1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "DwarfExpression.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/ArrayRef.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/DenseSet.h" 19 #include "llvm/ADT/MapVector.h" 20 #include "llvm/ADT/None.h" 21 #include "llvm/ADT/Optional.h" 22 #include "llvm/ADT/STLExtras.h" 23 #include "llvm/ADT/SmallString.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/StringRef.h" 26 #include "llvm/ADT/TinyPtrVector.h" 27 #include "llvm/ADT/Triple.h" 28 #include "llvm/ADT/Twine.h" 29 #include "llvm/BinaryFormat/COFF.h" 30 #include "llvm/BinaryFormat/Dwarf.h" 31 #include "llvm/CodeGen/AsmPrinter.h" 32 #include "llvm/CodeGen/LexicalScopes.h" 33 #include "llvm/CodeGen/MachineFrameInfo.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineInstr.h" 36 #include "llvm/CodeGen/MachineModuleInfo.h" 37 #include "llvm/CodeGen/MachineOperand.h" 38 #include "llvm/CodeGen/TargetFrameLowering.h" 39 #include "llvm/CodeGen/TargetRegisterInfo.h" 40 #include "llvm/CodeGen/TargetSubtargetInfo.h" 41 #include "llvm/Config/llvm-config.h" 42 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 43 #include "llvm/DebugInfo/CodeView/CodeView.h" 44 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 45 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 46 #include "llvm/DebugInfo/CodeView/EnumTables.h" 47 #include "llvm/DebugInfo/CodeView/Line.h" 48 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 49 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 50 #include "llvm/DebugInfo/CodeView/TypeIndex.h" 51 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 52 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 53 #include "llvm/IR/Constants.h" 54 #include "llvm/IR/DataLayout.h" 55 #include "llvm/IR/DebugInfoMetadata.h" 56 #include "llvm/IR/DebugLoc.h" 57 #include "llvm/IR/Function.h" 58 #include "llvm/IR/GlobalValue.h" 59 #include "llvm/IR/GlobalVariable.h" 60 #include "llvm/IR/Metadata.h" 61 #include "llvm/IR/Module.h" 62 #include "llvm/MC/MCAsmInfo.h" 63 #include "llvm/MC/MCContext.h" 64 #include "llvm/MC/MCSectionCOFF.h" 65 #include "llvm/MC/MCStreamer.h" 66 #include "llvm/MC/MCSymbol.h" 67 #include "llvm/Support/BinaryByteStream.h" 68 #include "llvm/Support/BinaryStreamReader.h" 69 #include "llvm/Support/Casting.h" 70 #include "llvm/Support/CommandLine.h" 71 #include "llvm/Support/Compiler.h" 72 #include "llvm/Support/Endian.h" 73 #include "llvm/Support/Error.h" 74 #include "llvm/Support/ErrorHandling.h" 75 #include "llvm/Support/FormatVariadic.h" 76 #include "llvm/Support/Path.h" 77 #include "llvm/Support/SMLoc.h" 78 #include "llvm/Support/ScopedPrinter.h" 79 #include "llvm/Target/TargetLoweringObjectFile.h" 80 #include "llvm/Target/TargetMachine.h" 81 #include <algorithm> 82 #include <cassert> 83 #include <cctype> 84 #include <cstddef> 85 #include <cstdint> 86 #include <iterator> 87 #include <limits> 88 #include <string> 89 #include <utility> 90 #include <vector> 91 92 using namespace llvm; 93 using namespace llvm::codeview; 94 95 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 96 switch (Type) { 97 case Triple::ArchType::x86: 98 return CPUType::Pentium3; 99 case Triple::ArchType::x86_64: 100 return CPUType::X64; 101 case Triple::ArchType::thumb: 102 return CPUType::Thumb; 103 case Triple::ArchType::aarch64: 104 return CPUType::ARM64; 105 default: 106 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 107 } 108 } 109 110 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 111 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 112 // If module doesn't have named metadata anchors or COFF debug section 113 // is not available, skip any debug info related stuff. 114 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 115 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 116 Asm = nullptr; 117 MMI->setDebugInfoAvailability(false); 118 return; 119 } 120 // Tell MMI that we have debug info. 121 MMI->setDebugInfoAvailability(true); 122 123 TheCPU = 124 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 125 126 collectGlobalVariableInfo(); 127 128 // Check if we should emit type record hashes. 129 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 130 MMI->getModule()->getModuleFlag("CodeViewGHash")); 131 EmitDebugGlobalHashes = GH && !GH->isZero(); 132 } 133 134 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 135 std::string &Filepath = FileToFilepathMap[File]; 136 if (!Filepath.empty()) 137 return Filepath; 138 139 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 140 141 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 142 // it textually because one of the path components could be a symlink. 143 if (Dir.startswith("/") || Filename.startswith("/")) { 144 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 145 return Filename; 146 Filepath = Dir; 147 if (Dir.back() != '/') 148 Filepath += '/'; 149 Filepath += Filename; 150 return Filepath; 151 } 152 153 // Clang emits directory and relative filename info into the IR, but CodeView 154 // operates on full paths. We could change Clang to emit full paths too, but 155 // that would increase the IR size and probably not needed for other users. 156 // For now, just concatenate and canonicalize the path here. 157 if (Filename.find(':') == 1) 158 Filepath = Filename; 159 else 160 Filepath = (Dir + "\\" + Filename).str(); 161 162 // Canonicalize the path. We have to do it textually because we may no longer 163 // have access the file in the filesystem. 164 // First, replace all slashes with backslashes. 165 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 166 167 // Remove all "\.\" with "\". 168 size_t Cursor = 0; 169 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 170 Filepath.erase(Cursor, 2); 171 172 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 173 // path should be well-formatted, e.g. start with a drive letter, etc. 174 Cursor = 0; 175 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 176 // Something's wrong if the path starts with "\..\", abort. 177 if (Cursor == 0) 178 break; 179 180 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 181 if (PrevSlash == std::string::npos) 182 // Something's wrong, abort. 183 break; 184 185 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 186 // The next ".." might be following the one we've just erased. 187 Cursor = PrevSlash; 188 } 189 190 // Remove all duplicate backslashes. 191 Cursor = 0; 192 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 193 Filepath.erase(Cursor, 1); 194 195 return Filepath; 196 } 197 198 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 199 StringRef FullPath = getFullFilepath(F); 200 unsigned NextId = FileIdMap.size() + 1; 201 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 202 if (Insertion.second) { 203 // We have to compute the full filepath and emit a .cv_file directive. 204 ArrayRef<uint8_t> ChecksumAsBytes; 205 FileChecksumKind CSKind = FileChecksumKind::None; 206 if (F->getChecksum()) { 207 std::string Checksum = fromHex(F->getChecksum()->Value); 208 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 209 memcpy(CKMem, Checksum.data(), Checksum.size()); 210 ChecksumAsBytes = ArrayRef<uint8_t>( 211 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 212 switch (F->getChecksum()->Kind) { 213 case DIFile::CSK_MD5: CSKind = FileChecksumKind::MD5; break; 214 case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break; 215 } 216 } 217 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 218 static_cast<unsigned>(CSKind)); 219 (void)Success; 220 assert(Success && ".cv_file directive failed"); 221 } 222 return Insertion.first->second; 223 } 224 225 CodeViewDebug::InlineSite & 226 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 227 const DISubprogram *Inlinee) { 228 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 229 InlineSite *Site = &SiteInsertion.first->second; 230 if (SiteInsertion.second) { 231 unsigned ParentFuncId = CurFn->FuncId; 232 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 233 ParentFuncId = 234 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 235 .SiteFuncId; 236 237 Site->SiteFuncId = NextFuncId++; 238 OS.EmitCVInlineSiteIdDirective( 239 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 240 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 241 Site->Inlinee = Inlinee; 242 InlinedSubprograms.insert(Inlinee); 243 getFuncIdForSubprogram(Inlinee); 244 } 245 return *Site; 246 } 247 248 static StringRef getPrettyScopeName(const DIScope *Scope) { 249 StringRef ScopeName = Scope->getName(); 250 if (!ScopeName.empty()) 251 return ScopeName; 252 253 switch (Scope->getTag()) { 254 case dwarf::DW_TAG_enumeration_type: 255 case dwarf::DW_TAG_class_type: 256 case dwarf::DW_TAG_structure_type: 257 case dwarf::DW_TAG_union_type: 258 return "<unnamed-tag>"; 259 case dwarf::DW_TAG_namespace: 260 return "`anonymous namespace'"; 261 } 262 263 return StringRef(); 264 } 265 266 static const DISubprogram *getQualifiedNameComponents( 267 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 268 const DISubprogram *ClosestSubprogram = nullptr; 269 while (Scope != nullptr) { 270 if (ClosestSubprogram == nullptr) 271 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 272 StringRef ScopeName = getPrettyScopeName(Scope); 273 if (!ScopeName.empty()) 274 QualifiedNameComponents.push_back(ScopeName); 275 Scope = Scope->getScope().resolve(); 276 } 277 return ClosestSubprogram; 278 } 279 280 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents, 281 StringRef TypeName) { 282 std::string FullyQualifiedName; 283 for (StringRef QualifiedNameComponent : 284 llvm::reverse(QualifiedNameComponents)) { 285 FullyQualifiedName.append(QualifiedNameComponent); 286 FullyQualifiedName.append("::"); 287 } 288 FullyQualifiedName.append(TypeName); 289 return FullyQualifiedName; 290 } 291 292 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) { 293 SmallVector<StringRef, 5> QualifiedNameComponents; 294 getQualifiedNameComponents(Scope, QualifiedNameComponents); 295 return getQualifiedName(QualifiedNameComponents, Name); 296 } 297 298 struct CodeViewDebug::TypeLoweringScope { 299 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 300 ~TypeLoweringScope() { 301 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 302 // inner TypeLoweringScopes don't attempt to emit deferred types. 303 if (CVD.TypeEmissionLevel == 1) 304 CVD.emitDeferredCompleteTypes(); 305 --CVD.TypeEmissionLevel; 306 } 307 CodeViewDebug &CVD; 308 }; 309 310 static std::string getFullyQualifiedName(const DIScope *Ty) { 311 const DIScope *Scope = Ty->getScope().resolve(); 312 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 313 } 314 315 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 316 // No scope means global scope and that uses the zero index. 317 if (!Scope || isa<DIFile>(Scope)) 318 return TypeIndex(); 319 320 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 321 322 // Check if we've already translated this scope. 323 auto I = TypeIndices.find({Scope, nullptr}); 324 if (I != TypeIndices.end()) 325 return I->second; 326 327 // Build the fully qualified name of the scope. 328 std::string ScopeName = getFullyQualifiedName(Scope); 329 StringIdRecord SID(TypeIndex(), ScopeName); 330 auto TI = TypeTable.writeLeafType(SID); 331 return recordTypeIndexForDINode(Scope, TI); 332 } 333 334 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 335 assert(SP); 336 337 // Check if we've already translated this subprogram. 338 auto I = TypeIndices.find({SP, nullptr}); 339 if (I != TypeIndices.end()) 340 return I->second; 341 342 // The display name includes function template arguments. Drop them to match 343 // MSVC. 344 StringRef DisplayName = SP->getName().split('<').first; 345 346 const DIScope *Scope = SP->getScope().resolve(); 347 TypeIndex TI; 348 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 349 // If the scope is a DICompositeType, then this must be a method. Member 350 // function types take some special handling, and require access to the 351 // subprogram. 352 TypeIndex ClassType = getTypeIndex(Class); 353 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 354 DisplayName); 355 TI = TypeTable.writeLeafType(MFuncId); 356 } else { 357 // Otherwise, this must be a free function. 358 TypeIndex ParentScope = getScopeIndex(Scope); 359 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 360 TI = TypeTable.writeLeafType(FuncId); 361 } 362 363 return recordTypeIndexForDINode(SP, TI); 364 } 365 366 static bool isNonTrivial(const DICompositeType *DCTy) { 367 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 368 } 369 370 static FunctionOptions 371 getFunctionOptions(const DISubroutineType *Ty, 372 const DICompositeType *ClassTy = nullptr, 373 StringRef SPName = StringRef("")) { 374 FunctionOptions FO = FunctionOptions::None; 375 const DIType *ReturnTy = nullptr; 376 if (auto TypeArray = Ty->getTypeArray()) { 377 if (TypeArray.size()) 378 ReturnTy = TypeArray[0].resolve(); 379 } 380 381 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) { 382 if (isNonTrivial(ReturnDCTy)) 383 FO |= FunctionOptions::CxxReturnUdt; 384 } 385 386 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 387 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 388 FO |= FunctionOptions::Constructor; 389 390 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 391 392 } 393 return FO; 394 } 395 396 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 397 const DICompositeType *Class) { 398 // Always use the method declaration as the key for the function type. The 399 // method declaration contains the this adjustment. 400 if (SP->getDeclaration()) 401 SP = SP->getDeclaration(); 402 assert(!SP->getDeclaration() && "should use declaration as key"); 403 404 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 405 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 406 auto I = TypeIndices.find({SP, Class}); 407 if (I != TypeIndices.end()) 408 return I->second; 409 410 // Make sure complete type info for the class is emitted *after* the member 411 // function type, as the complete class type is likely to reference this 412 // member function type. 413 TypeLoweringScope S(*this); 414 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 415 416 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 417 TypeIndex TI = lowerTypeMemberFunction( 418 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 419 return recordTypeIndexForDINode(SP, TI, Class); 420 } 421 422 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 423 TypeIndex TI, 424 const DIType *ClassTy) { 425 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 426 (void)InsertResult; 427 assert(InsertResult.second && "DINode was already assigned a type index"); 428 return TI; 429 } 430 431 unsigned CodeViewDebug::getPointerSizeInBytes() { 432 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 433 } 434 435 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 436 const LexicalScope *LS) { 437 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 438 // This variable was inlined. Associate it with the InlineSite. 439 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 440 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 441 Site.InlinedLocals.emplace_back(Var); 442 } else { 443 // This variable goes into the corresponding lexical scope. 444 ScopeVariables[LS].emplace_back(Var); 445 } 446 } 447 448 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 449 const DILocation *Loc) { 450 auto B = Locs.begin(), E = Locs.end(); 451 if (std::find(B, E, Loc) == E) 452 Locs.push_back(Loc); 453 } 454 455 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 456 const MachineFunction *MF) { 457 // Skip this instruction if it has the same location as the previous one. 458 if (!DL || DL == PrevInstLoc) 459 return; 460 461 const DIScope *Scope = DL.get()->getScope(); 462 if (!Scope) 463 return; 464 465 // Skip this line if it is longer than the maximum we can record. 466 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 467 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 468 LI.isNeverStepInto()) 469 return; 470 471 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 472 if (CI.getStartColumn() != DL.getCol()) 473 return; 474 475 if (!CurFn->HaveLineInfo) 476 CurFn->HaveLineInfo = true; 477 unsigned FileId = 0; 478 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 479 FileId = CurFn->LastFileId; 480 else 481 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 482 PrevInstLoc = DL; 483 484 unsigned FuncId = CurFn->FuncId; 485 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 486 const DILocation *Loc = DL.get(); 487 488 // If this location was actually inlined from somewhere else, give it the ID 489 // of the inline call site. 490 FuncId = 491 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 492 493 // Ensure we have links in the tree of inline call sites. 494 bool FirstLoc = true; 495 while ((SiteLoc = Loc->getInlinedAt())) { 496 InlineSite &Site = 497 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 498 if (!FirstLoc) 499 addLocIfNotPresent(Site.ChildSites, Loc); 500 FirstLoc = false; 501 Loc = SiteLoc; 502 } 503 addLocIfNotPresent(CurFn->ChildSites, Loc); 504 } 505 506 OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 507 /*PrologueEnd=*/false, /*IsStmt=*/false, 508 DL->getFilename(), SMLoc()); 509 } 510 511 void CodeViewDebug::emitCodeViewMagicVersion() { 512 OS.EmitValueToAlignment(4); 513 OS.AddComment("Debug section magic"); 514 OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4); 515 } 516 517 void CodeViewDebug::endModule() { 518 if (!Asm || !MMI->hasDebugInfo()) 519 return; 520 521 assert(Asm != nullptr); 522 523 // The COFF .debug$S section consists of several subsections, each starting 524 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 525 // of the payload followed by the payload itself. The subsections are 4-byte 526 // aligned. 527 528 // Use the generic .debug$S section, and make a subsection for all the inlined 529 // subprograms. 530 switchToDebugSectionForSymbol(nullptr); 531 532 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 533 emitCompilerInformation(); 534 endCVSubsection(CompilerInfo); 535 536 emitInlineeLinesSubsection(); 537 538 // Emit per-function debug information. 539 for (auto &P : FnDebugInfo) 540 if (!P.first->isDeclarationForLinker()) 541 emitDebugInfoForFunction(P.first, *P.second); 542 543 // Emit global variable debug information. 544 setCurrentSubprogram(nullptr); 545 emitDebugInfoForGlobals(); 546 547 // Emit retained types. 548 emitDebugInfoForRetainedTypes(); 549 550 // Switch back to the generic .debug$S section after potentially processing 551 // comdat symbol sections. 552 switchToDebugSectionForSymbol(nullptr); 553 554 // Emit UDT records for any types used by global variables. 555 if (!GlobalUDTs.empty()) { 556 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 557 emitDebugInfoForUDTs(GlobalUDTs); 558 endCVSubsection(SymbolsEnd); 559 } 560 561 // This subsection holds a file index to offset in string table table. 562 OS.AddComment("File index to string table offset subsection"); 563 OS.EmitCVFileChecksumsDirective(); 564 565 // This subsection holds the string table. 566 OS.AddComment("String table"); 567 OS.EmitCVStringTableDirective(); 568 569 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 570 // subsection in the generic .debug$S section at the end. There is no 571 // particular reason for this ordering other than to match MSVC. 572 emitBuildInfo(); 573 574 // Emit type information and hashes last, so that any types we translate while 575 // emitting function info are included. 576 emitTypeInformation(); 577 578 if (EmitDebugGlobalHashes) 579 emitTypeGlobalHashes(); 580 581 clear(); 582 } 583 584 static void emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 585 unsigned MaxFixedRecordLength = 0xF00) { 586 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 587 // after a fixed length portion of the record. The fixed length portion should 588 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 589 // overall record size is less than the maximum allowed. 590 SmallString<32> NullTerminatedString( 591 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 592 NullTerminatedString.push_back('\0'); 593 OS.EmitBytes(NullTerminatedString); 594 } 595 596 void CodeViewDebug::emitTypeInformation() { 597 if (TypeTable.empty()) 598 return; 599 600 // Start the .debug$T or .debug$P section with 0x4. 601 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 602 emitCodeViewMagicVersion(); 603 604 SmallString<8> CommentPrefix; 605 if (OS.isVerboseAsm()) { 606 CommentPrefix += '\t'; 607 CommentPrefix += Asm->MAI->getCommentString(); 608 CommentPrefix += ' '; 609 } 610 611 TypeTableCollection Table(TypeTable.records()); 612 Optional<TypeIndex> B = Table.getFirst(); 613 while (B) { 614 // This will fail if the record data is invalid. 615 CVType Record = Table.getType(*B); 616 617 if (OS.isVerboseAsm()) { 618 // Emit a block comment describing the type record for readability. 619 SmallString<512> CommentBlock; 620 raw_svector_ostream CommentOS(CommentBlock); 621 ScopedPrinter SP(CommentOS); 622 SP.setPrefix(CommentPrefix); 623 TypeDumpVisitor TDV(Table, &SP, false); 624 625 Error E = codeview::visitTypeRecord(Record, *B, TDV); 626 if (E) { 627 logAllUnhandledErrors(std::move(E), errs(), "error: "); 628 llvm_unreachable("produced malformed type record"); 629 } 630 // emitRawComment will insert its own tab and comment string before 631 // the first line, so strip off our first one. It also prints its own 632 // newline. 633 OS.emitRawComment( 634 CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim()); 635 } 636 OS.EmitBinaryData(Record.str_data()); 637 B = Table.getNext(*B); 638 } 639 } 640 641 void CodeViewDebug::emitTypeGlobalHashes() { 642 if (TypeTable.empty()) 643 return; 644 645 // Start the .debug$H section with the version and hash algorithm, currently 646 // hardcoded to version 0, SHA1. 647 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 648 649 OS.EmitValueToAlignment(4); 650 OS.AddComment("Magic"); 651 OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4); 652 OS.AddComment("Section Version"); 653 OS.EmitIntValue(0, 2); 654 OS.AddComment("Hash Algorithm"); 655 OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2); 656 657 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 658 for (const auto &GHR : TypeTable.hashes()) { 659 if (OS.isVerboseAsm()) { 660 // Emit an EOL-comment describing which TypeIndex this hash corresponds 661 // to, as well as the stringified SHA1 hash. 662 SmallString<32> Comment; 663 raw_svector_ostream CommentOS(Comment); 664 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 665 OS.AddComment(Comment); 666 ++TI; 667 } 668 assert(GHR.Hash.size() == 8); 669 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 670 GHR.Hash.size()); 671 OS.EmitBinaryData(S); 672 } 673 } 674 675 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 676 switch (DWLang) { 677 case dwarf::DW_LANG_C: 678 case dwarf::DW_LANG_C89: 679 case dwarf::DW_LANG_C99: 680 case dwarf::DW_LANG_C11: 681 case dwarf::DW_LANG_ObjC: 682 return SourceLanguage::C; 683 case dwarf::DW_LANG_C_plus_plus: 684 case dwarf::DW_LANG_C_plus_plus_03: 685 case dwarf::DW_LANG_C_plus_plus_11: 686 case dwarf::DW_LANG_C_plus_plus_14: 687 return SourceLanguage::Cpp; 688 case dwarf::DW_LANG_Fortran77: 689 case dwarf::DW_LANG_Fortran90: 690 case dwarf::DW_LANG_Fortran03: 691 case dwarf::DW_LANG_Fortran08: 692 return SourceLanguage::Fortran; 693 case dwarf::DW_LANG_Pascal83: 694 return SourceLanguage::Pascal; 695 case dwarf::DW_LANG_Cobol74: 696 case dwarf::DW_LANG_Cobol85: 697 return SourceLanguage::Cobol; 698 case dwarf::DW_LANG_Java: 699 return SourceLanguage::Java; 700 case dwarf::DW_LANG_D: 701 return SourceLanguage::D; 702 case dwarf::DW_LANG_Swift: 703 return SourceLanguage::Swift; 704 default: 705 // There's no CodeView representation for this language, and CV doesn't 706 // have an "unknown" option for the language field, so we'll use MASM, 707 // as it's very low level. 708 return SourceLanguage::Masm; 709 } 710 } 711 712 namespace { 713 struct Version { 714 int Part[4]; 715 }; 716 } // end anonymous namespace 717 718 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 719 // the version number. 720 static Version parseVersion(StringRef Name) { 721 Version V = {{0}}; 722 int N = 0; 723 for (const char C : Name) { 724 if (isdigit(C)) { 725 V.Part[N] *= 10; 726 V.Part[N] += C - '0'; 727 } else if (C == '.') { 728 ++N; 729 if (N >= 4) 730 return V; 731 } else if (N > 0) 732 return V; 733 } 734 return V; 735 } 736 737 void CodeViewDebug::emitCompilerInformation() { 738 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 739 uint32_t Flags = 0; 740 741 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 742 const MDNode *Node = *CUs->operands().begin(); 743 const auto *CU = cast<DICompileUnit>(Node); 744 745 // The low byte of the flags indicates the source language. 746 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 747 // TODO: Figure out which other flags need to be set. 748 749 OS.AddComment("Flags and language"); 750 OS.EmitIntValue(Flags, 4); 751 752 OS.AddComment("CPUType"); 753 OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2); 754 755 StringRef CompilerVersion = CU->getProducer(); 756 Version FrontVer = parseVersion(CompilerVersion); 757 OS.AddComment("Frontend version"); 758 for (int N = 0; N < 4; ++N) 759 OS.EmitIntValue(FrontVer.Part[N], 2); 760 761 // Some Microsoft tools, like Binscope, expect a backend version number of at 762 // least 8.something, so we'll coerce the LLVM version into a form that 763 // guarantees it'll be big enough without really lying about the version. 764 int Major = 1000 * LLVM_VERSION_MAJOR + 765 10 * LLVM_VERSION_MINOR + 766 LLVM_VERSION_PATCH; 767 // Clamp it for builds that use unusually large version numbers. 768 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 769 Version BackVer = {{ Major, 0, 0, 0 }}; 770 OS.AddComment("Backend version"); 771 for (int N = 0; N < 4; ++N) 772 OS.EmitIntValue(BackVer.Part[N], 2); 773 774 OS.AddComment("Null-terminated compiler version string"); 775 emitNullTerminatedSymbolName(OS, CompilerVersion); 776 777 endSymbolRecord(CompilerEnd); 778 } 779 780 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 781 StringRef S) { 782 StringIdRecord SIR(TypeIndex(0x0), S); 783 return TypeTable.writeLeafType(SIR); 784 } 785 786 void CodeViewDebug::emitBuildInfo() { 787 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 788 // build info. The known prefix is: 789 // - Absolute path of current directory 790 // - Compiler path 791 // - Main source file path, relative to CWD or absolute 792 // - Type server PDB file 793 // - Canonical compiler command line 794 // If frontend and backend compilation are separated (think llc or LTO), it's 795 // not clear if the compiler path should refer to the executable for the 796 // frontend or the backend. Leave it blank for now. 797 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 798 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 799 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 800 const auto *CU = cast<DICompileUnit>(Node); 801 const DIFile *MainSourceFile = CU->getFile(); 802 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 803 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 804 BuildInfoArgs[BuildInfoRecord::SourceFile] = 805 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 806 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 807 // we implement /Zi type servers. 808 BuildInfoRecord BIR(BuildInfoArgs); 809 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 810 811 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 812 // from the module symbols into the type stream. 813 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 814 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 815 OS.AddComment("LF_BUILDINFO index"); 816 OS.EmitIntValue(BuildInfoIndex.getIndex(), 4); 817 endSymbolRecord(BIEnd); 818 endCVSubsection(BISubsecEnd); 819 } 820 821 void CodeViewDebug::emitInlineeLinesSubsection() { 822 if (InlinedSubprograms.empty()) 823 return; 824 825 OS.AddComment("Inlinee lines subsection"); 826 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 827 828 // We emit the checksum info for files. This is used by debuggers to 829 // determine if a pdb matches the source before loading it. Visual Studio, 830 // for instance, will display a warning that the breakpoints are not valid if 831 // the pdb does not match the source. 832 OS.AddComment("Inlinee lines signature"); 833 OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4); 834 835 for (const DISubprogram *SP : InlinedSubprograms) { 836 assert(TypeIndices.count({SP, nullptr})); 837 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 838 839 OS.AddBlankLine(); 840 unsigned FileId = maybeRecordFile(SP->getFile()); 841 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 842 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 843 OS.AddBlankLine(); 844 OS.AddComment("Type index of inlined function"); 845 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 846 OS.AddComment("Offset into filechecksum table"); 847 OS.EmitCVFileChecksumOffsetDirective(FileId); 848 OS.AddComment("Starting line number"); 849 OS.EmitIntValue(SP->getLine(), 4); 850 } 851 852 endCVSubsection(InlineEnd); 853 } 854 855 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 856 const DILocation *InlinedAt, 857 const InlineSite &Site) { 858 assert(TypeIndices.count({Site.Inlinee, nullptr})); 859 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 860 861 // SymbolRecord 862 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 863 864 OS.AddComment("PtrParent"); 865 OS.EmitIntValue(0, 4); 866 OS.AddComment("PtrEnd"); 867 OS.EmitIntValue(0, 4); 868 OS.AddComment("Inlinee type index"); 869 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 870 871 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 872 unsigned StartLineNum = Site.Inlinee->getLine(); 873 874 OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 875 FI.Begin, FI.End); 876 877 endSymbolRecord(InlineEnd); 878 879 emitLocalVariableList(FI, Site.InlinedLocals); 880 881 // Recurse on child inlined call sites before closing the scope. 882 for (const DILocation *ChildSite : Site.ChildSites) { 883 auto I = FI.InlineSites.find(ChildSite); 884 assert(I != FI.InlineSites.end() && 885 "child site not in function inline site map"); 886 emitInlinedCallSite(FI, ChildSite, I->second); 887 } 888 889 // Close the scope. 890 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 891 } 892 893 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 894 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 895 // comdat key. A section may be comdat because of -ffunction-sections or 896 // because it is comdat in the IR. 897 MCSectionCOFF *GVSec = 898 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 899 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 900 901 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 902 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 903 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 904 905 OS.SwitchSection(DebugSec); 906 907 // Emit the magic version number if this is the first time we've switched to 908 // this section. 909 if (ComdatDebugSections.insert(DebugSec).second) 910 emitCodeViewMagicVersion(); 911 } 912 913 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 914 // The only supported thunk ordinal is currently the standard type. 915 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 916 FunctionInfo &FI, 917 const MCSymbol *Fn) { 918 std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 919 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 920 921 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 922 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 923 924 // Emit S_THUNK32 925 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 926 OS.AddComment("PtrParent"); 927 OS.EmitIntValue(0, 4); 928 OS.AddComment("PtrEnd"); 929 OS.EmitIntValue(0, 4); 930 OS.AddComment("PtrNext"); 931 OS.EmitIntValue(0, 4); 932 OS.AddComment("Thunk section relative address"); 933 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 934 OS.AddComment("Thunk section index"); 935 OS.EmitCOFFSectionIndex(Fn); 936 OS.AddComment("Code size"); 937 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 938 OS.AddComment("Ordinal"); 939 OS.EmitIntValue(unsigned(ordinal), 1); 940 OS.AddComment("Function name"); 941 emitNullTerminatedSymbolName(OS, FuncName); 942 // Additional fields specific to the thunk ordinal would go here. 943 endSymbolRecord(ThunkRecordEnd); 944 945 // Local variables/inlined routines are purposely omitted here. The point of 946 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 947 948 // Emit S_PROC_ID_END 949 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 950 951 endCVSubsection(SymbolsEnd); 952 } 953 954 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 955 FunctionInfo &FI) { 956 // For each function there is a separate subsection which holds the PC to 957 // file:line table. 958 const MCSymbol *Fn = Asm->getSymbol(GV); 959 assert(Fn); 960 961 // Switch to the to a comdat section, if appropriate. 962 switchToDebugSectionForSymbol(Fn); 963 964 std::string FuncName; 965 auto *SP = GV->getSubprogram(); 966 assert(SP); 967 setCurrentSubprogram(SP); 968 969 if (SP->isThunk()) { 970 emitDebugInfoForThunk(GV, FI, Fn); 971 return; 972 } 973 974 // If we have a display name, build the fully qualified name by walking the 975 // chain of scopes. 976 if (!SP->getName().empty()) 977 FuncName = 978 getFullyQualifiedName(SP->getScope().resolve(), SP->getName()); 979 980 // If our DISubprogram name is empty, use the mangled name. 981 if (FuncName.empty()) 982 FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 983 984 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 985 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 986 OS.EmitCVFPOData(Fn); 987 988 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 989 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 990 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 991 { 992 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 993 : SymbolKind::S_GPROC32_ID; 994 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 995 996 // These fields are filled in by tools like CVPACK which run after the fact. 997 OS.AddComment("PtrParent"); 998 OS.EmitIntValue(0, 4); 999 OS.AddComment("PtrEnd"); 1000 OS.EmitIntValue(0, 4); 1001 OS.AddComment("PtrNext"); 1002 OS.EmitIntValue(0, 4); 1003 // This is the important bit that tells the debugger where the function 1004 // code is located and what's its size: 1005 OS.AddComment("Code size"); 1006 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1007 OS.AddComment("Offset after prologue"); 1008 OS.EmitIntValue(0, 4); 1009 OS.AddComment("Offset before epilogue"); 1010 OS.EmitIntValue(0, 4); 1011 OS.AddComment("Function type index"); 1012 OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4); 1013 OS.AddComment("Function section relative address"); 1014 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1015 OS.AddComment("Function section index"); 1016 OS.EmitCOFFSectionIndex(Fn); 1017 OS.AddComment("Flags"); 1018 OS.EmitIntValue(0, 1); 1019 // Emit the function display name as a null-terminated string. 1020 OS.AddComment("Function name"); 1021 // Truncate the name so we won't overflow the record length field. 1022 emitNullTerminatedSymbolName(OS, FuncName); 1023 endSymbolRecord(ProcRecordEnd); 1024 1025 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1026 // Subtract out the CSR size since MSVC excludes that and we include it. 1027 OS.AddComment("FrameSize"); 1028 OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4); 1029 OS.AddComment("Padding"); 1030 OS.EmitIntValue(0, 4); 1031 OS.AddComment("Offset of padding"); 1032 OS.EmitIntValue(0, 4); 1033 OS.AddComment("Bytes of callee saved registers"); 1034 OS.EmitIntValue(FI.CSRSize, 4); 1035 OS.AddComment("Exception handler offset"); 1036 OS.EmitIntValue(0, 4); 1037 OS.AddComment("Exception handler section"); 1038 OS.EmitIntValue(0, 2); 1039 OS.AddComment("Flags (defines frame register)"); 1040 OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4); 1041 endSymbolRecord(FrameProcEnd); 1042 1043 emitLocalVariableList(FI, FI.Locals); 1044 emitGlobalVariableList(FI.Globals); 1045 emitLexicalBlockList(FI.ChildBlocks, FI); 1046 1047 // Emit inlined call site information. Only emit functions inlined directly 1048 // into the parent function. We'll emit the other sites recursively as part 1049 // of their parent inline site. 1050 for (const DILocation *InlinedAt : FI.ChildSites) { 1051 auto I = FI.InlineSites.find(InlinedAt); 1052 assert(I != FI.InlineSites.end() && 1053 "child site not in function inline site map"); 1054 emitInlinedCallSite(FI, InlinedAt, I->second); 1055 } 1056 1057 for (auto Annot : FI.Annotations) { 1058 MCSymbol *Label = Annot.first; 1059 MDTuple *Strs = cast<MDTuple>(Annot.second); 1060 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1061 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1062 // FIXME: Make sure we don't overflow the max record size. 1063 OS.EmitCOFFSectionIndex(Label); 1064 OS.EmitIntValue(Strs->getNumOperands(), 2); 1065 for (Metadata *MD : Strs->operands()) { 1066 // MDStrings are null terminated, so we can do EmitBytes and get the 1067 // nice .asciz directive. 1068 StringRef Str = cast<MDString>(MD)->getString(); 1069 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1070 OS.EmitBytes(StringRef(Str.data(), Str.size() + 1)); 1071 } 1072 endSymbolRecord(AnnotEnd); 1073 } 1074 1075 if (SP != nullptr) 1076 emitDebugInfoForUDTs(LocalUDTs); 1077 1078 // We're done with this function. 1079 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1080 } 1081 endCVSubsection(SymbolsEnd); 1082 1083 // We have an assembler directive that takes care of the whole line table. 1084 OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1085 } 1086 1087 CodeViewDebug::LocalVarDefRange 1088 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1089 LocalVarDefRange DR; 1090 DR.InMemory = -1; 1091 DR.DataOffset = Offset; 1092 assert(DR.DataOffset == Offset && "truncation"); 1093 DR.IsSubfield = 0; 1094 DR.StructOffset = 0; 1095 DR.CVRegister = CVRegister; 1096 return DR; 1097 } 1098 1099 void CodeViewDebug::collectVariableInfoFromMFTable( 1100 DenseSet<InlinedEntity> &Processed) { 1101 const MachineFunction &MF = *Asm->MF; 1102 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1103 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1104 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1105 1106 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1107 if (!VI.Var) 1108 continue; 1109 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1110 "Expected inlined-at fields to agree"); 1111 1112 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1113 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1114 1115 // If variable scope is not found then skip this variable. 1116 if (!Scope) 1117 continue; 1118 1119 // If the variable has an attached offset expression, extract it. 1120 // FIXME: Try to handle DW_OP_deref as well. 1121 int64_t ExprOffset = 0; 1122 if (VI.Expr) 1123 if (!VI.Expr->extractIfOffset(ExprOffset)) 1124 continue; 1125 1126 // Get the frame register used and the offset. 1127 unsigned FrameReg = 0; 1128 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1129 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1130 1131 // Calculate the label ranges. 1132 LocalVarDefRange DefRange = 1133 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1134 for (const InsnRange &Range : Scope->getRanges()) { 1135 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1136 const MCSymbol *End = getLabelAfterInsn(Range.second); 1137 End = End ? End : Asm->getFunctionEnd(); 1138 DefRange.Ranges.emplace_back(Begin, End); 1139 } 1140 1141 LocalVariable Var; 1142 Var.DIVar = VI.Var; 1143 Var.DefRanges.emplace_back(std::move(DefRange)); 1144 recordLocalVariable(std::move(Var), Scope); 1145 } 1146 } 1147 1148 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1149 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1150 } 1151 1152 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1153 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1154 } 1155 1156 void CodeViewDebug::calculateRanges( 1157 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1158 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1159 1160 // Calculate the definition ranges. 1161 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1162 const auto &Entry = *I; 1163 if (!Entry.isDbgValue()) 1164 continue; 1165 const MachineInstr *DVInst = Entry.getInstr(); 1166 assert(DVInst->isDebugValue() && "Invalid History entry"); 1167 // FIXME: Find a way to represent constant variables, since they are 1168 // relatively common. 1169 Optional<DbgVariableLocation> Location = 1170 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1171 if (!Location) 1172 continue; 1173 1174 // CodeView can only express variables in register and variables in memory 1175 // at a constant offset from a register. However, for variables passed 1176 // indirectly by pointer, it is common for that pointer to be spilled to a 1177 // stack location. For the special case of one offseted load followed by a 1178 // zero offset load (a pointer spilled to the stack), we change the type of 1179 // the local variable from a value type to a reference type. This tricks the 1180 // debugger into doing the load for us. 1181 if (Var.UseReferenceType) { 1182 // We're using a reference type. Drop the last zero offset load. 1183 if (canUseReferenceType(*Location)) 1184 Location->LoadChain.pop_back(); 1185 else 1186 continue; 1187 } else if (needsReferenceType(*Location)) { 1188 // This location can't be expressed without switching to a reference type. 1189 // Start over using that. 1190 Var.UseReferenceType = true; 1191 Var.DefRanges.clear(); 1192 calculateRanges(Var, Entries); 1193 return; 1194 } 1195 1196 // We can only handle a register or an offseted load of a register. 1197 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1198 continue; 1199 { 1200 LocalVarDefRange DR; 1201 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1202 DR.InMemory = !Location->LoadChain.empty(); 1203 DR.DataOffset = 1204 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1205 if (Location->FragmentInfo) { 1206 DR.IsSubfield = true; 1207 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1208 } else { 1209 DR.IsSubfield = false; 1210 DR.StructOffset = 0; 1211 } 1212 1213 if (Var.DefRanges.empty() || 1214 Var.DefRanges.back().isDifferentLocation(DR)) { 1215 Var.DefRanges.emplace_back(std::move(DR)); 1216 } 1217 } 1218 1219 // Compute the label range. 1220 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1221 const MCSymbol *End; 1222 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1223 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1224 End = EndingEntry.isDbgValue() 1225 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1226 : getLabelAfterInsn(EndingEntry.getInstr()); 1227 } else 1228 End = Asm->getFunctionEnd(); 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 &Entries = 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, Entries); 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 && 1341 !GV.hasOptSize() && !GV.hasOptNone()) 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 = TypeIndex::Void(); 1837 if (ReturnAndArgs.size() > Index) { 1838 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 1839 } 1840 1841 // If the first argument is a pointer type and this isn't a static method, 1842 // treat it as the special 'this' parameter, which is encoded separately from 1843 // the arguments. 1844 TypeIndex ThisTypeIndex; 1845 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 1846 if (const DIDerivedType *PtrTy = 1847 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index].resolve())) { 1848 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 1849 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 1850 Index++; 1851 } 1852 } 1853 } 1854 1855 while (Index < ReturnAndArgs.size()) 1856 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 1857 1858 // MSVC uses type none for variadic argument. 1859 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 1860 ArgTypeIndices.back() = TypeIndex::None(); 1861 1862 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1863 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1864 1865 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1866 1867 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1868 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1869 return TypeTable.writeLeafType(MFR); 1870 } 1871 1872 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1873 unsigned VSlotCount = 1874 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1875 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1876 1877 VFTableShapeRecord VFTSR(Slots); 1878 return TypeTable.writeLeafType(VFTSR); 1879 } 1880 1881 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1882 switch (Flags & DINode::FlagAccessibility) { 1883 case DINode::FlagPrivate: return MemberAccess::Private; 1884 case DINode::FlagPublic: return MemberAccess::Public; 1885 case DINode::FlagProtected: return MemberAccess::Protected; 1886 case 0: 1887 // If there was no explicit access control, provide the default for the tag. 1888 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1889 : MemberAccess::Public; 1890 } 1891 llvm_unreachable("access flags are exclusive"); 1892 } 1893 1894 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1895 if (SP->isArtificial()) 1896 return MethodOptions::CompilerGenerated; 1897 1898 // FIXME: Handle other MethodOptions. 1899 1900 return MethodOptions::None; 1901 } 1902 1903 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 1904 bool Introduced) { 1905 if (SP->getFlags() & DINode::FlagStaticMember) 1906 return MethodKind::Static; 1907 1908 switch (SP->getVirtuality()) { 1909 case dwarf::DW_VIRTUALITY_none: 1910 break; 1911 case dwarf::DW_VIRTUALITY_virtual: 1912 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 1913 case dwarf::DW_VIRTUALITY_pure_virtual: 1914 return Introduced ? MethodKind::PureIntroducingVirtual 1915 : MethodKind::PureVirtual; 1916 default: 1917 llvm_unreachable("unhandled virtuality case"); 1918 } 1919 1920 return MethodKind::Vanilla; 1921 } 1922 1923 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 1924 switch (Ty->getTag()) { 1925 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 1926 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 1927 } 1928 llvm_unreachable("unexpected tag"); 1929 } 1930 1931 /// Return ClassOptions that should be present on both the forward declaration 1932 /// and the defintion of a tag type. 1933 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 1934 ClassOptions CO = ClassOptions::None; 1935 1936 // MSVC always sets this flag, even for local types. Clang doesn't always 1937 // appear to give every type a linkage name, which may be problematic for us. 1938 // FIXME: Investigate the consequences of not following them here. 1939 if (!Ty->getIdentifier().empty()) 1940 CO |= ClassOptions::HasUniqueName; 1941 1942 // Put the Nested flag on a type if it appears immediately inside a tag type. 1943 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 1944 // here. That flag is only set on definitions, and not forward declarations. 1945 const DIScope *ImmediateScope = Ty->getScope().resolve(); 1946 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 1947 CO |= ClassOptions::Nested; 1948 1949 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 1950 // type only when it has an immediate function scope. Clang never puts enums 1951 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 1952 // always in function, class, or file scopes. 1953 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 1954 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 1955 CO |= ClassOptions::Scoped; 1956 } else { 1957 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 1958 Scope = Scope->getScope().resolve()) { 1959 if (isa<DISubprogram>(Scope)) { 1960 CO |= ClassOptions::Scoped; 1961 break; 1962 } 1963 } 1964 } 1965 1966 return CO; 1967 } 1968 1969 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 1970 switch (Ty->getTag()) { 1971 case dwarf::DW_TAG_class_type: 1972 case dwarf::DW_TAG_structure_type: 1973 case dwarf::DW_TAG_union_type: 1974 case dwarf::DW_TAG_enumeration_type: 1975 break; 1976 default: 1977 return; 1978 } 1979 1980 if (const auto *File = Ty->getFile()) { 1981 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 1982 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 1983 1984 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 1985 TypeTable.writeLeafType(USLR); 1986 } 1987 } 1988 1989 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 1990 ClassOptions CO = getCommonClassOptions(Ty); 1991 TypeIndex FTI; 1992 unsigned EnumeratorCount = 0; 1993 1994 if (Ty->isForwardDecl()) { 1995 CO |= ClassOptions::ForwardReference; 1996 } else { 1997 ContinuationRecordBuilder ContinuationBuilder; 1998 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 1999 for (const DINode *Element : Ty->getElements()) { 2000 // We assume that the frontend provides all members in source declaration 2001 // order, which is what MSVC does. 2002 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2003 EnumeratorRecord ER(MemberAccess::Public, 2004 APSInt::getUnsigned(Enumerator->getValue()), 2005 Enumerator->getName()); 2006 ContinuationBuilder.writeMemberType(ER); 2007 EnumeratorCount++; 2008 } 2009 } 2010 FTI = TypeTable.insertRecord(ContinuationBuilder); 2011 } 2012 2013 std::string FullName = getFullyQualifiedName(Ty); 2014 2015 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2016 getTypeIndex(Ty->getBaseType())); 2017 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2018 2019 addUDTSrcLine(Ty, EnumTI); 2020 2021 return EnumTI; 2022 } 2023 2024 //===----------------------------------------------------------------------===// 2025 // ClassInfo 2026 //===----------------------------------------------------------------------===// 2027 2028 struct llvm::ClassInfo { 2029 struct MemberInfo { 2030 const DIDerivedType *MemberTypeNode; 2031 uint64_t BaseOffset; 2032 }; 2033 // [MemberInfo] 2034 using MemberList = std::vector<MemberInfo>; 2035 2036 using MethodsList = TinyPtrVector<const DISubprogram *>; 2037 // MethodName -> MethodsList 2038 using MethodsMap = MapVector<MDString *, MethodsList>; 2039 2040 /// Base classes. 2041 std::vector<const DIDerivedType *> Inheritance; 2042 2043 /// Direct members. 2044 MemberList Members; 2045 // Direct overloaded methods gathered by name. 2046 MethodsMap Methods; 2047 2048 TypeIndex VShapeTI; 2049 2050 std::vector<const DIType *> NestedTypes; 2051 }; 2052 2053 void CodeViewDebug::clear() { 2054 assert(CurFn == nullptr); 2055 FileIdMap.clear(); 2056 FnDebugInfo.clear(); 2057 FileToFilepathMap.clear(); 2058 LocalUDTs.clear(); 2059 GlobalUDTs.clear(); 2060 TypeIndices.clear(); 2061 CompleteTypeIndices.clear(); 2062 ScopeGlobals.clear(); 2063 } 2064 2065 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2066 const DIDerivedType *DDTy) { 2067 if (!DDTy->getName().empty()) { 2068 Info.Members.push_back({DDTy, 0}); 2069 return; 2070 } 2071 2072 // An unnamed member may represent a nested struct or union. Attempt to 2073 // interpret the unnamed member as a DICompositeType possibly wrapped in 2074 // qualifier types. Add all the indirect fields to the current record if that 2075 // succeeds, and drop the member if that fails. 2076 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2077 uint64_t Offset = DDTy->getOffsetInBits(); 2078 const DIType *Ty = DDTy->getBaseType().resolve(); 2079 bool FullyResolved = false; 2080 while (!FullyResolved) { 2081 switch (Ty->getTag()) { 2082 case dwarf::DW_TAG_const_type: 2083 case dwarf::DW_TAG_volatile_type: 2084 // FIXME: we should apply the qualifier types to the indirect fields 2085 // rather than dropping them. 2086 Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve(); 2087 break; 2088 default: 2089 FullyResolved = true; 2090 break; 2091 } 2092 } 2093 2094 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2095 if (!DCTy) 2096 return; 2097 2098 ClassInfo NestedInfo = collectClassInfo(DCTy); 2099 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2100 Info.Members.push_back( 2101 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2102 } 2103 2104 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2105 ClassInfo Info; 2106 // Add elements to structure type. 2107 DINodeArray Elements = Ty->getElements(); 2108 for (auto *Element : Elements) { 2109 // We assume that the frontend provides all members in source declaration 2110 // order, which is what MSVC does. 2111 if (!Element) 2112 continue; 2113 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2114 Info.Methods[SP->getRawName()].push_back(SP); 2115 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2116 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2117 collectMemberInfo(Info, DDTy); 2118 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2119 Info.Inheritance.push_back(DDTy); 2120 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2121 DDTy->getName() == "__vtbl_ptr_type") { 2122 Info.VShapeTI = getTypeIndex(DDTy); 2123 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2124 Info.NestedTypes.push_back(DDTy); 2125 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2126 // Ignore friend members. It appears that MSVC emitted info about 2127 // friends in the past, but modern versions do not. 2128 } 2129 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2130 Info.NestedTypes.push_back(Composite); 2131 } 2132 // Skip other unrecognized kinds of elements. 2133 } 2134 return Info; 2135 } 2136 2137 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2138 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2139 // if a complete type should be emitted instead of a forward reference. 2140 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2141 !Ty->isForwardDecl(); 2142 } 2143 2144 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2145 // Emit the complete type for unnamed structs. C++ classes with methods 2146 // which have a circular reference back to the class type are expected to 2147 // be named by the front-end and should not be "unnamed". C unnamed 2148 // structs should not have circular references. 2149 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2150 // If this unnamed complete type is already in the process of being defined 2151 // then the description of the type is malformed and cannot be emitted 2152 // into CodeView correctly so report a fatal error. 2153 auto I = CompleteTypeIndices.find(Ty); 2154 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2155 report_fatal_error("cannot debug circular reference to unnamed type"); 2156 return getCompleteTypeIndex(Ty); 2157 } 2158 2159 // First, construct the forward decl. Don't look into Ty to compute the 2160 // forward decl options, since it might not be available in all TUs. 2161 TypeRecordKind Kind = getRecordKind(Ty); 2162 ClassOptions CO = 2163 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2164 std::string FullName = getFullyQualifiedName(Ty); 2165 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2166 FullName, Ty->getIdentifier()); 2167 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2168 if (!Ty->isForwardDecl()) 2169 DeferredCompleteTypes.push_back(Ty); 2170 return FwdDeclTI; 2171 } 2172 2173 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2174 // Construct the field list and complete type record. 2175 TypeRecordKind Kind = getRecordKind(Ty); 2176 ClassOptions CO = getCommonClassOptions(Ty); 2177 TypeIndex FieldTI; 2178 TypeIndex VShapeTI; 2179 unsigned FieldCount; 2180 bool ContainsNestedClass; 2181 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2182 lowerRecordFieldList(Ty); 2183 2184 if (ContainsNestedClass) 2185 CO |= ClassOptions::ContainsNestedClass; 2186 2187 // MSVC appears to set this flag by searching any destructor or method with 2188 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2189 // the members, however special member functions are not yet emitted into 2190 // debug information. For now checking a class's non-triviality seems enough. 2191 // FIXME: not true for a nested unnamed struct. 2192 if (isNonTrivial(Ty)) 2193 CO |= ClassOptions::HasConstructorOrDestructor; 2194 2195 std::string FullName = getFullyQualifiedName(Ty); 2196 2197 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2198 2199 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2200 SizeInBytes, FullName, Ty->getIdentifier()); 2201 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2202 2203 addUDTSrcLine(Ty, ClassTI); 2204 2205 addToUDTs(Ty); 2206 2207 return ClassTI; 2208 } 2209 2210 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2211 // Emit the complete type for unnamed unions. 2212 if (shouldAlwaysEmitCompleteClassType(Ty)) 2213 return getCompleteTypeIndex(Ty); 2214 2215 ClassOptions CO = 2216 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2217 std::string FullName = getFullyQualifiedName(Ty); 2218 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2219 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2220 if (!Ty->isForwardDecl()) 2221 DeferredCompleteTypes.push_back(Ty); 2222 return FwdDeclTI; 2223 } 2224 2225 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2226 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2227 TypeIndex FieldTI; 2228 unsigned FieldCount; 2229 bool ContainsNestedClass; 2230 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2231 lowerRecordFieldList(Ty); 2232 2233 if (ContainsNestedClass) 2234 CO |= ClassOptions::ContainsNestedClass; 2235 2236 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2237 std::string FullName = getFullyQualifiedName(Ty); 2238 2239 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2240 Ty->getIdentifier()); 2241 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2242 2243 addUDTSrcLine(Ty, UnionTI); 2244 2245 addToUDTs(Ty); 2246 2247 return UnionTI; 2248 } 2249 2250 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2251 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2252 // Manually count members. MSVC appears to count everything that generates a 2253 // field list record. Each individual overload in a method overload group 2254 // contributes to this count, even though the overload group is a single field 2255 // list record. 2256 unsigned MemberCount = 0; 2257 ClassInfo Info = collectClassInfo(Ty); 2258 ContinuationRecordBuilder ContinuationBuilder; 2259 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2260 2261 // Create base classes. 2262 for (const DIDerivedType *I : Info.Inheritance) { 2263 if (I->getFlags() & DINode::FlagVirtual) { 2264 // Virtual base. 2265 unsigned VBPtrOffset = I->getVBPtrOffset(); 2266 // FIXME: Despite the accessor name, the offset is really in bytes. 2267 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2268 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2269 ? TypeRecordKind::IndirectVirtualBaseClass 2270 : TypeRecordKind::VirtualBaseClass; 2271 VirtualBaseClassRecord VBCR( 2272 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2273 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2274 VBTableIndex); 2275 2276 ContinuationBuilder.writeMemberType(VBCR); 2277 MemberCount++; 2278 } else { 2279 assert(I->getOffsetInBits() % 8 == 0 && 2280 "bases must be on byte boundaries"); 2281 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2282 getTypeIndex(I->getBaseType()), 2283 I->getOffsetInBits() / 8); 2284 ContinuationBuilder.writeMemberType(BCR); 2285 MemberCount++; 2286 } 2287 } 2288 2289 // Create members. 2290 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2291 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2292 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2293 StringRef MemberName = Member->getName(); 2294 MemberAccess Access = 2295 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2296 2297 if (Member->isStaticMember()) { 2298 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2299 ContinuationBuilder.writeMemberType(SDMR); 2300 MemberCount++; 2301 continue; 2302 } 2303 2304 // Virtual function pointer member. 2305 if ((Member->getFlags() & DINode::FlagArtificial) && 2306 Member->getName().startswith("_vptr$")) { 2307 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2308 ContinuationBuilder.writeMemberType(VFPR); 2309 MemberCount++; 2310 continue; 2311 } 2312 2313 // Data member. 2314 uint64_t MemberOffsetInBits = 2315 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2316 if (Member->isBitField()) { 2317 uint64_t StartBitOffset = MemberOffsetInBits; 2318 if (const auto *CI = 2319 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2320 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2321 } 2322 StartBitOffset -= MemberOffsetInBits; 2323 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2324 StartBitOffset); 2325 MemberBaseType = TypeTable.writeLeafType(BFR); 2326 } 2327 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2328 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2329 MemberName); 2330 ContinuationBuilder.writeMemberType(DMR); 2331 MemberCount++; 2332 } 2333 2334 // Create methods 2335 for (auto &MethodItr : Info.Methods) { 2336 StringRef Name = MethodItr.first->getString(); 2337 2338 std::vector<OneMethodRecord> Methods; 2339 for (const DISubprogram *SP : MethodItr.second) { 2340 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2341 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2342 2343 unsigned VFTableOffset = -1; 2344 if (Introduced) 2345 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2346 2347 Methods.push_back(OneMethodRecord( 2348 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2349 translateMethodKindFlags(SP, Introduced), 2350 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2351 MemberCount++; 2352 } 2353 assert(!Methods.empty() && "Empty methods map entry"); 2354 if (Methods.size() == 1) 2355 ContinuationBuilder.writeMemberType(Methods[0]); 2356 else { 2357 // FIXME: Make this use its own ContinuationBuilder so that 2358 // MethodOverloadList can be split correctly. 2359 MethodOverloadListRecord MOLR(Methods); 2360 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2361 2362 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2363 ContinuationBuilder.writeMemberType(OMR); 2364 } 2365 } 2366 2367 // Create nested classes. 2368 for (const DIType *Nested : Info.NestedTypes) { 2369 NestedTypeRecord R(getTypeIndex(DITypeRef(Nested)), Nested->getName()); 2370 ContinuationBuilder.writeMemberType(R); 2371 MemberCount++; 2372 } 2373 2374 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2375 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2376 !Info.NestedTypes.empty()); 2377 } 2378 2379 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2380 if (!VBPType.getIndex()) { 2381 // Make a 'const int *' type. 2382 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2383 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2384 2385 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2386 : PointerKind::Near32; 2387 PointerMode PM = PointerMode::Pointer; 2388 PointerOptions PO = PointerOptions::None; 2389 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2390 VBPType = TypeTable.writeLeafType(PR); 2391 } 2392 2393 return VBPType; 2394 } 2395 2396 TypeIndex CodeViewDebug::getTypeIndex(DITypeRef TypeRef, DITypeRef ClassTyRef) { 2397 const DIType *Ty = TypeRef.resolve(); 2398 const DIType *ClassTy = ClassTyRef.resolve(); 2399 2400 // The null DIType is the void type. Don't try to hash it. 2401 if (!Ty) 2402 return TypeIndex::Void(); 2403 2404 // Check if we've already translated this type. Don't try to do a 2405 // get-or-create style insertion that caches the hash lookup across the 2406 // lowerType call. It will update the TypeIndices map. 2407 auto I = TypeIndices.find({Ty, ClassTy}); 2408 if (I != TypeIndices.end()) 2409 return I->second; 2410 2411 TypeLoweringScope S(*this); 2412 TypeIndex TI = lowerType(Ty, ClassTy); 2413 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2414 } 2415 2416 codeview::TypeIndex 2417 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2418 const DISubroutineType *SubroutineTy) { 2419 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2420 "this type must be a pointer type"); 2421 2422 PointerOptions Options = PointerOptions::None; 2423 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2424 Options = PointerOptions::LValueRefThisPointer; 2425 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2426 Options = PointerOptions::RValueRefThisPointer; 2427 2428 // Check if we've already translated this type. If there is no ref qualifier 2429 // on the function then we look up this pointer type with no associated class 2430 // so that the TypeIndex for the this pointer can be shared with the type 2431 // index for other pointers to this class type. If there is a ref qualifier 2432 // then we lookup the pointer using the subroutine as the parent type. 2433 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2434 if (I != TypeIndices.end()) 2435 return I->second; 2436 2437 TypeLoweringScope S(*this); 2438 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2439 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2440 } 2441 2442 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(DITypeRef TypeRef) { 2443 DIType *Ty = TypeRef.resolve(); 2444 PointerRecord PR(getTypeIndex(Ty), 2445 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2446 : PointerKind::Near32, 2447 PointerMode::LValueReference, PointerOptions::None, 2448 Ty->getSizeInBits() / 8); 2449 return TypeTable.writeLeafType(PR); 2450 } 2451 2452 TypeIndex CodeViewDebug::getCompleteTypeIndex(DITypeRef TypeRef) { 2453 const DIType *Ty = TypeRef.resolve(); 2454 2455 // The null DIType is the void type. Don't try to hash it. 2456 if (!Ty) 2457 return TypeIndex::Void(); 2458 2459 // Look through typedefs when getting the complete type index. Call 2460 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2461 // emitted only once. 2462 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2463 (void)getTypeIndex(Ty); 2464 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2465 Ty = cast<DIDerivedType>(Ty)->getBaseType().resolve(); 2466 2467 // If this is a non-record type, the complete type index is the same as the 2468 // normal type index. Just call getTypeIndex. 2469 switch (Ty->getTag()) { 2470 case dwarf::DW_TAG_class_type: 2471 case dwarf::DW_TAG_structure_type: 2472 case dwarf::DW_TAG_union_type: 2473 break; 2474 default: 2475 return getTypeIndex(Ty); 2476 } 2477 2478 // Check if we've already translated the complete record type. 2479 const auto *CTy = cast<DICompositeType>(Ty); 2480 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2481 if (!InsertResult.second) 2482 return InsertResult.first->second; 2483 2484 TypeLoweringScope S(*this); 2485 2486 // Make sure the forward declaration is emitted first. It's unclear if this 2487 // is necessary, but MSVC does it, and we should follow suit until we can show 2488 // otherwise. 2489 // We only emit a forward declaration for named types. 2490 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2491 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2492 2493 // Just use the forward decl if we don't have complete type info. This 2494 // might happen if the frontend is using modules and expects the complete 2495 // definition to be emitted elsewhere. 2496 if (CTy->isForwardDecl()) 2497 return FwdDeclTI; 2498 } 2499 2500 TypeIndex TI; 2501 switch (CTy->getTag()) { 2502 case dwarf::DW_TAG_class_type: 2503 case dwarf::DW_TAG_structure_type: 2504 TI = lowerCompleteTypeClass(CTy); 2505 break; 2506 case dwarf::DW_TAG_union_type: 2507 TI = lowerCompleteTypeUnion(CTy); 2508 break; 2509 default: 2510 llvm_unreachable("not a record"); 2511 } 2512 2513 // Update the type index associated with this CompositeType. This cannot 2514 // use the 'InsertResult' iterator above because it is potentially 2515 // invalidated by map insertions which can occur while lowering the class 2516 // type above. 2517 CompleteTypeIndices[CTy] = TI; 2518 return TI; 2519 } 2520 2521 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2522 /// and do this until fixpoint, as each complete record type typically 2523 /// references 2524 /// many other record types. 2525 void CodeViewDebug::emitDeferredCompleteTypes() { 2526 SmallVector<const DICompositeType *, 4> TypesToEmit; 2527 while (!DeferredCompleteTypes.empty()) { 2528 std::swap(DeferredCompleteTypes, TypesToEmit); 2529 for (const DICompositeType *RecordTy : TypesToEmit) 2530 getCompleteTypeIndex(RecordTy); 2531 TypesToEmit.clear(); 2532 } 2533 } 2534 2535 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2536 ArrayRef<LocalVariable> Locals) { 2537 // Get the sorted list of parameters and emit them first. 2538 SmallVector<const LocalVariable *, 6> Params; 2539 for (const LocalVariable &L : Locals) 2540 if (L.DIVar->isParameter()) 2541 Params.push_back(&L); 2542 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2543 return L->DIVar->getArg() < R->DIVar->getArg(); 2544 }); 2545 for (const LocalVariable *L : Params) 2546 emitLocalVariable(FI, *L); 2547 2548 // Next emit all non-parameters in the order that we found them. 2549 for (const LocalVariable &L : Locals) 2550 if (!L.DIVar->isParameter()) 2551 emitLocalVariable(FI, L); 2552 } 2553 2554 /// Only call this on endian-specific types like ulittle16_t and little32_t, or 2555 /// structs composed of them. 2556 template <typename T> 2557 static void copyBytesForDefRange(SmallString<20> &BytePrefix, 2558 SymbolKind SymKind, const T &DefRangeHeader) { 2559 BytePrefix.resize(2 + sizeof(T)); 2560 ulittle16_t SymKindLE = ulittle16_t(SymKind); 2561 memcpy(&BytePrefix[0], &SymKindLE, 2); 2562 memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T)); 2563 } 2564 2565 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2566 const LocalVariable &Var) { 2567 // LocalSym record, see SymbolRecord.h for more info. 2568 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2569 2570 LocalSymFlags Flags = LocalSymFlags::None; 2571 if (Var.DIVar->isParameter()) 2572 Flags |= LocalSymFlags::IsParameter; 2573 if (Var.DefRanges.empty()) 2574 Flags |= LocalSymFlags::IsOptimizedOut; 2575 2576 OS.AddComment("TypeIndex"); 2577 TypeIndex TI = Var.UseReferenceType 2578 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2579 : getCompleteTypeIndex(Var.DIVar->getType()); 2580 OS.EmitIntValue(TI.getIndex(), 4); 2581 OS.AddComment("Flags"); 2582 OS.EmitIntValue(static_cast<uint16_t>(Flags), 2); 2583 // Truncate the name so we won't overflow the record length field. 2584 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2585 endSymbolRecord(LocalEnd); 2586 2587 // Calculate the on disk prefix of the appropriate def range record. The 2588 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2589 // should be big enough to hold all forms without memory allocation. 2590 SmallString<20> BytePrefix; 2591 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2592 BytePrefix.clear(); 2593 if (DefRange.InMemory) { 2594 int Offset = DefRange.DataOffset; 2595 unsigned Reg = DefRange.CVRegister; 2596 2597 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2598 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2599 // instead. In frames without stack realignment, $T0 will be the CFA. 2600 if (RegisterId(Reg) == RegisterId::ESP) { 2601 Reg = unsigned(RegisterId::VFRAME); 2602 Offset += FI.OffsetAdjustment; 2603 } 2604 2605 // If we can use the chosen frame pointer for the frame and this isn't a 2606 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2607 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2608 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2609 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2610 (bool(Flags & LocalSymFlags::IsParameter) 2611 ? (EncFP == FI.EncodedParamFramePtrReg) 2612 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2613 little32_t FPOffset = little32_t(Offset); 2614 copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset); 2615 } else { 2616 uint16_t RegRelFlags = 0; 2617 if (DefRange.IsSubfield) { 2618 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2619 (DefRange.StructOffset 2620 << DefRangeRegisterRelSym::OffsetInParentShift); 2621 } 2622 DefRangeRegisterRelSym::Header DRHdr; 2623 DRHdr.Register = Reg; 2624 DRHdr.Flags = RegRelFlags; 2625 DRHdr.BasePointerOffset = Offset; 2626 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr); 2627 } 2628 } else { 2629 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2630 if (DefRange.IsSubfield) { 2631 DefRangeSubfieldRegisterSym::Header DRHdr; 2632 DRHdr.Register = DefRange.CVRegister; 2633 DRHdr.MayHaveNoName = 0; 2634 DRHdr.OffsetInParent = DefRange.StructOffset; 2635 copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr); 2636 } else { 2637 DefRangeRegisterSym::Header DRHdr; 2638 DRHdr.Register = DefRange.CVRegister; 2639 DRHdr.MayHaveNoName = 0; 2640 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr); 2641 } 2642 } 2643 OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix); 2644 } 2645 } 2646 2647 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2648 const FunctionInfo& FI) { 2649 for (LexicalBlock *Block : Blocks) 2650 emitLexicalBlock(*Block, FI); 2651 } 2652 2653 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2654 /// lexical block scope. 2655 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2656 const FunctionInfo& FI) { 2657 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2658 OS.AddComment("PtrParent"); 2659 OS.EmitIntValue(0, 4); // PtrParent 2660 OS.AddComment("PtrEnd"); 2661 OS.EmitIntValue(0, 4); // PtrEnd 2662 OS.AddComment("Code size"); 2663 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2664 OS.AddComment("Function section relative address"); 2665 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2666 OS.AddComment("Function section index"); 2667 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2668 OS.AddComment("Lexical block name"); 2669 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2670 endSymbolRecord(RecordEnd); 2671 2672 // Emit variables local to this lexical block. 2673 emitLocalVariableList(FI, Block.Locals); 2674 emitGlobalVariableList(Block.Globals); 2675 2676 // Emit lexical blocks contained within this block. 2677 emitLexicalBlockList(Block.Children, FI); 2678 2679 // Close the lexical block scope. 2680 emitEndSymbolRecord(SymbolKind::S_END); 2681 } 2682 2683 /// Convenience routine for collecting lexical block information for a list 2684 /// of lexical scopes. 2685 void CodeViewDebug::collectLexicalBlockInfo( 2686 SmallVectorImpl<LexicalScope *> &Scopes, 2687 SmallVectorImpl<LexicalBlock *> &Blocks, 2688 SmallVectorImpl<LocalVariable> &Locals, 2689 SmallVectorImpl<CVGlobalVariable> &Globals) { 2690 for (LexicalScope *Scope : Scopes) 2691 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2692 } 2693 2694 /// Populate the lexical blocks and local variable lists of the parent with 2695 /// information about the specified lexical scope. 2696 void CodeViewDebug::collectLexicalBlockInfo( 2697 LexicalScope &Scope, 2698 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2699 SmallVectorImpl<LocalVariable> &ParentLocals, 2700 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2701 if (Scope.isAbstractScope()) 2702 return; 2703 2704 // Gather information about the lexical scope including local variables, 2705 // global variables, and address ranges. 2706 bool IgnoreScope = false; 2707 auto LI = ScopeVariables.find(&Scope); 2708 SmallVectorImpl<LocalVariable> *Locals = 2709 LI != ScopeVariables.end() ? &LI->second : nullptr; 2710 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2711 SmallVectorImpl<CVGlobalVariable> *Globals = 2712 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2713 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2714 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2715 2716 // Ignore lexical scopes which do not contain variables. 2717 if (!Locals && !Globals) 2718 IgnoreScope = true; 2719 2720 // Ignore lexical scopes which are not lexical blocks. 2721 if (!DILB) 2722 IgnoreScope = true; 2723 2724 // Ignore scopes which have too many address ranges to represent in the 2725 // current CodeView format or do not have a valid address range. 2726 // 2727 // For lexical scopes with multiple address ranges you may be tempted to 2728 // construct a single range covering every instruction where the block is 2729 // live and everything in between. Unfortunately, Visual Studio only 2730 // displays variables from the first matching lexical block scope. If the 2731 // first lexical block contains exception handling code or cold code which 2732 // is moved to the bottom of the routine creating a single range covering 2733 // nearly the entire routine, then it will hide all other lexical blocks 2734 // and the variables they contain. 2735 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2736 IgnoreScope = true; 2737 2738 if (IgnoreScope) { 2739 // This scope can be safely ignored and eliminating it will reduce the 2740 // size of the debug information. Be sure to collect any variable and scope 2741 // information from the this scope or any of its children and collapse them 2742 // into the parent scope. 2743 if (Locals) 2744 ParentLocals.append(Locals->begin(), Locals->end()); 2745 if (Globals) 2746 ParentGlobals.append(Globals->begin(), Globals->end()); 2747 collectLexicalBlockInfo(Scope.getChildren(), 2748 ParentBlocks, 2749 ParentLocals, 2750 ParentGlobals); 2751 return; 2752 } 2753 2754 // Create a new CodeView lexical block for this lexical scope. If we've 2755 // seen this DILexicalBlock before then the scope tree is malformed and 2756 // we can handle this gracefully by not processing it a second time. 2757 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2758 if (!BlockInsertion.second) 2759 return; 2760 2761 // Create a lexical block containing the variables and collect the the 2762 // lexical block information for the children. 2763 const InsnRange &Range = Ranges.front(); 2764 assert(Range.first && Range.second); 2765 LexicalBlock &Block = BlockInsertion.first->second; 2766 Block.Begin = getLabelBeforeInsn(Range.first); 2767 Block.End = getLabelAfterInsn(Range.second); 2768 assert(Block.Begin && "missing label for scope begin"); 2769 assert(Block.End && "missing label for scope end"); 2770 Block.Name = DILB->getName(); 2771 if (Locals) 2772 Block.Locals = std::move(*Locals); 2773 if (Globals) 2774 Block.Globals = std::move(*Globals); 2775 ParentBlocks.push_back(&Block); 2776 collectLexicalBlockInfo(Scope.getChildren(), 2777 Block.Children, 2778 Block.Locals, 2779 Block.Globals); 2780 } 2781 2782 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2783 const Function &GV = MF->getFunction(); 2784 assert(FnDebugInfo.count(&GV)); 2785 assert(CurFn == FnDebugInfo[&GV].get()); 2786 2787 collectVariableInfo(GV.getSubprogram()); 2788 2789 // Build the lexical block structure to emit for this routine. 2790 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2791 collectLexicalBlockInfo(*CFS, 2792 CurFn->ChildBlocks, 2793 CurFn->Locals, 2794 CurFn->Globals); 2795 2796 // Clear the scope and variable information from the map which will not be 2797 // valid after we have finished processing this routine. This also prepares 2798 // the map for the subsequent routine. 2799 ScopeVariables.clear(); 2800 2801 // Don't emit anything if we don't have any line tables. 2802 // Thunks are compiler-generated and probably won't have source correlation. 2803 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2804 FnDebugInfo.erase(&GV); 2805 CurFn = nullptr; 2806 return; 2807 } 2808 2809 CurFn->Annotations = MF->getCodeViewAnnotations(); 2810 2811 CurFn->End = Asm->getFunctionEnd(); 2812 2813 CurFn = nullptr; 2814 } 2815 2816 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2817 DebugHandlerBase::beginInstruction(MI); 2818 2819 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2820 if (!Asm || !CurFn || MI->isDebugInstr() || 2821 MI->getFlag(MachineInstr::FrameSetup)) 2822 return; 2823 2824 // If the first instruction of a new MBB has no location, find the first 2825 // instruction with a location and use that. 2826 DebugLoc DL = MI->getDebugLoc(); 2827 if (!DL && MI->getParent() != PrevInstBB) { 2828 for (const auto &NextMI : *MI->getParent()) { 2829 if (NextMI.isDebugInstr()) 2830 continue; 2831 DL = NextMI.getDebugLoc(); 2832 if (DL) 2833 break; 2834 } 2835 } 2836 PrevInstBB = MI->getParent(); 2837 2838 // If we still don't have a debug location, don't record a location. 2839 if (!DL) 2840 return; 2841 2842 maybeRecordLocation(DL, Asm->MF); 2843 } 2844 2845 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2846 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2847 *EndLabel = MMI->getContext().createTempSymbol(); 2848 OS.EmitIntValue(unsigned(Kind), 4); 2849 OS.AddComment("Subsection size"); 2850 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2851 OS.EmitLabel(BeginLabel); 2852 return EndLabel; 2853 } 2854 2855 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2856 OS.EmitLabel(EndLabel); 2857 // Every subsection must be aligned to a 4-byte boundary. 2858 OS.EmitValueToAlignment(4); 2859 } 2860 2861 static StringRef getSymbolName(SymbolKind SymKind) { 2862 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 2863 if (EE.Value == SymKind) 2864 return EE.Name; 2865 return ""; 2866 } 2867 2868 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 2869 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2870 *EndLabel = MMI->getContext().createTempSymbol(); 2871 OS.AddComment("Record length"); 2872 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 2873 OS.EmitLabel(BeginLabel); 2874 if (OS.isVerboseAsm()) 2875 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 2876 OS.EmitIntValue(unsigned(SymKind), 2); 2877 return EndLabel; 2878 } 2879 2880 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 2881 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 2882 // an extra copy of every symbol record in LLD. This increases object file 2883 // size by less than 1% in the clang build, and is compatible with the Visual 2884 // C++ linker. 2885 OS.EmitValueToAlignment(4); 2886 OS.EmitLabel(SymEnd); 2887 } 2888 2889 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 2890 OS.AddComment("Record length"); 2891 OS.EmitIntValue(2, 2); 2892 if (OS.isVerboseAsm()) 2893 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 2894 OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind 2895 } 2896 2897 void CodeViewDebug::emitDebugInfoForUDTs( 2898 ArrayRef<std::pair<std::string, const DIType *>> UDTs) { 2899 for (const auto &UDT : UDTs) { 2900 const DIType *T = UDT.second; 2901 assert(shouldEmitUdt(T)); 2902 2903 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 2904 OS.AddComment("Type"); 2905 OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4); 2906 emitNullTerminatedSymbolName(OS, UDT.first); 2907 endSymbolRecord(UDTRecordEnd); 2908 } 2909 } 2910 2911 void CodeViewDebug::collectGlobalVariableInfo() { 2912 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 2913 GlobalMap; 2914 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 2915 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 2916 GV.getDebugInfo(GVEs); 2917 for (const auto *GVE : GVEs) 2918 GlobalMap[GVE] = &GV; 2919 } 2920 2921 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2922 for (const MDNode *Node : CUs->operands()) { 2923 const auto *CU = cast<DICompileUnit>(Node); 2924 for (const auto *GVE : CU->getGlobalVariables()) { 2925 const auto *GV = GlobalMap.lookup(GVE); 2926 if (!GV || GV->isDeclarationForLinker()) 2927 continue; 2928 const DIGlobalVariable *DIGV = GVE->getVariable(); 2929 DIScope *Scope = DIGV->getScope(); 2930 SmallVector<CVGlobalVariable, 1> *VariableList; 2931 if (Scope && isa<DILocalScope>(Scope)) { 2932 // Locate a global variable list for this scope, creating one if 2933 // necessary. 2934 auto Insertion = ScopeGlobals.insert( 2935 {Scope, std::unique_ptr<GlobalVariableList>()}); 2936 if (Insertion.second) 2937 Insertion.first->second = llvm::make_unique<GlobalVariableList>(); 2938 VariableList = Insertion.first->second.get(); 2939 } else if (GV->hasComdat()) 2940 // Emit this global variable into a COMDAT section. 2941 VariableList = &ComdatVariables; 2942 else 2943 // Emit this globla variable in a single global symbol section. 2944 VariableList = &GlobalVariables; 2945 CVGlobalVariable CVGV = {DIGV, GV}; 2946 VariableList->emplace_back(std::move(CVGV)); 2947 } 2948 } 2949 } 2950 2951 void CodeViewDebug::emitDebugInfoForGlobals() { 2952 // First, emit all globals that are not in a comdat in a single symbol 2953 // substream. MSVC doesn't like it if the substream is empty, so only open 2954 // it if we have at least one global to emit. 2955 switchToDebugSectionForSymbol(nullptr); 2956 if (!GlobalVariables.empty()) { 2957 OS.AddComment("Symbol subsection for globals"); 2958 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 2959 emitGlobalVariableList(GlobalVariables); 2960 endCVSubsection(EndLabel); 2961 } 2962 2963 // Second, emit each global that is in a comdat into its own .debug$S 2964 // section along with its own symbol substream. 2965 for (const CVGlobalVariable &CVGV : ComdatVariables) { 2966 MCSymbol *GVSym = Asm->getSymbol(CVGV.GV); 2967 OS.AddComment("Symbol subsection for " + 2968 Twine(GlobalValue::dropLLVMManglingEscape(CVGV.GV->getName()))); 2969 switchToDebugSectionForSymbol(GVSym); 2970 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 2971 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 2972 emitDebugInfoForGlobal(CVGV.DIGV, CVGV.GV, GVSym); 2973 endCVSubsection(EndLabel); 2974 } 2975 } 2976 2977 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 2978 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2979 for (const MDNode *Node : CUs->operands()) { 2980 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 2981 if (DIType *RT = dyn_cast<DIType>(Ty)) { 2982 getTypeIndex(RT); 2983 // FIXME: Add to global/local DTU list. 2984 } 2985 } 2986 } 2987 } 2988 2989 // Emit each global variable in the specified array. 2990 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 2991 for (const CVGlobalVariable &CVGV : Globals) { 2992 MCSymbol *GVSym = Asm->getSymbol(CVGV.GV); 2993 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 2994 emitDebugInfoForGlobal(CVGV.DIGV, CVGV.GV, GVSym); 2995 } 2996 } 2997 2998 void CodeViewDebug::emitDebugInfoForGlobal(const DIGlobalVariable *DIGV, 2999 const GlobalVariable *GV, 3000 MCSymbol *GVSym) { 3001 // DataSym record, see SymbolRecord.h for more info. Thread local data 3002 // happens to have the same format as global data. 3003 SymbolKind DataSym = GV->isThreadLocal() 3004 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3005 : SymbolKind::S_GTHREAD32) 3006 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3007 : SymbolKind::S_GDATA32); 3008 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3009 OS.AddComment("Type"); 3010 OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4); 3011 OS.AddComment("DataOffset"); 3012 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 3013 OS.AddComment("Segment"); 3014 OS.EmitCOFFSectionIndex(GVSym); 3015 OS.AddComment("Name"); 3016 const unsigned LengthOfDataRecord = 12; 3017 emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord); 3018 endSymbolRecord(DataEnd); 3019 } 3020