unrecognized kinds of elements.
2137   }
2138   return Info;
2139 }
2140 
2141 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) {
2142   // This routine is used by lowerTypeClass and lowerTypeUnion to determine
2143   // if a complete type should be emitted instead of a forward reference.
2144   return Ty->getName().empty() && Ty->getIdentifier().empty() &&
2145       !Ty->isForwardDecl();
2146 }
2147 
2148 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) {
2149   // Emit the complete type for unnamed structs.  C++ classes with methods
2150   // which have a circular reference back to the class type are expected to
2151   // be named by the front-end and should not be "unnamed".  C unnamed
2152   // structs should not have circular references.
2153   if (shouldAlwaysEmitCompleteClassType(Ty)) {
2154     // If this unnamed complete type is already in the process of being defined
2155     // then the description of the type is malformed and cannot be emitted
2156     // into CodeView correctly so report a fatal error.
2157     auto I = CompleteTypeIndices.find(Ty);
2158     if (I != CompleteTypeIndices.end() && I->second == TypeIndex())
2159       report_fatal_error("cannot debug circular reference to unnamed type");
2160     return getCompleteTypeIndex(Ty);
2161   }
2162 
2163   // First, construct the forward decl.  Don't look into Ty to compute the
2164   // forward decl options, since it might not be available in all TUs.
2165   TypeRecordKind Kind = getRecordKind(Ty);
2166   ClassOptions CO =
2167       ClassOptions::ForwardReference | getCommonClassOptions(Ty);
2168   std::string FullName = getFullyQualifiedName(Ty);
2169   ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0,
2170                  FullName, Ty->getIdentifier());
2171   TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR);
2172   if (!Ty->isForwardDecl())
2173     DeferredCompleteTypes.push_back(Ty);
2174   return FwdDeclTI;
2175 }
2176 
2177 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) {
2178   // Construct the field list and complete type record.
2179   TypeRecordKind Kind = getRecordKind(Ty);
2180   ClassOptions CO = getCommonClassOptions(Ty);
2181   TypeIndex FieldTI;
2182   TypeIndex VShapeTI;
2183   unsigned FieldCount;
2184   bool ContainsNestedClass;
2185   std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) =
2186       lowerRecordFieldList(Ty);
2187 
2188   if (ContainsNestedClass)
2189     CO |= ClassOptions::ContainsNestedClass;
2190 
2191   // MSVC appears to set this flag by searching any destructor or method with
2192   // FunctionOptions::Constructor among the emitted members. Clang AST has all
2193   // the members, however special member functions are not yet emitted into
2194   // debug information. For now checking a class's non-triviality seems enough.
2195   // FIXME: not true for a nested unnamed struct.
2196   if (isNonTrivial(Ty))
2197     CO |= ClassOptions::HasConstructorOrDestructor;
2198 
2199   std::string FullName = getFullyQualifiedName(Ty);
2200 
2201   uint64_t SizeInBytes = Ty->getSizeInBits() / 8;
2202 
2203   ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI,
2204                  SizeInBytes, FullName, Ty->getIdentifier());
2205   TypeIndex ClassTI = TypeTable.writeLeafType(CR);
2206 
2207   addUDTSrcLine(Ty, ClassTI);
2208 
2209   addToUDTs(Ty);
2210 
2211   return ClassTI;
2212 }
2213 
2214 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) {
2215   // Emit the complete type for unnamed unions.
2216   if (shouldAlwaysEmitCompleteClassType(Ty))
2217     return getCompleteTypeIndex(Ty);
2218 
2219   ClassOptions CO =
2220       ClassOptions::ForwardReference | getCommonClassOptions(Ty);
2221   std::string FullName = getFullyQualifiedName(Ty);
2222   UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier());
2223   TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR);
2224   if (!Ty->isForwardDecl())
2225     DeferredCompleteTypes.push_back(Ty);
2226   return FwdDeclTI;
2227 }
2228 
2229 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) {
2230   ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty);
2231   TypeIndex FieldTI;
2232   unsigned FieldCount;
2233   bool ContainsNestedClass;
2234   std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) =
2235       lowerRecordFieldList(Ty);
2236 
2237   if (ContainsNestedClass)
2238     CO |= ClassOptions::ContainsNestedClass;
2239 
2240   uint64_t SizeInBytes = Ty->getSizeInBits() / 8;
2241   std::string FullName = getFullyQualifiedName(Ty);
2242 
2243   UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName,
2244                  Ty->getIdentifier());
2245   TypeIndex UnionTI = TypeTable.writeLeafType(UR);
2246 
2247   addUDTSrcLine(Ty, UnionTI);
2248 
2249   addToUDTs(Ty);
2250 
2251   return UnionTI;
2252 }
2253 
2254 std::tuple<TypeIndex, TypeIndex, unsigned, bool>
2255 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) {
2256   // Manually count members. MSVC appears to count everything that generates a
2257   // field list record. Each individual overload in a method overload group
2258   // contributes to this count, even though the overload group is a single field
2259   // list record.
2260   unsigned MemberCount = 0;
2261   ClassInfo Info = collectClassInfo(Ty);
2262   ContinuationRecordBuilder ContinuationBuilder;
2263   ContinuationBuilder.begin(ContinuationRecordKind::FieldList);
2264 
2265   // Create base classes.
2266   for (const DIDerivedType *I : Info.Inheritance) {
2267     if (I->getFlags() & DINode::FlagVirtual) {
2268       // Virtual base.
2269       unsigned VBPtrOffset = I->getVBPtrOffset();
2270       // FIXME: Despite the accessor name, the offset is really in bytes.
2271       unsigned VBTableIndex = I->getOffsetInBits() / 4;
2272       auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase
2273                             ? TypeRecordKind::IndirectVirtualBaseClass
2274                             : TypeRecordKind::VirtualBaseClass;
2275       VirtualBaseClassRecord VBCR(
2276           RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()),
2277           getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset,
2278           VBTableIndex);
2279 
2280       ContinuationBuilder.writeMemberType(VBCR);
2281       MemberCount++;
2282     } else {
2283       assert(I->getOffsetInBits() % 8 == 0 &&
2284              "bases must be on byte boundaries");
2285       BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()),
2286                           getTypeIndex(I->getBaseType()),
2287                           I->getOffsetInBits() / 8);
2288       ContinuationBuilder.writeMemberType(BCR);
2289       MemberCount++;
2290     }
2291   }
2292 
2293   // Create members.
2294   for (ClassInfo::MemberInfo &MemberInfo : Info.Members) {
2295     const DIDerivedType *Member = MemberInfo.MemberTypeNode;
2296     TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType());
2297     StringRef MemberName = Member->getName();
2298     MemberAccess Access =
2299         translateAccessFlags(Ty->getTag(), Member->getFlags());
2300 
2301     if (Member->isStaticMember()) {
2302       StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName);
2303       ContinuationBuilder.writeMemberType(SDMR);
2304       MemberCount++;
2305       continue;
2306     }
2307 
2308     // Virtual function pointer member.
2309     if ((Member->getFlags() & DINode::FlagArtificial) &&
2310         Member->getName().startswith("_vptr$")) {
2311       VFPtrRecord VFPR(getTypeIndex(Member->getBaseType()));
2312       ContinuationBuilder.writeMemberType(VFPR);
2313       MemberCount++;
2314       continue;
2315     }
2316 
2317     // Data member.
2318     uint64_t MemberOffsetInBits =
2319         Member->getOffsetInBits() + MemberInfo.BaseOffset;
2320     if (Member->isBitField()) {
2321       uint64_t StartBitOffset = MemberOffsetInBits;
2322       if (const auto *CI =
2323               dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) {
2324         MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset;
2325       }
2326       StartBitOffset -= MemberOffsetInBits;
2327       BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(),
2328                          StartBitOffset);
2329       MemberBaseType = TypeTable.writeLeafType(BFR);
2330     }
2331     uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8;
2332     DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes,
2333                          MemberName);
2334     ContinuationBuilder.writeMemberType(DMR);
2335     MemberCount++;
2336   }
2337 
2338   // Create methods
2339   for (auto &MethodItr : Info.Methods) {
2340     StringRef Name = MethodItr.first->getString();
2341 
2342     std::vector<OneMethodRecord> Methods;
2343     for (const DISubprogram *SP : MethodItr.second) {
2344       TypeIndex MethodType = getMemberFunctionType(SP, Ty);
2345       bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual;
2346 
2347       unsigned VFTableOffset = -1;
2348       if (Introduced)
2349         VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes();
2350 
2351       Methods.push_back(OneMethodRecord(
2352           MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()),
2353           translateMethodKindFlags(SP, Introduced),
2354           translateMethodOptionFlags(SP), VFTableOffset, Name));
2355       MemberCount++;
2356     }
2357     assert(!Methods.empty() && "Empty methods map entry");
2358     if (Methods.size() == 1)
2359       ContinuationBuilder.writeMemberType(Methods[0]);
2360     else {
2361       // FIXME: Make this use its own ContinuationBuilder so that
2362       // MethodOverloadList can be split correctly.
2363       MethodOverloadListRecord MOLR(Methods);
2364       TypeIndex MethodList = TypeTable.writeLeafType(MOLR);
2365 
2366       OverloadedMethodRecord OMR(Methods.size(), MethodList, Name);
2367       ContinuationBuilder.writeMemberType(OMR);
2368     }
2369   }
2370 
2371   // Create nested classes.
2372   for (const DIType *Nested : Info.NestedTypes) {
2373     NestedTypeRecord R(getTypeIndex(DITypeRef(Nested)), Nested->getName());
2374     ContinuationBuilder.writeMemberType(R);
2375     MemberCount++;
2376   }
2377 
2378   TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder);
2379   return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount,
2380                          !Info.NestedTypes.empty());
2381 }
2382 
2383 TypeIndex CodeViewDebug::getVBPTypeIndex() {
2384   if (!VBPType.getIndex()) {
2385     // Make a 'const int *' type.
2386     ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const);
2387     TypeIndex ModifiedTI = TypeTable.writeLeafType(MR);
2388 
2389     PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64
2390                                                   : PointerKind::Near32;
2391     PointerMode PM = PointerMode::Pointer;
2392     PointerOptions PO = PointerOptions::None;
2393     PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes());
2394     VBPType = TypeTable.writeLeafType(PR);
2395   }
2396 
2397   return VBPType;
2398 }
2399 
2400 TypeIndex CodeViewDebug::getTypeIndex(DITypeRef TypeRef, DITypeRef ClassTyRef) {
2401   const DIType *Ty = TypeRef.resolve();
2402   const DIType *ClassTy = ClassTyRef.resolve();
2403 
2404   // The null DIType is the void type. Don't try to hash it.
2405   if (!Ty)
2406     return TypeIndex::Void();
2407 
2408   // Check if we've already translated this type. Don't try to do a
2409   // get-or-create style insertion that caches the hash lookup across the
2410   // lowerType call. It will update the TypeIndices map.
2411   auto I = TypeIndices.find({Ty, ClassTy});
2412   if (I != TypeIndices.end())
2413     return I->second;
2414 
2415   TypeLoweringScope S(*this);
2416   TypeIndex TI = lowerType(Ty, ClassTy);
2417   return recordTypeIndexForDINode(Ty, TI, ClassTy);
2418 }
2419 
2420 codeview::TypeIndex
2421 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy,
2422                                       const DISubroutineType *SubroutineTy) {
2423   assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type &&
2424          "this type must be a pointer type");
2425 
2426   PointerOptions Options = PointerOptions::None;
2427   if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference)
2428     Options = PointerOptions::LValueRefThisPointer;
2429   else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference)
2430     Options = PointerOptions::RValueRefThisPointer;
2431 
2432   // Check if we've already translated this type.  If there is no ref qualifier
2433   // on the function then we look up this pointer type with no associated class
2434   // so that the TypeIndex for the this pointer can be shared with the type
2435   // index for other pointers to this class type.  If there is a ref qualifier
2436   // then we lookup the pointer using the subroutine as the parent type.
2437   auto I = TypeIndices.find({PtrTy, SubroutineTy});
2438   if (I != TypeIndices.end())
2439     return I->second;
2440 
2441   TypeLoweringScope S(*this);
2442   TypeIndex TI = lowerTypePointer(PtrTy, Options);
2443   return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy);
2444 }
2445 
2446 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(DITypeRef TypeRef) {
2447   DIType *Ty = TypeRef.resolve();
2448   PointerRecord PR(getTypeIndex(Ty),
2449                    getPointerSizeInBytes() == 8 ? PointerKind::Near64
2450                                                 : PointerKind::Near32,
2451                    PointerMode::LValueReference, PointerOptions::None,
2452                    Ty->getSizeInBits() / 8);
2453   return TypeTable.writeLeafType(PR);
2454 }
2455 
2456 TypeIndex CodeViewDebug::getCompleteTypeIndex(DITypeRef TypeRef) {
2457   const DIType *Ty = TypeRef.resolve();
2458 
2459   // The null DIType is the void type. Don't try to hash it.
2460   if (!Ty)
2461     return TypeIndex::Void();
2462 
2463   // Look through typedefs when getting the complete type index. Call
2464   // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are
2465   // emitted only once.
2466   if (Ty->getTag() == dwarf::DW_TAG_typedef)
2467     (void)getTypeIndex(Ty);
2468   while (Ty->getTag() == dwarf::DW_TAG_typedef)
2469     Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve();
2470 
2471   // If this is a non-record type, the complete type index is the same as the
2472   // normal type index. Just call getTypeIndex.
2473   switch (Ty->getTag()) {
2474   case dwarf::DW_TAG_class_type:
2475   case dwarf::DW_TAG_structure_type:
2476   case dwarf::DW_TAG_union_type:
2477     break;
2478   default:
2479     return getTypeIndex(Ty);
2480   }
2481 
2482   // Check if we've already translated the complete record type.
2483   const auto *CTy = cast<DICompositeType>(Ty);
2484   auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()});
2485   if (!InsertResult.second)
2486     return InsertResult.first->second;
2487 
2488   TypeLoweringScope S(*this);
2489 
2490   // Make sure the forward declaration is emitted first. It's unclear if this
2491   // is necessary, but MSVC does it, and we should follow suit until we can show
2492   // otherwise.
2493   // We only emit a forward declaration for named types.
2494   if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) {
2495     TypeIndex FwdDeclTI = getTypeIndex(CTy);
2496 
2497     // Just use the forward decl if we don't have complete type info. This
2498     // might happen if the frontend is using modules and expects the complete
2499     // definition to be emitted elsewhere.
2500     if (CTy->isForwardDecl())
2501       return FwdDeclTI;
2502   }
2503 
2504   TypeIndex TI;
2505   switch (CTy->getTag()) {
2506   case dwarf::DW_TAG_class_type:
2507   case dwarf::DW_TAG_structure_type:
2508     TI = lowerCompleteTypeClass(CTy);
2509     break;
2510   case dwarf::DW_TAG_union_type:
2511     TI = lowerCompleteTypeUnion(CTy);
2512     break;
2513   default:
2514     llvm_unreachable("not a record");
2515   }
2516 
2517   // Update the type index associated with this CompositeType.  This cannot
2518   // use the 'InsertResult' iterator above because it is potentially
2519   // invalidated by map insertions which can occur while lowering the class
2520   // type above.
2521   CompleteTypeIndices[CTy] = TI;
2522   return TI;
2523 }
2524 
2525 /// Emit all the deferred complete record types. Try to do this in FIFO order,
2526 /// and do this until fixpoint, as each complete record type typically
2527 /// references
2528 /// many other record types.
2529 void CodeViewDebug::emitDeferredCompleteTypes() {
2530   SmallVector<const DICompositeType *, 4> TypesToEmit;
2531   while (!DeferredCompleteTypes.empty()) {
2532     std::swap(DeferredCompleteTypes, TypesToEmit);
2533     for (const DICompositeType *RecordTy : TypesToEmit)
2534       getCompleteTypeIndex(RecordTy);
2535     TypesToEmit.clear();
2536   }
2537 }
2538 
2539 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI,
2540                                           ArrayRef<LocalVariable> Locals) {
2541   // Get the sorted list of parameters and emit them first.
2542   SmallVector<const LocalVariable *, 6> Params;
2543   for (const LocalVariable &L : Locals)
2544     if (L.DIVar->isParameter())
2545       Params.push_back(&L);
2546   llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) {
2547     return L->DIVar->getArg() < R->DIVar->getArg();
2548   });
2549   for (const LocalVariable *L : Params)
2550     emitLocalVariable(FI, *L);
2551 
2552   // Next emit all non-parameters in the order that we found them.
2553   for (const LocalVariable &L : Locals)
2554     if (!L.DIVar->isParameter())
2555       emitLocalVariable(FI, L);
2556 }
2557 
2558 /// Only call this on endian-specific types like ulittle16_t and little32_t, or
2559 /// structs composed of them.
2560 template <typename T>
2561 static void copyBytesForDefRange(SmallString<20> &BytePrefix,
2562                                  SymbolKind SymKind, const T &DefRangeHeader) {
2563   BytePrefix.resize(2 + sizeof(T));
2564   ulittle16_t SymKindLE = ulittle16_t(SymKind);
2565   memcpy(&BytePrefix[0], &SymKindLE, 2);
2566   memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T));
2567 }
2568 
2569 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI,
2570                                       const LocalVariable &Var) {
2571   // LocalSym record, see SymbolRecord.h for more info.
2572   MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL);
2573 
2574   LocalSymFlags Flags = LocalSymFlags::None;
2575   if (Var.DIVar->isParameter())
2576     Flags |= LocalSymFlags::IsParameter;
2577   if (Var.DefRanges.empty())
2578     Flags |= LocalSymFlags::IsOptimizedOut;
2579 
2580   OS.AddComment("TypeIndex");
2581   TypeIndex TI = Var.UseReferenceType
2582                      ? getTypeIndexForReferenceTo(Var.DIVar->getType())
2583                      : getCompleteTypeIndex(Var.DIVar->getType());
2584   OS.EmitIntValue(TI.getIndex(), 4);
2585   OS.AddComment("Flags");
2586   OS.EmitIntValue(static_cast<uint16_t>(Flags), 2);
2587   // Truncate the name so we won't overflow the record length field.
2588   emitNullTerminatedSymbolName(OS, Var.DIVar->getName());
2589   endSymbolRecord(LocalEnd);
2590 
2591   // Calculate the on disk prefix of the appropriate def range record. The
2592   // records and on disk formats are described in SymbolRecords.h. BytePrefix
2593   // should be big enough to hold all forms without memory allocation.
2594   SmallString<20> BytePrefix;
2595   for (const LocalVarDefRange &DefRange : Var.DefRanges) {
2596     BytePrefix.clear();
2597     if (DefRange.InMemory) {
2598       int Offset = DefRange.DataOffset;
2599       unsigned Reg = DefRange.CVRegister;
2600 
2601       // 32-bit x86 call sequences often use PUSH instructions, which disrupt
2602       // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0,
2603       // instead. In frames without stack realignment, $T0 will be the CFA.
2604       if (RegisterId(Reg) == RegisterId::ESP) {
2605         Reg = unsigned(RegisterId::VFRAME);
2606         Offset += FI.OffsetAdjustment;
2607       }
2608 
2609       // If we can use the chosen frame pointer for the frame and this isn't a
2610       // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record.
2611       // Otherwise, use S_DEFRANGE_REGISTER_REL.
2612       EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU);
2613       if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None &&
2614           (bool(Flags & LocalSymFlags::IsParameter)
2615                ? (EncFP == FI.EncodedParamFramePtrReg)
2616                : (EncFP == FI.EncodedLocalFramePtrReg))) {
2617         little32_t FPOffset = little32_t(Offset);
2618         copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset);
2619       } else {
2620         uint16_t RegRelFlags = 0;
2621         if (DefRange.IsSubfield) {
2622           RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag |
2623                         (DefRange.StructOffset
2624                          << DefRangeRegisterRelSym::OffsetInParentShift);
2625         }
2626         DefRangeRegisterRelSym::Header DRHdr;
2627         DRHdr.Register = Reg;
2628         DRHdr.Flags = RegRelFlags;
2629         DRHdr.BasePointerOffset = Offset;
2630         copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr);
2631       }
2632     } else {
2633       assert(DefRange.DataOffset == 0 && "unexpected offset into register");
2634       if (DefRange.IsSubfield) {
2635         DefRangeSubfieldRegisterSym::Header DRHdr;
2636         DRHdr.Register = DefRange.CVRegister;
2637         DRHdr.MayHaveNoName = 0;
2638         DRHdr.OffsetInParent = DefRange.StructOffset;
2639         copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr);
2640       } else {
2641         DefRangeRegisterSym::Header DRHdr;
2642         DRHdr.Register = DefRange.CVRegister;
2643         DRHdr.MayHaveNoName = 0;
2644         copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr);
2645       }
2646     }
2647     OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix);
2648   }
2649 }
2650 
2651 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks,
2652                                          const FunctionInfo& FI) {
2653   for (LexicalBlock *Block : Blocks)
2654     emitLexicalBlock(*Block, FI);
2655 }
2656 
2657 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a
2658 /// lexical block scope.
2659 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block,
2660                                      const FunctionInfo& FI) {
2661   MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32);
2662   OS.AddComment("PtrParent");
2663   OS.EmitIntValue(0, 4);                                  // PtrParent
2664   OS.AddComment("PtrEnd");
2665   OS.EmitIntValue(0, 4);                                  // PtrEnd
2666   OS.AddComment("Code size");
2667   OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4);   // Code Size
2668   OS.AddComment("Function section relative address");
2669   OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0);         // Func Offset
2670   OS.AddComment("Function section index");
2671   OS.EmitCOFFSectionIndex(FI.Begin);                      // Func Symbol
2672   OS.AddComment("Lexical block name");
2673   emitNullTerminatedSymbolName(OS, Block.Name);           // Name
2674   endSymbolRecord(RecordEnd);
2675 
2676   // Emit variables local to this lexical block.
2677   emitLocalVariableList(FI, Block.Locals);
2678   emitGlobalVariableList(Block.Globals);
2679 
2680   // Emit lexical blocks contained within this block.
2681   emitLexicalBlockList(Block.Children, FI);
2682 
2683   // Close the lexical block scope.
2684   emitEndSymbolRecord(SymbolKind::S_END);
2685 }
2686 
2687 /// Convenience routine for collecting lexical block information for a list
2688 /// of lexical scopes.
2689 void CodeViewDebug::collectLexicalBlockInfo(
2690         SmallVectorImpl<LexicalScope *> &Scopes,
2691         SmallVectorImpl<LexicalBlock *> &Blocks,
2692         SmallVectorImpl<LocalVariable> &Locals,
2693         SmallVectorImpl<CVGlobalVariable> &Globals) {
2694   for (LexicalScope *Scope : Scopes)
2695     collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals);
2696 }
2697 
2698 /// Populate the lexical blocks and local variable lists of the parent with
2699 /// information about the specified lexical scope.
2700 void CodeViewDebug::collectLexicalBlockInfo(
2701     LexicalScope &Scope,
2702     SmallVectorImpl<LexicalBlock *> &ParentBlocks,
2703     SmallVectorImpl<LocalVariable> &ParentLocals,
2704     SmallVectorImpl<CVGlobalVariable> &ParentGlobals) {
2705   if (Scope.isAbstractScope())
2706     return;
2707 
2708   // Gather information about the lexical scope including local variables,
2709   // global variables, and address ranges.
2710   bool IgnoreScope = false;
2711   auto LI = ScopeVariables.find(&Scope);
2712   SmallVectorImpl<LocalVariable> *Locals =
2713       LI != ScopeVariables.end() ? &LI->second : nullptr;
2714   auto GI = ScopeGlobals.find(Scope.getScopeNode());
2715   SmallVectorImpl<CVGlobalVariable> *Globals =
2716       GI != ScopeGlobals.end() ? GI->second.get() : nullptr;
2717   const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode());
2718   const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges();
2719 
2720   // Ignore lexical scopes which do not contain variables.
2721   if (!Locals && !Globals)
2722     IgnoreScope = true;
2723 
2724   // Ignore lexical scopes which are not lexical blocks.
2725   if (!DILB)
2726     IgnoreScope = true;
2727 
2728   // Ignore scopes which have too many address ranges to represent in the
2729   // current CodeView format or do not have a valid address range.
2730   //
2731   // For lexical scopes with multiple address ranges you may be tempted to
2732   // construct a single range covering every instruction where the block is
2733   // live and everything in between.  Unfortunately, Visual Studio only
2734   // displays variables from the first matching lexical block scope.  If the
2735   // first lexical block contains exception handling code or cold code which
2736   // is moved to the bottom of the routine creating a single range covering
2737   // nearly the entire routine, then it will hide all other lexical blocks
2738   // and the variables they contain.
2739   if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second))
2740     IgnoreScope = true;
2741 
2742   if (IgnoreScope) {
2743     // This scope can be safely ignored and eliminating it will reduce the
2744     // size of the debug information. Be sure to collect any variable and scope
2745     // information from the this scope or any of its children and collapse them
2746     // into the parent scope.
2747     if (Locals)
2748       ParentLocals.append(Locals->begin(), Locals->end());
2749     if (Globals)
2750       ParentGlobals.append(Globals->begin(), Globals->end());
2751     collectLexicalBlockInfo(Scope.getChildren(),
2752                             ParentBlocks,
2753                             ParentLocals,
2754                             ParentGlobals);
2755     return;
2756   }
2757 
2758   // Create a new CodeView lexical block for this lexical scope.  If we've
2759   // seen this DILexicalBlock before then the scope tree is malformed and
2760   // we can handle this gracefully by not processing it a second time.
2761   auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()});
2762   if (!BlockInsertion.second)
2763     return;
2764 
2765   // Create a lexical block containing the variables and collect the the
2766   // lexical block information for the children.
2767   const InsnRange &Range = Ranges.front();
2768   assert(Range.first && Range.second);
2769   LexicalBlock &Block = BlockInsertion.first->second;
2770   Block.Begin = getLabelBeforeInsn(Range.first);
2771   Block.End = getLabelAfterInsn(Range.second);
2772   assert(Block.Begin && "missing label for scope begin");
2773   assert(Block.End && "missing label for scope end");
2774   Block.Name = DILB->getName();
2775   if (Locals)
2776     Block.Locals = std::move(*Locals);
2777   if (Globals)
2778     Block.Globals = std::move(*Globals);
2779   ParentBlocks.push_back(&Block);
2780   collectLexicalBlockInfo(Scope.getChildren(),
2781                           Block.Children,
2782                           Block.Locals,
2783                           Block.Globals);
2784 }
2785 
2786 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) {
2787   const Function &GV = MF->getFunction();
2788   assert(FnDebugInfo.count(&GV));
2789   assert(CurFn == FnDebugInfo[&GV].get());
2790 
2791   collectVariableInfo(GV.getSubprogram());
2792 
2793   // Build the lexical block structure to emit for this routine.
2794   if (LexicalScope *CFS = LScopes.getCurrentFunctionScope())
2795     collectLexicalBlockInfo(*CFS,
2796                             CurFn->ChildBlocks,
2797                             CurFn->Locals,
2798                             CurFn->Globals);
2799 
2800   // Clear the scope and variable information from the map which will not be
2801   // valid after we have finished processing this routine.  This also prepares
2802   // the map for the subsequent routine.
2803   ScopeVariables.clear();
2804 
2805   // Don't emit anything if we don't have any line tables.
2806   // Thunks are compiler-generated and probably won't have source correlation.
2807   if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) {
2808     FnDebugInfo.erase(&GV);
2809     CurFn = nullptr;
2810     return;
2811   }
2812 
2813   CurFn->Annotations = MF->getCodeViewAnnotations();
2814 
2815   CurFn->End = Asm->getFunctionEnd();
2816 
2817   CurFn = nullptr;
2818 }
2819 
2820 void CodeViewDebug::beginInstruction(const MachineInstr *MI) {
2821   DebugHandlerBase::beginInstruction(MI);
2822 
2823   // Ignore DBG_VALUE and DBG_LABEL locations and function prologue.
2824   if (!Asm || !CurFn || MI->isDebugInstr() ||
2825       MI->getFlag(MachineInstr::FrameSetup))
2826     return;
2827 
2828   // If the first instruction of a new MBB has no location, find the first
2829   // instruction with a location and use that.
2830   DebugLoc DL = MI->getDebugLoc();
2831   if (!DL && MI->getParent() != PrevInstBB) {
2832     for (const auto &NextMI : *MI->getParent()) {
2833       if (NextMI.isDebugInstr())
2834         continue;
2835       DL = NextMI.getDebugLoc();
2836       if (DL)
2837         break;
2838     }
2839   }
2840   PrevInstBB = MI->getParent();
2841 
2842   // If we still don't have a debug location, don't record a location.
2843   if (!DL)
2844     return;
2845 
2846   maybeRecordLocation(DL, Asm->MF);
2847 }
2848 
2849 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) {
2850   MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(),
2851            *EndLabel = MMI->getContext().createTempSymbol();
2852   OS.EmitIntValue(unsigned(Kind), 4);
2853   OS.AddComment("Subsection size");
2854   OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4);
2855   OS.EmitLabel(BeginLabel);
2856   return EndLabel;
2857 }
2858 
2859 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) {
2860   OS.EmitLabel(EndLabel);
2861   // Every subsection must be aligned to a 4-byte boundary.
2862   OS.EmitValueToAlignment(4);
2863 }
2864 
2865 static StringRef getSymbolName(SymbolKind SymKind) {
2866   for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames())
2867     if (EE.Value == SymKind)
2868       return EE.Name;
2869   return "";
2870 }
2871 
2872 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) {
2873   MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(),
2874            *EndLabel = MMI->getContext().createTempSymbol();
2875   OS.AddComment("Record length");
2876   OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2);
2877   OS.EmitLabel(BeginLabel);
2878   if (OS.isVerboseAsm())
2879     OS.AddComment("Record kind: " + getSymbolName(SymKind));
2880   OS.EmitIntValue(unsigned(SymKind), 2);
2881   return EndLabel;
2882 }
2883 
2884 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) {
2885   // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid
2886   // an extra copy of every symbol record in LLD. This increases object file
2887   // size by less than 1% in the clang build, and is compatible with the Visual
2888   // C++ linker.
2889   OS.EmitValueToAlignment(4);
2890   OS.EmitLabel(SymEnd);
2891 }
2892 
2893 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) {
2894   OS.AddComment("Record length");
2895   OS.EmitIntValue(2, 2);
2896   if (OS.isVerboseAsm())
2897     OS.AddComment("Record kind: " + getSymbolName(EndKind));
2898   OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind
2899 }
2900 
2901 void CodeViewDebug::emitDebugInfoForUDTs(
2902     ArrayRef<std::pair<std::string, const DIType *>> UDTs) {
2903   for (const auto &UDT : UDTs) {
2904     const DIType *T = UDT.second;
2905     assert(shouldEmitUdt(T));
2906 
2907     MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT);
2908     OS.AddComment("Type");
2909     OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4);
2910     emitNullTerminatedSymbolName(OS, UDT.first);
2911     endSymbolRecord(UDTRecordEnd);
2912   }
2913 }
2914 
2915 void CodeViewDebug::collectGlobalVariableInfo() {
2916   DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *>
2917       GlobalMap;
2918   for (const GlobalVariable &GV : MMI->getModule()->globals()) {
2919     SmallVector<DIGlobalVariableExpression *, 1> GVEs;
2920     GV.getDebugInfo(GVEs);
2921     for (const auto *GVE : GVEs)
2922       GlobalMap[GVE] = &GV;
2923   }
2924 
2925   NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
2926   for (const MDNode *Node : CUs->operands()) {
2927     const auto *CU = cast<DICompileUnit>(Node);
2928     for (const auto *GVE : CU->getGlobalVariables()) {
2929       const auto *GV = GlobalMap.lookup(GVE);
2930       if (!GV || GV->isDeclarationForLinker())
2931         continue;
2932       const DIGlobalVariable *DIGV = GVE->getVariable();
2933       DIScope *Scope = DIGV->getScope();
2934       SmallVector<CVGlobalVariable, 1> *VariableList;
2935       if (Scope && isa<DILocalScope>(Scope)) {
2936         // Locate a global variable list for this scope, creating one if
2937         // necessary.
2938         auto Insertion = ScopeGlobals.insert(
2939             {Scope, std::unique_ptr<GlobalVariableList>()});
2940         if (Insertion.second)
2941           Insertion.first->second = llvm::make_unique<GlobalVariableList>();
2942         VariableList = Insertion.first->second.get();
2943       } else if (GV->hasComdat())
2944         // Emit this global variable into a COMDAT section.
2945         VariableList = &ComdatVariables;
2946       else
2947         // Emit this globla variable in a single global symbol section.
2948         VariableList = &GlobalVariables;
2949       CVGlobalVariable CVGV = {DIGV, GV};
2950       VariableList->emplace_back(std::move(CVGV));
2951     }
2952   }
2953 }
2954 
2955 void CodeViewDebug::emitDebugInfoForGlobals() {
2956   // First, emit all globals that are not in a comdat in a single symbol
2957   // substream. MSVC doesn't like it if the substream is empty, so only open
2958   // it if we have at least one global to emit.
2959   switchToDebugSectionForSymbol(nullptr);
2960   if (!GlobalVariables.empty()) {
2961     OS.AddComment("Symbol subsection for globals");
2962     MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols);
2963     emitGlobalVariableList(GlobalVariables);
2964     endCVSubsection(EndLabel);
2965   }
2966 
2967   // Second, emit each global that is in a comdat into its own .debug$S
2968   // section along with its own symbol substream.
2969   for (const CVGlobalVariable &CVGV : ComdatVariables) {
2970     MCSymbol *GVSym = Asm->getSymbol(CVGV.GV);
2971     OS.AddComment("Symbol subsection for " +
2972             Twine(GlobalValue::dropLLVMManglingEscape(CVGV.GV->getName())));
2973     switchToDebugSectionForSymbol(GVSym);
2974     MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols);
2975     // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions.
2976     emitDebugInfoForGlobal(CVGV.DIGV, CVGV.GV, GVSym);
2977     endCVSubsection(EndLabel);
2978   }
2979 }
2980 
2981 void CodeViewDebug::emitDebugInfoForRetainedTypes() {
2982   NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
2983   for (const MDNode *Node : CUs->operands()) {
2984     for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) {
2985       if (DIType *RT = dyn_cast<DIType>(Ty)) {
2986         getTypeIndex(RT);
2987         // FIXME: Add to global/local DTU list.
2988       }
2989     }
2990   }
2991 }
2992 
2993 // Emit each global variable in the specified array.
2994 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) {
2995   for (const CVGlobalVariable &CVGV : Globals) {
2996     MCSymbol *GVSym = Asm->getSymbol(CVGV.GV);
2997     // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions.
2998     emitDebugInfoForGlobal(CVGV.DIGV, CVGV.GV, GVSym);
2999   }
3000 }
3001 
3002 void CodeViewDebug::emitDebugInfoForGlobal(const DIGlobalVariable *DIGV,
3003                                            const GlobalVariable *GV,
3004                                            MCSymbol *GVSym) {
3005   // DataSym record, see SymbolRecord.h for more info. Thread local data
3006   // happens to have the same format as global data.
3007   SymbolKind DataSym = GV->isThreadLocal()
3008                            ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32
3009                                                     : SymbolKind::S_GTHREAD32)
3010                            : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32
3011                                                     : SymbolKind::S_GDATA32);
3012   MCSymbol *DataEnd = beginSymbolRecord(DataSym);
3013   OS.AddComment("Type");
3014   OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4);
3015   OS.AddComment("DataOffset");
3016   OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0);
3017   OS.AddComment("Segment");
3018   OS.EmitCOFFSectionIndex(GVSym);
3019   OS.AddComment("Name");
3020   const unsigned LengthOfDataRecord = 12;
3021   emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord);
3022   endSymbolRecord(DataEnd);
3023 }
3024