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