1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains support for writing Microsoft CodeView debug info.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "CodeViewDebug.h"
15 #include "DwarfExpression.h"
16 #include "llvm/ADT/APSInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/DenseSet.h"
20 #include "llvm/ADT/MapVector.h"
21 #include "llvm/ADT/None.h"
22 #include "llvm/ADT/Optional.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/ADT/SmallString.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/ADT/TinyPtrVector.h"
28 #include "llvm/ADT/Triple.h"
29 #include "llvm/ADT/Twine.h"
30 #include "llvm/BinaryFormat/COFF.h"
31 #include "llvm/BinaryFormat/Dwarf.h"
32 #include "llvm/CodeGen/AsmPrinter.h"
33 #include "llvm/CodeGen/LexicalScopes.h"
34 #include "llvm/CodeGen/MachineFrameInfo.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineInstr.h"
37 #include "llvm/CodeGen/MachineModuleInfo.h"
38 #include "llvm/CodeGen/MachineOperand.h"
39 #include "llvm/CodeGen/TargetFrameLowering.h"
40 #include "llvm/CodeGen/TargetRegisterInfo.h"
41 #include "llvm/CodeGen/TargetSubtargetInfo.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h"
44 #include "llvm/DebugInfo/CodeView/CodeView.h"
45 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h"
46 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h"
47 #include "llvm/DebugInfo/CodeView/EnumTables.h"
48 #include "llvm/DebugInfo/CodeView/Line.h"
49 #include "llvm/DebugInfo/CodeView/SymbolRecord.h"
50 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h"
51 #include "llvm/DebugInfo/CodeView/TypeIndex.h"
52 #include "llvm/DebugInfo/CodeView/TypeRecord.h"
53 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h"
54 #include "llvm/IR/Constants.h"
55 #include "llvm/IR/DataLayout.h"
56 #include "llvm/IR/DebugInfoMetadata.h"
57 #include "llvm/IR/DebugLoc.h"
58 #include "llvm/IR/Function.h"
59 #include "llvm/IR/GlobalValue.h"
60 #include "llvm/IR/GlobalVariable.h"
61 #include "llvm/IR/Metadata.h"
62 #include "llvm/IR/Module.h"
63 #include "llvm/MC/MCAsmInfo.h"
64 #include "llvm/MC/MCContext.h"
65 #include "llvm/MC/MCSectionCOFF.h"
66 #include "llvm/MC/MCStreamer.h"
67 #include "llvm/MC/MCSymbol.h"
68 #include "llvm/Support/BinaryByteStream.h"
69 #include "llvm/Support/BinaryStreamReader.h"
70 #include "llvm/Support/Casting.h"
71 #include "llvm/Support/CommandLine.h"
72 #include "llvm/Support/Compiler.h"
73 #include "llvm/Support/Endian.h"
74 #include "llvm/Support/Error.h"
75 #include "llvm/Support/ErrorHandling.h"
76 #include "llvm/Support/FormatVariadic.h"
77 #include "llvm/Support/Path.h"
78 #include "llvm/Support/SMLoc.h"
79 #include "llvm/Support/ScopedPrinter.h"
80 #include "llvm/Target/TargetLoweringObjectFile.h"
81 #include "llvm/Target/TargetMachine.h"
82 #include <algorithm>
83 #include <cassert>
84 #include <cctype>
85 #include <cstddef>
86 #include <cstdint>
87 #include <iterator>
88 #include <limits>
89 #include <string>
90 #include <utility>
91 #include <vector>
92 
93 using namespace llvm;
94 using namespace llvm::codeview;
95 
96 static CPUType mapArchToCVCPUType(Triple::ArchType Type) {
97   switch (Type) {
98   case Triple::ArchType::x86:
99     return CPUType::Pentium3;
100   case Triple::ArchType::x86_64:
101     return CPUType::X64;
102   case Triple::ArchType::thumb:
103     return CPUType::Thumb;
104   case Triple::ArchType::aarch64:
105     return CPUType::ARM64;
106   default:
107     report_fatal_error("target architecture doesn't map to a CodeView CPUType");
108   }
109 }
110 
111 CodeViewDebug::CodeViewDebug(AsmPrinter *AP)
112     : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {
113   // If module doesn't have named metadata anchors or COFF debug section
114   // is not available, skip any debug info related stuff.
115   if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") ||
116       !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) {
117     Asm = nullptr;
118     MMI->setDebugInfoAvailability(false);
119     return;
120   }
121   // Tell MMI that we have debug info.
122   MMI->setDebugInfoAvailability(true);
123 
124   TheCPU =
125       mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch());
126 
127   // Check if we should emit type record hashes.
128   ConstantInt *GH = mdconst::extract_or_null<ConstantInt>(
129       MMI->getModule()->getModuleFlag("CodeViewGHash"));
130   EmitDebugGlobalHashes = GH && !GH->isZero();
131 }
132 
133 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) {
134   std::string &Filepath = FileToFilepathMap[File];
135   if (!Filepath.empty())
136     return Filepath;
137 
138   StringRef Dir = File->getDirectory(), Filename = File->getFilename();
139 
140   // If this is a Unix-style path, just use it as is. Don't try to canonicalize
141   // it textually because one of the path components could be a symlink.
142   if (Dir.startswith("/") || Filename.startswith("/")) {
143     if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix))
144       return Filename;
145     Filepath = Dir;
146     if (Dir.back() != '/')
147       Filepath += '/';
148     Filepath += Filename;
149     return Filepath;
150   }
151 
152   // Clang emits directory and relative filename info into the IR, but CodeView
153   // operates on full paths.  We could change Clang to emit full paths too, but
154   // that would increase the IR size and probably not needed for other users.
155   // For now, just concatenate and canonicalize the path here.
156   if (Filename.find(':') == 1)
157     Filepath = Filename;
158   else
159     Filepath = (Dir + "\\" + Filename).str();
160 
161   // Canonicalize the path.  We have to do it textually because we may no longer
162   // have access the file in the filesystem.
163   // First, replace all slashes with backslashes.
164   std::replace(Filepath.begin(), Filepath.end(), '/', '\\');
165 
166   // Remove all "\.\" with "\".
167   size_t Cursor = 0;
168   while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos)
169     Filepath.erase(Cursor, 2);
170 
171   // Replace all "\XXX\..\" with "\".  Don't try too hard though as the original
172   // path should be well-formatted, e.g. start with a drive letter, etc.
173   Cursor = 0;
174   while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) {
175     // Something's wrong if the path starts with "\..\", abort.
176     if (Cursor == 0)
177       break;
178 
179     size_t PrevSlash = Filepath.rfind('\\', Cursor - 1);
180     if (PrevSlash == std::string::npos)
181       // Something's wrong, abort.
182       break;
183 
184     Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash);
185     // The next ".." might be following the one we've just erased.
186     Cursor = PrevSlash;
187   }
188 
189   // Remove all duplicate backslashes.
190   Cursor = 0;
191   while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos)
192     Filepath.erase(Cursor, 1);
193 
194   return Filepath;
195 }
196 
197 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) {
198   StringRef FullPath = getFullFilepath(F);
199   unsigned NextId = FileIdMap.size() + 1;
200   auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId));
201   if (Insertion.second) {
202     // We have to compute the full filepath and emit a .cv_file directive.
203     ArrayRef<uint8_t> ChecksumAsBytes;
204     FileChecksumKind CSKind = FileChecksumKind::None;
205     if (F->getChecksum()) {
206       std::string Checksum = fromHex(F->getChecksum()->Value);
207       void *CKMem = OS.getContext().allocate(Checksum.size(), 1);
208       memcpy(CKMem, Checksum.data(), Checksum.size());
209       ChecksumAsBytes = ArrayRef<uint8_t>(
210           reinterpret_cast<const uint8_t *>(CKMem), Checksum.size());
211       switch (F->getChecksum()->Kind) {
212       case DIFile::CSK_MD5:  CSKind = FileChecksumKind::MD5; break;
213       case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break;
214       }
215     }
216     bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes,
217                                           static_cast<unsigned>(CSKind));
218     (void)Success;
219     assert(Success && ".cv_file directive failed");
220   }
221   return Insertion.first->second;
222 }
223 
224 CodeViewDebug::InlineSite &
225 CodeViewDebug::getInlineSite(const DILocation *InlinedAt,
226                              const DISubprogram *Inlinee) {
227   auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()});
228   InlineSite *Site = &SiteInsertion.first->second;
229   if (SiteInsertion.second) {
230     unsigned ParentFuncId = CurFn->FuncId;
231     if (const DILocation *OuterIA = InlinedAt->getInlinedAt())
232       ParentFuncId =
233           getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram())
234               .SiteFuncId;
235 
236     Site->SiteFuncId = NextFuncId++;
237     OS.EmitCVInlineSiteIdDirective(
238         Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()),
239         InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc());
240     Site->Inlinee = Inlinee;
241     InlinedSubprograms.insert(Inlinee);
242     getFuncIdForSubprogram(Inlinee);
243   }
244   return *Site;
245 }
246 
247 static StringRef getPrettyScopeName(const DIScope *Scope) {
248   StringRef ScopeName = Scope->getName();
249   if (!ScopeName.empty())
250     return ScopeName;
251 
252   switch (Scope->getTag()) {
253   case dwarf::DW_TAG_enumeration_type:
254   case dwarf::DW_TAG_class_type:
255   case dwarf::DW_TAG_structure_type:
256   case dwarf::DW_TAG_union_type:
257     return "<unnamed-tag>";
258   case dwarf::DW_TAG_namespace:
259     return "`anonymous namespace'";
260   }
261 
262   return StringRef();
263 }
264 
265 static const DISubprogram *getQualifiedNameComponents(
266     const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) {
267   const DISubprogram *ClosestSubprogram = nullptr;
268   while (Scope != nullptr) {
269     if (ClosestSubprogram == nullptr)
270       ClosestSubprogram = dyn_cast<DISubprogram>(Scope);
271     StringRef ScopeName = getPrettyScopeName(Scope);
272     if (!ScopeName.empty())
273       QualifiedNameComponents.push_back(ScopeName);
274     Scope = Scope->getScope().resolve();
275   }
276   return ClosestSubprogram;
277 }
278 
279 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents,
280                                     StringRef TypeName) {
281   std::string FullyQualifiedName;
282   for (StringRef QualifiedNameComponent :
283        llvm::reverse(QualifiedNameComponents)) {
284     FullyQualifiedName.append(QualifiedNameComponent);
285     FullyQualifiedName.append("::");
286   }
287   FullyQualifiedName.append(TypeName);
288   return FullyQualifiedName;
289 }
290 
291 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) {
292   SmallVector<StringRef, 5> QualifiedNameComponents;
293   getQualifiedNameComponents(Scope, QualifiedNameComponents);
294   return getQualifiedName(QualifiedNameComponents, Name);
295 }
296 
297 struct CodeViewDebug::TypeLoweringScope {
298   TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; }
299   ~TypeLoweringScope() {
300     // Don't decrement TypeEmissionLevel until after emitting deferred types, so
301     // inner TypeLoweringScopes don't attempt to emit deferred types.
302     if (CVD.TypeEmissionLevel == 1)
303       CVD.emitDeferredCompleteTypes();
304     --CVD.TypeEmissionLevel;
305   }
306   CodeViewDebug &CVD;
307 };
308 
309 static std::string getFullyQualifiedName(const DIScope *Ty) {
310   const DIScope *Scope = Ty->getScope().resolve();
311   return getFullyQualifiedName(Scope, getPrettyScopeName(Ty));
312 }
313 
314 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) {
315   // No scope means global scope and that uses the zero index.
316   if (!Scope || isa<DIFile>(Scope))
317     return TypeIndex();
318 
319   assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type");
320 
321   // Check if we've already translated this scope.
322   auto I = TypeIndices.find({Scope, nullptr});
323   if (I != TypeIndices.end())
324     return I->second;
325 
326   // Build the fully qualified name of the scope.
327   std::string ScopeName = getFullyQualifiedName(Scope);
328   StringIdRecord SID(TypeIndex(), ScopeName);
329   auto TI = TypeTable.writeLeafType(SID);
330   return recordTypeIndexForDINode(Scope, TI);
331 }
332 
333 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) {
334   assert(SP);
335 
336   // Check if we've already translated this subprogram.
337   auto I = TypeIndices.find({SP, nullptr});
338   if (I != TypeIndices.end())
339     return I->second;
340 
341   // The display name includes function template arguments. Drop them to match
342   // MSVC.
343   StringRef DisplayName = SP->getName().split('<').first;
344 
345   const DIScope *Scope = SP->getScope().resolve();
346   TypeIndex TI;
347   if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) {
348     // If the scope is a DICompositeType, then this must be a method. Member
349     // function types take some special handling, and require access to the
350     // subprogram.
351     TypeIndex ClassType = getTypeIndex(Class);
352     MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class),
353                                DisplayName);
354     TI = TypeTable.writeLeafType(MFuncId);
355   } else {
356     // Otherwise, this must be a free function.
357     TypeIndex ParentScope = getScopeIndex(Scope);
358     FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName);
359     TI = TypeTable.writeLeafType(FuncId);
360   }
361 
362   return recordTypeIndexForDINode(SP, TI);
363 }
364 
365 static bool isTrivial(const DICompositeType *DCTy) {
366   return ((DCTy->getFlags() & DINode::FlagTrivial) == DINode::FlagTrivial);
367 }
368 
369 static FunctionOptions
370 getFunctionOptions(const DISubroutineType *Ty,
371                    const DICompositeType *ClassTy = nullptr,
372                    StringRef SPName = StringRef("")) {
373   FunctionOptions FO = FunctionOptions::None;
374   const DIType *ReturnTy = nullptr;
375   if (auto TypeArray = Ty->getTypeArray()) {
376     if (TypeArray.size())
377       ReturnTy = TypeArray[0].resolve();
378   }
379 
380   if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) {
381     if (!isTrivial(ReturnDCTy))
382       FO |= FunctionOptions::CxxReturnUdt;
383   }
384 
385   // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison.
386   if (ClassTy && !isTrivial(ClassTy) && SPName == ClassTy->getName()) {
387     FO |= FunctionOptions::Constructor;
388 
389   // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag.
390 
391   }
392   return FO;
393 }
394 
395 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP,
396                                                const DICompositeType *Class) {
397   // Always use the method declaration as the key for the function type. The
398   // method declaration contains the this adjustment.
399   if (SP->getDeclaration())
400     SP = SP->getDeclaration();
401   assert(!SP->getDeclaration() && "should use declaration as key");
402 
403   // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide
404   // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}.
405   auto I = TypeIndices.find({SP, Class});
406   if (I != TypeIndices.end())
407     return I->second;
408 
409   // Make sure complete type info for the class is emitted *after* the member
410   // function type, as the complete class type is likely to reference this
411   // member function type.
412   TypeLoweringScope S(*this);
413   const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0;
414 
415   FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName());
416   TypeIndex TI = lowerTypeMemberFunction(
417       SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO);
418   return recordTypeIndexForDINode(SP, TI, Class);
419 }
420 
421 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node,
422                                                   TypeIndex TI,
423                                                   const DIType *ClassTy) {
424   auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI});
425   (void)InsertResult;
426   assert(InsertResult.second && "DINode was already assigned a type index");
427   return TI;
428 }
429 
430 unsigned CodeViewDebug::getPointerSizeInBytes() {
431   return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8;
432 }
433 
434 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var,
435                                         const LexicalScope *LS) {
436   if (const DILocation *InlinedAt = LS->getInlinedAt()) {
437     // This variable was inlined. Associate it with the InlineSite.
438     const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram();
439     InlineSite &Site = getInlineSite(InlinedAt, Inlinee);
440     Site.InlinedLocals.emplace_back(Var);
441   } else {
442     // This variable goes into the corresponding lexical scope.
443     ScopeVariables[LS].emplace_back(Var);
444   }
445 }
446 
447 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs,
448                                const DILocation *Loc) {
449   auto B = Locs.begin(), E = Locs.end();
450   if (std::find(B, E, Loc) == E)
451     Locs.push_back(Loc);
452 }
453 
454 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL,
455                                         const MachineFunction *MF) {
456   // Skip this instruction if it has the same location as the previous one.
457   if (!DL || DL == PrevInstLoc)
458     return;
459 
460   const DIScope *Scope = DL.get()->getScope();
461   if (!Scope)
462     return;
463 
464   // Skip this line if it is longer than the maximum we can record.
465   LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true);
466   if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() ||
467       LI.isNeverStepInto())
468     return;
469 
470   ColumnInfo CI(DL.getCol(), /*EndColumn=*/0);
471   if (CI.getStartColumn() != DL.getCol())
472     return;
473 
474   if (!CurFn->HaveLineInfo)
475     CurFn->HaveLineInfo = true;
476   unsigned FileId = 0;
477   if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile())
478     FileId = CurFn->LastFileId;
479   else
480     FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile());
481   PrevInstLoc = DL;
482 
483   unsigned FuncId = CurFn->FuncId;
484   if (const DILocation *SiteLoc = DL->getInlinedAt()) {
485     const DILocation *Loc = DL.get();
486 
487     // If this location was actually inlined from somewhere else, give it the ID
488     // of the inline call site.
489     FuncId =
490         getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId;
491 
492     // Ensure we have links in the tree of inline call sites.
493     bool FirstLoc = true;
494     while ((SiteLoc = Loc->getInlinedAt())) {
495       InlineSite &Site =
496           getInlineSite(SiteLoc, Loc->getScope()->getSubprogram());
497       if (!FirstLoc)
498         addLocIfNotPresent(Site.ChildSites, Loc);
499       FirstLoc = false;
500       Loc = SiteLoc;
501     }
502     addLocIfNotPresent(CurFn->ChildSites, Loc);
503   }
504 
505   OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(),
506                         /*PrologueEnd=*/false, /*IsStmt=*/false,
507                         DL->getFilename(), SMLoc());
508 }
509 
510 void CodeViewDebug::emitCodeViewMagicVersion() {
511   OS.EmitValueToAlignment(4);
512   OS.AddComment("Debug section magic");
513   OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4);
514 }
515 
516 void CodeViewDebug::endModule() {
517   if (!Asm || !MMI->hasDebugInfo())
518     return;
519 
520   assert(Asm != nullptr);
521 
522   // The COFF .debug$S section consists of several subsections, each starting
523   // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length
524   // of the payload followed by the payload itself.  The subsections are 4-byte
525   // aligned.
526 
527   // Use the generic .debug$S section, and make a subsection for all the inlined
528   // subprograms.
529   switchToDebugSectionForSymbol(nullptr);
530 
531   MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols);
532   emitCompilerInformation();
533   endCVSubsection(CompilerInfo);
534 
535   emitInlineeLinesSubsection();
536 
537   // Emit per-function debug information.
538   for (auto &P : FnDebugInfo)
539     if (!P.first->isDeclarationForLinker())
540       emitDebugInfoForFunction(P.first, *P.second);
541 
542   // Emit global variable debug information.
543   setCurrentSubprogram(nullptr);
544   emitDebugInfoForGlobals();
545 
546   // Emit retained types.
547   emitDebugInfoForRetainedTypes();
548 
549   // Switch back to the generic .debug$S section after potentially processing
550   // comdat symbol sections.
551   switchToDebugSectionForSymbol(nullptr);
552 
553   // Emit UDT records for any types used by global variables.
554   if (!GlobalUDTs.empty()) {
555     MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
556     emitDebugInfoForUDTs(GlobalUDTs);
557     endCVSubsection(SymbolsEnd);
558   }
559 
560   // This subsection holds a file index to offset in string table table.
561   OS.AddComment("File index to string table offset subsection");
562   OS.EmitCVFileChecksumsDirective();
563 
564   // This subsection holds the string table.
565   OS.AddComment("String table");
566   OS.EmitCVStringTableDirective();
567 
568   // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol
569   // subsection in the generic .debug$S section at the end. There is no
570   // particular reason for this ordering other than to match MSVC.
571   emitBuildInfo();
572 
573   // Emit type information and hashes last, so that any types we translate while
574   // emitting function info are included.
575   emitTypeInformation();
576 
577   if (EmitDebugGlobalHashes)
578     emitTypeGlobalHashes();
579 
580   clear();
581 }
582 
583 static void emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S,
584     unsigned MaxFixedRecordLength = 0xF00) {
585   // The maximum CV record length is 0xFF00. Most of the strings we emit appear
586   // after a fixed length portion of the record. The fixed length portion should
587   // always be less than 0xF00 (3840) bytes, so truncate the string so that the
588   // overall record size is less than the maximum allowed.
589   SmallString<32> NullTerminatedString(
590       S.take_front(MaxRecordLength - MaxFixedRecordLength - 1));
591   NullTerminatedString.push_back('\0');
592   OS.EmitBytes(NullTerminatedString);
593 }
594 
595 void CodeViewDebug::emitTypeInformation() {
596   if (TypeTable.empty())
597     return;
598 
599   // Start the .debug$T or .debug$P section with 0x4.
600   OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection());
601   emitCodeViewMagicVersion();
602 
603   SmallString<8> CommentPrefix;
604   if (OS.isVerboseAsm()) {
605     CommentPrefix += '\t';
606     CommentPrefix += Asm->MAI->getCommentString();
607     CommentPrefix += ' ';
608   }
609 
610   TypeTableCollection Table(TypeTable.records());
611   Optional<TypeIndex> B = Table.getFirst();
612   while (B) {
613     // This will fail if the record data is invalid.
614     CVType Record = Table.getType(*B);
615 
616     if (OS.isVerboseAsm()) {
617       // Emit a block comment describing the type record for readability.
618       SmallString<512> CommentBlock;
619       raw_svector_ostream CommentOS(CommentBlock);
620       ScopedPrinter SP(CommentOS);
621       SP.setPrefix(CommentPrefix);
622       TypeDumpVisitor TDV(Table, &SP, false);
623 
624       Error E = codeview::visitTypeRecord(Record, *B, TDV);
625       if (E) {
626         logAllUnhandledErrors(std::move(E), errs(), "error: ");
627         llvm_unreachable("produced malformed type record");
628       }
629       // emitRawComment will insert its own tab and comment string before
630       // the first line, so strip off our first one. It also prints its own
631       // newline.
632       OS.emitRawComment(
633           CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim());
634     }
635     OS.EmitBinaryData(Record.str_data());
636     B = Table.getNext(*B);
637   }
638 }
639 
640 void CodeViewDebug::emitTypeGlobalHashes() {
641   if (TypeTable.empty())
642     return;
643 
644   // Start the .debug$H section with the version and hash algorithm, currently
645   // hardcoded to version 0, SHA1.
646   OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection());
647 
648   OS.EmitValueToAlignment(4);
649   OS.AddComment("Magic");
650   OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4);
651   OS.AddComment("Section Version");
652   OS.EmitIntValue(0, 2);
653   OS.AddComment("Hash Algorithm");
654   OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2);
655 
656   TypeIndex TI(TypeIndex::FirstNonSimpleIndex);
657   for (const auto &GHR : TypeTable.hashes()) {
658     if (OS.isVerboseAsm()) {
659       // Emit an EOL-comment describing which TypeIndex this hash corresponds
660       // to, as well as the stringified SHA1 hash.
661       SmallString<32> Comment;
662       raw_svector_ostream CommentOS(Comment);
663       CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR);
664       OS.AddComment(Comment);
665       ++TI;
666     }
667     assert(GHR.Hash.size() == 8);
668     StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()),
669                 GHR.Hash.size());
670     OS.EmitBinaryData(S);
671   }
672 }
673 
674 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) {
675   switch (DWLang) {
676   case dwarf::DW_LANG_C:
677   case dwarf::DW_LANG_C89:
678   case dwarf::DW_LANG_C99:
679   case dwarf::DW_LANG_C11:
680   case dwarf::DW_LANG_ObjC:
681     return SourceLanguage::C;
682   case dwarf::DW_LANG_C_plus_plus:
683   case dwarf::DW_LANG_C_plus_plus_03:
684   case dwarf::DW_LANG_C_plus_plus_11:
685   case dwarf::DW_LANG_C_plus_plus_14:
686     return SourceLanguage::Cpp;
687   case dwarf::DW_LANG_Fortran77:
688   case dwarf::DW_LANG_Fortran90:
689   case dwarf::DW_LANG_Fortran03:
690   case dwarf::DW_LANG_Fortran08:
691     return SourceLanguage::Fortran;
692   case dwarf::DW_LANG_Pascal83:
693     return SourceLanguage::Pascal;
694   case dwarf::DW_LANG_Cobol74:
695   case dwarf::DW_LANG_Cobol85:
696     return SourceLanguage::Cobol;
697   case dwarf::DW_LANG_Java:
698     return SourceLanguage::Java;
699   case dwarf::DW_LANG_D:
700     return SourceLanguage::D;
701   default:
702     // There's no CodeView representation for this language, and CV doesn't
703     // have an "unknown" option for the language field, so we'll use MASM,
704     // as it's very low level.
705     return SourceLanguage::Masm;
706   }
707 }
708 
709 namespace {
710 struct Version {
711   int Part[4];
712 };
713 } // end anonymous namespace
714 
715 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out
716 // the version number.
717 static Version parseVersion(StringRef Name) {
718   Version V = {{0}};
719   int N = 0;
720   for (const char C : Name) {
721     if (isdigit(C)) {
722       V.Part[N] *= 10;
723       V.Part[N] += C - '0';
724     } else if (C == '.') {
725       ++N;
726       if (N >= 4)
727         return V;
728     } else if (N > 0)
729       return V;
730   }
731   return V;
732 }
733 
734 void CodeViewDebug::emitCompilerInformation() {
735   MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3);
736   uint32_t Flags = 0;
737 
738   NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
739   const MDNode *Node = *CUs->operands().begin();
740   const auto *CU = cast<DICompileUnit>(Node);
741 
742   // The low byte of the flags indicates the source language.
743   Flags = MapDWLangToCVLang(CU->getSourceLanguage());
744   // TODO:  Figure out which other flags need to be set.
745 
746   OS.AddComment("Flags and language");
747   OS.EmitIntValue(Flags, 4);
748 
749   OS.AddComment("CPUType");
750   OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2);
751 
752   StringRef CompilerVersion = CU->getProducer();
753   Version FrontVer = parseVersion(CompilerVersion);
754   OS.AddComment("Frontend version");
755   for (int N = 0; N < 4; ++N)
756     OS.EmitIntValue(FrontVer.Part[N], 2);
757 
758   // Some Microsoft tools, like Binscope, expect a backend version number of at
759   // least 8.something, so we'll coerce the LLVM version into a form that
760   // guarantees it'll be big enough without really lying about the version.
761   int Major = 1000 * LLVM_VERSION_MAJOR +
762               10 * LLVM_VERSION_MINOR +
763               LLVM_VERSION_PATCH;
764   // Clamp it for builds that use unusually large version numbers.
765   Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max());
766   Version BackVer = {{ Major, 0, 0, 0 }};
767   OS.AddComment("Backend version");
768   for (int N = 0; N < 4; ++N)
769     OS.EmitIntValue(BackVer.Part[N], 2);
770 
771   OS.AddComment("Null-terminated compiler version string");
772   emitNullTerminatedSymbolName(OS, CompilerVersion);
773 
774   endSymbolRecord(CompilerEnd);
775 }
776 
777 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable,
778                                     StringRef S) {
779   StringIdRecord SIR(TypeIndex(0x0), S);
780   return TypeTable.writeLeafType(SIR);
781 }
782 
783 void CodeViewDebug::emitBuildInfo() {
784   // First, make LF_BUILDINFO. It's a sequence of strings with various bits of
785   // build info. The known prefix is:
786   // - Absolute path of current directory
787   // - Compiler path
788   // - Main source file path, relative to CWD or absolute
789   // - Type server PDB file
790   // - Canonical compiler command line
791   // If frontend and backend compilation are separated (think llc or LTO), it's
792   // not clear if the compiler path should refer to the executable for the
793   // frontend or the backend. Leave it blank for now.
794   TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {};
795   NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
796   const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs.
797   const auto *CU = cast<DICompileUnit>(Node);
798   const DIFile *MainSourceFile = CU->getFile();
799   BuildInfoArgs[BuildInfoRecord::CurrentDirectory] =
800       getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory());
801   BuildInfoArgs[BuildInfoRecord::SourceFile] =
802       getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename());
803   // FIXME: Path to compiler and command line. PDB is intentionally blank unless
804   // we implement /Zi type servers.
805   BuildInfoRecord BIR(BuildInfoArgs);
806   TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR);
807 
808   // Make a new .debug$S subsection for the S_BUILDINFO record, which points
809   // from the module symbols into the type stream.
810   MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
811   MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO);
812   OS.AddComment("LF_BUILDINFO index");
813   OS.EmitIntValue(BuildInfoIndex.getIndex(), 4);
814   endSymbolRecord(BIEnd);
815   endCVSubsection(BISubsecEnd);
816 }
817 
818 void CodeViewDebug::emitInlineeLinesSubsection() {
819   if (InlinedSubprograms.empty())
820     return;
821 
822   OS.AddComment("Inlinee lines subsection");
823   MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines);
824 
825   // We emit the checksum info for files.  This is used by debuggers to
826   // determine if a pdb matches the source before loading it.  Visual Studio,
827   // for instance, will display a warning that the breakpoints are not valid if
828   // the pdb does not match the source.
829   OS.AddComment("Inlinee lines signature");
830   OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4);
831 
832   for (const DISubprogram *SP : InlinedSubprograms) {
833     assert(TypeIndices.count({SP, nullptr}));
834     TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}];
835 
836     OS.AddBlankLine();
837     unsigned FileId = maybeRecordFile(SP->getFile());
838     OS.AddComment("Inlined function " + SP->getName() + " starts at " +
839                   SP->getFilename() + Twine(':') + Twine(SP->getLine()));
840     OS.AddBlankLine();
841     OS.AddComment("Type index of inlined function");
842     OS.EmitIntValue(InlineeIdx.getIndex(), 4);
843     OS.AddComment("Offset into filechecksum table");
844     OS.EmitCVFileChecksumOffsetDirective(FileId);
845     OS.AddComment("Starting line number");
846     OS.EmitIntValue(SP->getLine(), 4);
847   }
848 
849   endCVSubsection(InlineEnd);
850 }
851 
852 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI,
853                                         const DILocation *InlinedAt,
854                                         const InlineSite &Site) {
855   assert(TypeIndices.count({Site.Inlinee, nullptr}));
856   TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}];
857 
858   // SymbolRecord
859   MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE);
860 
861   OS.AddComment("PtrParent");
862   OS.EmitIntValue(0, 4);
863   OS.AddComment("PtrEnd");
864   OS.EmitIntValue(0, 4);
865   OS.AddComment("Inlinee type index");
866   OS.EmitIntValue(InlineeIdx.getIndex(), 4);
867 
868   unsigned FileId = maybeRecordFile(Site.Inlinee->getFile());
869   unsigned StartLineNum = Site.Inlinee->getLine();
870 
871   OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum,
872                                     FI.Begin, FI.End);
873 
874   endSymbolRecord(InlineEnd);
875 
876   emitLocalVariableList(FI, Site.InlinedLocals);
877 
878   // Recurse on child inlined call sites before closing the scope.
879   for (const DILocation *ChildSite : Site.ChildSites) {
880     auto I = FI.InlineSites.find(ChildSite);
881     assert(I != FI.InlineSites.end() &&
882            "child site not in function inline site map");
883     emitInlinedCallSite(FI, ChildSite, I->second);
884   }
885 
886   // Close the scope.
887   emitEndSymbolRecord(SymbolKind::S_INLINESITE_END);
888 }
889 
890 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) {
891   // If we have a symbol, it may be in a section that is COMDAT. If so, find the
892   // comdat key. A section may be comdat because of -ffunction-sections or
893   // because it is comdat in the IR.
894   MCSectionCOFF *GVSec =
895       GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr;
896   const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr;
897 
898   MCSectionCOFF *DebugSec = cast<MCSectionCOFF>(
899       Asm->getObjFileLowering().getCOFFDebugSymbolsSection());
900   DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym);
901 
902   OS.SwitchSection(DebugSec);
903 
904   // Emit the magic version number if this is the first time we've switched to
905   // this section.
906   if (ComdatDebugSections.insert(DebugSec).second)
907     emitCodeViewMagicVersion();
908 }
909 
910 // Emit an S_THUNK32/S_END symbol pair for a thunk routine.
911 // The only supported thunk ordinal is currently the standard type.
912 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV,
913                                           FunctionInfo &FI,
914                                           const MCSymbol *Fn) {
915   std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName());
916   const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind.
917 
918   OS.AddComment("Symbol subsection for " + Twine(FuncName));
919   MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
920 
921   // Emit S_THUNK32
922   MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32);
923   OS.AddComment("PtrParent");
924   OS.EmitIntValue(0, 4);
925   OS.AddComment("PtrEnd");
926   OS.EmitIntValue(0, 4);
927   OS.AddComment("PtrNext");
928   OS.EmitIntValue(0, 4);
929   OS.AddComment("Thunk section relative address");
930   OS.EmitCOFFSecRel32(Fn, /*Offset=*/0);
931   OS.AddComment("Thunk section index");
932   OS.EmitCOFFSectionIndex(Fn);
933   OS.AddComment("Code size");
934   OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2);
935   OS.AddComment("Ordinal");
936   OS.EmitIntValue(unsigned(ordinal), 1);
937   OS.AddComment("Function name");
938   emitNullTerminatedSymbolName(OS, FuncName);
939   // Additional fields specific to the thunk ordinal would go here.
940   endSymbolRecord(ThunkRecordEnd);
941 
942   // Local variables/inlined routines are purposely omitted here.  The point of
943   // marking this as a thunk is so Visual Studio will NOT stop in this routine.
944 
945   // Emit S_PROC_ID_END
946   emitEndSymbolRecord(SymbolKind::S_PROC_ID_END);
947 
948   endCVSubsection(SymbolsEnd);
949 }
950 
951 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV,
952                                              FunctionInfo &FI) {
953   // For each function there is a separate subsection which holds the PC to
954   // file:line table.
955   const MCSymbol *Fn = Asm->getSymbol(GV);
956   assert(Fn);
957 
958   // Switch to the to a comdat section, if appropriate.
959   switchToDebugSectionForSymbol(Fn);
960 
961   std::string FuncName;
962   auto *SP = GV->getSubprogram();
963   assert(SP);
964   setCurrentSubprogram(SP);
965 
966   if (SP->isThunk()) {
967     emitDebugInfoForThunk(GV, FI, Fn);
968     return;
969   }
970 
971   // If we have a display name, build the fully qualified name by walking the
972   // chain of scopes.
973   if (!SP->getName().empty())
974     FuncName =
975         getFullyQualifiedName(SP->getScope().resolve(), SP->getName());
976 
977   // If our DISubprogram name is empty, use the mangled name.
978   if (FuncName.empty())
979     FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName());
980 
981   // Emit FPO data, but only on 32-bit x86. No other platforms use it.
982   if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86)
983     OS.EmitCVFPOData(Fn);
984 
985   // Emit a symbol subsection, required by VS2012+ to find function boundaries.
986   OS.AddComment("Symbol subsection for " + Twine(FuncName));
987   MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
988   {
989     SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID
990                                                 : SymbolKind::S_GPROC32_ID;
991     MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind);
992 
993     // These fields are filled in by tools like CVPACK which run after the fact.
994     OS.AddComment("PtrParent");
995     OS.EmitIntValue(0, 4);
996     OS.AddComment("PtrEnd");
997     OS.EmitIntValue(0, 4);
998     OS.AddComment("PtrNext");
999     OS.EmitIntValue(0, 4);
1000     // This is the important bit that tells the debugger where the function
1001     // code is located and what's its size:
1002     OS.AddComment("Code size");
1003     OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4);
1004     OS.AddComment("Offset after prologue");
1005     OS.EmitIntValue(0, 4);
1006     OS.AddComment("Offset before epilogue");
1007     OS.EmitIntValue(0, 4);
1008     OS.AddComment("Function type index");
1009     OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4);
1010     OS.AddComment("Function section relative address");
1011     OS.EmitCOFFSecRel32(Fn, /*Offset=*/0);
1012     OS.AddComment("Function section index");
1013     OS.EmitCOFFSectionIndex(Fn);
1014     OS.AddComment("Flags");
1015     OS.EmitIntValue(0, 1);
1016     // Emit the function display name as a null-terminated string.
1017     OS.AddComment("Function name");
1018     // Truncate the name so we won't overflow the record length field.
1019     emitNullTerminatedSymbolName(OS, FuncName);
1020     endSymbolRecord(ProcRecordEnd);
1021 
1022     MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC);
1023     // Subtract out the CSR size since MSVC excludes that and we include it.
1024     OS.AddComment("FrameSize");
1025     OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4);
1026     OS.AddComment("Padding");
1027     OS.EmitIntValue(0, 4);
1028     OS.AddComment("Offset of padding");
1029     OS.EmitIntValue(0, 4);
1030     OS.AddComment("Bytes of callee saved registers");
1031     OS.EmitIntValue(FI.CSRSize, 4);
1032     OS.AddComment("Exception handler offset");
1033     OS.EmitIntValue(0, 4);
1034     OS.AddComment("Exception handler section");
1035     OS.EmitIntValue(0, 2);
1036     OS.AddComment("Flags (defines frame register)");
1037     OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4);
1038     endSymbolRecord(FrameProcEnd);
1039 
1040     emitLocalVariableList(FI, FI.Locals);
1041     emitLexicalBlockList(FI.ChildBlocks, FI);
1042 
1043     // Emit inlined call site information. Only emit functions inlined directly
1044     // into the parent function. We'll emit the other sites recursively as part
1045     // of their parent inline site.
1046     for (const DILocation *InlinedAt : FI.ChildSites) {
1047       auto I = FI.InlineSites.find(InlinedAt);
1048       assert(I != FI.InlineSites.end() &&
1049              "child site not in function inline site map");
1050       emitInlinedCallSite(FI, InlinedAt, I->second);
1051     }
1052 
1053     for (auto Annot : FI.Annotations) {
1054       MCSymbol *Label = Annot.first;
1055       MDTuple *Strs = cast<MDTuple>(Annot.second);
1056       MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION);
1057       OS.EmitCOFFSecRel32(Label, /*Offset=*/0);
1058       // FIXME: Make sure we don't overflow the max record size.
1059       OS.EmitCOFFSectionIndex(Label);
1060       OS.EmitIntValue(Strs->getNumOperands(), 2);
1061       for (Metadata *MD : Strs->operands()) {
1062         // MDStrings are null terminated, so we can do EmitBytes and get the
1063         // nice .asciz directive.
1064         StringRef Str = cast<MDString>(MD)->getString();
1065         assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString");
1066         OS.EmitBytes(StringRef(Str.data(), Str.size() + 1));
1067       }
1068       endSymbolRecord(AnnotEnd);
1069     }
1070 
1071     if (SP != nullptr)
1072       emitDebugInfoForUDTs(LocalUDTs);
1073 
1074     // We're done with this function.
1075     emitEndSymbolRecord(SymbolKind::S_PROC_ID_END);
1076   }
1077   endCVSubsection(SymbolsEnd);
1078 
1079   // We have an assembler directive that takes care of the whole line table.
1080   OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End);
1081 }
1082 
1083 CodeViewDebug::LocalVarDefRange
1084 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) {
1085   LocalVarDefRange DR;
1086   DR.InMemory = -1;
1087   DR.DataOffset = Offset;
1088   assert(DR.DataOffset == Offset && "truncation");
1089   DR.IsSubfield = 0;
1090   DR.StructOffset = 0;
1091   DR.CVRegister = CVRegister;
1092   return DR;
1093 }
1094 
1095 void CodeViewDebug::collectVariableInfoFromMFTable(
1096     DenseSet<InlinedEntity> &Processed) {
1097   const MachineFunction &MF = *Asm->MF;
1098   const TargetSubtargetInfo &TSI = MF.getSubtarget();
1099   const TargetFrameLowering *TFI = TSI.getFrameLowering();
1100   const TargetRegisterInfo *TRI = TSI.getRegisterInfo();
1101 
1102   for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) {
1103     if (!VI.Var)
1104       continue;
1105     assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) &&
1106            "Expected inlined-at fields to agree");
1107 
1108     Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt()));
1109     LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc);
1110 
1111     // If variable scope is not found then skip this variable.
1112     if (!Scope)
1113       continue;
1114 
1115     // If the variable has an attached offset expression, extract it.
1116     // FIXME: Try to handle DW_OP_deref as well.
1117     int64_t ExprOffset = 0;
1118     if (VI.Expr)
1119       if (!VI.Expr->extractIfOffset(ExprOffset))
1120         continue;
1121 
1122     // Get the frame register used and the offset.
1123     unsigned FrameReg = 0;
1124     int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg);
1125     uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg);
1126 
1127     // Calculate the label ranges.
1128     LocalVarDefRange DefRange =
1129         createDefRangeMem(CVReg, FrameOffset + ExprOffset);
1130     for (const InsnRange &Range : Scope->getRanges()) {
1131       const MCSymbol *Begin = getLabelBeforeInsn(Range.first);
1132       const MCSymbol *End = getLabelAfterInsn(Range.second);
1133       End = End ? End : Asm->getFunctionEnd();
1134       DefRange.Ranges.emplace_back(Begin, End);
1135     }
1136 
1137     LocalVariable Var;
1138     Var.DIVar = VI.Var;
1139     Var.DefRanges.emplace_back(std::move(DefRange));
1140     recordLocalVariable(std::move(Var), Scope);
1141   }
1142 }
1143 
1144 static bool canUseReferenceType(const DbgVariableLocation &Loc) {
1145   return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0;
1146 }
1147 
1148 static bool needsReferenceType(const DbgVariableLocation &Loc) {
1149   return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0;
1150 }
1151 
1152 void CodeViewDebug::calculateRanges(
1153     LocalVariable &Var, const DbgValueHistoryMap::InstrRanges &Ranges) {
1154   const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo();
1155 
1156   // Calculate the definition ranges.
1157   for (auto I = Ranges.begin(), E = Ranges.end(); I != E; ++I) {
1158     const InsnRange &Range = *I;
1159     const MachineInstr *DVInst = Range.first;
1160     assert(DVInst->isDebugValue() && "Invalid History entry");
1161     // FIXME: Find a way to represent constant variables, since they are
1162     // relatively common.
1163     Optional<DbgVariableLocation> Location =
1164         DbgVariableLocation::extractFromMachineInstruction(*DVInst);
1165     if (!Location)
1166       continue;
1167 
1168     // CodeView can only express variables in register and variables in memory
1169     // at a constant offset from a register. However, for variables passed
1170     // indirectly by pointer, it is common for that pointer to be spilled to a
1171     // stack location. For the special case of one offseted load followed by a
1172     // zero offset load (a pointer spilled to the stack), we change the type of
1173     // the local variable from a value type to a reference type. This tricks the
1174     // debugger into doing the load for us.
1175     if (Var.UseReferenceType) {
1176       // We're using a reference type. Drop the last zero offset load.
1177       if (canUseReferenceType(*Location))
1178         Location->LoadChain.pop_back();
1179       else
1180         continue;
1181     } else if (needsReferenceType(*Location)) {
1182       // This location can't be expressed without switching to a reference type.
1183       // Start over using that.
1184       Var.UseReferenceType = true;
1185       Var.DefRanges.clear();
1186       calculateRanges(Var, Ranges);
1187       return;
1188     }
1189 
1190     // We can only handle a register or an offseted load of a register.
1191     if (Location->Register == 0 || Location->LoadChain.size() > 1)
1192       continue;
1193     {
1194       LocalVarDefRange DR;
1195       DR.CVRegister = TRI->getCodeViewRegNum(Location->Register);
1196       DR.InMemory = !Location->LoadChain.empty();
1197       DR.DataOffset =
1198           !Location->LoadChain.empty() ? Location->LoadChain.back() : 0;
1199       if (Location->FragmentInfo) {
1200         DR.IsSubfield = true;
1201         DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8;
1202       } else {
1203         DR.IsSubfield = false;
1204         DR.StructOffset = 0;
1205       }
1206 
1207       if (Var.DefRanges.empty() ||
1208           Var.DefRanges.back().isDifferentLocation(DR)) {
1209         Var.DefRanges.emplace_back(std::move(DR));
1210       }
1211     }
1212 
1213     // Compute the label range.
1214     const MCSymbol *Begin = getLabelBeforeInsn(Range.first);
1215     const MCSymbol *End = getLabelAfterInsn(Range.second);
1216     if (!End) {
1217       // This range is valid until the next overlapping bitpiece. In the
1218       // common case, ranges will not be bitpieces, so they will overlap.
1219       auto J = std::next(I);
1220       const DIExpression *DIExpr = DVInst->getDebugExpression();
1221       while (J != E &&
1222              !DIExpr->fragmentsOverlap(J->first->getDebugExpression()))
1223         ++J;
1224       if (J != E)
1225         End = getLabelBeforeInsn(J->first);
1226       else
1227         End = Asm->getFunctionEnd();
1228     }
1229 
1230     // If the last range end is our begin, just extend the last range.
1231     // Otherwise make a new range.
1232     SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R =
1233         Var.DefRanges.back().Ranges;
1234     if (!R.empty() && R.back().second == Begin)
1235       R.back().second = End;
1236     else
1237       R.emplace_back(Begin, End);
1238 
1239     // FIXME: Do more range combining.
1240   }
1241 }
1242 
1243 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) {
1244   DenseSet<InlinedEntity> Processed;
1245   // Grab the variable info that was squirreled away in the MMI side-table.
1246   collectVariableInfoFromMFTable(Processed);
1247 
1248   for (const auto &I : DbgValues) {
1249     InlinedEntity IV = I.first;
1250     if (Processed.count(IV))
1251       continue;
1252     const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first);
1253     const DILocation *InlinedAt = IV.second;
1254 
1255     // Instruction ranges, specifying where IV is accessible.
1256     const auto &Ranges = I.second;
1257 
1258     LexicalScope *Scope = nullptr;
1259     if (InlinedAt)
1260       Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt);
1261     else
1262       Scope = LScopes.findLexicalScope(DIVar->getScope());
1263     // If variable scope is not found then skip this variable.
1264     if (!Scope)
1265       continue;
1266 
1267     LocalVariable Var;
1268     Var.DIVar = DIVar;
1269 
1270     calculateRanges(Var, Ranges);
1271     recordLocalVariable(std::move(Var), Scope);
1272   }
1273 }
1274 
1275 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) {
1276   const TargetSubtargetInfo &TSI = MF->getSubtarget();
1277   const TargetRegisterInfo *TRI = TSI.getRegisterInfo();
1278   const MachineFrameInfo &MFI = MF->getFrameInfo();
1279   const Function &GV = MF->getFunction();
1280   auto Insertion = FnDebugInfo.insert({&GV, llvm::make_unique<FunctionInfo>()});
1281   assert(Insertion.second && "function already has info");
1282   CurFn = Insertion.first->second.get();
1283   CurFn->FuncId = NextFuncId++;
1284   CurFn->Begin = Asm->getFunctionBegin();
1285 
1286   // The S_FRAMEPROC record reports the stack size, and how many bytes of
1287   // callee-saved registers were used. For targets that don't use a PUSH
1288   // instruction (AArch64), this will be zero.
1289   CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters();
1290   CurFn->FrameSize = MFI.getStackSize();
1291   CurFn->OffsetAdjustment = MFI.getOffsetAdjustment();
1292   CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF);
1293 
1294   // For this function S_FRAMEPROC record, figure out which codeview register
1295   // will be the frame pointer.
1296   CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None.
1297   CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None.
1298   if (CurFn->FrameSize > 0) {
1299     if (!TSI.getFrameLowering()->hasFP(*MF)) {
1300       CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr;
1301       CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr;
1302     } else {
1303       // If there is an FP, parameters are always relative to it.
1304       CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr;
1305       if (CurFn->HasStackRealignment) {
1306         // If the stack needs realignment, locals are relative to SP or VFRAME.
1307         CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr;
1308       } else {
1309         // Otherwise, locals are relative to EBP, and we probably have VLAs or
1310         // other stack adjustments.
1311         CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr;
1312       }
1313     }
1314   }
1315 
1316   // Compute other frame procedure options.
1317   FrameProcedureOptions FPO = FrameProcedureOptions::None;
1318   if (MFI.hasVarSizedObjects())
1319     FPO |= FrameProcedureOptions::HasAlloca;
1320   if (MF->exposesReturnsTwice())
1321     FPO |= FrameProcedureOptions::HasSetJmp;
1322   // FIXME: Set HasLongJmp if we ever track that info.
1323   if (MF->hasInlineAsm())
1324     FPO |= FrameProcedureOptions::HasInlineAssembly;
1325   if (GV.hasPersonalityFn()) {
1326     if (isAsynchronousEHPersonality(
1327             classifyEHPersonality(GV.getPersonalityFn())))
1328       FPO |= FrameProcedureOptions::HasStructuredExceptionHandling;
1329     else
1330       FPO |= FrameProcedureOptions::HasExceptionHandling;
1331   }
1332   if (GV.hasFnAttribute(Attribute::InlineHint))
1333     FPO |= FrameProcedureOptions::MarkedInline;
1334   if (GV.hasFnAttribute(Attribute::Naked))
1335     FPO |= FrameProcedureOptions::Naked;
1336   if (MFI.hasStackProtectorIndex())
1337     FPO |= FrameProcedureOptions::SecurityChecks;
1338   FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U);
1339   FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U);
1340   if (Asm->TM.getOptLevel() != CodeGenOpt::None && !GV.optForSize() &&
1341       !GV.hasFnAttribute(Attribute::OptimizeNone))
1342     FPO |= FrameProcedureOptions::OptimizedForSpeed;
1343   // FIXME: Set GuardCfg when it is implemented.
1344   CurFn->FrameProcOpts = FPO;
1345 
1346   OS.EmitCVFuncIdDirective(CurFn->FuncId);
1347 
1348   // Find the end of the function prolog.  First known non-DBG_VALUE and
1349   // non-frame setup location marks the beginning of the function body.
1350   // FIXME: is there a simpler a way to do this? Can we just search
1351   // for the first instruction of the function, not the last of the prolog?
1352   DebugLoc PrologEndLoc;
1353   bool EmptyPrologue = true;
1354   for (const auto &MBB : *MF) {
1355     for (const auto &MI : MBB) {
1356       if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) &&
1357           MI.getDebugLoc()) {
1358         PrologEndLoc = MI.getDebugLoc();
1359         break;
1360       } else if (!MI.isMetaInstruction()) {
1361         EmptyPrologue = false;
1362       }
1363     }
1364   }
1365 
1366   // Record beginning of function if we have a non-empty prologue.
1367   if (PrologEndLoc && !EmptyPrologue) {
1368     DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc();
1369     maybeRecordLocation(FnStartDL, MF);
1370   }
1371 }
1372 
1373 static bool shouldEmitUdt(const DIType *T) {
1374   if (!T)
1375     return false;
1376 
1377   // MSVC does not emit UDTs for typedefs that are scoped to classes.
1378   if (T->getTag() == dwarf::DW_TAG_typedef) {
1379     if (DIScope *Scope = T->getScope().resolve()) {
1380       switch (Scope->getTag()) {
1381       case dwarf::DW_TAG_structure_type:
1382       case dwarf::DW_TAG_class_type:
1383       case dwarf::DW_TAG_union_type:
1384         return false;
1385       }
1386     }
1387   }
1388 
1389   while (true) {
1390     if (!T || T->isForwardDecl())
1391       return false;
1392 
1393     const DIDerivedType *DT = dyn_cast<DIDerivedType>(T);
1394     if (!DT)
1395       return true;
1396     T = DT->getBaseType().resolve();
1397   }
1398   return true;
1399 }
1400 
1401 void CodeViewDebug::addToUDTs(const DIType *Ty) {
1402   // Don't record empty UDTs.
1403   if (Ty->getName().empty())
1404     return;
1405   if (!shouldEmitUdt(Ty))
1406     return;
1407 
1408   SmallVector<StringRef, 5> QualifiedNameComponents;
1409   const DISubprogram *ClosestSubprogram = getQualifiedNameComponents(
1410       Ty->getScope().resolve(), QualifiedNameComponents);
1411 
1412   std::string FullyQualifiedName =
1413       getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty));
1414 
1415   if (ClosestSubprogram == nullptr) {
1416     GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty);
1417   } else if (ClosestSubprogram == CurrentSubprogram) {
1418     LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty);
1419   }
1420 
1421   // TODO: What if the ClosestSubprogram is neither null or the current
1422   // subprogram?  Currently, the UDT just gets dropped on the floor.
1423   //
1424   // The current behavior is not desirable.  To get maximal fidelity, we would
1425   // need to perform all type translation before beginning emission of .debug$S
1426   // and then make LocalUDTs a member of FunctionInfo
1427 }
1428 
1429 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) {
1430   // Generic dispatch for lowering an unknown type.
1431   switch (Ty->getTag()) {
1432   case dwarf::DW_TAG_array_type:
1433     return lowerTypeArray(cast<DICompositeType>(Ty));
1434   case dwarf::DW_TAG_typedef:
1435     return lowerTypeAlias(cast<DIDerivedType>(Ty));
1436   case dwarf::DW_TAG_base_type:
1437     return lowerTypeBasic(cast<DIBasicType>(Ty));
1438   case dwarf::DW_TAG_pointer_type:
1439     if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type")
1440       return lowerTypeVFTableShape(cast<DIDerivedType>(Ty));
1441     LLVM_FALLTHROUGH;
1442   case dwarf::DW_TAG_reference_type:
1443   case dwarf::DW_TAG_rvalue_reference_type:
1444     return lowerTypePointer(cast<DIDerivedType>(Ty));
1445   case dwarf::DW_TAG_ptr_to_member_type:
1446     return lowerTypeMemberPointer(cast<DIDerivedType>(Ty));
1447   case dwarf::DW_TAG_restrict_type:
1448   case dwarf::DW_TAG_const_type:
1449   case dwarf::DW_TAG_volatile_type:
1450   // TODO: add support for DW_TAG_atomic_type here
1451     return lowerTypeModifier(cast<DIDerivedType>(Ty));
1452   case dwarf::DW_TAG_subroutine_type:
1453     if (ClassTy) {
1454       // The member function type of a member function pointer has no
1455       // ThisAdjustment.
1456       return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy,
1457                                      /*ThisAdjustment=*/0,
1458                                      /*IsStaticMethod=*/false);
1459     }
1460     return lowerTypeFunction(cast<DISubroutineType>(Ty));
1461   case dwarf::DW_TAG_enumeration_type:
1462     return lowerTypeEnum(cast<DICompositeType>(Ty));
1463   case dwarf::DW_TAG_class_type:
1464   case dwarf::DW_TAG_structure_type:
1465     return lowerTypeClass(cast<DICompositeType>(Ty));
1466   case dwarf::DW_TAG_union_type:
1467     return lowerTypeUnion(cast<DICompositeType>(Ty));
1468   case dwarf::DW_TAG_unspecified_type:
1469     if (Ty->getName() == "decltype(nullptr)")
1470       return TypeIndex::NullptrT();
1471     return TypeIndex::None();
1472   default:
1473     // Use the null type index.
1474     return TypeIndex();
1475   }
1476 }
1477 
1478 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) {
1479   DITypeRef UnderlyingTypeRef = Ty->getBaseType();
1480   TypeIndex UnderlyingTypeIndex = getTypeIndex(UnderlyingTypeRef);
1481   StringRef TypeName = Ty->getName();
1482 
1483   addToUDTs(Ty);
1484 
1485   if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) &&
1486       TypeName == "HRESULT")
1487     return TypeIndex(SimpleTypeKind::HResult);
1488   if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) &&
1489       TypeName == "wchar_t")
1490     return TypeIndex(SimpleTypeKind::WideCharacter);
1491 
1492   return UnderlyingTypeIndex;
1493 }
1494 
1495 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) {
1496   DITypeRef ElementTypeRef = Ty->getBaseType();
1497   TypeIndex ElementTypeIndex = getTypeIndex(ElementTypeRef);
1498   // IndexType is size_t, which depends on the bitness of the target.
1499   TypeIndex IndexType = getPointerSizeInBytes() == 8
1500                             ? TypeIndex(SimpleTypeKind::UInt64Quad)
1501                             : TypeIndex(SimpleTypeKind::UInt32Long);
1502 
1503   uint64_t ElementSize = getBaseTypeSize(ElementTypeRef) / 8;
1504 
1505   // Add subranges to array type.
1506   DINodeArray Elements = Ty->getElements();
1507   for (int i = Elements.size() - 1; i >= 0; --i) {
1508     const DINode *Element = Elements[i];
1509     assert(Element->getTag() == dwarf::DW_TAG_subrange_type);
1510 
1511     const DISubrange *Subrange = cast<DISubrange>(Element);
1512     assert(Subrange->getLowerBound() == 0 &&
1513            "codeview doesn't support subranges with lower bounds");
1514     int64_t Count = -1;
1515     if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>())
1516       Count = CI->getSExtValue();
1517 
1518     // Forward declarations of arrays without a size and VLAs use a count of -1.
1519     // Emit a count of zero in these cases to match what MSVC does for arrays
1520     // without a size. MSVC doesn't support VLAs, so it's not clear what we
1521     // should do for them even if we could distinguish them.
1522     if (Count == -1)
1523       Count = 0;
1524 
1525     // Update the element size and element type index for subsequent subranges.
1526     ElementSize *= Count;
1527 
1528     // If this is the outermost array, use the size from the array. It will be
1529     // more accurate if we had a VLA or an incomplete element type size.
1530     uint64_t ArraySize =
1531         (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize;
1532 
1533     StringRef Name = (i == 0) ? Ty->getName() : "";
1534     ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name);
1535     ElementTypeIndex = TypeTable.writeLeafType(AR);
1536   }
1537 
1538   return ElementTypeIndex;
1539 }
1540 
1541 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) {
1542   TypeIndex Index;
1543   dwarf::TypeKind Kind;
1544   uint32_t ByteSize;
1545 
1546   Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding());
1547   ByteSize = Ty->getSizeInBits() / 8;
1548 
1549   SimpleTypeKind STK = SimpleTypeKind::None;
1550   switch (Kind) {
1551   case dwarf::DW_ATE_address:
1552     // FIXME: Translate
1553     break;
1554   case dwarf::DW_ATE_boolean:
1555     switch (ByteSize) {
1556     case 1:  STK = SimpleTypeKind::Boolean8;   break;
1557     case 2:  STK = SimpleTypeKind::Boolean16;  break;
1558     case 4:  STK = SimpleTypeKind::Boolean32;  break;
1559     case 8:  STK = SimpleTypeKind::Boolean64;  break;
1560     case 16: STK = SimpleTypeKind::Boolean128; break;
1561     }
1562     break;
1563   case dwarf::DW_ATE_complex_float:
1564     switch (ByteSize) {
1565     case 2:  STK = SimpleTypeKind::Complex16;  break;
1566     case 4:  STK = SimpleTypeKind::Complex32;  break;
1567     case 8:  STK = SimpleTypeKind::Complex64;  break;
1568     case 10: STK = SimpleTypeKind::Complex80;  break;
1569     case 16: STK = SimpleTypeKind::Complex128; break;
1570     }
1571     break;
1572   case dwarf::DW_ATE_float:
1573     switch (ByteSize) {
1574     case 2:  STK = SimpleTypeKind::Float16;  break;
1575     case 4:  STK = SimpleTypeKind::Float32;  break;
1576     case 6:  STK = SimpleTypeKind::Float48;  break;
1577     case 8:  STK = SimpleTypeKind::Float64;  break;
1578     case 10: STK = SimpleTypeKind::Float80;  break;
1579     case 16: STK = SimpleTypeKind::Float128; break;
1580     }
1581     break;
1582   case dwarf::DW_ATE_signed:
1583     switch (ByteSize) {
1584     case 1:  STK = SimpleTypeKind::SignedCharacter; break;
1585     case 2:  STK = SimpleTypeKind::Int16Short;      break;
1586     case 4:  STK = SimpleTypeKind::Int32;           break;
1587     case 8:  STK = SimpleTypeKind::Int64Quad;       break;
1588     case 16: STK = SimpleTypeKind::Int128Oct;       break;
1589     }
1590     break;
1591   case dwarf::DW_ATE_unsigned:
1592     switch (ByteSize) {
1593     case 1:  STK = SimpleTypeKind::UnsignedCharacter; break;
1594     case 2:  STK = SimpleTypeKind::UInt16Short;       break;
1595     case 4:  STK = SimpleTypeKind::UInt32;            break;
1596     case 8:  STK = SimpleTypeKind::UInt64Quad;        break;
1597     case 16: STK = SimpleTypeKind::UInt128Oct;        break;
1598     }
1599     break;
1600   case dwarf::DW_ATE_UTF:
1601     switch (ByteSize) {
1602     case 2: STK = SimpleTypeKind::Character16; break;
1603     case 4: STK = SimpleTypeKind::Character32; break;
1604     }
1605     break;
1606   case dwarf::DW_ATE_signed_char:
1607     if (ByteSize == 1)
1608       STK = SimpleTypeKind::SignedCharacter;
1609     break;
1610   case dwarf::DW_ATE_unsigned_char:
1611     if (ByteSize == 1)
1612       STK = SimpleTypeKind::UnsignedCharacter;
1613     break;
1614   default:
1615     break;
1616   }
1617 
1618   // Apply some fixups based on the source-level type name.
1619   if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int")
1620     STK = SimpleTypeKind::Int32Long;
1621   if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int")
1622     STK = SimpleTypeKind::UInt32Long;
1623   if (STK == SimpleTypeKind::UInt16Short &&
1624       (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t"))
1625     STK = SimpleTypeKind::WideCharacter;
1626   if ((STK == SimpleTypeKind::SignedCharacter ||
1627        STK == SimpleTypeKind::UnsignedCharacter) &&
1628       Ty->getName() == "char")
1629     STK = SimpleTypeKind::NarrowCharacter;
1630 
1631   return TypeIndex(STK);
1632 }
1633 
1634 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty,
1635                                           PointerOptions PO) {
1636   TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType());
1637 
1638   // Pointers to simple types without any options can use SimpleTypeMode, rather
1639   // than having a dedicated pointer type record.
1640   if (PointeeTI.isSimple() && PO == PointerOptions::None &&
1641       PointeeTI.getSimpleMode() == SimpleTypeMode::Direct &&
1642       Ty->getTag() == dwarf::DW_TAG_pointer_type) {
1643     SimpleTypeMode Mode = Ty->getSizeInBits() == 64
1644                               ? SimpleTypeMode::NearPointer64
1645                               : SimpleTypeMode::NearPointer32;
1646     return TypeIndex(PointeeTI.getSimpleKind(), Mode);
1647   }
1648 
1649   PointerKind PK =
1650       Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32;
1651   PointerMode PM = PointerMode::Pointer;
1652   switch (Ty->getTag()) {
1653   default: llvm_unreachable("not a pointer tag type");
1654   case dwarf::DW_TAG_pointer_type:
1655     PM = PointerMode::Pointer;
1656     break;
1657   case dwarf::DW_TAG_reference_type:
1658     PM = PointerMode::LValueReference;
1659     break;
1660   case dwarf::DW_TAG_rvalue_reference_type:
1661     PM = PointerMode::RValueReference;
1662     break;
1663   }
1664 
1665   if (Ty->isObjectPointer())
1666     PO |= PointerOptions::Const;
1667 
1668   PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8);
1669   return TypeTable.writeLeafType(PR);
1670 }
1671 
1672 static PointerToMemberRepresentation
1673 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) {
1674   // SizeInBytes being zero generally implies that the member pointer type was
1675   // incomplete, which can happen if it is part of a function prototype. In this
1676   // case, use the unknown model instead of the general model.
1677   if (IsPMF) {
1678     switch (Flags & DINode::FlagPtrToMemberRep) {
1679     case 0:
1680       return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown
1681                               : PointerToMemberRepresentation::GeneralFunction;
1682     case DINode::FlagSingleInheritance:
1683       return PointerToMemberRepresentation::SingleInheritanceFunction;
1684     case DINode::FlagMultipleInheritance:
1685       return PointerToMemberRepresentation::MultipleInheritanceFunction;
1686     case DINode::FlagVirtualInheritance:
1687       return PointerToMemberRepresentation::VirtualInheritanceFunction;
1688     }
1689   } else {
1690     switch (Flags & DINode::FlagPtrToMemberRep) {
1691     case 0:
1692       return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown
1693                               : PointerToMemberRepresentation::GeneralData;
1694     case DINode::FlagSingleInheritance:
1695       return PointerToMemberRepresentation::SingleInheritanceData;
1696     case DINode::FlagMultipleInheritance:
1697       return PointerToMemberRepresentation::MultipleInheritanceData;
1698     case DINode::FlagVirtualInheritance:
1699       return PointerToMemberRepresentation::VirtualInheritanceData;
1700     }
1701   }
1702   llvm_unreachable("invalid ptr to member representation");
1703 }
1704 
1705 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty,
1706                                                 PointerOptions PO) {
1707   assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type);
1708   TypeIndex ClassTI = getTypeIndex(Ty->getClassType());
1709   TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType());
1710   PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64
1711                                                 : PointerKind::Near32;
1712   bool IsPMF = isa<DISubroutineType>(Ty->getBaseType());
1713   PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction
1714                          : PointerMode::PointerToDataMember;
1715 
1716   assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big");
1717   uint8_t SizeInBytes = Ty->getSizeInBits() / 8;
1718   MemberPointerInfo MPI(
1719       ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags()));
1720   PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI);
1721   return TypeTable.writeLeafType(PR);
1722 }
1723 
1724 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't
1725 /// have a translation, use the NearC convention.
1726 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) {
1727   switch (DwarfCC) {
1728   case dwarf::DW_CC_normal:             return CallingConvention::NearC;
1729   case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast;
1730   case dwarf::DW_CC_BORLAND_thiscall:   return CallingConvention::ThisCall;
1731   case dwarf::DW_CC_BORLAND_stdcall:    return CallingConvention::NearStdCall;
1732   case dwarf::DW_CC_BORLAND_pascal:     return CallingConvention::NearPascal;
1733   case dwarf::DW_CC_LLVM_vectorcall:    return CallingConvention::NearVector;
1734   }
1735   return CallingConvention::NearC;
1736 }
1737 
1738 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) {
1739   ModifierOptions Mods = ModifierOptions::None;
1740   PointerOptions PO = PointerOptions::None;
1741   bool IsModifier = true;
1742   const DIType *BaseTy = Ty;
1743   while (IsModifier && BaseTy) {
1744     // FIXME: Need to add DWARF tags for __unaligned and _Atomic
1745     switch (BaseTy->getTag()) {
1746     case dwarf::DW_TAG_const_type:
1747       Mods |= ModifierOptions::Const;
1748       PO |= PointerOptions::Const;
1749       break;
1750     case dwarf::DW_TAG_volatile_type:
1751       Mods |= ModifierOptions::Volatile;
1752       PO |= PointerOptions::Volatile;
1753       break;
1754     case dwarf::DW_TAG_restrict_type:
1755       // Only pointer types be marked with __restrict. There is no known flag
1756       // for __restrict in LF_MODIFIER records.
1757       PO |= PointerOptions::Restrict;
1758       break;
1759     default:
1760       IsModifier = false;
1761       break;
1762     }
1763     if (IsModifier)
1764       BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType().resolve();
1765   }
1766 
1767   // Check if the inner type will use an LF_POINTER record. If so, the
1768   // qualifiers will go in the LF_POINTER record. This comes up for types like
1769   // 'int *const' and 'int *__restrict', not the more common cases like 'const
1770   // char *'.
1771   if (BaseTy) {
1772     switch (BaseTy->getTag()) {
1773     case dwarf::DW_TAG_pointer_type:
1774     case dwarf::DW_TAG_reference_type:
1775     case dwarf::DW_TAG_rvalue_reference_type:
1776       return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO);
1777     case dwarf::DW_TAG_ptr_to_member_type:
1778       return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO);
1779     default:
1780       break;
1781     }
1782   }
1783 
1784   TypeIndex ModifiedTI = getTypeIndex(BaseTy);
1785 
1786   // Return the base type index if there aren't any modifiers. For example, the
1787   // metadata could contain restrict wrappers around non-pointer types.
1788   if (Mods == ModifierOptions::None)
1789     return ModifiedTI;
1790 
1791   ModifierRecord MR(ModifiedTI, Mods);
1792   return TypeTable.writeLeafType(MR);
1793 }
1794 
1795 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) {
1796   SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices;
1797   for (DITypeRef ArgTypeRef : Ty->getTypeArray())
1798     ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgTypeRef));
1799 
1800   // MSVC uses type none for variadic argument.
1801   if (ReturnAndArgTypeIndices.size() > 1 &&
1802       ReturnAndArgTypeIndices.back() == TypeIndex::Void()) {
1803     ReturnAndArgTypeIndices.back() = TypeIndex::None();
1804   }
1805   TypeIndex ReturnTypeIndex = TypeIndex::Void();
1806   ArrayRef<TypeIndex> ArgTypeIndices = None;
1807   if (!ReturnAndArgTypeIndices.empty()) {
1808     auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices);
1809     ReturnTypeIndex = ReturnAndArgTypesRef.front();
1810     ArgTypeIndices = ReturnAndArgTypesRef.drop_front();
1811   }
1812 
1813   ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices);
1814   TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec);
1815 
1816   CallingConvention CC = dwarfCCToCodeView(Ty->getCC());
1817 
1818   FunctionOptions FO = getFunctionOptions(Ty);
1819   ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(),
1820                             ArgListIndex);
1821   return TypeTable.writeLeafType(Procedure);
1822 }
1823 
1824 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty,
1825                                                  const DIType *ClassTy,
1826                                                  int ThisAdjustment,
1827                                                  bool IsStaticMethod,
1828                                                  FunctionOptions FO) {
1829   // Lower the containing class type.
1830   TypeIndex ClassType = getTypeIndex(ClassTy);
1831 
1832   DITypeRefArray ReturnAndArgs = Ty->getTypeArray();
1833 
1834   unsigned Index = 0;
1835   SmallVector<TypeIndex, 8> ArgTypeIndices;
1836   TypeIndex ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]);
1837 
1838   TypeIndex ThisTypeIndex;
1839   if (!IsStaticMethod && ReturnAndArgs.size() > 1)
1840     ThisTypeIndex = getTypeIndexForThisPtr(ReturnAndArgs[Index++], Ty);
1841 
1842   while (Index < ReturnAndArgs.size())
1843     ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++]));
1844 
1845   // MSVC uses type none for variadic argument.
1846   if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void())
1847     ArgTypeIndices.back() = TypeIndex::None();
1848 
1849   ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices);
1850   TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec);
1851 
1852   CallingConvention CC = dwarfCCToCodeView(Ty->getCC());
1853 
1854   MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO,
1855                            ArgTypeIndices.size(), ArgListIndex, ThisAdjustment);
1856   return TypeTable.writeLeafType(MFR);
1857 }
1858 
1859 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) {
1860   unsigned VSlotCount =
1861       Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize());
1862   SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near);
1863 
1864   VFTableShapeRecord VFTSR(Slots);
1865   return TypeTable.writeLeafType(VFTSR);
1866 }
1867 
1868 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) {
1869   switch (Flags & DINode::FlagAccessibility) {
1870   case DINode::FlagPrivate:   return MemberAccess::Private;
1871   case DINode::FlagPublic:    return MemberAccess::Public;
1872   case DINode::FlagProtected: return MemberAccess::Protected;
1873   case 0:
1874     // If there was no explicit access control, provide the default for the tag.
1875     return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private
1876                                                  : MemberAccess::Public;
1877   }
1878   llvm_unreachable("access flags are exclusive");
1879 }
1880 
1881 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) {
1882   if (SP->isArtificial())
1883     return MethodOptions::CompilerGenerated;
1884 
1885   // FIXME: Handle other MethodOptions.
1886 
1887   return MethodOptions::None;
1888 }
1889 
1890 static MethodKind translateMethodKindFlags(const DISubprogram *SP,
1891                                            bool Introduced) {
1892   if (SP->getFlags() & DINode::FlagStaticMember)
1893     return MethodKind::Static;
1894 
1895   switch (SP->getVirtuality()) {
1896   case dwarf::DW_VIRTUALITY_none:
1897     break;
1898   case dwarf::DW_VIRTUALITY_virtual:
1899     return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual;
1900   case dwarf::DW_VIRTUALITY_pure_virtual:
1901     return Introduced ? MethodKind::PureIntroducingVirtual
1902                       : MethodKind::PureVirtual;
1903   default:
1904     llvm_unreachable("unhandled virtuality case");
1905   }
1906 
1907   return MethodKind::Vanilla;
1908 }
1909 
1910 static TypeRecordKind getRecordKind(const DICompositeType *Ty) {
1911   switch (Ty->getTag()) {
1912   case dwarf::DW_TAG_class_type:     return TypeRecordKind::Class;
1913   case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct;
1914   }
1915   llvm_unreachable("unexpected tag");
1916 }
1917 
1918 /// Return ClassOptions that should be present on both the forward declaration
1919 /// and the defintion of a tag type.
1920 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) {
1921   ClassOptions CO = ClassOptions::None;
1922 
1923   // MSVC always sets this flag, even for local types. Clang doesn't always
1924   // appear to give every type a linkage name, which may be problematic for us.
1925   // FIXME: Investigate the consequences of not following them here.
1926   if (!Ty->getIdentifier().empty())
1927     CO |= ClassOptions::HasUniqueName;
1928 
1929   // Put the Nested flag on a type if it appears immediately inside a tag type.
1930   // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass
1931   // here. That flag is only set on definitions, and not forward declarations.
1932   const DIScope *ImmediateScope = Ty->getScope().resolve();
1933   if (ImmediateScope && isa<DICompositeType>(ImmediateScope))
1934     CO |= ClassOptions::Nested;
1935 
1936   // Put the Scoped flag on function-local types. MSVC puts this flag for enum
1937   // type only when it has an immediate function scope. Clang never puts enums
1938   // inside DILexicalBlock scopes. Enum types, as generated by clang, are
1939   // always in function, class, or file scopes.
1940   if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) {
1941     if (ImmediateScope && isa<DISubprogram>(ImmediateScope))
1942       CO |= ClassOptions::Scoped;
1943   } else {
1944     for (const DIScope *Scope = ImmediateScope; Scope != nullptr;
1945          Scope = Scope->getScope().resolve()) {
1946       if (isa<DISubprogram>(Scope)) {
1947         CO |= ClassOptions::Scoped;
1948         break;
1949       }
1950     }
1951   }
1952 
1953   return CO;
1954 }
1955 
1956 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) {
1957   switch (Ty->getTag()) {
1958   case dwarf::DW_TAG_class_type:
1959   case dwarf::DW_TAG_structure_type:
1960   case dwarf::DW_TAG_union_type:
1961   case dwarf::DW_TAG_enumeration_type:
1962     break;
1963   default:
1964     return;
1965   }
1966 
1967   if (const auto *File = Ty->getFile()) {
1968     StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File));
1969     TypeIndex SIDI = TypeTable.writeLeafType(SIDR);
1970 
1971     UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine());
1972     TypeTable.writeLeafType(USLR);
1973   }
1974 }
1975 
1976 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) {
1977   ClassOptions CO = getCommonClassOptions(Ty);
1978   TypeIndex FTI;
1979   unsigned EnumeratorCount = 0;
1980 
1981   if (Ty->isForwardDecl()) {
1982     CO |= ClassOptions::ForwardReference;
1983   } else {
1984     ContinuationRecordBuilder ContinuationBuilder;
1985     ContinuationBuilder.begin(ContinuationRecordKind::FieldList);
1986     for (const DINode *Element : Ty->getElements()) {
1987       // We assume that the frontend provides all members in source declaration
1988       // order, which is what MSVC does.
1989       if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) {
1990         EnumeratorRecord ER(MemberAccess::Public,
1991                             APSInt::getUnsigned(Enumerator->getValue()),
1992                             Enumerator->getName());
1993         ContinuationBuilder.writeMemberType(ER);
1994         EnumeratorCount++;
1995       }
1996     }
1997     FTI = TypeTable.insertRecord(ContinuationBuilder);
1998   }
1999 
2000   std::string FullName = getFullyQualifiedName(Ty);
2001 
2002   EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(),
2003                 getTypeIndex(Ty->getBaseType()));
2004   TypeIndex EnumTI = TypeTable.writeLeafType(ER);
2005 
2006   addUDTSrcLine(Ty, EnumTI);
2007 
2008   return EnumTI;
2009 }
2010 
2011 //===----------------------------------------------------------------------===//
2012 // ClassInfo
2013 //===----------------------------------------------------------------------===//
2014 
2015 struct llvm::ClassInfo {
2016   struct MemberInfo {
2017     const DIDerivedType *MemberTypeNode;
2018     uint64_t BaseOffset;
2019   };
2020   // [MemberInfo]
2021   using MemberList = std::vector<MemberInfo>;
2022 
2023   using MethodsList = TinyPtrVector<const DISubprogram *>;
2024   // MethodName -> MethodsList
2025   using MethodsMap = MapVector<MDString *, MethodsList>;
2026 
2027   /// Base classes.
2028   std::vector<const DIDerivedType *> Inheritance;
2029 
2030   /// Direct members.
2031   MemberList Members;
2032   // Direct overloaded methods gathered by name.
2033   MethodsMap Methods;
2034 
2035   TypeIndex VShapeTI;
2036 
2037   std::vector<const DIType *> NestedTypes;
2038 };
2039 
2040 void CodeViewDebug::clear() {
2041   assert(CurFn == nullptr);
2042   FileIdMap.clear();
2043   FnDebugInfo.clear();
2044   FileToFilepathMap.clear();
2045   LocalUDTs.clear();
2046   GlobalUDTs.clear();
2047   TypeIndices.clear();
2048   CompleteTypeIndices.clear();
2049 }
2050 
2051 void CodeViewDebug::collectMemberInfo(ClassInfo &Info,
2052                                       const DIDerivedType *DDTy) {
2053   if (!DDTy->getName().empty()) {
2054     Info.Members.push_back({DDTy, 0});
2055     return;
2056   }
2057 
2058   // An unnamed member may represent a nested struct or union. Attempt to
2059   // interpret the unnamed member as a DICompositeType possibly wrapped in
2060   // qualifier types. Add all the indirect fields to the current record if that
2061   // succeeds, and drop the member if that fails.
2062   assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!");
2063   uint64_t Offset = DDTy->getOffsetInBits();
2064   const DIType *Ty = DDTy->getBaseType().resolve();
2065   bool FullyResolved = false;
2066   while (!FullyResolved) {
2067     switch (Ty->getTag()) {
2068     case dwarf::DW_TAG_const_type:
2069     case dwarf::DW_TAG_volatile_type:
2070       // FIXME: we should apply the qualifier types to the indirect fields
2071       // rather than dropping them.
2072       Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve();
2073       break;
2074     default:
2075       FullyResolved = true;
2076       break;
2077     }
2078   }
2079 
2080   const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty);
2081   if (!DCTy)
2082     return;
2083 
2084   ClassInfo NestedInfo = collectClassInfo(DCTy);
2085   for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members)
2086     Info.Members.push_back(
2087         {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset});
2088 }
2089 
2090 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) {
2091   ClassInfo Info;
2092   // Add elements to structure type.
2093   DINodeArray Elements = Ty->getElements();
2094   for (auto *Element : Elements) {
2095     // We assume that the frontend provides all members in source declaration
2096     // order, which is what MSVC does.
2097     if (!Element)
2098       continue;
2099     if (auto *SP = dyn_cast<DISubprogram>(Element)) {
2100       Info.Methods[SP->getRawName()].push_back(SP);
2101     } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) {
2102       if (DDTy->getTag() == dwarf::DW_TAG_member) {
2103         collectMemberInfo(Info, DDTy);
2104       } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) {
2105         Info.Inheritance.push_back(DDTy);
2106       } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type &&
2107                  DDTy->getName() == "__vtbl_ptr_type") {
2108         Info.VShapeTI = getTypeIndex(DDTy);
2109       } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) {
2110         Info.NestedTypes.push_back(DDTy);
2111       } else if (DDTy->getTag() == dwarf::DW_TAG_friend) {
2112         // Ignore friend members. It appears that MSVC emitted info about
2113         // friends in the past, but modern versions do not.
2114       }
2115     } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) {
2116       Info.NestedTypes.push_back(Composite);
2117     }
2118     // Skip other unrecognized kinds of elements.
2119   }
2120   return Info;
2121 }
2122 
2123 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) {
2124   // This routine is used by lowerTypeClass and lowerTypeUnion to determine
2125   // if a complete type should be emitted instead of a forward reference.
2126   return Ty->getName().empty() && Ty->getIdentifier().empty() &&
2127       !Ty->isForwardDecl();
2128 }
2129 
2130 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) {
2131   // Emit the complete type for unnamed structs.  C++ classes with methods
2132   // which have a circular reference back to the class type are expected to
2133   // be named by the front-end and should not be "unnamed".  C unnamed
2134   // structs should not have circular references.
2135   if (shouldAlwaysEmitCompleteClassType(Ty)) {
2136     // If this unnamed complete type is already in the process of being defined
2137     // then the description of the type is malformed and cannot be emitted
2138     // into CodeView correctly so report a fatal error.
2139     auto I = CompleteTypeIndices.find(Ty);
2140     if (I != CompleteTypeIndices.end() && I->second == TypeIndex())
2141       report_fatal_error("cannot debug circular reference to unnamed type");
2142     return getCompleteTypeIndex(Ty);
2143   }
2144 
2145   // First, construct the forward decl.  Don't look into Ty to compute the
2146   // forward decl options, since it might not be available in all TUs.
2147   TypeRecordKind Kind = getRecordKind(Ty);
2148   ClassOptions CO =
2149       ClassOptions::ForwardReference | getCommonClassOptions(Ty);
2150   std::string FullName = getFullyQualifiedName(Ty);
2151   ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0,
2152                  FullName, Ty->getIdentifier());
2153   TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR);
2154   if (!Ty->isForwardDecl())
2155     DeferredCompleteTypes.push_back(Ty);
2156   return FwdDeclTI;
2157 }
2158 
2159 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) {
2160   // Construct the field list and complete type record.
2161   TypeRecordKind Kind = getRecordKind(Ty);
2162   ClassOptions CO = getCommonClassOptions(Ty);
2163   TypeIndex FieldTI;
2164   TypeIndex VShapeTI;
2165   unsigned FieldCount;
2166   bool ContainsNestedClass;
2167   std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) =
2168       lowerRecordFieldList(Ty);
2169 
2170   if (ContainsNestedClass)
2171     CO |= ClassOptions::ContainsNestedClass;
2172 
2173   std::string FullName = getFullyQualifiedName(Ty);
2174 
2175   uint64_t SizeInBytes = Ty->getSizeInBits() / 8;
2176 
2177   ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI,
2178                  SizeInBytes, FullName, Ty->getIdentifier());
2179   TypeIndex ClassTI = TypeTable.writeLeafType(CR);
2180 
2181   addUDTSrcLine(Ty, ClassTI);
2182 
2183   addToUDTs(Ty);
2184 
2185   return ClassTI;
2186 }
2187 
2188 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) {
2189   // Emit the complete type for unnamed unions.
2190   if (shouldAlwaysEmitCompleteClassType(Ty))
2191     return getCompleteTypeIndex(Ty);
2192 
2193   ClassOptions CO =
2194       ClassOptions::ForwardReference | getCommonClassOptions(Ty);
2195   std::string FullName = getFullyQualifiedName(Ty);
2196   UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier());
2197   TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR);
2198   if (!Ty->isForwardDecl())
2199     DeferredCompleteTypes.push_back(Ty);
2200   return FwdDeclTI;
2201 }
2202 
2203 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) {
2204   ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty);
2205   TypeIndex FieldTI;
2206   unsigned FieldCount;
2207   bool ContainsNestedClass;
2208   std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) =
2209       lowerRecordFieldList(Ty);
2210 
2211   if (ContainsNestedClass)
2212     CO |= ClassOptions::ContainsNestedClass;
2213 
2214   uint64_t SizeInBytes = Ty->getSizeInBits() / 8;
2215   std::string FullName = getFullyQualifiedName(Ty);
2216 
2217   UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName,
2218                  Ty->getIdentifier());
2219   TypeIndex UnionTI = TypeTable.writeLeafType(UR);
2220 
2221   addUDTSrcLine(Ty, UnionTI);
2222 
2223   addToUDTs(Ty);
2224 
2225   return UnionTI;
2226 }
2227 
2228 std::tuple<TypeIndex, TypeIndex, unsigned, bool>
2229 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) {
2230   // Manually count members. MSVC appears to count everything that generates a
2231   // field list record. Each individual overload in a method overload group
2232   // contributes to this count, even though the overload group is a single field
2233   // list record.
2234   unsigned MemberCount = 0;
2235   ClassInfo Info = collectClassInfo(Ty);
2236   ContinuationRecordBuilder ContinuationBuilder;
2237   ContinuationBuilder.begin(ContinuationRecordKind::FieldList);
2238 
2239   // Create base classes.
2240   for (const DIDerivedType *I : Info.Inheritance) {
2241     if (I->getFlags() & DINode::FlagVirtual) {
2242       // Virtual base.
2243       unsigned VBPtrOffset = I->getVBPtrOffset();
2244       // FIXME: Despite the accessor name, the offset is really in bytes.
2245       unsigned VBTableIndex = I->getOffsetInBits() / 4;
2246       auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase
2247                             ? TypeRecordKind::IndirectVirtualBaseClass
2248                             : TypeRecordKind::VirtualBaseClass;
2249       VirtualBaseClassRecord VBCR(
2250           RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()),
2251           getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset,
2252           VBTableIndex);
2253 
2254       ContinuationBuilder.writeMemberType(VBCR);
2255       MemberCount++;
2256     } else {
2257       assert(I->getOffsetInBits() % 8 == 0 &&
2258              "bases must be on byte boundaries");
2259       BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()),
2260                           getTypeIndex(I->getBaseType()),
2261                           I->getOffsetInBits() / 8);
2262       ContinuationBuilder.writeMemberType(BCR);
2263       MemberCount++;
2264     }
2265   }
2266 
2267   // Create members.
2268   for (ClassInfo::MemberInfo &MemberInfo : Info.Members) {
2269     const DIDerivedType *Member = MemberInfo.MemberTypeNode;
2270     TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType());
2271     StringRef MemberName = Member->getName();
2272     MemberAccess Access =
2273         translateAccessFlags(Ty->getTag(), Member->getFlags());
2274 
2275     if (Member->isStaticMember()) {
2276       StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName);
2277       ContinuationBuilder.writeMemberType(SDMR);
2278       MemberCount++;
2279       continue;
2280     }
2281 
2282     // Virtual function pointer member.
2283     if ((Member->getFlags() & DINode::FlagArtificial) &&
2284         Member->getName().startswith("_vptr$")) {
2285       VFPtrRecord VFPR(getTypeIndex(Member->getBaseType()));
2286       ContinuationBuilder.writeMemberType(VFPR);
2287       MemberCount++;
2288       continue;
2289     }
2290 
2291     // Data member.
2292     uint64_t MemberOffsetInBits =
2293         Member->getOffsetInBits() + MemberInfo.BaseOffset;
2294     if (Member->isBitField()) {
2295       uint64_t StartBitOffset = MemberOffsetInBits;
2296       if (const auto *CI =
2297               dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) {
2298         MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset;
2299       }
2300       StartBitOffset -= MemberOffsetInBits;
2301       BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(),
2302                          StartBitOffset);
2303       MemberBaseType = TypeTable.writeLeafType(BFR);
2304     }
2305     uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8;
2306     DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes,
2307                          MemberName);
2308     ContinuationBuilder.writeMemberType(DMR);
2309     MemberCount++;
2310   }
2311 
2312   // Create methods
2313   for (auto &MethodItr : Info.Methods) {
2314     StringRef Name = MethodItr.first->getString();
2315 
2316     std::vector<OneMethodRecord> Methods;
2317     for (const DISubprogram *SP : MethodItr.second) {
2318       TypeIndex MethodType = getMemberFunctionType(SP, Ty);
2319       bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual;
2320 
2321       unsigned VFTableOffset = -1;
2322       if (Introduced)
2323         VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes();
2324 
2325       Methods.push_back(OneMethodRecord(
2326           MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()),
2327           translateMethodKindFlags(SP, Introduced),
2328           translateMethodOptionFlags(SP), VFTableOffset, Name));
2329       MemberCount++;
2330     }
2331     assert(!Methods.empty() && "Empty methods map entry");
2332     if (Methods.size() == 1)
2333       ContinuationBuilder.writeMemberType(Methods[0]);
2334     else {
2335       // FIXME: Make this use its own ContinuationBuilder so that
2336       // MethodOverloadList can be split correctly.
2337       MethodOverloadListRecord MOLR(Methods);
2338       TypeIndex MethodList = TypeTable.writeLeafType(MOLR);
2339 
2340       OverloadedMethodRecord OMR(Methods.size(), MethodList, Name);
2341       ContinuationBuilder.writeMemberType(OMR);
2342     }
2343   }
2344 
2345   // Create nested classes.
2346   for (const DIType *Nested : Info.NestedTypes) {
2347     NestedTypeRecord R(getTypeIndex(DITypeRef(Nested)), Nested->getName());
2348     ContinuationBuilder.writeMemberType(R);
2349     MemberCount++;
2350   }
2351 
2352   TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder);
2353   return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount,
2354                          !Info.NestedTypes.empty());
2355 }
2356 
2357 TypeIndex CodeViewDebug::getVBPTypeIndex() {
2358   if (!VBPType.getIndex()) {
2359     // Make a 'const int *' type.
2360     ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const);
2361     TypeIndex ModifiedTI = TypeTable.writeLeafType(MR);
2362 
2363     PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64
2364                                                   : PointerKind::Near32;
2365     PointerMode PM = PointerMode::Pointer;
2366     PointerOptions PO = PointerOptions::None;
2367     PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes());
2368     VBPType = TypeTable.writeLeafType(PR);
2369   }
2370 
2371   return VBPType;
2372 }
2373 
2374 TypeIndex CodeViewDebug::getTypeIndex(DITypeRef TypeRef, DITypeRef ClassTyRef) {
2375   const DIType *Ty = TypeRef.resolve();
2376   const DIType *ClassTy = ClassTyRef.resolve();
2377 
2378   // The null DIType is the void type. Don't try to hash it.
2379   if (!Ty)
2380     return TypeIndex::Void();
2381 
2382   // Check if we've already translated this type. Don't try to do a
2383   // get-or-create style insertion that caches the hash lookup across the
2384   // lowerType call. It will update the TypeIndices map.
2385   auto I = TypeIndices.find({Ty, ClassTy});
2386   if (I != TypeIndices.end())
2387     return I->second;
2388 
2389   TypeLoweringScope S(*this);
2390   TypeIndex TI = lowerType(Ty, ClassTy);
2391   return recordTypeIndexForDINode(Ty, TI, ClassTy);
2392 }
2393 
2394 codeview::TypeIndex
2395 CodeViewDebug::getTypeIndexForThisPtr(DITypeRef TypeRef,
2396                                       const DISubroutineType *SubroutineTy) {
2397   const DIType *Ty = TypeRef.resolve();
2398 
2399   PointerOptions Options = PointerOptions::None;
2400   if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference)
2401     Options = PointerOptions::LValueRefThisPointer;
2402   else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference)
2403     Options = PointerOptions::RValueRefThisPointer;
2404 
2405   // Check if we've already translated this type.  If there is no ref qualifier
2406   // on the function then we look up this pointer type with no associated class
2407   // so that the TypeIndex for the this pointer can be shared with the type
2408   // index for other pointers to this class type.  If there is a ref qualifier
2409   // then we lookup the pointer using the subroutine as the parent type.
2410   auto I = TypeIndices.find({Ty, SubroutineTy});
2411   if (I != TypeIndices.end())
2412     return I->second;
2413 
2414   TypeLoweringScope S(*this);
2415   TypeIndex TI = lowerTypePointer(cast<DIDerivedType>(Ty), Options);
2416   return recordTypeIndexForDINode(Ty, TI, SubroutineTy);
2417 }
2418 
2419 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(DITypeRef TypeRef) {
2420   DIType *Ty = TypeRef.resolve();
2421   PointerRecord PR(getTypeIndex(Ty),
2422                    getPointerSizeInBytes() == 8 ? PointerKind::Near64
2423                                                 : PointerKind::Near32,
2424                    PointerMode::LValueReference, PointerOptions::None,
2425                    Ty->getSizeInBits() / 8);
2426   return TypeTable.writeLeafType(PR);
2427 }
2428 
2429 TypeIndex CodeViewDebug::getCompleteTypeIndex(DITypeRef TypeRef) {
2430   const DIType *Ty = TypeRef.resolve();
2431 
2432   // The null DIType is the void type. Don't try to hash it.
2433   if (!Ty)
2434     return TypeIndex::Void();
2435 
2436   // Look through typedefs when getting the complete type index. Call
2437   // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are
2438   // emitted only once.
2439   if (Ty->getTag() == dwarf::DW_TAG_typedef)
2440     (void)getTypeIndex(Ty);
2441   while (Ty->getTag() == dwarf::DW_TAG_typedef)
2442     Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve();
2443 
2444   // If this is a non-record type, the complete type index is the same as the
2445   // normal type index. Just call getTypeIndex.
2446   switch (Ty->getTag()) {
2447   case dwarf::DW_TAG_class_type:
2448   case dwarf::DW_TAG_structure_type:
2449   case dwarf::DW_TAG_union_type:
2450     break;
2451   default:
2452     return getTypeIndex(Ty);
2453   }
2454 
2455   // Check if we've already translated the complete record type.
2456   const auto *CTy = cast<DICompositeType>(Ty);
2457   auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()});
2458   if (!InsertResult.second)
2459     return InsertResult.first->second;
2460 
2461   TypeLoweringScope S(*this);
2462 
2463   // Make sure the forward declaration is emitted first. It's unclear if this
2464   // is necessary, but MSVC does it, and we should follow suit until we can show
2465   // otherwise.
2466   // We only emit a forward declaration for named types.
2467   if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) {
2468     TypeIndex FwdDeclTI = getTypeIndex(CTy);
2469 
2470     // Just use the forward decl if we don't have complete type info. This
2471     // might happen if the frontend is using modules and expects the complete
2472     // definition to be emitted elsewhere.
2473     if (CTy->isForwardDecl())
2474       return FwdDeclTI;
2475   }
2476 
2477   TypeIndex TI;
2478   switch (CTy->getTag()) {
2479   case dwarf::DW_TAG_class_type:
2480   case dwarf::DW_TAG_structure_type:
2481     TI = lowerCompleteTypeClass(CTy);
2482     break;
2483   case dwarf::DW_TAG_union_type:
2484     TI = lowerCompleteTypeUnion(CTy);
2485     break;
2486   default:
2487     llvm_unreachable("not a record");
2488   }
2489 
2490   // Update the type index associated with this CompositeType.  This cannot
2491   // use the 'InsertResult' iterator above because it is potentially
2492   // invalidated by map insertions which can occur while lowering the class
2493   // type above.
2494   CompleteTypeIndices[CTy] = TI;
2495   return TI;
2496 }
2497 
2498 /// Emit all the deferred complete record types. Try to do this in FIFO order,
2499 /// and do this until fixpoint, as each complete record type typically
2500 /// references
2501 /// many other record types.
2502 void CodeViewDebug::emitDeferredCompleteTypes() {
2503   SmallVector<const DICompositeType *, 4> TypesToEmit;
2504   while (!DeferredCompleteTypes.empty()) {
2505     std::swap(DeferredCompleteTypes, TypesToEmit);
2506     for (const DICompositeType *RecordTy : TypesToEmit)
2507       getCompleteTypeIndex(RecordTy);
2508     TypesToEmit.clear();
2509   }
2510 }
2511 
2512 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI,
2513                                           ArrayRef<LocalVariable> Locals) {
2514   // Get the sorted list of parameters and emit them first.
2515   SmallVector<const LocalVariable *, 6> Params;
2516   for (const LocalVariable &L : Locals)
2517     if (L.DIVar->isParameter())
2518       Params.push_back(&L);
2519   llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) {
2520     return L->DIVar->getArg() < R->DIVar->getArg();
2521   });
2522   for (const LocalVariable *L : Params)
2523     emitLocalVariable(FI, *L);
2524 
2525   // Next emit all non-parameters in the order that we found them.
2526   for (const LocalVariable &L : Locals)
2527     if (!L.DIVar->isParameter())
2528       emitLocalVariable(FI, L);
2529 }
2530 
2531 /// Only call this on endian-specific types like ulittle16_t and little32_t, or
2532 /// structs composed of them.
2533 template <typename T>
2534 static void copyBytesForDefRange(SmallString<20> &BytePrefix,
2535                                  SymbolKind SymKind, const T &DefRangeHeader) {
2536   BytePrefix.resize(2 + sizeof(T));
2537   ulittle16_t SymKindLE = ulittle16_t(SymKind);
2538   memcpy(&BytePrefix[0], &SymKindLE, 2);
2539   memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T));
2540 }
2541 
2542 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI,
2543                                       const LocalVariable &Var) {
2544   // LocalSym record, see SymbolRecord.h for more info.
2545   MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL);
2546 
2547   LocalSymFlags Flags = LocalSymFlags::None;
2548   if (Var.DIVar->isParameter())
2549     Flags |= LocalSymFlags::IsParameter;
2550   if (Var.DefRanges.empty())
2551     Flags |= LocalSymFlags::IsOptimizedOut;
2552 
2553   OS.AddComment("TypeIndex");
2554   TypeIndex TI = Var.UseReferenceType
2555                      ? getTypeIndexForReferenceTo(Var.DIVar->getType())
2556                      : getCompleteTypeIndex(Var.DIVar->getType());
2557   OS.EmitIntValue(TI.getIndex(), 4);
2558   OS.AddComment("Flags");
2559   OS.EmitIntValue(static_cast<uint16_t>(Flags), 2);
2560   // Truncate the name so we won't overflow the record length field.
2561   emitNullTerminatedSymbolName(OS, Var.DIVar->getName());
2562   endSymbolRecord(LocalEnd);
2563 
2564   // Calculate the on disk prefix of the appropriate def range record. The
2565   // records and on disk formats are described in SymbolRecords.h. BytePrefix
2566   // should be big enough to hold all forms without memory allocation.
2567   SmallString<20> BytePrefix;
2568   for (const LocalVarDefRange &DefRange : Var.DefRanges) {
2569     BytePrefix.clear();
2570     if (DefRange.InMemory) {
2571       int Offset = DefRange.DataOffset;
2572       unsigned Reg = DefRange.CVRegister;
2573 
2574       // 32-bit x86 call sequences often use PUSH instructions, which disrupt
2575       // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0,
2576       // instead. In frames without stack realignment, $T0 will be the CFA.
2577       if (RegisterId(Reg) == RegisterId::ESP) {
2578         Reg = unsigned(RegisterId::VFRAME);
2579         Offset += FI.OffsetAdjustment;
2580       }
2581 
2582       // If we can use the chosen frame pointer for the frame and this isn't a
2583       // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record.
2584       // Otherwise, use S_DEFRANGE_REGISTER_REL.
2585       EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU);
2586       if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None &&
2587           (bool(Flags & LocalSymFlags::IsParameter)
2588                ? (EncFP == FI.EncodedParamFramePtrReg)
2589                : (EncFP == FI.EncodedLocalFramePtrReg))) {
2590         little32_t FPOffset = little32_t(Offset);
2591         copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset);
2592       } else {
2593         uint16_t RegRelFlags = 0;
2594         if (DefRange.IsSubfield) {
2595           RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag |
2596                         (DefRange.StructOffset
2597                          << DefRangeRegisterRelSym::OffsetInParentShift);
2598         }
2599         DefRangeRegisterRelSym::Header DRHdr;
2600         DRHdr.Register = Reg;
2601         DRHdr.Flags = RegRelFlags;
2602         DRHdr.BasePointerOffset = Offset;
2603         copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr);
2604       }
2605     } else {
2606       assert(DefRange.DataOffset == 0 && "unexpected offset into register");
2607       if (DefRange.IsSubfield) {
2608         DefRangeSubfieldRegisterSym::Header DRHdr;
2609         DRHdr.Register = DefRange.CVRegister;
2610         DRHdr.MayHaveNoName = 0;
2611         DRHdr.OffsetInParent = DefRange.StructOffset;
2612         copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr);
2613       } else {
2614         DefRangeRegisterSym::Header DRHdr;
2615         DRHdr.Register = DefRange.CVRegister;
2616         DRHdr.MayHaveNoName = 0;
2617         copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr);
2618       }
2619     }
2620     OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix);
2621   }
2622 }
2623 
2624 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks,
2625                                          const FunctionInfo& FI) {
2626   for (LexicalBlock *Block : Blocks)
2627     emitLexicalBlock(*Block, FI);
2628 }
2629 
2630 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a
2631 /// lexical block scope.
2632 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block,
2633                                      const FunctionInfo& FI) {
2634   MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32);
2635   OS.AddComment("PtrParent");
2636   OS.EmitIntValue(0, 4);                                  // PtrParent
2637   OS.AddComment("PtrEnd");
2638   OS.EmitIntValue(0, 4);                                  // PtrEnd
2639   OS.AddComment("Code size");
2640   OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4);   // Code Size
2641   OS.AddComment("Function section relative address");
2642   OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0);         // Func Offset
2643   OS.AddComment("Function section index");
2644   OS.EmitCOFFSectionIndex(FI.Begin);                      // Func Symbol
2645   OS.AddComment("Lexical block name");
2646   emitNullTerminatedSymbolName(OS, Block.Name);           // Name
2647   endSymbolRecord(RecordEnd);
2648 
2649   // Emit variables local to this lexical block.
2650   emitLocalVariableList(FI, Block.Locals);
2651 
2652   // Emit lexical blocks contained within this block.
2653   emitLexicalBlockList(Block.Children, FI);
2654 
2655   // Close the lexical block scope.
2656   emitEndSymbolRecord(SymbolKind::S_END);
2657 }
2658 
2659 /// Convenience routine for collecting lexical block information for a list
2660 /// of lexical scopes.
2661 void CodeViewDebug::collectLexicalBlockInfo(
2662         SmallVectorImpl<LexicalScope *> &Scopes,
2663         SmallVectorImpl<LexicalBlock *> &Blocks,
2664         SmallVectorImpl<LocalVariable> &Locals) {
2665   for (LexicalScope *Scope : Scopes)
2666     collectLexicalBlockInfo(*Scope, Blocks, Locals);
2667 }
2668 
2669 /// Populate the lexical blocks and local variable lists of the parent with
2670 /// information about the specified lexical scope.
2671 void CodeViewDebug::collectLexicalBlockInfo(
2672     LexicalScope &Scope,
2673     SmallVectorImpl<LexicalBlock *> &ParentBlocks,
2674     SmallVectorImpl<LocalVariable> &ParentLocals) {
2675   if (Scope.isAbstractScope())
2676     return;
2677 
2678   auto LocalsIter = ScopeVariables.find(&Scope);
2679   if (LocalsIter == ScopeVariables.end()) {
2680     // This scope does not contain variables and can be eliminated.
2681     collectLexicalBlockInfo(Scope.getChildren(), ParentBlocks, ParentLocals);
2682     return;
2683   }
2684   SmallVectorImpl<LocalVariable> &Locals = LocalsIter->second;
2685 
2686   const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode());
2687   if (!DILB) {
2688     // This scope is not a lexical block and can be eliminated, but keep any
2689     // local variables it contains.
2690     ParentLocals.append(Locals.begin(), Locals.end());
2691     collectLexicalBlockInfo(Scope.getChildren(), ParentBlocks, ParentLocals);
2692     return;
2693   }
2694 
2695   const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges();
2696   if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) {
2697     // This lexical block scope has too many address ranges to represent in the
2698     // current CodeView format or does not have a valid address range.
2699     // Eliminate this lexical scope and promote any locals it contains to the
2700     // parent scope.
2701     //
2702     // For lexical scopes with multiple address ranges you may be tempted to
2703     // construct a single range covering every instruction where the block is
2704     // live and everything in between.  Unfortunately, Visual Studio only
2705     // displays variables from the first matching lexical block scope.  If the
2706     // first lexical block contains exception handling code or cold code which
2707     // is moved to the bottom of the routine creating a single range covering
2708     // nearly the entire routine, then it will hide all other lexical blocks
2709     // and the variables they contain.
2710     //
2711     ParentLocals.append(Locals.begin(), Locals.end());
2712     collectLexicalBlockInfo(Scope.getChildren(), ParentBlocks, ParentLocals);
2713     return;
2714   }
2715 
2716   // Create a new CodeView lexical block for this lexical scope.  If we've
2717   // seen this DILexicalBlock before then the scope tree is malformed and
2718   // we can handle this gracefully by not processing it a second time.
2719   auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()});
2720   if (!BlockInsertion.second)
2721     return;
2722 
2723   // Create a lexical block containing the local variables and collect the
2724   // the lexical block information for the children.
2725   const InsnRange &Range = Ranges.front();
2726   assert(Range.first && Range.second);
2727   LexicalBlock &Block = BlockInsertion.first->second;
2728   Block.Begin = getLabelBeforeInsn(Range.first);
2729   Block.End = getLabelAfterInsn(Range.second);
2730   assert(Block.Begin && "missing label for scope begin");
2731   assert(Block.End && "missing label for scope end");
2732   Block.Name = DILB->getName();
2733   Block.Locals = std::move(Locals);
2734   ParentBlocks.push_back(&Block);
2735   collectLexicalBlockInfo(Scope.getChildren(), Block.Children, Block.Locals);
2736 }
2737 
2738 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) {
2739   const Function &GV = MF->getFunction();
2740   assert(FnDebugInfo.count(&GV));
2741   assert(CurFn == FnDebugInfo[&GV].get());
2742 
2743   collectVariableInfo(GV.getSubprogram());
2744 
2745   // Build the lexical block structure to emit for this routine.
2746   if (LexicalScope *CFS = LScopes.getCurrentFunctionScope())
2747     collectLexicalBlockInfo(*CFS, CurFn->ChildBlocks, CurFn->Locals);
2748 
2749   // Clear the scope and variable information from the map which will not be
2750   // valid after we have finished processing this routine.  This also prepares
2751   // the map for the subsequent routine.
2752   ScopeVariables.clear();
2753 
2754   // Don't emit anything if we don't have any line tables.
2755   // Thunks are compiler-generated and probably won't have source correlation.
2756   if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) {
2757     FnDebugInfo.erase(&GV);
2758     CurFn = nullptr;
2759     return;
2760   }
2761 
2762   CurFn->Annotations = MF->getCodeViewAnnotations();
2763 
2764   CurFn->End = Asm->getFunctionEnd();
2765 
2766   CurFn = nullptr;
2767 }
2768 
2769 void CodeViewDebug::beginInstruction(const MachineInstr *MI) {
2770   DebugHandlerBase::beginInstruction(MI);
2771 
2772   // Ignore DBG_VALUE and DBG_LABEL locations and function prologue.
2773   if (!Asm || !CurFn || MI->isDebugInstr() ||
2774       MI->getFlag(MachineInstr::FrameSetup))
2775     return;
2776 
2777   // If the first instruction of a new MBB has no location, find the first
2778   // instruction with a location and use that.
2779   DebugLoc DL = MI->getDebugLoc();
2780   if (!DL && MI->getParent() != PrevInstBB) {
2781     for (const auto &NextMI : *MI->getParent()) {
2782       if (NextMI.isDebugInstr())
2783         continue;
2784       DL = NextMI.getDebugLoc();
2785       if (DL)
2786         break;
2787     }
2788   }
2789   PrevInstBB = MI->getParent();
2790 
2791   // If we still don't have a debug location, don't record a location.
2792   if (!DL)
2793     return;
2794 
2795   maybeRecordLocation(DL, Asm->MF);
2796 }
2797 
2798 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) {
2799   MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(),
2800            *EndLabel = MMI->getContext().createTempSymbol();
2801   OS.EmitIntValue(unsigned(Kind), 4);
2802   OS.AddComment("Subsection size");
2803   OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4);
2804   OS.EmitLabel(BeginLabel);
2805   return EndLabel;
2806 }
2807 
2808 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) {
2809   OS.EmitLabel(EndLabel);
2810   // Every subsection must be aligned to a 4-byte boundary.
2811   OS.EmitValueToAlignment(4);
2812 }
2813 
2814 static StringRef getSymbolName(SymbolKind SymKind) {
2815   for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames())
2816     if (EE.Value == SymKind)
2817       return EE.Name;
2818   return "";
2819 }
2820 
2821 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) {
2822   MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(),
2823            *EndLabel = MMI->getContext().createTempSymbol();
2824   OS.AddComment("Record length");
2825   OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2);
2826   OS.EmitLabel(BeginLabel);
2827   if (OS.isVerboseAsm())
2828     OS.AddComment("Record kind: " + getSymbolName(SymKind));
2829   OS.EmitIntValue(unsigned(SymKind), 2);
2830   return EndLabel;
2831 }
2832 
2833 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) {
2834   // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid
2835   // an extra copy of every symbol record in LLD. This increases object file
2836   // size by less than 1% in the clang build, and is compatible with the Visual
2837   // C++ linker.
2838   OS.EmitValueToAlignment(4);
2839   OS.EmitLabel(SymEnd);
2840 }
2841 
2842 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) {
2843   OS.AddComment("Record length");
2844   OS.EmitIntValue(2, 2);
2845   if (OS.isVerboseAsm())
2846     OS.AddComment("Record kind: " + getSymbolName(EndKind));
2847   OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind
2848 }
2849 
2850 void CodeViewDebug::emitDebugInfoForUDTs(
2851     ArrayRef<std::pair<std::string, const DIType *>> UDTs) {
2852   for (const auto &UDT : UDTs) {
2853     const DIType *T = UDT.second;
2854     assert(shouldEmitUdt(T));
2855 
2856     MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT);
2857     OS.AddComment("Type");
2858     OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4);
2859     emitNullTerminatedSymbolName(OS, UDT.first);
2860     endSymbolRecord(UDTRecordEnd);
2861   }
2862 }
2863 
2864 void CodeViewDebug::emitDebugInfoForGlobals() {
2865   DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *>
2866       GlobalMap;
2867   for (const GlobalVariable &GV : MMI->getModule()->globals()) {
2868     SmallVector<DIGlobalVariableExpression *, 1> GVEs;
2869     GV.getDebugInfo(GVEs);
2870     for (const auto *GVE : GVEs)
2871       GlobalMap[GVE] = &GV;
2872   }
2873 
2874   NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
2875   for (const MDNode *Node : CUs->operands()) {
2876     const auto *CU = cast<DICompileUnit>(Node);
2877 
2878     // First, emit all globals that are not in a comdat in a single symbol
2879     // substream. MSVC doesn't like it if the substream is empty, so only open
2880     // it if we have at least one global to emit.
2881     switchToDebugSectionForSymbol(nullptr);
2882     MCSymbol *EndLabel = nullptr;
2883     for (const auto *GVE : CU->getGlobalVariables()) {
2884       if (const auto *GV = GlobalMap.lookup(GVE))
2885         if (!GV->hasComdat() && !GV->isDeclarationForLinker()) {
2886           if (!EndLabel) {
2887             OS.AddComment("Symbol subsection for globals");
2888             EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols);
2889           }
2890           // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions.
2891           emitDebugInfoForGlobal(GVE->getVariable(), GV, Asm->getSymbol(GV));
2892         }
2893     }
2894     if (EndLabel)
2895       endCVSubsection(EndLabel);
2896 
2897     // Second, emit each global that is in a comdat into its own .debug$S
2898     // section along with its own symbol substream.
2899     for (const auto *GVE : CU->getGlobalVariables()) {
2900       if (const auto *GV = GlobalMap.lookup(GVE)) {
2901         if (GV->hasComdat()) {
2902           MCSymbol *GVSym = Asm->getSymbol(GV);
2903           OS.AddComment("Symbol subsection for " +
2904                         Twine(GlobalValue::dropLLVMManglingEscape(GV->getName())));
2905           switchToDebugSectionForSymbol(GVSym);
2906           EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols);
2907           // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions.
2908           emitDebugInfoForGlobal(GVE->getVariable(), GV, GVSym);
2909           endCVSubsection(EndLabel);
2910         }
2911       }
2912     }
2913   }
2914 }
2915 
2916 void CodeViewDebug::emitDebugInfoForRetainedTypes() {
2917   NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
2918   for (const MDNode *Node : CUs->operands()) {
2919     for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) {
2920       if (DIType *RT = dyn_cast<DIType>(Ty)) {
2921         getTypeIndex(RT);
2922         // FIXME: Add to global/local DTU list.
2923       }
2924     }
2925   }
2926 }
2927 
2928 void CodeViewDebug::emitDebugInfoForGlobal(const DIGlobalVariable *DIGV,
2929                                            const GlobalVariable *GV,
2930                                            MCSymbol *GVSym) {
2931   // DataSym record, see SymbolRecord.h for more info. Thread local data
2932   // happens to have the same format as global data.
2933   SymbolKind DataSym = GV->isThreadLocal()
2934                            ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32
2935                                                     : SymbolKind::S_GTHREAD32)
2936                            : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32
2937                                                     : SymbolKind::S_GDATA32);
2938   MCSymbol *DataEnd = beginSymbolRecord(DataSym);
2939   OS.AddComment("Type");
2940   OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4);
2941   OS.AddComment("DataOffset");
2942   OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0);
2943   OS.AddComment("Segment");
2944   OS.EmitCOFFSectionIndex(GVSym);
2945   OS.AddComment("Name");
2946   const unsigned LengthOfDataRecord = 12;
2947   emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord);
2948   endSymbolRecord(DataEnd);
2949 }
2950