1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "DwarfExpression.h" 15 #include "llvm/ADT/APSInt.h" 16 #include "llvm/ADT/ArrayRef.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/DenseSet.h" 19 #include "llvm/ADT/MapVector.h" 20 #include "llvm/ADT/None.h" 21 #include "llvm/ADT/Optional.h" 22 #include "llvm/ADT/STLExtras.h" 23 #include "llvm/ADT/SmallString.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/StringRef.h" 26 #include "llvm/ADT/TinyPtrVector.h" 27 #include "llvm/ADT/Triple.h" 28 #include "llvm/ADT/Twine.h" 29 #include "llvm/BinaryFormat/COFF.h" 30 #include "llvm/BinaryFormat/Dwarf.h" 31 #include "llvm/CodeGen/AsmPrinter.h" 32 #include "llvm/CodeGen/LexicalScopes.h" 33 #include "llvm/CodeGen/MachineFrameInfo.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineInstr.h" 36 #include "llvm/CodeGen/MachineModuleInfo.h" 37 #include "llvm/CodeGen/MachineOperand.h" 38 #include "llvm/CodeGen/TargetFrameLowering.h" 39 #include "llvm/CodeGen/TargetRegisterInfo.h" 40 #include "llvm/CodeGen/TargetSubtargetInfo.h" 41 #include "llvm/Config/llvm-config.h" 42 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 43 #include "llvm/DebugInfo/CodeView/CodeView.h" 44 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 45 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 46 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 47 #include "llvm/DebugInfo/CodeView/EnumTables.h" 48 #include "llvm/DebugInfo/CodeView/Line.h" 49 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 50 #include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h" 51 #include "llvm/DebugInfo/CodeView/TypeIndex.h" 52 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 53 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 54 #include "llvm/IR/Constants.h" 55 #include "llvm/IR/DataLayout.h" 56 #include "llvm/IR/DebugInfoMetadata.h" 57 #include "llvm/IR/DebugLoc.h" 58 #include "llvm/IR/Function.h" 59 #include "llvm/IR/GlobalValue.h" 60 #include "llvm/IR/GlobalVariable.h" 61 #include "llvm/IR/Metadata.h" 62 #include "llvm/IR/Module.h" 63 #include "llvm/MC/MCAsmInfo.h" 64 #include "llvm/MC/MCContext.h" 65 #include "llvm/MC/MCSectionCOFF.h" 66 #include "llvm/MC/MCStreamer.h" 67 #include "llvm/MC/MCSymbol.h" 68 #include "llvm/Support/BinaryByteStream.h" 69 #include "llvm/Support/BinaryStreamReader.h" 70 #include "llvm/Support/BinaryStreamWriter.h" 71 #include "llvm/Support/Casting.h" 72 #include "llvm/Support/CommandLine.h" 73 #include "llvm/Support/Compiler.h" 74 #include "llvm/Support/Endian.h" 75 #include "llvm/Support/Error.h" 76 #include "llvm/Support/ErrorHandling.h" 77 #include "llvm/Support/FormatVariadic.h" 78 #include "llvm/Support/Path.h" 79 #include "llvm/Support/SMLoc.h" 80 #include "llvm/Support/ScopedPrinter.h" 81 #include "llvm/Target/TargetLoweringObjectFile.h" 82 #include "llvm/Target/TargetMachine.h" 83 #include <algorithm> 84 #include <cassert> 85 #include <cctype> 86 #include <cstddef> 87 #include <cstdint> 88 #include <iterator> 89 #include <limits> 90 #include <string> 91 #include <utility> 92 #include <vector> 93 94 using namespace llvm; 95 using namespace llvm::codeview; 96 97 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 98 switch (Type) { 99 case Triple::ArchType::x86: 100 return CPUType::Pentium3; 101 case Triple::ArchType::x86_64: 102 return CPUType::X64; 103 case Triple::ArchType::thumb: 104 return CPUType::Thumb; 105 case Triple::ArchType::aarch64: 106 return CPUType::ARM64; 107 default: 108 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 109 } 110 } 111 112 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 113 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 114 // If module doesn't have named metadata anchors or COFF debug section 115 // is not available, skip any debug info related stuff. 116 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 117 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 118 Asm = nullptr; 119 MMI->setDebugInfoAvailability(false); 120 return; 121 } 122 // Tell MMI that we have debug info. 123 MMI->setDebugInfoAvailability(true); 124 125 TheCPU = 126 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 127 128 collectGlobalVariableInfo(); 129 130 // Check if we should emit type record hashes. 131 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 132 MMI->getModule()->getModuleFlag("CodeViewGHash")); 133 EmitDebugGlobalHashes = GH && !GH->isZero(); 134 } 135 136 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 137 std::string &Filepath = FileToFilepathMap[File]; 138 if (!Filepath.empty()) 139 return Filepath; 140 141 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 142 143 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 144 // it textually because one of the path components could be a symlink. 145 if (Dir.startswith("/") || Filename.startswith("/")) { 146 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 147 return Filename; 148 Filepath = Dir; 149 if (Dir.back() != '/') 150 Filepath += '/'; 151 Filepath += Filename; 152 return Filepath; 153 } 154 155 // Clang emits directory and relative filename info into the IR, but CodeView 156 // operates on full paths. We could change Clang to emit full paths too, but 157 // that would increase the IR size and probably not needed for other users. 158 // For now, just concatenate and canonicalize the path here. 159 if (Filename.find(':') == 1) 160 Filepath = Filename; 161 else 162 Filepath = (Dir + "\\" + Filename).str(); 163 164 // Canonicalize the path. We have to do it textually because we may no longer 165 // have access the file in the filesystem. 166 // First, replace all slashes with backslashes. 167 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 168 169 // Remove all "\.\" with "\". 170 size_t Cursor = 0; 171 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 172 Filepath.erase(Cursor, 2); 173 174 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 175 // path should be well-formatted, e.g. start with a drive letter, etc. 176 Cursor = 0; 177 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 178 // Something's wrong if the path starts with "\..\", abort. 179 if (Cursor == 0) 180 break; 181 182 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 183 if (PrevSlash == std::string::npos) 184 // Something's wrong, abort. 185 break; 186 187 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 188 // The next ".." might be following the one we've just erased. 189 Cursor = PrevSlash; 190 } 191 192 // Remove all duplicate backslashes. 193 Cursor = 0; 194 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 195 Filepath.erase(Cursor, 1); 196 197 return Filepath; 198 } 199 200 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 201 StringRef FullPath = getFullFilepath(F); 202 unsigned NextId = FileIdMap.size() + 1; 203 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 204 if (Insertion.second) { 205 // We have to compute the full filepath and emit a .cv_file directive. 206 ArrayRef<uint8_t> ChecksumAsBytes; 207 FileChecksumKind CSKind = FileChecksumKind::None; 208 if (F->getChecksum()) { 209 std::string Checksum = fromHex(F->getChecksum()->Value); 210 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 211 memcpy(CKMem, Checksum.data(), Checksum.size()); 212 ChecksumAsBytes = ArrayRef<uint8_t>( 213 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 214 switch (F->getChecksum()->Kind) { 215 case DIFile::CSK_MD5: CSKind = FileChecksumKind::MD5; break; 216 case DIFile::CSK_SHA1: CSKind = FileChecksumKind::SHA1; break; 217 } 218 } 219 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 220 static_cast<unsigned>(CSKind)); 221 (void)Success; 222 assert(Success && ".cv_file directive failed"); 223 } 224 return Insertion.first->second; 225 } 226 227 CodeViewDebug::InlineSite & 228 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 229 const DISubprogram *Inlinee) { 230 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 231 InlineSite *Site = &SiteInsertion.first->second; 232 if (SiteInsertion.second) { 233 unsigned ParentFuncId = CurFn->FuncId; 234 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 235 ParentFuncId = 236 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 237 .SiteFuncId; 238 239 Site->SiteFuncId = NextFuncId++; 240 OS.EmitCVInlineSiteIdDirective( 241 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 242 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 243 Site->Inlinee = Inlinee; 244 InlinedSubprograms.insert(Inlinee); 245 getFuncIdForSubprogram(Inlinee); 246 } 247 return *Site; 248 } 249 250 static StringRef getPrettyScopeName(const DIScope *Scope) { 251 StringRef ScopeName = Scope->getName(); 252 if (!ScopeName.empty()) 253 return ScopeName; 254 255 switch (Scope->getTag()) { 256 case dwarf::DW_TAG_enumeration_type: 257 case dwarf::DW_TAG_class_type: 258 case dwarf::DW_TAG_structure_type: 259 case dwarf::DW_TAG_union_type: 260 return "<unnamed-tag>"; 261 case dwarf::DW_TAG_namespace: 262 return "`anonymous namespace'"; 263 } 264 265 return StringRef(); 266 } 267 268 static const DISubprogram *getQualifiedNameComponents( 269 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 270 const DISubprogram *ClosestSubprogram = nullptr; 271 while (Scope != nullptr) { 272 if (ClosestSubprogram == nullptr) 273 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 274 StringRef ScopeName = getPrettyScopeName(Scope); 275 if (!ScopeName.empty()) 276 QualifiedNameComponents.push_back(ScopeName); 277 Scope = Scope->getScope(); 278 } 279 return ClosestSubprogram; 280 } 281 282 static std::string getQualifiedName(ArrayRef<StringRef> QualifiedNameComponents, 283 StringRef TypeName) { 284 std::string FullyQualifiedName; 285 for (StringRef QualifiedNameComponent : 286 llvm::reverse(QualifiedNameComponents)) { 287 FullyQualifiedName.append(QualifiedNameComponent); 288 FullyQualifiedName.append("::"); 289 } 290 FullyQualifiedName.append(TypeName); 291 return FullyQualifiedName; 292 } 293 294 static std::string getFullyQualifiedName(const DIScope *Scope, StringRef Name) { 295 SmallVector<StringRef, 5> QualifiedNameComponents; 296 getQualifiedNameComponents(Scope, QualifiedNameComponents); 297 return getQualifiedName(QualifiedNameComponents, Name); 298 } 299 300 struct CodeViewDebug::TypeLoweringScope { 301 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 302 ~TypeLoweringScope() { 303 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 304 // inner TypeLoweringScopes don't attempt to emit deferred types. 305 if (CVD.TypeEmissionLevel == 1) 306 CVD.emitDeferredCompleteTypes(); 307 --CVD.TypeEmissionLevel; 308 } 309 CodeViewDebug &CVD; 310 }; 311 312 static std::string getFullyQualifiedName(const DIScope *Ty) { 313 const DIScope *Scope = Ty->getScope(); 314 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 315 } 316 317 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 318 // No scope means global scope and that uses the zero index. 319 if (!Scope || isa<DIFile>(Scope)) 320 return TypeIndex(); 321 322 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 323 324 // Check if we've already translated this scope. 325 auto I = TypeIndices.find({Scope, nullptr}); 326 if (I != TypeIndices.end()) 327 return I->second; 328 329 // Build the fully qualified name of the scope. 330 std::string ScopeName = getFullyQualifiedName(Scope); 331 StringIdRecord SID(TypeIndex(), ScopeName); 332 auto TI = TypeTable.writeLeafType(SID); 333 return recordTypeIndexForDINode(Scope, TI); 334 } 335 336 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 337 assert(SP); 338 339 // Check if we've already translated this subprogram. 340 auto I = TypeIndices.find({SP, nullptr}); 341 if (I != TypeIndices.end()) 342 return I->second; 343 344 // The display name includes function template arguments. Drop them to match 345 // MSVC. 346 StringRef DisplayName = SP->getName().split('<').first; 347 348 const DIScope *Scope = SP->getScope(); 349 TypeIndex TI; 350 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 351 // If the scope is a DICompositeType, then this must be a method. Member 352 // function types take some special handling, and require access to the 353 // subprogram. 354 TypeIndex ClassType = getTypeIndex(Class); 355 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 356 DisplayName); 357 TI = TypeTable.writeLeafType(MFuncId); 358 } else { 359 // Otherwise, this must be a free function. 360 TypeIndex ParentScope = getScopeIndex(Scope); 361 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 362 TI = TypeTable.writeLeafType(FuncId); 363 } 364 365 return recordTypeIndexForDINode(SP, TI); 366 } 367 368 static bool isNonTrivial(const DICompositeType *DCTy) { 369 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 370 } 371 372 static FunctionOptions 373 getFunctionOptions(const DISubroutineType *Ty, 374 const DICompositeType *ClassTy = nullptr, 375 StringRef SPName = StringRef("")) { 376 FunctionOptions FO = FunctionOptions::None; 377 const DIType *ReturnTy = nullptr; 378 if (auto TypeArray = Ty->getTypeArray()) { 379 if (TypeArray.size()) 380 ReturnTy = TypeArray[0]; 381 } 382 383 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) { 384 if (isNonTrivial(ReturnDCTy)) 385 FO |= FunctionOptions::CxxReturnUdt; 386 } 387 388 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 389 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 390 FO |= FunctionOptions::Constructor; 391 392 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 393 394 } 395 return FO; 396 } 397 398 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 399 const DICompositeType *Class) { 400 // Always use the method declaration as the key for the function type. The 401 // method declaration contains the this adjustment. 402 if (SP->getDeclaration()) 403 SP = SP->getDeclaration(); 404 assert(!SP->getDeclaration() && "should use declaration as key"); 405 406 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 407 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 408 auto I = TypeIndices.find({SP, Class}); 409 if (I != TypeIndices.end()) 410 return I->second; 411 412 // Make sure complete type info for the class is emitted *after* the member 413 // function type, as the complete class type is likely to reference this 414 // member function type. 415 TypeLoweringScope S(*this); 416 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 417 418 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 419 TypeIndex TI = lowerTypeMemberFunction( 420 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 421 return recordTypeIndexForDINode(SP, TI, Class); 422 } 423 424 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 425 TypeIndex TI, 426 const DIType *ClassTy) { 427 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 428 (void)InsertResult; 429 assert(InsertResult.second && "DINode was already assigned a type index"); 430 return TI; 431 } 432 433 unsigned CodeViewDebug::getPointerSizeInBytes() { 434 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 435 } 436 437 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 438 const LexicalScope *LS) { 439 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 440 // This variable was inlined. Associate it with the InlineSite. 441 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 442 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 443 Site.InlinedLocals.emplace_back(Var); 444 } else { 445 // This variable goes into the corresponding lexical scope. 446 ScopeVariables[LS].emplace_back(Var); 447 } 448 } 449 450 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 451 const DILocation *Loc) { 452 auto B = Locs.begin(), E = Locs.end(); 453 if (std::find(B, E, Loc) == E) 454 Locs.push_back(Loc); 455 } 456 457 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 458 const MachineFunction *MF) { 459 // Skip this instruction if it has the same location as the previous one. 460 if (!DL || DL == PrevInstLoc) 461 return; 462 463 const DIScope *Scope = DL.get()->getScope(); 464 if (!Scope) 465 return; 466 467 // Skip this line if it is longer than the maximum we can record. 468 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 469 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 470 LI.isNeverStepInto()) 471 return; 472 473 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 474 if (CI.getStartColumn() != DL.getCol()) 475 return; 476 477 if (!CurFn->HaveLineInfo) 478 CurFn->HaveLineInfo = true; 479 unsigned FileId = 0; 480 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 481 FileId = CurFn->LastFileId; 482 else 483 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 484 PrevInstLoc = DL; 485 486 unsigned FuncId = CurFn->FuncId; 487 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 488 const DILocation *Loc = DL.get(); 489 490 // If this location was actually inlined from somewhere else, give it the ID 491 // of the inline call site. 492 FuncId = 493 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 494 495 // Ensure we have links in the tree of inline call sites. 496 bool FirstLoc = true; 497 while ((SiteLoc = Loc->getInlinedAt())) { 498 InlineSite &Site = 499 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 500 if (!FirstLoc) 501 addLocIfNotPresent(Site.ChildSites, Loc); 502 FirstLoc = false; 503 Loc = SiteLoc; 504 } 505 addLocIfNotPresent(CurFn->ChildSites, Loc); 506 } 507 508 OS.EmitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 509 /*PrologueEnd=*/false, /*IsStmt=*/false, 510 DL->getFilename(), SMLoc()); 511 } 512 513 void CodeViewDebug::emitCodeViewMagicVersion() { 514 OS.EmitValueToAlignment(4); 515 OS.AddComment("Debug section magic"); 516 OS.EmitIntValue(COFF::DEBUG_SECTION_MAGIC, 4); 517 } 518 519 void CodeViewDebug::endModule() { 520 if (!Asm || !MMI->hasDebugInfo()) 521 return; 522 523 assert(Asm != nullptr); 524 525 // The COFF .debug$S section consists of several subsections, each starting 526 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 527 // of the payload followed by the payload itself. The subsections are 4-byte 528 // aligned. 529 530 // Use the generic .debug$S section, and make a subsection for all the inlined 531 // subprograms. 532 switchToDebugSectionForSymbol(nullptr); 533 534 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 535 emitCompilerInformation(); 536 endCVSubsection(CompilerInfo); 537 538 emitInlineeLinesSubsection(); 539 540 // Emit per-function debug information. 541 for (auto &P : FnDebugInfo) 542 if (!P.first->isDeclarationForLinker()) 543 emitDebugInfoForFunction(P.first, *P.second); 544 545 // Emit global variable debug information. 546 setCurrentSubprogram(nullptr); 547 emitDebugInfoForGlobals(); 548 549 // Emit retained types. 550 emitDebugInfoForRetainedTypes(); 551 552 // Switch back to the generic .debug$S section after potentially processing 553 // comdat symbol sections. 554 switchToDebugSectionForSymbol(nullptr); 555 556 // Emit UDT records for any types used by global variables. 557 if (!GlobalUDTs.empty()) { 558 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 559 emitDebugInfoForUDTs(GlobalUDTs); 560 endCVSubsection(SymbolsEnd); 561 } 562 563 // This subsection holds a file index to offset in string table table. 564 OS.AddComment("File index to string table offset subsection"); 565 OS.EmitCVFileChecksumsDirective(); 566 567 // This subsection holds the string table. 568 OS.AddComment("String table"); 569 OS.EmitCVStringTableDirective(); 570 571 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 572 // subsection in the generic .debug$S section at the end. There is no 573 // particular reason for this ordering other than to match MSVC. 574 emitBuildInfo(); 575 576 // Emit type information and hashes last, so that any types we translate while 577 // emitting function info are included. 578 emitTypeInformation(); 579 580 if (EmitDebugGlobalHashes) 581 emitTypeGlobalHashes(); 582 583 clear(); 584 } 585 586 static void 587 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 588 unsigned MaxFixedRecordLength = 0xF00) { 589 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 590 // after a fixed length portion of the record. The fixed length portion should 591 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 592 // overall record size is less than the maximum allowed. 593 SmallString<32> NullTerminatedString( 594 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 595 NullTerminatedString.push_back('\0'); 596 OS.EmitBytes(NullTerminatedString); 597 } 598 599 void CodeViewDebug::emitTypeInformation() { 600 if (TypeTable.empty()) 601 return; 602 603 // Start the .debug$T or .debug$P section with 0x4. 604 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 605 emitCodeViewMagicVersion(); 606 607 SmallString<8> CommentPrefix; 608 if (OS.isVerboseAsm()) { 609 CommentPrefix += '\t'; 610 CommentPrefix += Asm->MAI->getCommentString(); 611 CommentPrefix += ' '; 612 } 613 614 TypeTableCollection Table(TypeTable.records()); 615 Optional<TypeIndex> B = Table.getFirst(); 616 while (B) { 617 // This will fail if the record data is invalid. 618 CVType Record = Table.getType(*B); 619 620 if (OS.isVerboseAsm()) { 621 // Emit a block comment describing the type record for readability. 622 SmallString<512> CommentBlock; 623 raw_svector_ostream CommentOS(CommentBlock); 624 ScopedPrinter SP(CommentOS); 625 SP.setPrefix(CommentPrefix); 626 TypeDumpVisitor TDV(Table, &SP, false); 627 628 Error E = codeview::visitTypeRecord(Record, *B, TDV); 629 if (E) { 630 logAllUnhandledErrors(std::move(E), errs(), "error: "); 631 llvm_unreachable("produced malformed type record"); 632 } 633 // emitRawComment will insert its own tab and comment string before 634 // the first line, so strip off our first one. It also prints its own 635 // newline. 636 OS.emitRawComment( 637 CommentOS.str().drop_front(CommentPrefix.size() - 1).rtrim()); 638 } 639 OS.EmitBinaryData(Record.str_data()); 640 B = Table.getNext(*B); 641 } 642 } 643 644 void CodeViewDebug::emitTypeGlobalHashes() { 645 if (TypeTable.empty()) 646 return; 647 648 // Start the .debug$H section with the version and hash algorithm, currently 649 // hardcoded to version 0, SHA1. 650 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 651 652 OS.EmitValueToAlignment(4); 653 OS.AddComment("Magic"); 654 OS.EmitIntValue(COFF::DEBUG_HASHES_SECTION_MAGIC, 4); 655 OS.AddComment("Section Version"); 656 OS.EmitIntValue(0, 2); 657 OS.AddComment("Hash Algorithm"); 658 OS.EmitIntValue(uint16_t(GlobalTypeHashAlg::SHA1_8), 2); 659 660 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 661 for (const auto &GHR : TypeTable.hashes()) { 662 if (OS.isVerboseAsm()) { 663 // Emit an EOL-comment describing which TypeIndex this hash corresponds 664 // to, as well as the stringified SHA1 hash. 665 SmallString<32> Comment; 666 raw_svector_ostream CommentOS(Comment); 667 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 668 OS.AddComment(Comment); 669 ++TI; 670 } 671 assert(GHR.Hash.size() == 8); 672 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 673 GHR.Hash.size()); 674 OS.EmitBinaryData(S); 675 } 676 } 677 678 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 679 switch (DWLang) { 680 case dwarf::DW_LANG_C: 681 case dwarf::DW_LANG_C89: 682 case dwarf::DW_LANG_C99: 683 case dwarf::DW_LANG_C11: 684 case dwarf::DW_LANG_ObjC: 685 return SourceLanguage::C; 686 case dwarf::DW_LANG_C_plus_plus: 687 case dwarf::DW_LANG_C_plus_plus_03: 688 case dwarf::DW_LANG_C_plus_plus_11: 689 case dwarf::DW_LANG_C_plus_plus_14: 690 return SourceLanguage::Cpp; 691 case dwarf::DW_LANG_Fortran77: 692 case dwarf::DW_LANG_Fortran90: 693 case dwarf::DW_LANG_Fortran03: 694 case dwarf::DW_LANG_Fortran08: 695 return SourceLanguage::Fortran; 696 case dwarf::DW_LANG_Pascal83: 697 return SourceLanguage::Pascal; 698 case dwarf::DW_LANG_Cobol74: 699 case dwarf::DW_LANG_Cobol85: 700 return SourceLanguage::Cobol; 701 case dwarf::DW_LANG_Java: 702 return SourceLanguage::Java; 703 case dwarf::DW_LANG_D: 704 return SourceLanguage::D; 705 case dwarf::DW_LANG_Swift: 706 return SourceLanguage::Swift; 707 default: 708 // There's no CodeView representation for this language, and CV doesn't 709 // have an "unknown" option for the language field, so we'll use MASM, 710 // as it's very low level. 711 return SourceLanguage::Masm; 712 } 713 } 714 715 namespace { 716 struct Version { 717 int Part[4]; 718 }; 719 } // end anonymous namespace 720 721 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 722 // the version number. 723 static Version parseVersion(StringRef Name) { 724 Version V = {{0}}; 725 int N = 0; 726 for (const char C : Name) { 727 if (isdigit(C)) { 728 V.Part[N] *= 10; 729 V.Part[N] += C - '0'; 730 } else if (C == '.') { 731 ++N; 732 if (N >= 4) 733 return V; 734 } else if (N > 0) 735 return V; 736 } 737 return V; 738 } 739 740 void CodeViewDebug::emitCompilerInformation() { 741 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 742 uint32_t Flags = 0; 743 744 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 745 const MDNode *Node = *CUs->operands().begin(); 746 const auto *CU = cast<DICompileUnit>(Node); 747 748 // The low byte of the flags indicates the source language. 749 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 750 // TODO: Figure out which other flags need to be set. 751 752 OS.AddComment("Flags and language"); 753 OS.EmitIntValue(Flags, 4); 754 755 OS.AddComment("CPUType"); 756 OS.EmitIntValue(static_cast<uint64_t>(TheCPU), 2); 757 758 StringRef CompilerVersion = CU->getProducer(); 759 Version FrontVer = parseVersion(CompilerVersion); 760 OS.AddComment("Frontend version"); 761 for (int N = 0; N < 4; ++N) 762 OS.EmitIntValue(FrontVer.Part[N], 2); 763 764 // Some Microsoft tools, like Binscope, expect a backend version number of at 765 // least 8.something, so we'll coerce the LLVM version into a form that 766 // guarantees it'll be big enough without really lying about the version. 767 int Major = 1000 * LLVM_VERSION_MAJOR + 768 10 * LLVM_VERSION_MINOR + 769 LLVM_VERSION_PATCH; 770 // Clamp it for builds that use unusually large version numbers. 771 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 772 Version BackVer = {{ Major, 0, 0, 0 }}; 773 OS.AddComment("Backend version"); 774 for (int N = 0; N < 4; ++N) 775 OS.EmitIntValue(BackVer.Part[N], 2); 776 777 OS.AddComment("Null-terminated compiler version string"); 778 emitNullTerminatedSymbolName(OS, CompilerVersion); 779 780 endSymbolRecord(CompilerEnd); 781 } 782 783 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 784 StringRef S) { 785 StringIdRecord SIR(TypeIndex(0x0), S); 786 return TypeTable.writeLeafType(SIR); 787 } 788 789 void CodeViewDebug::emitBuildInfo() { 790 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 791 // build info. The known prefix is: 792 // - Absolute path of current directory 793 // - Compiler path 794 // - Main source file path, relative to CWD or absolute 795 // - Type server PDB file 796 // - Canonical compiler command line 797 // If frontend and backend compilation are separated (think llc or LTO), it's 798 // not clear if the compiler path should refer to the executable for the 799 // frontend or the backend. Leave it blank for now. 800 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 801 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 802 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 803 const auto *CU = cast<DICompileUnit>(Node); 804 const DIFile *MainSourceFile = CU->getFile(); 805 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 806 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 807 BuildInfoArgs[BuildInfoRecord::SourceFile] = 808 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 809 // FIXME: Path to compiler and command line. PDB is intentionally blank unless 810 // we implement /Zi type servers. 811 BuildInfoRecord BIR(BuildInfoArgs); 812 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 813 814 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 815 // from the module symbols into the type stream. 816 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 817 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 818 OS.AddComment("LF_BUILDINFO index"); 819 OS.EmitIntValue(BuildInfoIndex.getIndex(), 4); 820 endSymbolRecord(BIEnd); 821 endCVSubsection(BISubsecEnd); 822 } 823 824 void CodeViewDebug::emitInlineeLinesSubsection() { 825 if (InlinedSubprograms.empty()) 826 return; 827 828 OS.AddComment("Inlinee lines subsection"); 829 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 830 831 // We emit the checksum info for files. This is used by debuggers to 832 // determine if a pdb matches the source before loading it. Visual Studio, 833 // for instance, will display a warning that the breakpoints are not valid if 834 // the pdb does not match the source. 835 OS.AddComment("Inlinee lines signature"); 836 OS.EmitIntValue(unsigned(InlineeLinesSignature::Normal), 4); 837 838 for (const DISubprogram *SP : InlinedSubprograms) { 839 assert(TypeIndices.count({SP, nullptr})); 840 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 841 842 OS.AddBlankLine(); 843 unsigned FileId = maybeRecordFile(SP->getFile()); 844 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 845 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 846 OS.AddBlankLine(); 847 OS.AddComment("Type index of inlined function"); 848 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 849 OS.AddComment("Offset into filechecksum table"); 850 OS.EmitCVFileChecksumOffsetDirective(FileId); 851 OS.AddComment("Starting line number"); 852 OS.EmitIntValue(SP->getLine(), 4); 853 } 854 855 endCVSubsection(InlineEnd); 856 } 857 858 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 859 const DILocation *InlinedAt, 860 const InlineSite &Site) { 861 assert(TypeIndices.count({Site.Inlinee, nullptr})); 862 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 863 864 // SymbolRecord 865 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 866 867 OS.AddComment("PtrParent"); 868 OS.EmitIntValue(0, 4); 869 OS.AddComment("PtrEnd"); 870 OS.EmitIntValue(0, 4); 871 OS.AddComment("Inlinee type index"); 872 OS.EmitIntValue(InlineeIdx.getIndex(), 4); 873 874 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 875 unsigned StartLineNum = Site.Inlinee->getLine(); 876 877 OS.EmitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 878 FI.Begin, FI.End); 879 880 endSymbolRecord(InlineEnd); 881 882 emitLocalVariableList(FI, Site.InlinedLocals); 883 884 // Recurse on child inlined call sites before closing the scope. 885 for (const DILocation *ChildSite : Site.ChildSites) { 886 auto I = FI.InlineSites.find(ChildSite); 887 assert(I != FI.InlineSites.end() && 888 "child site not in function inline site map"); 889 emitInlinedCallSite(FI, ChildSite, I->second); 890 } 891 892 // Close the scope. 893 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 894 } 895 896 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 897 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 898 // comdat key. A section may be comdat because of -ffunction-sections or 899 // because it is comdat in the IR. 900 MCSectionCOFF *GVSec = 901 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 902 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 903 904 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 905 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 906 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 907 908 OS.SwitchSection(DebugSec); 909 910 // Emit the magic version number if this is the first time we've switched to 911 // this section. 912 if (ComdatDebugSections.insert(DebugSec).second) 913 emitCodeViewMagicVersion(); 914 } 915 916 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 917 // The only supported thunk ordinal is currently the standard type. 918 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 919 FunctionInfo &FI, 920 const MCSymbol *Fn) { 921 std::string FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 922 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 923 924 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 925 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 926 927 // Emit S_THUNK32 928 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 929 OS.AddComment("PtrParent"); 930 OS.EmitIntValue(0, 4); 931 OS.AddComment("PtrEnd"); 932 OS.EmitIntValue(0, 4); 933 OS.AddComment("PtrNext"); 934 OS.EmitIntValue(0, 4); 935 OS.AddComment("Thunk section relative address"); 936 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 937 OS.AddComment("Thunk section index"); 938 OS.EmitCOFFSectionIndex(Fn); 939 OS.AddComment("Code size"); 940 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 941 OS.AddComment("Ordinal"); 942 OS.EmitIntValue(unsigned(ordinal), 1); 943 OS.AddComment("Function name"); 944 emitNullTerminatedSymbolName(OS, FuncName); 945 // Additional fields specific to the thunk ordinal would go here. 946 endSymbolRecord(ThunkRecordEnd); 947 948 // Local variables/inlined routines are purposely omitted here. The point of 949 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 950 951 // Emit S_PROC_ID_END 952 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 953 954 endCVSubsection(SymbolsEnd); 955 } 956 957 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 958 FunctionInfo &FI) { 959 // For each function there is a separate subsection which holds the PC to 960 // file:line table. 961 const MCSymbol *Fn = Asm->getSymbol(GV); 962 assert(Fn); 963 964 // Switch to the to a comdat section, if appropriate. 965 switchToDebugSectionForSymbol(Fn); 966 967 std::string FuncName; 968 auto *SP = GV->getSubprogram(); 969 assert(SP); 970 setCurrentSubprogram(SP); 971 972 if (SP->isThunk()) { 973 emitDebugInfoForThunk(GV, FI, Fn); 974 return; 975 } 976 977 // If we have a display name, build the fully qualified name by walking the 978 // chain of scopes. 979 if (!SP->getName().empty()) 980 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 981 982 // If our DISubprogram name is empty, use the mangled name. 983 if (FuncName.empty()) 984 FuncName = GlobalValue::dropLLVMManglingEscape(GV->getName()); 985 986 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 987 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 988 OS.EmitCVFPOData(Fn); 989 990 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 991 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 992 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 993 { 994 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 995 : SymbolKind::S_GPROC32_ID; 996 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 997 998 // These fields are filled in by tools like CVPACK which run after the fact. 999 OS.AddComment("PtrParent"); 1000 OS.EmitIntValue(0, 4); 1001 OS.AddComment("PtrEnd"); 1002 OS.EmitIntValue(0, 4); 1003 OS.AddComment("PtrNext"); 1004 OS.EmitIntValue(0, 4); 1005 // This is the important bit that tells the debugger where the function 1006 // code is located and what's its size: 1007 OS.AddComment("Code size"); 1008 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1009 OS.AddComment("Offset after prologue"); 1010 OS.EmitIntValue(0, 4); 1011 OS.AddComment("Offset before epilogue"); 1012 OS.EmitIntValue(0, 4); 1013 OS.AddComment("Function type index"); 1014 OS.EmitIntValue(getFuncIdForSubprogram(GV->getSubprogram()).getIndex(), 4); 1015 OS.AddComment("Function section relative address"); 1016 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1017 OS.AddComment("Function section index"); 1018 OS.EmitCOFFSectionIndex(Fn); 1019 OS.AddComment("Flags"); 1020 OS.EmitIntValue(0, 1); 1021 // Emit the function display name as a null-terminated string. 1022 OS.AddComment("Function name"); 1023 // Truncate the name so we won't overflow the record length field. 1024 emitNullTerminatedSymbolName(OS, FuncName); 1025 endSymbolRecord(ProcRecordEnd); 1026 1027 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1028 // Subtract out the CSR size since MSVC excludes that and we include it. 1029 OS.AddComment("FrameSize"); 1030 OS.EmitIntValue(FI.FrameSize - FI.CSRSize, 4); 1031 OS.AddComment("Padding"); 1032 OS.EmitIntValue(0, 4); 1033 OS.AddComment("Offset of padding"); 1034 OS.EmitIntValue(0, 4); 1035 OS.AddComment("Bytes of callee saved registers"); 1036 OS.EmitIntValue(FI.CSRSize, 4); 1037 OS.AddComment("Exception handler offset"); 1038 OS.EmitIntValue(0, 4); 1039 OS.AddComment("Exception handler section"); 1040 OS.EmitIntValue(0, 2); 1041 OS.AddComment("Flags (defines frame register)"); 1042 OS.EmitIntValue(uint32_t(FI.FrameProcOpts), 4); 1043 endSymbolRecord(FrameProcEnd); 1044 1045 emitLocalVariableList(FI, FI.Locals); 1046 emitGlobalVariableList(FI.Globals); 1047 emitLexicalBlockList(FI.ChildBlocks, FI); 1048 1049 // Emit inlined call site information. Only emit functions inlined directly 1050 // into the parent function. We'll emit the other sites recursively as part 1051 // of their parent inline site. 1052 for (const DILocation *InlinedAt : FI.ChildSites) { 1053 auto I = FI.InlineSites.find(InlinedAt); 1054 assert(I != FI.InlineSites.end() && 1055 "child site not in function inline site map"); 1056 emitInlinedCallSite(FI, InlinedAt, I->second); 1057 } 1058 1059 for (auto Annot : FI.Annotations) { 1060 MCSymbol *Label = Annot.first; 1061 MDTuple *Strs = cast<MDTuple>(Annot.second); 1062 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1063 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1064 // FIXME: Make sure we don't overflow the max record size. 1065 OS.EmitCOFFSectionIndex(Label); 1066 OS.EmitIntValue(Strs->getNumOperands(), 2); 1067 for (Metadata *MD : Strs->operands()) { 1068 // MDStrings are null terminated, so we can do EmitBytes and get the 1069 // nice .asciz directive. 1070 StringRef Str = cast<MDString>(MD)->getString(); 1071 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1072 OS.EmitBytes(StringRef(Str.data(), Str.size() + 1)); 1073 } 1074 endSymbolRecord(AnnotEnd); 1075 } 1076 1077 for (auto HeapAllocSite : FI.HeapAllocSites) { 1078 MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1079 MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1080 1081 // The labels might not be defined if the instruction was replaced 1082 // somewhere in the codegen pipeline. 1083 if (!BeginLabel->isDefined() || !EndLabel->isDefined()) 1084 continue; 1085 1086 DIType *DITy = std::get<2>(HeapAllocSite); 1087 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1088 OS.AddComment("Call site offset"); 1089 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1090 OS.AddComment("Call site section index"); 1091 OS.EmitCOFFSectionIndex(BeginLabel); 1092 OS.AddComment("Call instruction length"); 1093 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1094 OS.AddComment("Type index"); 1095 OS.EmitIntValue(getCompleteTypeIndex(DITy).getIndex(), 4); 1096 endSymbolRecord(HeapAllocEnd); 1097 } 1098 1099 if (SP != nullptr) 1100 emitDebugInfoForUDTs(LocalUDTs); 1101 1102 // We're done with this function. 1103 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1104 } 1105 endCVSubsection(SymbolsEnd); 1106 1107 // We have an assembler directive that takes care of the whole line table. 1108 OS.EmitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1109 } 1110 1111 CodeViewDebug::LocalVarDefRange 1112 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1113 LocalVarDefRange DR; 1114 DR.InMemory = -1; 1115 DR.DataOffset = Offset; 1116 assert(DR.DataOffset == Offset && "truncation"); 1117 DR.IsSubfield = 0; 1118 DR.StructOffset = 0; 1119 DR.CVRegister = CVRegister; 1120 return DR; 1121 } 1122 1123 void CodeViewDebug::collectVariableInfoFromMFTable( 1124 DenseSet<InlinedEntity> &Processed) { 1125 const MachineFunction &MF = *Asm->MF; 1126 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1127 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1128 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1129 1130 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1131 if (!VI.Var) 1132 continue; 1133 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1134 "Expected inlined-at fields to agree"); 1135 1136 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1137 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1138 1139 // If variable scope is not found then skip this variable. 1140 if (!Scope) 1141 continue; 1142 1143 // If the variable has an attached offset expression, extract it. 1144 // FIXME: Try to handle DW_OP_deref as well. 1145 int64_t ExprOffset = 0; 1146 bool Deref = false; 1147 if (VI.Expr) { 1148 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1149 if (VI.Expr->getNumElements() == 1 && 1150 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1151 Deref = true; 1152 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1153 continue; 1154 } 1155 1156 // Get the frame register used and the offset. 1157 unsigned FrameReg = 0; 1158 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1159 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1160 1161 // Calculate the label ranges. 1162 LocalVarDefRange DefRange = 1163 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1164 1165 for (const InsnRange &Range : Scope->getRanges()) { 1166 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1167 const MCSymbol *End = getLabelAfterInsn(Range.second); 1168 End = End ? End : Asm->getFunctionEnd(); 1169 DefRange.Ranges.emplace_back(Begin, End); 1170 } 1171 1172 LocalVariable Var; 1173 Var.DIVar = VI.Var; 1174 Var.DefRanges.emplace_back(std::move(DefRange)); 1175 if (Deref) 1176 Var.UseReferenceType = true; 1177 1178 recordLocalVariable(std::move(Var), Scope); 1179 } 1180 } 1181 1182 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1183 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1184 } 1185 1186 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1187 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1188 } 1189 1190 void CodeViewDebug::calculateRanges( 1191 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1192 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1193 1194 // Calculate the definition ranges. 1195 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1196 const auto &Entry = *I; 1197 if (!Entry.isDbgValue()) 1198 continue; 1199 const MachineInstr *DVInst = Entry.getInstr(); 1200 assert(DVInst->isDebugValue() && "Invalid History entry"); 1201 // FIXME: Find a way to represent constant variables, since they are 1202 // relatively common. 1203 Optional<DbgVariableLocation> Location = 1204 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1205 if (!Location) 1206 continue; 1207 1208 // CodeView can only express variables in register and variables in memory 1209 // at a constant offset from a register. However, for variables passed 1210 // indirectly by pointer, it is common for that pointer to be spilled to a 1211 // stack location. For the special case of one offseted load followed by a 1212 // zero offset load (a pointer spilled to the stack), we change the type of 1213 // the local variable from a value type to a reference type. This tricks the 1214 // debugger into doing the load for us. 1215 if (Var.UseReferenceType) { 1216 // We're using a reference type. Drop the last zero offset load. 1217 if (canUseReferenceType(*Location)) 1218 Location->LoadChain.pop_back(); 1219 else 1220 continue; 1221 } else if (needsReferenceType(*Location)) { 1222 // This location can't be expressed without switching to a reference type. 1223 // Start over using that. 1224 Var.UseReferenceType = true; 1225 Var.DefRanges.clear(); 1226 calculateRanges(Var, Entries); 1227 return; 1228 } 1229 1230 // We can only handle a register or an offseted load of a register. 1231 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1232 continue; 1233 { 1234 LocalVarDefRange DR; 1235 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1236 DR.InMemory = !Location->LoadChain.empty(); 1237 DR.DataOffset = 1238 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1239 if (Location->FragmentInfo) { 1240 DR.IsSubfield = true; 1241 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1242 } else { 1243 DR.IsSubfield = false; 1244 DR.StructOffset = 0; 1245 } 1246 1247 if (Var.DefRanges.empty() || 1248 Var.DefRanges.back().isDifferentLocation(DR)) { 1249 Var.DefRanges.emplace_back(std::move(DR)); 1250 } 1251 } 1252 1253 // Compute the label range. 1254 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1255 const MCSymbol *End; 1256 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1257 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1258 End = EndingEntry.isDbgValue() 1259 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1260 : getLabelAfterInsn(EndingEntry.getInstr()); 1261 } else 1262 End = Asm->getFunctionEnd(); 1263 1264 // If the last range end is our begin, just extend the last range. 1265 // Otherwise make a new range. 1266 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1267 Var.DefRanges.back().Ranges; 1268 if (!R.empty() && R.back().second == Begin) 1269 R.back().second = End; 1270 else 1271 R.emplace_back(Begin, End); 1272 1273 // FIXME: Do more range combining. 1274 } 1275 } 1276 1277 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1278 DenseSet<InlinedEntity> Processed; 1279 // Grab the variable info that was squirreled away in the MMI side-table. 1280 collectVariableInfoFromMFTable(Processed); 1281 1282 for (const auto &I : DbgValues) { 1283 InlinedEntity IV = I.first; 1284 if (Processed.count(IV)) 1285 continue; 1286 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1287 const DILocation *InlinedAt = IV.second; 1288 1289 // Instruction ranges, specifying where IV is accessible. 1290 const auto &Entries = I.second; 1291 1292 LexicalScope *Scope = nullptr; 1293 if (InlinedAt) 1294 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1295 else 1296 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1297 // If variable scope is not found then skip this variable. 1298 if (!Scope) 1299 continue; 1300 1301 LocalVariable Var; 1302 Var.DIVar = DIVar; 1303 1304 calculateRanges(Var, Entries); 1305 recordLocalVariable(std::move(Var), Scope); 1306 } 1307 } 1308 1309 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1310 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1311 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1312 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1313 const Function &GV = MF->getFunction(); 1314 auto Insertion = FnDebugInfo.insert({&GV, llvm::make_unique<FunctionInfo>()}); 1315 assert(Insertion.second && "function already has info"); 1316 CurFn = Insertion.first->second.get(); 1317 CurFn->FuncId = NextFuncId++; 1318 CurFn->Begin = Asm->getFunctionBegin(); 1319 1320 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1321 // callee-saved registers were used. For targets that don't use a PUSH 1322 // instruction (AArch64), this will be zero. 1323 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1324 CurFn->FrameSize = MFI.getStackSize(); 1325 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1326 CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF); 1327 1328 // For this function S_FRAMEPROC record, figure out which codeview register 1329 // will be the frame pointer. 1330 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1331 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1332 if (CurFn->FrameSize > 0) { 1333 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1334 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1335 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1336 } else { 1337 // If there is an FP, parameters are always relative to it. 1338 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1339 if (CurFn->HasStackRealignment) { 1340 // If the stack needs realignment, locals are relative to SP or VFRAME. 1341 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1342 } else { 1343 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1344 // other stack adjustments. 1345 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1346 } 1347 } 1348 } 1349 1350 // Compute other frame procedure options. 1351 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1352 if (MFI.hasVarSizedObjects()) 1353 FPO |= FrameProcedureOptions::HasAlloca; 1354 if (MF->exposesReturnsTwice()) 1355 FPO |= FrameProcedureOptions::HasSetJmp; 1356 // FIXME: Set HasLongJmp if we ever track that info. 1357 if (MF->hasInlineAsm()) 1358 FPO |= FrameProcedureOptions::HasInlineAssembly; 1359 if (GV.hasPersonalityFn()) { 1360 if (isAsynchronousEHPersonality( 1361 classifyEHPersonality(GV.getPersonalityFn()))) 1362 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1363 else 1364 FPO |= FrameProcedureOptions::HasExceptionHandling; 1365 } 1366 if (GV.hasFnAttribute(Attribute::InlineHint)) 1367 FPO |= FrameProcedureOptions::MarkedInline; 1368 if (GV.hasFnAttribute(Attribute::Naked)) 1369 FPO |= FrameProcedureOptions::Naked; 1370 if (MFI.hasStackProtectorIndex()) 1371 FPO |= FrameProcedureOptions::SecurityChecks; 1372 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1373 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1374 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1375 !GV.hasOptSize() && !GV.hasOptNone()) 1376 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1377 // FIXME: Set GuardCfg when it is implemented. 1378 CurFn->FrameProcOpts = FPO; 1379 1380 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1381 1382 // Find the end of the function prolog. First known non-DBG_VALUE and 1383 // non-frame setup location marks the beginning of the function body. 1384 // FIXME: is there a simpler a way to do this? Can we just search 1385 // for the first instruction of the function, not the last of the prolog? 1386 DebugLoc PrologEndLoc; 1387 bool EmptyPrologue = true; 1388 for (const auto &MBB : *MF) { 1389 for (const auto &MI : MBB) { 1390 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1391 MI.getDebugLoc()) { 1392 PrologEndLoc = MI.getDebugLoc(); 1393 break; 1394 } else if (!MI.isMetaInstruction()) { 1395 EmptyPrologue = false; 1396 } 1397 } 1398 } 1399 1400 // Record beginning of function if we have a non-empty prologue. 1401 if (PrologEndLoc && !EmptyPrologue) { 1402 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1403 maybeRecordLocation(FnStartDL, MF); 1404 } 1405 } 1406 1407 static bool shouldEmitUdt(const DIType *T) { 1408 if (!T) 1409 return false; 1410 1411 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1412 if (T->getTag() == dwarf::DW_TAG_typedef) { 1413 if (DIScope *Scope = T->getScope()) { 1414 switch (Scope->getTag()) { 1415 case dwarf::DW_TAG_structure_type: 1416 case dwarf::DW_TAG_class_type: 1417 case dwarf::DW_TAG_union_type: 1418 return false; 1419 } 1420 } 1421 } 1422 1423 while (true) { 1424 if (!T || T->isForwardDecl()) 1425 return false; 1426 1427 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1428 if (!DT) 1429 return true; 1430 T = DT->getBaseType(); 1431 } 1432 return true; 1433 } 1434 1435 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1436 // Don't record empty UDTs. 1437 if (Ty->getName().empty()) 1438 return; 1439 if (!shouldEmitUdt(Ty)) 1440 return; 1441 1442 SmallVector<StringRef, 5> QualifiedNameComponents; 1443 const DISubprogram *ClosestSubprogram = 1444 getQualifiedNameComponents(Ty->getScope(), QualifiedNameComponents); 1445 1446 std::string FullyQualifiedName = 1447 getQualifiedName(QualifiedNameComponents, getPrettyScopeName(Ty)); 1448 1449 if (ClosestSubprogram == nullptr) { 1450 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1451 } else if (ClosestSubprogram == CurrentSubprogram) { 1452 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1453 } 1454 1455 // TODO: What if the ClosestSubprogram is neither null or the current 1456 // subprogram? Currently, the UDT just gets dropped on the floor. 1457 // 1458 // The current behavior is not desirable. To get maximal fidelity, we would 1459 // need to perform all type translation before beginning emission of .debug$S 1460 // and then make LocalUDTs a member of FunctionInfo 1461 } 1462 1463 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1464 // Generic dispatch for lowering an unknown type. 1465 switch (Ty->getTag()) { 1466 case dwarf::DW_TAG_array_type: 1467 return lowerTypeArray(cast<DICompositeType>(Ty)); 1468 case dwarf::DW_TAG_typedef: 1469 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1470 case dwarf::DW_TAG_base_type: 1471 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1472 case dwarf::DW_TAG_pointer_type: 1473 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1474 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1475 LLVM_FALLTHROUGH; 1476 case dwarf::DW_TAG_reference_type: 1477 case dwarf::DW_TAG_rvalue_reference_type: 1478 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1479 case dwarf::DW_TAG_ptr_to_member_type: 1480 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1481 case dwarf::DW_TAG_restrict_type: 1482 case dwarf::DW_TAG_const_type: 1483 case dwarf::DW_TAG_volatile_type: 1484 // TODO: add support for DW_TAG_atomic_type here 1485 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1486 case dwarf::DW_TAG_subroutine_type: 1487 if (ClassTy) { 1488 // The member function type of a member function pointer has no 1489 // ThisAdjustment. 1490 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1491 /*ThisAdjustment=*/0, 1492 /*IsStaticMethod=*/false); 1493 } 1494 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1495 case dwarf::DW_TAG_enumeration_type: 1496 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1497 case dwarf::DW_TAG_class_type: 1498 case dwarf::DW_TAG_structure_type: 1499 return lowerTypeClass(cast<DICompositeType>(Ty)); 1500 case dwarf::DW_TAG_union_type: 1501 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1502 case dwarf::DW_TAG_unspecified_type: 1503 if (Ty->getName() == "decltype(nullptr)") 1504 return TypeIndex::NullptrT(); 1505 return TypeIndex::None(); 1506 default: 1507 // Use the null type index. 1508 return TypeIndex(); 1509 } 1510 } 1511 1512 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1513 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1514 StringRef TypeName = Ty->getName(); 1515 1516 addToUDTs(Ty); 1517 1518 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1519 TypeName == "HRESULT") 1520 return TypeIndex(SimpleTypeKind::HResult); 1521 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1522 TypeName == "wchar_t") 1523 return TypeIndex(SimpleTypeKind::WideCharacter); 1524 1525 return UnderlyingTypeIndex; 1526 } 1527 1528 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1529 const DIType *ElementType = Ty->getBaseType(); 1530 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1531 // IndexType is size_t, which depends on the bitness of the target. 1532 TypeIndex IndexType = getPointerSizeInBytes() == 8 1533 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1534 : TypeIndex(SimpleTypeKind::UInt32Long); 1535 1536 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1537 1538 // Add subranges to array type. 1539 DINodeArray Elements = Ty->getElements(); 1540 for (int i = Elements.size() - 1; i >= 0; --i) { 1541 const DINode *Element = Elements[i]; 1542 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1543 1544 const DISubrange *Subrange = cast<DISubrange>(Element); 1545 assert(Subrange->getLowerBound() == 0 && 1546 "codeview doesn't support subranges with lower bounds"); 1547 int64_t Count = -1; 1548 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>()) 1549 Count = CI->getSExtValue(); 1550 1551 // Forward declarations of arrays without a size and VLAs use a count of -1. 1552 // Emit a count of zero in these cases to match what MSVC does for arrays 1553 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1554 // should do for them even if we could distinguish them. 1555 if (Count == -1) 1556 Count = 0; 1557 1558 // Update the element size and element type index for subsequent subranges. 1559 ElementSize *= Count; 1560 1561 // If this is the outermost array, use the size from the array. It will be 1562 // more accurate if we had a VLA or an incomplete element type size. 1563 uint64_t ArraySize = 1564 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1565 1566 StringRef Name = (i == 0) ? Ty->getName() : ""; 1567 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1568 ElementTypeIndex = TypeTable.writeLeafType(AR); 1569 } 1570 1571 return ElementTypeIndex; 1572 } 1573 1574 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1575 TypeIndex Index; 1576 dwarf::TypeKind Kind; 1577 uint32_t ByteSize; 1578 1579 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1580 ByteSize = Ty->getSizeInBits() / 8; 1581 1582 SimpleTypeKind STK = SimpleTypeKind::None; 1583 switch (Kind) { 1584 case dwarf::DW_ATE_address: 1585 // FIXME: Translate 1586 break; 1587 case dwarf::DW_ATE_boolean: 1588 switch (ByteSize) { 1589 case 1: STK = SimpleTypeKind::Boolean8; break; 1590 case 2: STK = SimpleTypeKind::Boolean16; break; 1591 case 4: STK = SimpleTypeKind::Boolean32; break; 1592 case 8: STK = SimpleTypeKind::Boolean64; break; 1593 case 16: STK = SimpleTypeKind::Boolean128; break; 1594 } 1595 break; 1596 case dwarf::DW_ATE_complex_float: 1597 switch (ByteSize) { 1598 case 2: STK = SimpleTypeKind::Complex16; break; 1599 case 4: STK = SimpleTypeKind::Complex32; break; 1600 case 8: STK = SimpleTypeKind::Complex64; break; 1601 case 10: STK = SimpleTypeKind::Complex80; break; 1602 case 16: STK = SimpleTypeKind::Complex128; break; 1603 } 1604 break; 1605 case dwarf::DW_ATE_float: 1606 switch (ByteSize) { 1607 case 2: STK = SimpleTypeKind::Float16; break; 1608 case 4: STK = SimpleTypeKind::Float32; break; 1609 case 6: STK = SimpleTypeKind::Float48; break; 1610 case 8: STK = SimpleTypeKind::Float64; break; 1611 case 10: STK = SimpleTypeKind::Float80; break; 1612 case 16: STK = SimpleTypeKind::Float128; break; 1613 } 1614 break; 1615 case dwarf::DW_ATE_signed: 1616 switch (ByteSize) { 1617 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1618 case 2: STK = SimpleTypeKind::Int16Short; break; 1619 case 4: STK = SimpleTypeKind::Int32; break; 1620 case 8: STK = SimpleTypeKind::Int64Quad; break; 1621 case 16: STK = SimpleTypeKind::Int128Oct; break; 1622 } 1623 break; 1624 case dwarf::DW_ATE_unsigned: 1625 switch (ByteSize) { 1626 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1627 case 2: STK = SimpleTypeKind::UInt16Short; break; 1628 case 4: STK = SimpleTypeKind::UInt32; break; 1629 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1630 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1631 } 1632 break; 1633 case dwarf::DW_ATE_UTF: 1634 switch (ByteSize) { 1635 case 2: STK = SimpleTypeKind::Character16; break; 1636 case 4: STK = SimpleTypeKind::Character32; break; 1637 } 1638 break; 1639 case dwarf::DW_ATE_signed_char: 1640 if (ByteSize == 1) 1641 STK = SimpleTypeKind::SignedCharacter; 1642 break; 1643 case dwarf::DW_ATE_unsigned_char: 1644 if (ByteSize == 1) 1645 STK = SimpleTypeKind::UnsignedCharacter; 1646 break; 1647 default: 1648 break; 1649 } 1650 1651 // Apply some fixups based on the source-level type name. 1652 if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int") 1653 STK = SimpleTypeKind::Int32Long; 1654 if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int") 1655 STK = SimpleTypeKind::UInt32Long; 1656 if (STK == SimpleTypeKind::UInt16Short && 1657 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1658 STK = SimpleTypeKind::WideCharacter; 1659 if ((STK == SimpleTypeKind::SignedCharacter || 1660 STK == SimpleTypeKind::UnsignedCharacter) && 1661 Ty->getName() == "char") 1662 STK = SimpleTypeKind::NarrowCharacter; 1663 1664 return TypeIndex(STK); 1665 } 1666 1667 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1668 PointerOptions PO) { 1669 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1670 1671 // Pointers to simple types without any options can use SimpleTypeMode, rather 1672 // than having a dedicated pointer type record. 1673 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1674 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1675 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1676 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1677 ? SimpleTypeMode::NearPointer64 1678 : SimpleTypeMode::NearPointer32; 1679 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1680 } 1681 1682 PointerKind PK = 1683 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1684 PointerMode PM = PointerMode::Pointer; 1685 switch (Ty->getTag()) { 1686 default: llvm_unreachable("not a pointer tag type"); 1687 case dwarf::DW_TAG_pointer_type: 1688 PM = PointerMode::Pointer; 1689 break; 1690 case dwarf::DW_TAG_reference_type: 1691 PM = PointerMode::LValueReference; 1692 break; 1693 case dwarf::DW_TAG_rvalue_reference_type: 1694 PM = PointerMode::RValueReference; 1695 break; 1696 } 1697 1698 if (Ty->isObjectPointer()) 1699 PO |= PointerOptions::Const; 1700 1701 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1702 return TypeTable.writeLeafType(PR); 1703 } 1704 1705 static PointerToMemberRepresentation 1706 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1707 // SizeInBytes being zero generally implies that the member pointer type was 1708 // incomplete, which can happen if it is part of a function prototype. In this 1709 // case, use the unknown model instead of the general model. 1710 if (IsPMF) { 1711 switch (Flags & DINode::FlagPtrToMemberRep) { 1712 case 0: 1713 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1714 : PointerToMemberRepresentation::GeneralFunction; 1715 case DINode::FlagSingleInheritance: 1716 return PointerToMemberRepresentation::SingleInheritanceFunction; 1717 case DINode::FlagMultipleInheritance: 1718 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1719 case DINode::FlagVirtualInheritance: 1720 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1721 } 1722 } else { 1723 switch (Flags & DINode::FlagPtrToMemberRep) { 1724 case 0: 1725 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1726 : PointerToMemberRepresentation::GeneralData; 1727 case DINode::FlagSingleInheritance: 1728 return PointerToMemberRepresentation::SingleInheritanceData; 1729 case DINode::FlagMultipleInheritance: 1730 return PointerToMemberRepresentation::MultipleInheritanceData; 1731 case DINode::FlagVirtualInheritance: 1732 return PointerToMemberRepresentation::VirtualInheritanceData; 1733 } 1734 } 1735 llvm_unreachable("invalid ptr to member representation"); 1736 } 1737 1738 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1739 PointerOptions PO) { 1740 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1741 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1742 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType(), Ty->getClassType()); 1743 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1744 : PointerKind::Near32; 1745 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1746 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1747 : PointerMode::PointerToDataMember; 1748 1749 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1750 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1751 MemberPointerInfo MPI( 1752 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1753 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1754 return TypeTable.writeLeafType(PR); 1755 } 1756 1757 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1758 /// have a translation, use the NearC convention. 1759 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1760 switch (DwarfCC) { 1761 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1762 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1763 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1764 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1765 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1766 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1767 } 1768 return CallingConvention::NearC; 1769 } 1770 1771 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1772 ModifierOptions Mods = ModifierOptions::None; 1773 PointerOptions PO = PointerOptions::None; 1774 bool IsModifier = true; 1775 const DIType *BaseTy = Ty; 1776 while (IsModifier && BaseTy) { 1777 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1778 switch (BaseTy->getTag()) { 1779 case dwarf::DW_TAG_const_type: 1780 Mods |= ModifierOptions::Const; 1781 PO |= PointerOptions::Const; 1782 break; 1783 case dwarf::DW_TAG_volatile_type: 1784 Mods |= ModifierOptions::Volatile; 1785 PO |= PointerOptions::Volatile; 1786 break; 1787 case dwarf::DW_TAG_restrict_type: 1788 // Only pointer types be marked with __restrict. There is no known flag 1789 // for __restrict in LF_MODIFIER records. 1790 PO |= PointerOptions::Restrict; 1791 break; 1792 default: 1793 IsModifier = false; 1794 break; 1795 } 1796 if (IsModifier) 1797 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1798 } 1799 1800 // Check if the inner type will use an LF_POINTER record. If so, the 1801 // qualifiers will go in the LF_POINTER record. This comes up for types like 1802 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1803 // char *'. 1804 if (BaseTy) { 1805 switch (BaseTy->getTag()) { 1806 case dwarf::DW_TAG_pointer_type: 1807 case dwarf::DW_TAG_reference_type: 1808 case dwarf::DW_TAG_rvalue_reference_type: 1809 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1810 case dwarf::DW_TAG_ptr_to_member_type: 1811 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1812 default: 1813 break; 1814 } 1815 } 1816 1817 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1818 1819 // Return the base type index if there aren't any modifiers. For example, the 1820 // metadata could contain restrict wrappers around non-pointer types. 1821 if (Mods == ModifierOptions::None) 1822 return ModifiedTI; 1823 1824 ModifierRecord MR(ModifiedTI, Mods); 1825 return TypeTable.writeLeafType(MR); 1826 } 1827 1828 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1829 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1830 for (const DIType *ArgType : Ty->getTypeArray()) 1831 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1832 1833 // MSVC uses type none for variadic argument. 1834 if (ReturnAndArgTypeIndices.size() > 1 && 1835 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1836 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1837 } 1838 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1839 ArrayRef<TypeIndex> ArgTypeIndices = None; 1840 if (!ReturnAndArgTypeIndices.empty()) { 1841 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1842 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1843 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1844 } 1845 1846 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1847 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1848 1849 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1850 1851 FunctionOptions FO = getFunctionOptions(Ty); 1852 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 1853 ArgListIndex); 1854 return TypeTable.writeLeafType(Procedure); 1855 } 1856 1857 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 1858 const DIType *ClassTy, 1859 int ThisAdjustment, 1860 bool IsStaticMethod, 1861 FunctionOptions FO) { 1862 // Lower the containing class type. 1863 TypeIndex ClassType = getTypeIndex(ClassTy); 1864 1865 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 1866 1867 unsigned Index = 0; 1868 SmallVector<TypeIndex, 8> ArgTypeIndices; 1869 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1870 if (ReturnAndArgs.size() > Index) { 1871 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 1872 } 1873 1874 // If the first argument is a pointer type and this isn't a static method, 1875 // treat it as the special 'this' parameter, which is encoded separately from 1876 // the arguments. 1877 TypeIndex ThisTypeIndex; 1878 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 1879 if (const DIDerivedType *PtrTy = 1880 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 1881 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 1882 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 1883 Index++; 1884 } 1885 } 1886 } 1887 1888 while (Index < ReturnAndArgs.size()) 1889 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 1890 1891 // MSVC uses type none for variadic argument. 1892 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 1893 ArgTypeIndices.back() = TypeIndex::None(); 1894 1895 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1896 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1897 1898 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1899 1900 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1901 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1902 return TypeTable.writeLeafType(MFR); 1903 } 1904 1905 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1906 unsigned VSlotCount = 1907 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1908 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1909 1910 VFTableShapeRecord VFTSR(Slots); 1911 return TypeTable.writeLeafType(VFTSR); 1912 } 1913 1914 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1915 switch (Flags & DINode::FlagAccessibility) { 1916 case DINode::FlagPrivate: return MemberAccess::Private; 1917 case DINode::FlagPublic: return MemberAccess::Public; 1918 case DINode::FlagProtected: return MemberAccess::Protected; 1919 case 0: 1920 // If there was no explicit access control, provide the default for the tag. 1921 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1922 : MemberAccess::Public; 1923 } 1924 llvm_unreachable("access flags are exclusive"); 1925 } 1926 1927 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1928 if (SP->isArtificial()) 1929 return MethodOptions::CompilerGenerated; 1930 1931 // FIXME: Handle other MethodOptions. 1932 1933 return MethodOptions::None; 1934 } 1935 1936 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 1937 bool Introduced) { 1938 if (SP->getFlags() & DINode::FlagStaticMember) 1939 return MethodKind::Static; 1940 1941 switch (SP->getVirtuality()) { 1942 case dwarf::DW_VIRTUALITY_none: 1943 break; 1944 case dwarf::DW_VIRTUALITY_virtual: 1945 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 1946 case dwarf::DW_VIRTUALITY_pure_virtual: 1947 return Introduced ? MethodKind::PureIntroducingVirtual 1948 : MethodKind::PureVirtual; 1949 default: 1950 llvm_unreachable("unhandled virtuality case"); 1951 } 1952 1953 return MethodKind::Vanilla; 1954 } 1955 1956 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 1957 switch (Ty->getTag()) { 1958 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 1959 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 1960 } 1961 llvm_unreachable("unexpected tag"); 1962 } 1963 1964 /// Return ClassOptions that should be present on both the forward declaration 1965 /// and the defintion of a tag type. 1966 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 1967 ClassOptions CO = ClassOptions::None; 1968 1969 // MSVC always sets this flag, even for local types. Clang doesn't always 1970 // appear to give every type a linkage name, which may be problematic for us. 1971 // FIXME: Investigate the consequences of not following them here. 1972 if (!Ty->getIdentifier().empty()) 1973 CO |= ClassOptions::HasUniqueName; 1974 1975 // Put the Nested flag on a type if it appears immediately inside a tag type. 1976 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 1977 // here. That flag is only set on definitions, and not forward declarations. 1978 const DIScope *ImmediateScope = Ty->getScope(); 1979 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 1980 CO |= ClassOptions::Nested; 1981 1982 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 1983 // type only when it has an immediate function scope. Clang never puts enums 1984 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 1985 // always in function, class, or file scopes. 1986 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 1987 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 1988 CO |= ClassOptions::Scoped; 1989 } else { 1990 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 1991 Scope = Scope->getScope()) { 1992 if (isa<DISubprogram>(Scope)) { 1993 CO |= ClassOptions::Scoped; 1994 break; 1995 } 1996 } 1997 } 1998 1999 return CO; 2000 } 2001 2002 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2003 switch (Ty->getTag()) { 2004 case dwarf::DW_TAG_class_type: 2005 case dwarf::DW_TAG_structure_type: 2006 case dwarf::DW_TAG_union_type: 2007 case dwarf::DW_TAG_enumeration_type: 2008 break; 2009 default: 2010 return; 2011 } 2012 2013 if (const auto *File = Ty->getFile()) { 2014 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2015 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2016 2017 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2018 TypeTable.writeLeafType(USLR); 2019 } 2020 } 2021 2022 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2023 ClassOptions CO = getCommonClassOptions(Ty); 2024 TypeIndex FTI; 2025 unsigned EnumeratorCount = 0; 2026 2027 if (Ty->isForwardDecl()) { 2028 CO |= ClassOptions::ForwardReference; 2029 } else { 2030 ContinuationRecordBuilder ContinuationBuilder; 2031 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2032 for (const DINode *Element : Ty->getElements()) { 2033 // We assume that the frontend provides all members in source declaration 2034 // order, which is what MSVC does. 2035 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2036 EnumeratorRecord ER(MemberAccess::Public, 2037 APSInt::getUnsigned(Enumerator->getValue()), 2038 Enumerator->getName()); 2039 ContinuationBuilder.writeMemberType(ER); 2040 EnumeratorCount++; 2041 } 2042 } 2043 FTI = TypeTable.insertRecord(ContinuationBuilder); 2044 } 2045 2046 std::string FullName = getFullyQualifiedName(Ty); 2047 2048 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2049 getTypeIndex(Ty->getBaseType())); 2050 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2051 2052 addUDTSrcLine(Ty, EnumTI); 2053 2054 return EnumTI; 2055 } 2056 2057 //===----------------------------------------------------------------------===// 2058 // ClassInfo 2059 //===----------------------------------------------------------------------===// 2060 2061 struct llvm::ClassInfo { 2062 struct MemberInfo { 2063 const DIDerivedType *MemberTypeNode; 2064 uint64_t BaseOffset; 2065 }; 2066 // [MemberInfo] 2067 using MemberList = std::vector<MemberInfo>; 2068 2069 using MethodsList = TinyPtrVector<const DISubprogram *>; 2070 // MethodName -> MethodsList 2071 using MethodsMap = MapVector<MDString *, MethodsList>; 2072 2073 /// Base classes. 2074 std::vector<const DIDerivedType *> Inheritance; 2075 2076 /// Direct members. 2077 MemberList Members; 2078 // Direct overloaded methods gathered by name. 2079 MethodsMap Methods; 2080 2081 TypeIndex VShapeTI; 2082 2083 std::vector<const DIType *> NestedTypes; 2084 }; 2085 2086 void CodeViewDebug::clear() { 2087 assert(CurFn == nullptr); 2088 FileIdMap.clear(); 2089 FnDebugInfo.clear(); 2090 FileToFilepathMap.clear(); 2091 LocalUDTs.clear(); 2092 GlobalUDTs.clear(); 2093 TypeIndices.clear(); 2094 CompleteTypeIndices.clear(); 2095 ScopeGlobals.clear(); 2096 } 2097 2098 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2099 const DIDerivedType *DDTy) { 2100 if (!DDTy->getName().empty()) { 2101 Info.Members.push_back({DDTy, 0}); 2102 return; 2103 } 2104 2105 // An unnamed member may represent a nested struct or union. Attempt to 2106 // interpret the unnamed member as a DICompositeType possibly wrapped in 2107 // qualifier types. Add all the indirect fields to the current record if that 2108 // succeeds, and drop the member if that fails. 2109 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2110 uint64_t Offset = DDTy->getOffsetInBits(); 2111 const DIType *Ty = DDTy->getBaseType(); 2112 bool FullyResolved = false; 2113 while (!FullyResolved) { 2114 switch (Ty->getTag()) { 2115 case dwarf::DW_TAG_const_type: 2116 case dwarf::DW_TAG_volatile_type: 2117 // FIXME: we should apply the qualifier types to the indirect fields 2118 // rather than dropping them. 2119 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2120 break; 2121 default: 2122 FullyResolved = true; 2123 break; 2124 } 2125 } 2126 2127 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2128 if (!DCTy) 2129 return; 2130 2131 ClassInfo NestedInfo = collectClassInfo(DCTy); 2132 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2133 Info.Members.push_back( 2134 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2135 } 2136 2137 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2138 ClassInfo Info; 2139 // Add elements to structure type. 2140 DINodeArray Elements = Ty->getElements(); 2141 for (auto *Element : Elements) { 2142 // We assume that the frontend provides all members in source declaration 2143 // order, which is what MSVC does. 2144 if (!Element) 2145 continue; 2146 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2147 Info.Methods[SP->getRawName()].push_back(SP); 2148 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2149 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2150 collectMemberInfo(Info, DDTy); 2151 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2152 Info.Inheritance.push_back(DDTy); 2153 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2154 DDTy->getName() == "__vtbl_ptr_type") { 2155 Info.VShapeTI = getTypeIndex(DDTy); 2156 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2157 Info.NestedTypes.push_back(DDTy); 2158 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2159 // Ignore friend members. It appears that MSVC emitted info about 2160 // friends in the past, but modern versions do not. 2161 } 2162 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2163 Info.NestedTypes.push_back(Composite); 2164 } 2165 // Skip other unrecognized kinds of elements. 2166 } 2167 return Info; 2168 } 2169 2170 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2171 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2172 // if a complete type should be emitted instead of a forward reference. 2173 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2174 !Ty->isForwardDecl(); 2175 } 2176 2177 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2178 // Emit the complete type for unnamed structs. C++ classes with methods 2179 // which have a circular reference back to the class type are expected to 2180 // be named by the front-end and should not be "unnamed". C unnamed 2181 // structs should not have circular references. 2182 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2183 // If this unnamed complete type is already in the process of being defined 2184 // then the description of the type is malformed and cannot be emitted 2185 // into CodeView correctly so report a fatal error. 2186 auto I = CompleteTypeIndices.find(Ty); 2187 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2188 report_fatal_error("cannot debug circular reference to unnamed type"); 2189 return getCompleteTypeIndex(Ty); 2190 } 2191 2192 // First, construct the forward decl. Don't look into Ty to compute the 2193 // forward decl options, since it might not be available in all TUs. 2194 TypeRecordKind Kind = getRecordKind(Ty); 2195 ClassOptions CO = 2196 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2197 std::string FullName = getFullyQualifiedName(Ty); 2198 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2199 FullName, Ty->getIdentifier()); 2200 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2201 if (!Ty->isForwardDecl()) 2202 DeferredCompleteTypes.push_back(Ty); 2203 return FwdDeclTI; 2204 } 2205 2206 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2207 // Construct the field list and complete type record. 2208 TypeRecordKind Kind = getRecordKind(Ty); 2209 ClassOptions CO = getCommonClassOptions(Ty); 2210 TypeIndex FieldTI; 2211 TypeIndex VShapeTI; 2212 unsigned FieldCount; 2213 bool ContainsNestedClass; 2214 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2215 lowerRecordFieldList(Ty); 2216 2217 if (ContainsNestedClass) 2218 CO |= ClassOptions::ContainsNestedClass; 2219 2220 // MSVC appears to set this flag by searching any destructor or method with 2221 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2222 // the members, however special member functions are not yet emitted into 2223 // debug information. For now checking a class's non-triviality seems enough. 2224 // FIXME: not true for a nested unnamed struct. 2225 if (isNonTrivial(Ty)) 2226 CO |= ClassOptions::HasConstructorOrDestructor; 2227 2228 std::string FullName = getFullyQualifiedName(Ty); 2229 2230 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2231 2232 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2233 SizeInBytes, FullName, Ty->getIdentifier()); 2234 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2235 2236 addUDTSrcLine(Ty, ClassTI); 2237 2238 addToUDTs(Ty); 2239 2240 return ClassTI; 2241 } 2242 2243 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2244 // Emit the complete type for unnamed unions. 2245 if (shouldAlwaysEmitCompleteClassType(Ty)) 2246 return getCompleteTypeIndex(Ty); 2247 2248 ClassOptions CO = 2249 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2250 std::string FullName = getFullyQualifiedName(Ty); 2251 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2252 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2253 if (!Ty->isForwardDecl()) 2254 DeferredCompleteTypes.push_back(Ty); 2255 return FwdDeclTI; 2256 } 2257 2258 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2259 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2260 TypeIndex FieldTI; 2261 unsigned FieldCount; 2262 bool ContainsNestedClass; 2263 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2264 lowerRecordFieldList(Ty); 2265 2266 if (ContainsNestedClass) 2267 CO |= ClassOptions::ContainsNestedClass; 2268 2269 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2270 std::string FullName = getFullyQualifiedName(Ty); 2271 2272 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2273 Ty->getIdentifier()); 2274 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2275 2276 addUDTSrcLine(Ty, UnionTI); 2277 2278 addToUDTs(Ty); 2279 2280 return UnionTI; 2281 } 2282 2283 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2284 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2285 // Manually count members. MSVC appears to count everything that generates a 2286 // field list record. Each individual overload in a method overload group 2287 // contributes to this count, even though the overload group is a single field 2288 // list record. 2289 unsigned MemberCount = 0; 2290 ClassInfo Info = collectClassInfo(Ty); 2291 ContinuationRecordBuilder ContinuationBuilder; 2292 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2293 2294 // Create base classes. 2295 for (const DIDerivedType *I : Info.Inheritance) { 2296 if (I->getFlags() & DINode::FlagVirtual) { 2297 // Virtual base. 2298 unsigned VBPtrOffset = I->getVBPtrOffset(); 2299 // FIXME: Despite the accessor name, the offset is really in bytes. 2300 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2301 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2302 ? TypeRecordKind::IndirectVirtualBaseClass 2303 : TypeRecordKind::VirtualBaseClass; 2304 VirtualBaseClassRecord VBCR( 2305 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2306 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2307 VBTableIndex); 2308 2309 ContinuationBuilder.writeMemberType(VBCR); 2310 MemberCount++; 2311 } else { 2312 assert(I->getOffsetInBits() % 8 == 0 && 2313 "bases must be on byte boundaries"); 2314 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2315 getTypeIndex(I->getBaseType()), 2316 I->getOffsetInBits() / 8); 2317 ContinuationBuilder.writeMemberType(BCR); 2318 MemberCount++; 2319 } 2320 } 2321 2322 // Create members. 2323 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2324 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2325 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2326 StringRef MemberName = Member->getName(); 2327 MemberAccess Access = 2328 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2329 2330 if (Member->isStaticMember()) { 2331 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2332 ContinuationBuilder.writeMemberType(SDMR); 2333 MemberCount++; 2334 continue; 2335 } 2336 2337 // Virtual function pointer member. 2338 if ((Member->getFlags() & DINode::FlagArtificial) && 2339 Member->getName().startswith("_vptr$")) { 2340 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2341 ContinuationBuilder.writeMemberType(VFPR); 2342 MemberCount++; 2343 continue; 2344 } 2345 2346 // Data member. 2347 uint64_t MemberOffsetInBits = 2348 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2349 if (Member->isBitField()) { 2350 uint64_t StartBitOffset = MemberOffsetInBits; 2351 if (const auto *CI = 2352 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2353 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2354 } 2355 StartBitOffset -= MemberOffsetInBits; 2356 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2357 StartBitOffset); 2358 MemberBaseType = TypeTable.writeLeafType(BFR); 2359 } 2360 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2361 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2362 MemberName); 2363 ContinuationBuilder.writeMemberType(DMR); 2364 MemberCount++; 2365 } 2366 2367 // Create methods 2368 for (auto &MethodItr : Info.Methods) { 2369 StringRef Name = MethodItr.first->getString(); 2370 2371 std::vector<OneMethodRecord> Methods; 2372 for (const DISubprogram *SP : MethodItr.second) { 2373 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2374 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2375 2376 unsigned VFTableOffset = -1; 2377 if (Introduced) 2378 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2379 2380 Methods.push_back(OneMethodRecord( 2381 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2382 translateMethodKindFlags(SP, Introduced), 2383 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2384 MemberCount++; 2385 } 2386 assert(!Methods.empty() && "Empty methods map entry"); 2387 if (Methods.size() == 1) 2388 ContinuationBuilder.writeMemberType(Methods[0]); 2389 else { 2390 // FIXME: Make this use its own ContinuationBuilder so that 2391 // MethodOverloadList can be split correctly. 2392 MethodOverloadListRecord MOLR(Methods); 2393 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2394 2395 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2396 ContinuationBuilder.writeMemberType(OMR); 2397 } 2398 } 2399 2400 // Create nested classes. 2401 for (const DIType *Nested : Info.NestedTypes) { 2402 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2403 ContinuationBuilder.writeMemberType(R); 2404 MemberCount++; 2405 } 2406 2407 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2408 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2409 !Info.NestedTypes.empty()); 2410 } 2411 2412 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2413 if (!VBPType.getIndex()) { 2414 // Make a 'const int *' type. 2415 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2416 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2417 2418 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2419 : PointerKind::Near32; 2420 PointerMode PM = PointerMode::Pointer; 2421 PointerOptions PO = PointerOptions::None; 2422 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2423 VBPType = TypeTable.writeLeafType(PR); 2424 } 2425 2426 return VBPType; 2427 } 2428 2429 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2430 // The null DIType is the void type. Don't try to hash it. 2431 if (!Ty) 2432 return TypeIndex::Void(); 2433 2434 // Check if we've already translated this type. Don't try to do a 2435 // get-or-create style insertion that caches the hash lookup across the 2436 // lowerType call. It will update the TypeIndices map. 2437 auto I = TypeIndices.find({Ty, ClassTy}); 2438 if (I != TypeIndices.end()) 2439 return I->second; 2440 2441 TypeLoweringScope S(*this); 2442 TypeIndex TI = lowerType(Ty, ClassTy); 2443 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2444 } 2445 2446 codeview::TypeIndex 2447 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2448 const DISubroutineType *SubroutineTy) { 2449 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2450 "this type must be a pointer type"); 2451 2452 PointerOptions Options = PointerOptions::None; 2453 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2454 Options = PointerOptions::LValueRefThisPointer; 2455 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2456 Options = PointerOptions::RValueRefThisPointer; 2457 2458 // Check if we've already translated this type. If there is no ref qualifier 2459 // on the function then we look up this pointer type with no associated class 2460 // so that the TypeIndex for the this pointer can be shared with the type 2461 // index for other pointers to this class type. If there is a ref qualifier 2462 // then we lookup the pointer using the subroutine as the parent type. 2463 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2464 if (I != TypeIndices.end()) 2465 return I->second; 2466 2467 TypeLoweringScope S(*this); 2468 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2469 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2470 } 2471 2472 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2473 PointerRecord PR(getTypeIndex(Ty), 2474 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2475 : PointerKind::Near32, 2476 PointerMode::LValueReference, PointerOptions::None, 2477 Ty->getSizeInBits() / 8); 2478 return TypeTable.writeLeafType(PR); 2479 } 2480 2481 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2482 // The null DIType is the void type. Don't try to hash it. 2483 if (!Ty) 2484 return TypeIndex::Void(); 2485 2486 // Look through typedefs when getting the complete type index. Call 2487 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2488 // emitted only once. 2489 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2490 (void)getTypeIndex(Ty); 2491 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2492 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2493 2494 // If this is a non-record type, the complete type index is the same as the 2495 // normal type index. Just call getTypeIndex. 2496 switch (Ty->getTag()) { 2497 case dwarf::DW_TAG_class_type: 2498 case dwarf::DW_TAG_structure_type: 2499 case dwarf::DW_TAG_union_type: 2500 break; 2501 default: 2502 return getTypeIndex(Ty); 2503 } 2504 2505 const auto *CTy = cast<DICompositeType>(Ty); 2506 2507 TypeLoweringScope S(*this); 2508 2509 // Make sure the forward declaration is emitted first. It's unclear if this 2510 // is necessary, but MSVC does it, and we should follow suit until we can show 2511 // otherwise. 2512 // We only emit a forward declaration for named types. 2513 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2514 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2515 2516 // Just use the forward decl if we don't have complete type info. This 2517 // might happen if the frontend is using modules and expects the complete 2518 // definition to be emitted elsewhere. 2519 if (CTy->isForwardDecl()) 2520 return FwdDeclTI; 2521 } 2522 2523 // Check if we've already translated the complete record type. 2524 // Insert the type with a null TypeIndex to signify that the type is currently 2525 // being lowered. 2526 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2527 if (!InsertResult.second) 2528 return InsertResult.first->second; 2529 2530 TypeIndex TI; 2531 switch (CTy->getTag()) { 2532 case dwarf::DW_TAG_class_type: 2533 case dwarf::DW_TAG_structure_type: 2534 TI = lowerCompleteTypeClass(CTy); 2535 break; 2536 case dwarf::DW_TAG_union_type: 2537 TI = lowerCompleteTypeUnion(CTy); 2538 break; 2539 default: 2540 llvm_unreachable("not a record"); 2541 } 2542 2543 // Update the type index associated with this CompositeType. This cannot 2544 // use the 'InsertResult' iterator above because it is potentially 2545 // invalidated by map insertions which can occur while lowering the class 2546 // type above. 2547 CompleteTypeIndices[CTy] = TI; 2548 return TI; 2549 } 2550 2551 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2552 /// and do this until fixpoint, as each complete record type typically 2553 /// references 2554 /// many other record types. 2555 void CodeViewDebug::emitDeferredCompleteTypes() { 2556 SmallVector<const DICompositeType *, 4> TypesToEmit; 2557 while (!DeferredCompleteTypes.empty()) { 2558 std::swap(DeferredCompleteTypes, TypesToEmit); 2559 for (const DICompositeType *RecordTy : TypesToEmit) 2560 getCompleteTypeIndex(RecordTy); 2561 TypesToEmit.clear(); 2562 } 2563 } 2564 2565 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2566 ArrayRef<LocalVariable> Locals) { 2567 // Get the sorted list of parameters and emit them first. 2568 SmallVector<const LocalVariable *, 6> Params; 2569 for (const LocalVariable &L : Locals) 2570 if (L.DIVar->isParameter()) 2571 Params.push_back(&L); 2572 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2573 return L->DIVar->getArg() < R->DIVar->getArg(); 2574 }); 2575 for (const LocalVariable *L : Params) 2576 emitLocalVariable(FI, *L); 2577 2578 // Next emit all non-parameters in the order that we found them. 2579 for (const LocalVariable &L : Locals) 2580 if (!L.DIVar->isParameter()) 2581 emitLocalVariable(FI, L); 2582 } 2583 2584 /// Only call this on endian-specific types like ulittle16_t and little32_t, or 2585 /// structs composed of them. 2586 template <typename T> 2587 static void copyBytesForDefRange(SmallString<20> &BytePrefix, 2588 SymbolKind SymKind, const T &DefRangeHeader) { 2589 BytePrefix.resize(2 + sizeof(T)); 2590 ulittle16_t SymKindLE = ulittle16_t(SymKind); 2591 memcpy(&BytePrefix[0], &SymKindLE, 2); 2592 memcpy(&BytePrefix[2], &DefRangeHeader, sizeof(T)); 2593 } 2594 2595 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2596 const LocalVariable &Var) { 2597 // LocalSym record, see SymbolRecord.h for more info. 2598 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2599 2600 LocalSymFlags Flags = LocalSymFlags::None; 2601 if (Var.DIVar->isParameter()) 2602 Flags |= LocalSymFlags::IsParameter; 2603 if (Var.DefRanges.empty()) 2604 Flags |= LocalSymFlags::IsOptimizedOut; 2605 2606 OS.AddComment("TypeIndex"); 2607 TypeIndex TI = Var.UseReferenceType 2608 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2609 : getCompleteTypeIndex(Var.DIVar->getType()); 2610 OS.EmitIntValue(TI.getIndex(), 4); 2611 OS.AddComment("Flags"); 2612 OS.EmitIntValue(static_cast<uint16_t>(Flags), 2); 2613 // Truncate the name so we won't overflow the record length field. 2614 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2615 endSymbolRecord(LocalEnd); 2616 2617 // Calculate the on disk prefix of the appropriate def range record. The 2618 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2619 // should be big enough to hold all forms without memory allocation. 2620 SmallString<20> BytePrefix; 2621 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2622 BytePrefix.clear(); 2623 if (DefRange.InMemory) { 2624 int Offset = DefRange.DataOffset; 2625 unsigned Reg = DefRange.CVRegister; 2626 2627 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2628 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2629 // instead. In frames without stack realignment, $T0 will be the CFA. 2630 if (RegisterId(Reg) == RegisterId::ESP) { 2631 Reg = unsigned(RegisterId::VFRAME); 2632 Offset += FI.OffsetAdjustment; 2633 } 2634 2635 // If we can use the chosen frame pointer for the frame and this isn't a 2636 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2637 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2638 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2639 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2640 (bool(Flags & LocalSymFlags::IsParameter) 2641 ? (EncFP == FI.EncodedParamFramePtrReg) 2642 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2643 little32_t FPOffset = little32_t(Offset); 2644 copyBytesForDefRange(BytePrefix, S_DEFRANGE_FRAMEPOINTER_REL, FPOffset); 2645 } else { 2646 uint16_t RegRelFlags = 0; 2647 if (DefRange.IsSubfield) { 2648 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2649 (DefRange.StructOffset 2650 << DefRangeRegisterRelSym::OffsetInParentShift); 2651 } 2652 DefRangeRegisterRelSym::Header DRHdr; 2653 DRHdr.Register = Reg; 2654 DRHdr.Flags = RegRelFlags; 2655 DRHdr.BasePointerOffset = Offset; 2656 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER_REL, DRHdr); 2657 } 2658 } else { 2659 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2660 if (DefRange.IsSubfield) { 2661 DefRangeSubfieldRegisterSym::Header DRHdr; 2662 DRHdr.Register = DefRange.CVRegister; 2663 DRHdr.MayHaveNoName = 0; 2664 DRHdr.OffsetInParent = DefRange.StructOffset; 2665 copyBytesForDefRange(BytePrefix, S_DEFRANGE_SUBFIELD_REGISTER, DRHdr); 2666 } else { 2667 DefRangeRegisterSym::Header DRHdr; 2668 DRHdr.Register = DefRange.CVRegister; 2669 DRHdr.MayHaveNoName = 0; 2670 copyBytesForDefRange(BytePrefix, S_DEFRANGE_REGISTER, DRHdr); 2671 } 2672 } 2673 OS.EmitCVDefRangeDirective(DefRange.Ranges, BytePrefix); 2674 } 2675 } 2676 2677 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2678 const FunctionInfo& FI) { 2679 for (LexicalBlock *Block : Blocks) 2680 emitLexicalBlock(*Block, FI); 2681 } 2682 2683 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2684 /// lexical block scope. 2685 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2686 const FunctionInfo& FI) { 2687 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2688 OS.AddComment("PtrParent"); 2689 OS.EmitIntValue(0, 4); // PtrParent 2690 OS.AddComment("PtrEnd"); 2691 OS.EmitIntValue(0, 4); // PtrEnd 2692 OS.AddComment("Code size"); 2693 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2694 OS.AddComment("Function section relative address"); 2695 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2696 OS.AddComment("Function section index"); 2697 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2698 OS.AddComment("Lexical block name"); 2699 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2700 endSymbolRecord(RecordEnd); 2701 2702 // Emit variables local to this lexical block. 2703 emitLocalVariableList(FI, Block.Locals); 2704 emitGlobalVariableList(Block.Globals); 2705 2706 // Emit lexical blocks contained within this block. 2707 emitLexicalBlockList(Block.Children, FI); 2708 2709 // Close the lexical block scope. 2710 emitEndSymbolRecord(SymbolKind::S_END); 2711 } 2712 2713 /// Convenience routine for collecting lexical block information for a list 2714 /// of lexical scopes. 2715 void CodeViewDebug::collectLexicalBlockInfo( 2716 SmallVectorImpl<LexicalScope *> &Scopes, 2717 SmallVectorImpl<LexicalBlock *> &Blocks, 2718 SmallVectorImpl<LocalVariable> &Locals, 2719 SmallVectorImpl<CVGlobalVariable> &Globals) { 2720 for (LexicalScope *Scope : Scopes) 2721 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2722 } 2723 2724 /// Populate the lexical blocks and local variable lists of the parent with 2725 /// information about the specified lexical scope. 2726 void CodeViewDebug::collectLexicalBlockInfo( 2727 LexicalScope &Scope, 2728 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2729 SmallVectorImpl<LocalVariable> &ParentLocals, 2730 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2731 if (Scope.isAbstractScope()) 2732 return; 2733 2734 // Gather information about the lexical scope including local variables, 2735 // global variables, and address ranges. 2736 bool IgnoreScope = false; 2737 auto LI = ScopeVariables.find(&Scope); 2738 SmallVectorImpl<LocalVariable> *Locals = 2739 LI != ScopeVariables.end() ? &LI->second : nullptr; 2740 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2741 SmallVectorImpl<CVGlobalVariable> *Globals = 2742 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2743 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2744 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2745 2746 // Ignore lexical scopes which do not contain variables. 2747 if (!Locals && !Globals) 2748 IgnoreScope = true; 2749 2750 // Ignore lexical scopes which are not lexical blocks. 2751 if (!DILB) 2752 IgnoreScope = true; 2753 2754 // Ignore scopes which have too many address ranges to represent in the 2755 // current CodeView format or do not have a valid address range. 2756 // 2757 // For lexical scopes with multiple address ranges you may be tempted to 2758 // construct a single range covering every instruction where the block is 2759 // live and everything in between. Unfortunately, Visual Studio only 2760 // displays variables from the first matching lexical block scope. If the 2761 // first lexical block contains exception handling code or cold code which 2762 // is moved to the bottom of the routine creating a single range covering 2763 // nearly the entire routine, then it will hide all other lexical blocks 2764 // and the variables they contain. 2765 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2766 IgnoreScope = true; 2767 2768 if (IgnoreScope) { 2769 // This scope can be safely ignored and eliminating it will reduce the 2770 // size of the debug information. Be sure to collect any variable and scope 2771 // information from the this scope or any of its children and collapse them 2772 // into the parent scope. 2773 if (Locals) 2774 ParentLocals.append(Locals->begin(), Locals->end()); 2775 if (Globals) 2776 ParentGlobals.append(Globals->begin(), Globals->end()); 2777 collectLexicalBlockInfo(Scope.getChildren(), 2778 ParentBlocks, 2779 ParentLocals, 2780 ParentGlobals); 2781 return; 2782 } 2783 2784 // Create a new CodeView lexical block for this lexical scope. If we've 2785 // seen this DILexicalBlock before then the scope tree is malformed and 2786 // we can handle this gracefully by not processing it a second time. 2787 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2788 if (!BlockInsertion.second) 2789 return; 2790 2791 // Create a lexical block containing the variables and collect the the 2792 // lexical block information for the children. 2793 const InsnRange &Range = Ranges.front(); 2794 assert(Range.first && Range.second); 2795 LexicalBlock &Block = BlockInsertion.first->second; 2796 Block.Begin = getLabelBeforeInsn(Range.first); 2797 Block.End = getLabelAfterInsn(Range.second); 2798 assert(Block.Begin && "missing label for scope begin"); 2799 assert(Block.End && "missing label for scope end"); 2800 Block.Name = DILB->getName(); 2801 if (Locals) 2802 Block.Locals = std::move(*Locals); 2803 if (Globals) 2804 Block.Globals = std::move(*Globals); 2805 ParentBlocks.push_back(&Block); 2806 collectLexicalBlockInfo(Scope.getChildren(), 2807 Block.Children, 2808 Block.Locals, 2809 Block.Globals); 2810 } 2811 2812 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2813 const Function &GV = MF->getFunction(); 2814 assert(FnDebugInfo.count(&GV)); 2815 assert(CurFn == FnDebugInfo[&GV].get()); 2816 2817 collectVariableInfo(GV.getSubprogram()); 2818 2819 // Build the lexical block structure to emit for this routine. 2820 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2821 collectLexicalBlockInfo(*CFS, 2822 CurFn->ChildBlocks, 2823 CurFn->Locals, 2824 CurFn->Globals); 2825 2826 // Clear the scope and variable information from the map which will not be 2827 // valid after we have finished processing this routine. This also prepares 2828 // the map for the subsequent routine. 2829 ScopeVariables.clear(); 2830 2831 // Don't emit anything if we don't have any line tables. 2832 // Thunks are compiler-generated and probably won't have source correlation. 2833 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2834 FnDebugInfo.erase(&GV); 2835 CurFn = nullptr; 2836 return; 2837 } 2838 2839 CurFn->Annotations = MF->getCodeViewAnnotations(); 2840 CurFn->HeapAllocSites = MF->getCodeViewHeapAllocSites(); 2841 2842 CurFn->End = Asm->getFunctionEnd(); 2843 2844 CurFn = nullptr; 2845 } 2846 2847 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2848 DebugHandlerBase::beginInstruction(MI); 2849 2850 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2851 if (!Asm || !CurFn || MI->isDebugInstr() || 2852 MI->getFlag(MachineInstr::FrameSetup)) 2853 return; 2854 2855 // If the first instruction of a new MBB has no location, find the first 2856 // instruction with a location and use that. 2857 DebugLoc DL = MI->getDebugLoc(); 2858 if (!DL && MI->getParent() != PrevInstBB) { 2859 for (const auto &NextMI : *MI->getParent()) { 2860 if (NextMI.isDebugInstr()) 2861 continue; 2862 DL = NextMI.getDebugLoc(); 2863 if (DL) 2864 break; 2865 } 2866 } 2867 PrevInstBB = MI->getParent(); 2868 2869 // If we still don't have a debug location, don't record a location. 2870 if (!DL) 2871 return; 2872 2873 maybeRecordLocation(DL, Asm->MF); 2874 } 2875 2876 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2877 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2878 *EndLabel = MMI->getContext().createTempSymbol(); 2879 OS.EmitIntValue(unsigned(Kind), 4); 2880 OS.AddComment("Subsection size"); 2881 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2882 OS.EmitLabel(BeginLabel); 2883 return EndLabel; 2884 } 2885 2886 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2887 OS.EmitLabel(EndLabel); 2888 // Every subsection must be aligned to a 4-byte boundary. 2889 OS.EmitValueToAlignment(4); 2890 } 2891 2892 static StringRef getSymbolName(SymbolKind SymKind) { 2893 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 2894 if (EE.Value == SymKind) 2895 return EE.Name; 2896 return ""; 2897 } 2898 2899 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 2900 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2901 *EndLabel = MMI->getContext().createTempSymbol(); 2902 OS.AddComment("Record length"); 2903 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 2904 OS.EmitLabel(BeginLabel); 2905 if (OS.isVerboseAsm()) 2906 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 2907 OS.EmitIntValue(unsigned(SymKind), 2); 2908 return EndLabel; 2909 } 2910 2911 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 2912 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 2913 // an extra copy of every symbol record in LLD. This increases object file 2914 // size by less than 1% in the clang build, and is compatible with the Visual 2915 // C++ linker. 2916 OS.EmitValueToAlignment(4); 2917 OS.EmitLabel(SymEnd); 2918 } 2919 2920 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 2921 OS.AddComment("Record length"); 2922 OS.EmitIntValue(2, 2); 2923 if (OS.isVerboseAsm()) 2924 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 2925 OS.EmitIntValue(unsigned(EndKind), 2); // Record Kind 2926 } 2927 2928 void CodeViewDebug::emitDebugInfoForUDTs( 2929 ArrayRef<std::pair<std::string, const DIType *>> UDTs) { 2930 for (const auto &UDT : UDTs) { 2931 const DIType *T = UDT.second; 2932 assert(shouldEmitUdt(T)); 2933 2934 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 2935 OS.AddComment("Type"); 2936 OS.EmitIntValue(getCompleteTypeIndex(T).getIndex(), 4); 2937 emitNullTerminatedSymbolName(OS, UDT.first); 2938 endSymbolRecord(UDTRecordEnd); 2939 } 2940 } 2941 2942 void CodeViewDebug::collectGlobalVariableInfo() { 2943 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 2944 GlobalMap; 2945 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 2946 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 2947 GV.getDebugInfo(GVEs); 2948 for (const auto *GVE : GVEs) 2949 GlobalMap[GVE] = &GV; 2950 } 2951 2952 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 2953 for (const MDNode *Node : CUs->operands()) { 2954 const auto *CU = cast<DICompileUnit>(Node); 2955 for (const auto *GVE : CU->getGlobalVariables()) { 2956 const DIGlobalVariable *DIGV = GVE->getVariable(); 2957 const DIExpression *DIE = GVE->getExpression(); 2958 2959 // Emit constant global variables in a global symbol section. 2960 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 2961 CVGlobalVariable CVGV = {DIGV, DIE}; 2962 GlobalVariables.emplace_back(std::move(CVGV)); 2963 } 2964 2965 const auto *GV = GlobalMap.lookup(GVE); 2966 if (!GV || GV->isDeclarationForLinker()) 2967 continue; 2968 2969 DIScope *Scope = DIGV->getScope(); 2970 SmallVector<CVGlobalVariable, 1> *VariableList; 2971 if (Scope && isa<DILocalScope>(Scope)) { 2972 // Locate a global variable list for this scope, creating one if 2973 // necessary. 2974 auto Insertion = ScopeGlobals.insert( 2975 {Scope, std::unique_ptr<GlobalVariableList>()}); 2976 if (Insertion.second) 2977 Insertion.first->second = llvm::make_unique<GlobalVariableList>(); 2978 VariableList = Insertion.first->second.get(); 2979 } else if (GV->hasComdat()) 2980 // Emit this global variable into a COMDAT section. 2981 VariableList = &ComdatVariables; 2982 else 2983 // Emit this global variable in a single global symbol section. 2984 VariableList = &GlobalVariables; 2985 CVGlobalVariable CVGV = {DIGV, GV}; 2986 VariableList->emplace_back(std::move(CVGV)); 2987 } 2988 } 2989 } 2990 2991 void CodeViewDebug::emitDebugInfoForGlobals() { 2992 // First, emit all globals that are not in a comdat in a single symbol 2993 // substream. MSVC doesn't like it if the substream is empty, so only open 2994 // it if we have at least one global to emit. 2995 switchToDebugSectionForSymbol(nullptr); 2996 if (!GlobalVariables.empty()) { 2997 OS.AddComment("Symbol subsection for globals"); 2998 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 2999 emitGlobalVariableList(GlobalVariables); 3000 endCVSubsection(EndLabel); 3001 } 3002 3003 // Second, emit each global that is in a comdat into its own .debug$S 3004 // section along with its own symbol substream. 3005 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3006 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3007 MCSymbol *GVSym = Asm->getSymbol(GV); 3008 OS.AddComment("Symbol subsection for " + 3009 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3010 switchToDebugSectionForSymbol(GVSym); 3011 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3012 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3013 emitDebugInfoForGlobal(CVGV); 3014 endCVSubsection(EndLabel); 3015 } 3016 } 3017 3018 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3019 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3020 for (const MDNode *Node : CUs->operands()) { 3021 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3022 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3023 getTypeIndex(RT); 3024 // FIXME: Add to global/local DTU list. 3025 } 3026 } 3027 } 3028 } 3029 3030 // Emit each global variable in the specified array. 3031 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3032 for (const CVGlobalVariable &CVGV : Globals) { 3033 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3034 emitDebugInfoForGlobal(CVGV); 3035 } 3036 } 3037 3038 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3039 const DIGlobalVariable *DIGV = CVGV.DIGV; 3040 if (const GlobalVariable *GV = 3041 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3042 // DataSym record, see SymbolRecord.h for more info. Thread local data 3043 // happens to have the same format as global data. 3044 MCSymbol *GVSym = Asm->getSymbol(GV); 3045 SymbolKind DataSym = GV->isThreadLocal() 3046 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3047 : SymbolKind::S_GTHREAD32) 3048 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3049 : SymbolKind::S_GDATA32); 3050 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3051 OS.AddComment("Type"); 3052 OS.EmitIntValue(getCompleteTypeIndex(DIGV->getType()).getIndex(), 4); 3053 OS.AddComment("DataOffset"); 3054 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 3055 OS.AddComment("Segment"); 3056 OS.EmitCOFFSectionIndex(GVSym); 3057 OS.AddComment("Name"); 3058 const unsigned LengthOfDataRecord = 12; 3059 emitNullTerminatedSymbolName(OS, DIGV->getName(), LengthOfDataRecord); 3060 endSymbolRecord(DataEnd); 3061 } else { 3062 // FIXME: Currently this only emits the global variables in the IR metadata. 3063 // This should also emit enums and static data members. 3064 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3065 assert(DIE->isConstant() && 3066 "Global constant variables must contain a constant expression."); 3067 uint64_t Val = DIE->getElement(1); 3068 3069 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3070 OS.AddComment("Type"); 3071 OS.EmitIntValue(getTypeIndex(DIGV->getType()).getIndex(), 4); 3072 OS.AddComment("Value"); 3073 3074 // Encoded integers shouldn't need more than 10 bytes. 3075 uint8_t data[10]; 3076 BinaryStreamWriter Writer(data, llvm::support::endianness::little); 3077 CodeViewRecordIO IO(Writer); 3078 cantFail(IO.mapEncodedInteger(Val)); 3079 StringRef SRef((char *)data, Writer.getOffset()); 3080 OS.EmitBinaryData(SRef); 3081 3082 OS.AddComment("Name"); 3083 const DIScope *Scope = DIGV->getScope(); 3084 // For static data members, get the scope from the declaration. 3085 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3086 DIGV->getRawStaticDataMemberDeclaration())) 3087 Scope = MemberDecl->getScope(); 3088 emitNullTerminatedSymbolName(OS, 3089 getFullyQualifiedName(Scope, DIGV->getName())); 3090 endSymbolRecord(SConstantEnd); 3091 } 3092 } 3093