[ARM] MVE predicate bitcast test and VPSEL adjustment. NFC
[llvm-core.git] / tools / dsymutil / DwarfLinker.cpp
blob18dbf6b87ef4a81fb84622068e7450353521349c
1 //===- tools/dsymutil/DwarfLinker.cpp - Dwarf debug info linker -----------===//
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 //===----------------------------------------------------------------------===//
9 #include "DwarfLinker.h"
10 #include "BinaryHolder.h"
11 #include "DebugMap.h"
12 #include "DeclContext.h"
13 #include "DwarfStreamer.h"
14 #include "MachOUtils.h"
15 #include "NonRelocatableStringpool.h"
16 #include "dsymutil.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/BitVector.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DenseMapInfo.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/FoldingSet.h"
23 #include "llvm/ADT/Hashing.h"
24 #include "llvm/ADT/IntervalMap.h"
25 #include "llvm/ADT/None.h"
26 #include "llvm/ADT/Optional.h"
27 #include "llvm/ADT/PointerIntPair.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SmallString.h"
30 #include "llvm/ADT/StringMap.h"
31 #include "llvm/ADT/StringRef.h"
32 #include "llvm/ADT/Triple.h"
33 #include "llvm/ADT/Twine.h"
34 #include "llvm/BinaryFormat/Dwarf.h"
35 #include "llvm/BinaryFormat/MachO.h"
36 #include "llvm/CodeGen/AccelTable.h"
37 #include "llvm/CodeGen/AsmPrinter.h"
38 #include "llvm/CodeGen/DIE.h"
39 #include "llvm/Config/config.h"
40 #include "llvm/DebugInfo/DIContext.h"
41 #include "llvm/DebugInfo/DWARF/DWARFAbbreviationDeclaration.h"
42 #include "llvm/DebugInfo/DWARF/DWARFContext.h"
43 #include "llvm/DebugInfo/DWARF/DWARFDataExtractor.h"
44 #include "llvm/DebugInfo/DWARF/DWARFDebugLine.h"
45 #include "llvm/DebugInfo/DWARF/DWARFDebugRangeList.h"
46 #include "llvm/DebugInfo/DWARF/DWARFDie.h"
47 #include "llvm/DebugInfo/DWARF/DWARFFormValue.h"
48 #include "llvm/DebugInfo/DWARF/DWARFSection.h"
49 #include "llvm/DebugInfo/DWARF/DWARFUnit.h"
50 #include "llvm/MC/MCAsmBackend.h"
51 #include "llvm/MC/MCAsmInfo.h"
52 #include "llvm/MC/MCCodeEmitter.h"
53 #include "llvm/MC/MCContext.h"
54 #include "llvm/MC/MCDwarf.h"
55 #include "llvm/MC/MCInstrInfo.h"
56 #include "llvm/MC/MCObjectFileInfo.h"
57 #include "llvm/MC/MCObjectWriter.h"
58 #include "llvm/MC/MCRegisterInfo.h"
59 #include "llvm/MC/MCSection.h"
60 #include "llvm/MC/MCStreamer.h"
61 #include "llvm/MC/MCSubtargetInfo.h"
62 #include "llvm/MC/MCTargetOptions.h"
63 #include "llvm/Object/MachO.h"
64 #include "llvm/Object/ObjectFile.h"
65 #include "llvm/Object/SymbolicFile.h"
66 #include "llvm/Support/Allocator.h"
67 #include "llvm/Support/Casting.h"
68 #include "llvm/Support/Compiler.h"
69 #include "llvm/Support/DJB.h"
70 #include "llvm/Support/DataExtractor.h"
71 #include "llvm/Support/Error.h"
72 #include "llvm/Support/ErrorHandling.h"
73 #include "llvm/Support/ErrorOr.h"
74 #include "llvm/Support/FileSystem.h"
75 #include "llvm/Support/Format.h"
76 #include "llvm/Support/LEB128.h"
77 #include "llvm/Support/MathExtras.h"
78 #include "llvm/Support/MemoryBuffer.h"
79 #include "llvm/Support/Path.h"
80 #include "llvm/Support/TargetRegistry.h"
81 #include "llvm/Support/ThreadPool.h"
82 #include "llvm/Support/ToolOutputFile.h"
83 #include "llvm/Support/WithColor.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/Target/TargetMachine.h"
86 #include "llvm/Target/TargetOptions.h"
87 #include <algorithm>
88 #include <cassert>
89 #include <cinttypes>
90 #include <climits>
91 #include <cstdint>
92 #include <cstdlib>
93 #include <cstring>
94 #include <limits>
95 #include <map>
96 #include <memory>
97 #include <string>
98 #include <system_error>
99 #include <tuple>
100 #include <utility>
101 #include <vector>
103 namespace llvm {
104 namespace dsymutil {
106 /// Similar to DWARFUnitSection::getUnitForOffset(), but returning our
107 /// CompileUnit object instead.
108 static CompileUnit *getUnitForOffset(const UnitListTy &Units, uint64_t Offset) {
109 auto CU = std::upper_bound(
110 Units.begin(), Units.end(), Offset,
111 [](uint64_t LHS, const std::unique_ptr<CompileUnit> &RHS) {
112 return LHS < RHS->getOrigUnit().getNextUnitOffset();
114 return CU != Units.end() ? CU->get() : nullptr;
117 /// Resolve the DIE attribute reference that has been extracted in \p RefValue.
118 /// The resulting DIE might be in another CompileUnit which is stored into \p
119 /// ReferencedCU. \returns null if resolving fails for any reason.
120 static DWARFDie resolveDIEReference(const DwarfLinker &Linker,
121 const DebugMapObject &DMO,
122 const UnitListTy &Units,
123 const DWARFFormValue &RefValue,
124 const DWARFDie &DIE, CompileUnit *&RefCU) {
125 assert(RefValue.isFormClass(DWARFFormValue::FC_Reference));
126 uint64_t RefOffset = *RefValue.getAsReference();
127 if ((RefCU = getUnitForOffset(Units, RefOffset)))
128 if (const auto RefDie = RefCU->getOrigUnit().getDIEForOffset(RefOffset)) {
129 // In a file with broken references, an attribute might point to a NULL
130 // DIE.
131 if (!RefDie.isNULL())
132 return RefDie;
135 Linker.reportWarning("could not find referenced DIE", DMO, &DIE);
136 return DWARFDie();
139 /// \returns whether the passed \a Attr type might contain a DIE reference
140 /// suitable for ODR uniquing.
141 static bool isODRAttribute(uint16_t Attr) {
142 switch (Attr) {
143 default:
144 return false;
145 case dwarf::DW_AT_type:
146 case dwarf::DW_AT_containing_type:
147 case dwarf::DW_AT_specification:
148 case dwarf::DW_AT_abstract_origin:
149 case dwarf::DW_AT_import:
150 return true;
152 llvm_unreachable("Improper attribute.");
155 static bool isTypeTag(uint16_t Tag) {
156 switch (Tag) {
157 case dwarf::DW_TAG_array_type:
158 case dwarf::DW_TAG_class_type:
159 case dwarf::DW_TAG_enumeration_type:
160 case dwarf::DW_TAG_pointer_type:
161 case dwarf::DW_TAG_reference_type:
162 case dwarf::DW_TAG_string_type:
163 case dwarf::DW_TAG_structure_type:
164 case dwarf::DW_TAG_subroutine_type:
165 case dwarf::DW_TAG_typedef:
166 case dwarf::DW_TAG_union_type:
167 case dwarf::DW_TAG_ptr_to_member_type:
168 case dwarf::DW_TAG_set_type:
169 case dwarf::DW_TAG_subrange_type:
170 case dwarf::DW_TAG_base_type:
171 case dwarf::DW_TAG_const_type:
172 case dwarf::DW_TAG_constant:
173 case dwarf::DW_TAG_file_type:
174 case dwarf::DW_TAG_namelist:
175 case dwarf::DW_TAG_packed_type:
176 case dwarf::DW_TAG_volatile_type:
177 case dwarf::DW_TAG_restrict_type:
178 case dwarf::DW_TAG_atomic_type:
179 case dwarf::DW_TAG_interface_type:
180 case dwarf::DW_TAG_unspecified_type:
181 case dwarf::DW_TAG_shared_type:
182 return true;
183 default:
184 break;
186 return false;
189 bool DwarfLinker::DIECloner::getDIENames(const DWARFDie &Die,
190 AttributesInfo &Info,
191 OffsetsStringPool &StringPool,
192 bool StripTemplate) {
193 // This function will be called on DIEs having low_pcs and
194 // ranges. As getting the name might be more expansive, filter out
195 // blocks directly.
196 if (Die.getTag() == dwarf::DW_TAG_lexical_block)
197 return false;
199 // FIXME: a bit wasteful as the first getName might return the
200 // short name.
201 if (!Info.MangledName)
202 if (const char *MangledName = Die.getName(DINameKind::LinkageName))
203 Info.MangledName = StringPool.getEntry(MangledName);
205 if (!Info.Name)
206 if (const char *Name = Die.getName(DINameKind::ShortName))
207 Info.Name = StringPool.getEntry(Name);
209 if (StripTemplate && Info.Name && Info.MangledName != Info.Name) {
210 // FIXME: dsymutil compatibility. This is wrong for operator<
211 auto Split = Info.Name.getString().split('<');
212 if (!Split.second.empty())
213 Info.NameWithoutTemplate = StringPool.getEntry(Split.first);
216 return Info.Name || Info.MangledName;
219 /// Report a warning to the user, optionally including information about a
220 /// specific \p DIE related to the warning.
221 void DwarfLinker::reportWarning(const Twine &Warning, const DebugMapObject &DMO,
222 const DWARFDie *DIE) const {
223 StringRef Context = DMO.getObjectFilename();
224 warn(Warning, Context);
226 if (!Options.Verbose || !DIE)
227 return;
229 DIDumpOptions DumpOpts;
230 DumpOpts.ChildRecurseDepth = 0;
231 DumpOpts.Verbose = Options.Verbose;
233 WithColor::note() << " in DIE:\n";
234 DIE->dump(errs(), 6 /* Indent */, DumpOpts);
237 bool DwarfLinker::createStreamer(const Triple &TheTriple,
238 raw_fd_ostream &OutFile) {
239 if (Options.NoOutput)
240 return true;
242 Streamer = std::make_unique<DwarfStreamer>(OutFile, Options);
243 return Streamer->init(TheTriple);
246 /// Resolve the relative path to a build artifact referenced by DWARF by
247 /// applying DW_AT_comp_dir.
248 static void resolveRelativeObjectPath(SmallVectorImpl<char> &Buf, DWARFDie CU) {
249 sys::path::append(Buf, dwarf::toString(CU.find(dwarf::DW_AT_comp_dir), ""));
252 /// Collect references to parseable Swift interfaces in imported
253 /// DW_TAG_module blocks.
254 static void analyzeImportedModule(
255 const DWARFDie &DIE, CompileUnit &CU,
256 std::map<std::string, std::string> &ParseableSwiftInterfaces,
257 std::function<void(const Twine &, const DWARFDie &)> ReportWarning) {
258 if (CU.getLanguage() != dwarf::DW_LANG_Swift)
259 return;
261 StringRef Path = dwarf::toStringRef(DIE.find(dwarf::DW_AT_LLVM_include_path));
262 if (!Path.endswith(".swiftinterface"))
263 return;
264 if (Optional<DWARFFormValue> Val = DIE.find(dwarf::DW_AT_name))
265 if (Optional<const char *> Name = Val->getAsCString()) {
266 auto &Entry = ParseableSwiftInterfaces[*Name];
267 // The prepend path is applied later when copying.
268 DWARFDie CUDie = CU.getOrigUnit().getUnitDIE();
269 SmallString<128> ResolvedPath;
270 if (sys::path::is_relative(Path))
271 resolveRelativeObjectPath(ResolvedPath, CUDie);
272 sys::path::append(ResolvedPath, Path);
273 if (!Entry.empty() && Entry != ResolvedPath)
274 ReportWarning(
275 Twine("Conflicting parseable interfaces for Swift Module ") +
276 *Name + ": " + Entry + " and " + Path,
277 DIE);
278 Entry = ResolvedPath.str();
282 /// Recursive helper to build the global DeclContext information and
283 /// gather the child->parent relationships in the original compile unit.
285 /// \return true when this DIE and all of its children are only
286 /// forward declarations to types defined in external clang modules
287 /// (i.e., forward declarations that are children of a DW_TAG_module).
288 static bool analyzeContextInfo(
289 const DWARFDie &DIE, unsigned ParentIdx, CompileUnit &CU,
290 DeclContext *CurrentDeclContext, UniquingStringPool &StringPool,
291 DeclContextTree &Contexts, uint64_t ModulesEndOffset,
292 std::map<std::string, std::string> &ParseableSwiftInterfaces,
293 std::function<void(const Twine &, const DWARFDie &)> ReportWarning,
294 bool InImportedModule = false) {
295 unsigned MyIdx = CU.getOrigUnit().getDIEIndex(DIE);
296 CompileUnit::DIEInfo &Info = CU.getInfo(MyIdx);
298 // Clang imposes an ODR on modules(!) regardless of the language:
299 // "The module-id should consist of only a single identifier,
300 // which provides the name of the module being defined. Each
301 // module shall have a single definition."
303 // This does not extend to the types inside the modules:
304 // "[I]n C, this implies that if two structs are defined in
305 // different submodules with the same name, those two types are
306 // distinct types (but may be compatible types if their
307 // definitions match)."
309 // We treat non-C++ modules like namespaces for this reason.
310 if (DIE.getTag() == dwarf::DW_TAG_module && ParentIdx == 0 &&
311 dwarf::toString(DIE.find(dwarf::DW_AT_name), "") !=
312 CU.getClangModuleName()) {
313 InImportedModule = true;
314 analyzeImportedModule(DIE, CU, ParseableSwiftInterfaces, ReportWarning);
317 Info.ParentIdx = ParentIdx;
318 bool InClangModule = CU.isClangModule() || InImportedModule;
319 if (CU.hasODR() || InClangModule) {
320 if (CurrentDeclContext) {
321 auto PtrInvalidPair = Contexts.getChildDeclContext(
322 *CurrentDeclContext, DIE, CU, StringPool, InClangModule);
323 CurrentDeclContext = PtrInvalidPair.getPointer();
324 Info.Ctxt =
325 PtrInvalidPair.getInt() ? nullptr : PtrInvalidPair.getPointer();
326 if (Info.Ctxt)
327 Info.Ctxt->setDefinedInClangModule(InClangModule);
328 } else
329 Info.Ctxt = CurrentDeclContext = nullptr;
332 Info.Prune = InImportedModule;
333 if (DIE.hasChildren())
334 for (auto Child : DIE.children())
335 Info.Prune &= analyzeContextInfo(Child, MyIdx, CU, CurrentDeclContext,
336 StringPool, Contexts, ModulesEndOffset,
337 ParseableSwiftInterfaces, ReportWarning,
338 InImportedModule);
340 // Prune this DIE if it is either a forward declaration inside a
341 // DW_TAG_module or a DW_TAG_module that contains nothing but
342 // forward declarations.
343 Info.Prune &= (DIE.getTag() == dwarf::DW_TAG_module) ||
344 (isTypeTag(DIE.getTag()) &&
345 dwarf::toUnsigned(DIE.find(dwarf::DW_AT_declaration), 0));
347 // Only prune forward declarations inside a DW_TAG_module for which a
348 // definition exists elsewhere.
349 if (ModulesEndOffset == 0)
350 Info.Prune &= Info.Ctxt && Info.Ctxt->getCanonicalDIEOffset();
351 else
352 Info.Prune &= Info.Ctxt && Info.Ctxt->getCanonicalDIEOffset() > 0 &&
353 Info.Ctxt->getCanonicalDIEOffset() <= ModulesEndOffset;
355 return Info.Prune;
356 } // namespace dsymutil
358 static bool dieNeedsChildrenToBeMeaningful(uint32_t Tag) {
359 switch (Tag) {
360 default:
361 return false;
362 case dwarf::DW_TAG_subprogram:
363 case dwarf::DW_TAG_lexical_block:
364 case dwarf::DW_TAG_subroutine_type:
365 case dwarf::DW_TAG_structure_type:
366 case dwarf::DW_TAG_class_type:
367 case dwarf::DW_TAG_union_type:
368 return true;
370 llvm_unreachable("Invalid Tag");
373 void DwarfLinker::startDebugObject(LinkContext &Context) {
374 // Iterate over the debug map entries and put all the ones that are
375 // functions (because they have a size) into the Ranges map. This map is
376 // very similar to the FunctionRanges that are stored in each unit, with 2
377 // notable differences:
379 // 1. Obviously this one is global, while the other ones are per-unit.
381 // 2. This one contains not only the functions described in the DIE
382 // tree, but also the ones that are only in the debug map.
384 // The latter information is required to reproduce dsymutil's logic while
385 // linking line tables. The cases where this information matters look like
386 // bugs that need to be investigated, but for now we need to reproduce
387 // dsymutil's behavior.
388 // FIXME: Once we understood exactly if that information is needed,
389 // maybe totally remove this (or try to use it to do a real
390 // -gline-tables-only on Darwin.
391 for (const auto &Entry : Context.DMO.symbols()) {
392 const auto &Mapping = Entry.getValue();
393 if (Mapping.Size && Mapping.ObjectAddress)
394 Context.Ranges[*Mapping.ObjectAddress] = DebugMapObjectRange(
395 *Mapping.ObjectAddress + Mapping.Size,
396 int64_t(Mapping.BinaryAddress) - *Mapping.ObjectAddress);
400 void DwarfLinker::endDebugObject(LinkContext &Context) {
401 Context.Clear();
403 for (auto I = DIEBlocks.begin(), E = DIEBlocks.end(); I != E; ++I)
404 (*I)->~DIEBlock();
405 for (auto I = DIELocs.begin(), E = DIELocs.end(); I != E; ++I)
406 (*I)->~DIELoc();
408 DIEBlocks.clear();
409 DIELocs.clear();
410 DIEAlloc.Reset();
413 static bool isMachOPairedReloc(uint64_t RelocType, uint64_t Arch) {
414 switch (Arch) {
415 case Triple::x86:
416 return RelocType == MachO::GENERIC_RELOC_SECTDIFF ||
417 RelocType == MachO::GENERIC_RELOC_LOCAL_SECTDIFF;
418 case Triple::x86_64:
419 return RelocType == MachO::X86_64_RELOC_SUBTRACTOR;
420 case Triple::arm:
421 case Triple::thumb:
422 return RelocType == MachO::ARM_RELOC_SECTDIFF ||
423 RelocType == MachO::ARM_RELOC_LOCAL_SECTDIFF ||
424 RelocType == MachO::ARM_RELOC_HALF ||
425 RelocType == MachO::ARM_RELOC_HALF_SECTDIFF;
426 case Triple::aarch64:
427 return RelocType == MachO::ARM64_RELOC_SUBTRACTOR;
428 default:
429 return false;
433 /// Iterate over the relocations of the given \p Section and
434 /// store the ones that correspond to debug map entries into the
435 /// ValidRelocs array.
436 void DwarfLinker::RelocationManager::findValidRelocsMachO(
437 const object::SectionRef &Section, const object::MachOObjectFile &Obj,
438 const DebugMapObject &DMO) {
439 Expected<StringRef> ContentsOrErr = Section.getContents();
440 if (!ContentsOrErr) {
441 consumeError(ContentsOrErr.takeError());
442 Linker.reportWarning("error reading section", DMO);
443 return;
445 DataExtractor Data(*ContentsOrErr, Obj.isLittleEndian(), 0);
446 bool SkipNext = false;
448 for (const object::RelocationRef &Reloc : Section.relocations()) {
449 if (SkipNext) {
450 SkipNext = false;
451 continue;
454 object::DataRefImpl RelocDataRef = Reloc.getRawDataRefImpl();
455 MachO::any_relocation_info MachOReloc = Obj.getRelocation(RelocDataRef);
457 if (isMachOPairedReloc(Obj.getAnyRelocationType(MachOReloc),
458 Obj.getArch())) {
459 SkipNext = true;
460 Linker.reportWarning("unsupported relocation in debug_info section.",
461 DMO);
462 continue;
465 unsigned RelocSize = 1 << Obj.getAnyRelocationLength(MachOReloc);
466 uint64_t Offset64 = Reloc.getOffset();
467 if ((RelocSize != 4 && RelocSize != 8)) {
468 Linker.reportWarning("unsupported relocation in debug_info section.",
469 DMO);
470 continue;
472 uint64_t OffsetCopy = Offset64;
473 // Mach-o uses REL relocations, the addend is at the relocation offset.
474 uint64_t Addend = Data.getUnsigned(&OffsetCopy, RelocSize);
475 uint64_t SymAddress;
476 int64_t SymOffset;
478 if (Obj.isRelocationScattered(MachOReloc)) {
479 // The address of the base symbol for scattered relocations is
480 // stored in the reloc itself. The actual addend will store the
481 // base address plus the offset.
482 SymAddress = Obj.getScatteredRelocationValue(MachOReloc);
483 SymOffset = int64_t(Addend) - SymAddress;
484 } else {
485 SymAddress = Addend;
486 SymOffset = 0;
489 auto Sym = Reloc.getSymbol();
490 if (Sym != Obj.symbol_end()) {
491 Expected<StringRef> SymbolName = Sym->getName();
492 if (!SymbolName) {
493 consumeError(SymbolName.takeError());
494 Linker.reportWarning("error getting relocation symbol name.", DMO);
495 continue;
497 if (const auto *Mapping = DMO.lookupSymbol(*SymbolName))
498 ValidRelocs.emplace_back(Offset64, RelocSize, Addend, Mapping);
499 } else if (const auto *Mapping = DMO.lookupObjectAddress(SymAddress)) {
500 // Do not store the addend. The addend was the address of the symbol in
501 // the object file, the address in the binary that is stored in the debug
502 // map doesn't need to be offset.
503 ValidRelocs.emplace_back(Offset64, RelocSize, SymOffset, Mapping);
508 /// Dispatch the valid relocation finding logic to the
509 /// appropriate handler depending on the object file format.
510 bool DwarfLinker::RelocationManager::findValidRelocs(
511 const object::SectionRef &Section, const object::ObjectFile &Obj,
512 const DebugMapObject &DMO) {
513 // Dispatch to the right handler depending on the file type.
514 if (auto *MachOObj = dyn_cast<object::MachOObjectFile>(&Obj))
515 findValidRelocsMachO(Section, *MachOObj, DMO);
516 else
517 Linker.reportWarning(
518 Twine("unsupported object file type: ") + Obj.getFileName(), DMO);
520 if (ValidRelocs.empty())
521 return false;
523 // Sort the relocations by offset. We will walk the DIEs linearly in
524 // the file, this allows us to just keep an index in the relocation
525 // array that we advance during our walk, rather than resorting to
526 // some associative container. See DwarfLinker::NextValidReloc.
527 llvm::sort(ValidRelocs);
528 return true;
531 /// Look for relocations in the debug_info section that match
532 /// entries in the debug map. These relocations will drive the Dwarf
533 /// link by indicating which DIEs refer to symbols present in the
534 /// linked binary.
535 /// \returns whether there are any valid relocations in the debug info.
536 bool DwarfLinker::RelocationManager::findValidRelocsInDebugInfo(
537 const object::ObjectFile &Obj, const DebugMapObject &DMO) {
538 // Find the debug_info section.
539 for (const object::SectionRef &Section : Obj.sections()) {
540 StringRef SectionName;
541 if (Expected<StringRef> NameOrErr = Section.getName())
542 SectionName = *NameOrErr;
543 else
544 consumeError(NameOrErr.takeError());
546 SectionName = SectionName.substr(SectionName.find_first_not_of("._"));
547 if (SectionName != "debug_info")
548 continue;
549 return findValidRelocs(Section, Obj, DMO);
551 return false;
554 /// Checks that there is a relocation against an actual debug
555 /// map entry between \p StartOffset and \p NextOffset.
557 /// This function must be called with offsets in strictly ascending
558 /// order because it never looks back at relocations it already 'went past'.
559 /// \returns true and sets Info.InDebugMap if it is the case.
560 bool DwarfLinker::RelocationManager::hasValidRelocation(
561 uint64_t StartOffset, uint64_t EndOffset, CompileUnit::DIEInfo &Info) {
562 assert(NextValidReloc == 0 ||
563 StartOffset > ValidRelocs[NextValidReloc - 1].Offset);
564 if (NextValidReloc >= ValidRelocs.size())
565 return false;
567 uint64_t RelocOffset = ValidRelocs[NextValidReloc].Offset;
569 // We might need to skip some relocs that we didn't consider. For
570 // example the high_pc of a discarded DIE might contain a reloc that
571 // is in the list because it actually corresponds to the start of a
572 // function that is in the debug map.
573 while (RelocOffset < StartOffset && NextValidReloc < ValidRelocs.size() - 1)
574 RelocOffset = ValidRelocs[++NextValidReloc].Offset;
576 if (RelocOffset < StartOffset || RelocOffset >= EndOffset)
577 return false;
579 const auto &ValidReloc = ValidRelocs[NextValidReloc++];
580 const auto &Mapping = ValidReloc.Mapping->getValue();
581 uint64_t ObjectAddress = Mapping.ObjectAddress
582 ? uint64_t(*Mapping.ObjectAddress)
583 : std::numeric_limits<uint64_t>::max();
584 if (Linker.Options.Verbose)
585 outs() << "Found valid debug map entry: " << ValidReloc.Mapping->getKey()
586 << " "
587 << format("\t%016" PRIx64 " => %016" PRIx64, ObjectAddress,
588 uint64_t(Mapping.BinaryAddress));
590 Info.AddrAdjust = int64_t(Mapping.BinaryAddress) + ValidReloc.Addend;
591 if (Mapping.ObjectAddress)
592 Info.AddrAdjust -= ObjectAddress;
593 Info.InDebugMap = true;
594 return true;
597 /// Get the starting and ending (exclusive) offset for the
598 /// attribute with index \p Idx descibed by \p Abbrev. \p Offset is
599 /// supposed to point to the position of the first attribute described
600 /// by \p Abbrev.
601 /// \return [StartOffset, EndOffset) as a pair.
602 static std::pair<uint64_t, uint64_t>
603 getAttributeOffsets(const DWARFAbbreviationDeclaration *Abbrev, unsigned Idx,
604 uint64_t Offset, const DWARFUnit &Unit) {
605 DataExtractor Data = Unit.getDebugInfoExtractor();
607 for (unsigned i = 0; i < Idx; ++i)
608 DWARFFormValue::skipValue(Abbrev->getFormByIndex(i), Data, &Offset,
609 Unit.getFormParams());
611 uint64_t End = Offset;
612 DWARFFormValue::skipValue(Abbrev->getFormByIndex(Idx), Data, &End,
613 Unit.getFormParams());
615 return std::make_pair(Offset, End);
618 /// Check if a variable describing DIE should be kept.
619 /// \returns updated TraversalFlags.
620 unsigned DwarfLinker::shouldKeepVariableDIE(RelocationManager &RelocMgr,
621 const DWARFDie &DIE,
622 CompileUnit &Unit,
623 CompileUnit::DIEInfo &MyInfo,
624 unsigned Flags) {
625 const auto *Abbrev = DIE.getAbbreviationDeclarationPtr();
627 // Global variables with constant value can always be kept.
628 if (!(Flags & TF_InFunctionScope) &&
629 Abbrev->findAttributeIndex(dwarf::DW_AT_const_value)) {
630 MyInfo.InDebugMap = true;
631 return Flags | TF_Keep;
634 Optional<uint32_t> LocationIdx =
635 Abbrev->findAttributeIndex(dwarf::DW_AT_location);
636 if (!LocationIdx)
637 return Flags;
639 uint64_t Offset = DIE.getOffset() + getULEB128Size(Abbrev->getCode());
640 const DWARFUnit &OrigUnit = Unit.getOrigUnit();
641 uint64_t LocationOffset, LocationEndOffset;
642 std::tie(LocationOffset, LocationEndOffset) =
643 getAttributeOffsets(Abbrev, *LocationIdx, Offset, OrigUnit);
645 // See if there is a relocation to a valid debug map entry inside
646 // this variable's location. The order is important here. We want to
647 // always check in the variable has a valid relocation, so that the
648 // DIEInfo is filled. However, we don't want a static variable in a
649 // function to force us to keep the enclosing function.
650 if (!RelocMgr.hasValidRelocation(LocationOffset, LocationEndOffset, MyInfo) ||
651 (Flags & TF_InFunctionScope))
652 return Flags;
654 if (Options.Verbose) {
655 DIDumpOptions DumpOpts;
656 DumpOpts.ChildRecurseDepth = 0;
657 DumpOpts.Verbose = Options.Verbose;
658 DIE.dump(outs(), 8 /* Indent */, DumpOpts);
661 return Flags | TF_Keep;
664 /// Check if a function describing DIE should be kept.
665 /// \returns updated TraversalFlags.
666 unsigned DwarfLinker::shouldKeepSubprogramDIE(
667 RelocationManager &RelocMgr, RangesTy &Ranges, const DWARFDie &DIE,
668 const DebugMapObject &DMO, CompileUnit &Unit, CompileUnit::DIEInfo &MyInfo,
669 unsigned Flags) {
670 const auto *Abbrev = DIE.getAbbreviationDeclarationPtr();
672 Flags |= TF_InFunctionScope;
674 Optional<uint32_t> LowPcIdx = Abbrev->findAttributeIndex(dwarf::DW_AT_low_pc);
675 if (!LowPcIdx)
676 return Flags;
678 uint64_t Offset = DIE.getOffset() + getULEB128Size(Abbrev->getCode());
679 DWARFUnit &OrigUnit = Unit.getOrigUnit();
680 uint64_t LowPcOffset, LowPcEndOffset;
681 std::tie(LowPcOffset, LowPcEndOffset) =
682 getAttributeOffsets(Abbrev, *LowPcIdx, Offset, OrigUnit);
684 auto LowPc = dwarf::toAddress(DIE.find(dwarf::DW_AT_low_pc));
685 assert(LowPc.hasValue() && "low_pc attribute is not an address.");
686 if (!LowPc ||
687 !RelocMgr.hasValidRelocation(LowPcOffset, LowPcEndOffset, MyInfo))
688 return Flags;
690 if (Options.Verbose) {
691 DIDumpOptions DumpOpts;
692 DumpOpts.ChildRecurseDepth = 0;
693 DumpOpts.Verbose = Options.Verbose;
694 DIE.dump(outs(), 8 /* Indent */, DumpOpts);
697 if (DIE.getTag() == dwarf::DW_TAG_label) {
698 if (Unit.hasLabelAt(*LowPc))
699 return Flags;
700 // FIXME: dsymutil-classic compat. dsymutil-classic doesn't consider labels
701 // that don't fall into the CU's aranges. This is wrong IMO. Debug info
702 // generation bugs aside, this is really wrong in the case of labels, where
703 // a label marking the end of a function will have a PC == CU's high_pc.
704 if (dwarf::toAddress(OrigUnit.getUnitDIE().find(dwarf::DW_AT_high_pc))
705 .getValueOr(UINT64_MAX) <= LowPc)
706 return Flags;
707 Unit.addLabelLowPc(*LowPc, MyInfo.AddrAdjust);
708 return Flags | TF_Keep;
711 Flags |= TF_Keep;
713 Optional<uint64_t> HighPc = DIE.getHighPC(*LowPc);
714 if (!HighPc) {
715 reportWarning("Function without high_pc. Range will be discarded.\n", DMO,
716 &DIE);
717 return Flags;
720 // Replace the debug map range with a more accurate one.
721 Ranges[*LowPc] = DebugMapObjectRange(*HighPc, MyInfo.AddrAdjust);
722 Unit.addFunctionRange(*LowPc, *HighPc, MyInfo.AddrAdjust);
723 return Flags;
726 /// Check if a DIE should be kept.
727 /// \returns updated TraversalFlags.
728 unsigned DwarfLinker::shouldKeepDIE(RelocationManager &RelocMgr,
729 RangesTy &Ranges, const DWARFDie &DIE,
730 const DebugMapObject &DMO,
731 CompileUnit &Unit,
732 CompileUnit::DIEInfo &MyInfo,
733 unsigned Flags) {
734 switch (DIE.getTag()) {
735 case dwarf::DW_TAG_constant:
736 case dwarf::DW_TAG_variable:
737 return shouldKeepVariableDIE(RelocMgr, DIE, Unit, MyInfo, Flags);
738 case dwarf::DW_TAG_subprogram:
739 case dwarf::DW_TAG_label:
740 return shouldKeepSubprogramDIE(RelocMgr, Ranges, DIE, DMO, Unit, MyInfo,
741 Flags);
742 case dwarf::DW_TAG_base_type:
743 // DWARF Expressions may reference basic types, but scanning them
744 // is expensive. Basic types are tiny, so just keep all of them.
745 case dwarf::DW_TAG_imported_module:
746 case dwarf::DW_TAG_imported_declaration:
747 case dwarf::DW_TAG_imported_unit:
748 // We always want to keep these.
749 return Flags | TF_Keep;
750 default:
751 break;
754 return Flags;
757 /// Mark the passed DIE as well as all the ones it depends on
758 /// as kept.
760 /// This function is called by lookForDIEsToKeep on DIEs that are
761 /// newly discovered to be needed in the link. It recursively calls
762 /// back to lookForDIEsToKeep while adding TF_DependencyWalk to the
763 /// TraversalFlags to inform it that it's not doing the primary DIE
764 /// tree walk.
765 void DwarfLinker::keepDIEAndDependencies(
766 RelocationManager &RelocMgr, RangesTy &Ranges, const UnitListTy &Units,
767 const DWARFDie &Die, CompileUnit::DIEInfo &MyInfo,
768 const DebugMapObject &DMO, CompileUnit &CU, bool UseODR) {
769 DWARFUnit &Unit = CU.getOrigUnit();
770 MyInfo.Keep = true;
772 // We're looking for incomplete types.
773 MyInfo.Incomplete = Die.getTag() != dwarf::DW_TAG_subprogram &&
774 Die.getTag() != dwarf::DW_TAG_member &&
775 dwarf::toUnsigned(Die.find(dwarf::DW_AT_declaration), 0);
777 // First mark all the parent chain as kept.
778 unsigned AncestorIdx = MyInfo.ParentIdx;
779 while (!CU.getInfo(AncestorIdx).Keep) {
780 unsigned ODRFlag = UseODR ? TF_ODR : 0;
781 lookForDIEsToKeep(RelocMgr, Ranges, Units, Unit.getDIEAtIndex(AncestorIdx),
782 DMO, CU,
783 TF_ParentWalk | TF_Keep | TF_DependencyWalk | ODRFlag);
784 AncestorIdx = CU.getInfo(AncestorIdx).ParentIdx;
787 // Then we need to mark all the DIEs referenced by this DIE's
788 // attributes as kept.
789 DWARFDataExtractor Data = Unit.getDebugInfoExtractor();
790 const auto *Abbrev = Die.getAbbreviationDeclarationPtr();
791 uint64_t Offset = Die.getOffset() + getULEB128Size(Abbrev->getCode());
793 // Mark all DIEs referenced through attributes as kept.
794 for (const auto &AttrSpec : Abbrev->attributes()) {
795 DWARFFormValue Val(AttrSpec.Form);
796 if (!Val.isFormClass(DWARFFormValue::FC_Reference) ||
797 AttrSpec.Attr == dwarf::DW_AT_sibling) {
798 DWARFFormValue::skipValue(AttrSpec.Form, Data, &Offset,
799 Unit.getFormParams());
800 continue;
803 Val.extractValue(Data, &Offset, Unit.getFormParams(), &Unit);
804 CompileUnit *ReferencedCU;
805 if (auto RefDie =
806 resolveDIEReference(*this, DMO, Units, Val, Die, ReferencedCU)) {
807 uint32_t RefIdx = ReferencedCU->getOrigUnit().getDIEIndex(RefDie);
808 CompileUnit::DIEInfo &Info = ReferencedCU->getInfo(RefIdx);
809 bool IsModuleRef = Info.Ctxt && Info.Ctxt->getCanonicalDIEOffset() &&
810 Info.Ctxt->isDefinedInClangModule();
811 // If the referenced DIE has a DeclContext that has already been
812 // emitted, then do not keep the one in this CU. We'll link to
813 // the canonical DIE in cloneDieReferenceAttribute.
814 // FIXME: compatibility with dsymutil-classic. UseODR shouldn't
815 // be necessary and could be advantageously replaced by
816 // ReferencedCU->hasODR() && CU.hasODR().
817 // FIXME: compatibility with dsymutil-classic. There is no
818 // reason not to unique ref_addr references.
819 if (AttrSpec.Form != dwarf::DW_FORM_ref_addr && (UseODR || IsModuleRef) &&
820 Info.Ctxt &&
821 Info.Ctxt != ReferencedCU->getInfo(Info.ParentIdx).Ctxt &&
822 Info.Ctxt->getCanonicalDIEOffset() && isODRAttribute(AttrSpec.Attr))
823 continue;
825 // Keep a module forward declaration if there is no definition.
826 if (!(isODRAttribute(AttrSpec.Attr) && Info.Ctxt &&
827 Info.Ctxt->getCanonicalDIEOffset()))
828 Info.Prune = false;
830 unsigned ODRFlag = UseODR ? TF_ODR : 0;
831 lookForDIEsToKeep(RelocMgr, Ranges, Units, RefDie, DMO, *ReferencedCU,
832 TF_Keep | TF_DependencyWalk | ODRFlag);
834 // The incomplete property is propagated if the current DIE is complete
835 // but references an incomplete DIE.
836 if (Info.Incomplete && !MyInfo.Incomplete &&
837 (Die.getTag() == dwarf::DW_TAG_typedef ||
838 Die.getTag() == dwarf::DW_TAG_member ||
839 Die.getTag() == dwarf::DW_TAG_reference_type ||
840 Die.getTag() == dwarf::DW_TAG_ptr_to_member_type ||
841 Die.getTag() == dwarf::DW_TAG_pointer_type))
842 MyInfo.Incomplete = true;
847 namespace {
848 /// This class represents an item in the work list. In addition to it's obvious
849 /// purpose of representing the state associated with a particular run of the
850 /// work loop, it also serves as a marker to indicate that we should run the
851 /// "continuation" code.
853 /// Originally, the latter was lambda which allowed arbitrary code to be run.
854 /// Because we always need to run the exact same code, it made more sense to
855 /// use a boolean and repurpose the already existing DIE field.
856 struct WorklistItem {
857 DWARFDie Die;
858 unsigned Flags;
859 bool IsContinuation;
860 CompileUnit::DIEInfo *ChildInfo = nullptr;
862 /// Construct a classic worklist item.
863 WorklistItem(DWARFDie Die, unsigned Flags)
864 : Die(Die), Flags(Flags), IsContinuation(false){};
866 /// Creates a continuation marker.
867 WorklistItem(DWARFDie Die) : Die(Die), IsContinuation(true){};
869 } // namespace
871 // Helper that updates the completeness of the current DIE. It depends on the
872 // fact that the incompletness of its children is already computed.
873 static void updateIncompleteness(const DWARFDie &Die,
874 CompileUnit::DIEInfo &ChildInfo,
875 CompileUnit &CU) {
876 // Only propagate incomplete members.
877 if (Die.getTag() != dwarf::DW_TAG_structure_type &&
878 Die.getTag() != dwarf::DW_TAG_class_type)
879 return;
881 unsigned Idx = CU.getOrigUnit().getDIEIndex(Die);
882 CompileUnit::DIEInfo &MyInfo = CU.getInfo(Idx);
884 if (MyInfo.Incomplete)
885 return;
887 if (ChildInfo.Incomplete || ChildInfo.Prune)
888 MyInfo.Incomplete = true;
891 /// Recursively walk the \p DIE tree and look for DIEs to
892 /// keep. Store that information in \p CU's DIEInfo.
894 /// This function is the entry point of the DIE selection
895 /// algorithm. It is expected to walk the DIE tree in file order and
896 /// (though the mediation of its helper) call hasValidRelocation() on
897 /// each DIE that might be a 'root DIE' (See DwarfLinker class
898 /// comment).
899 /// While walking the dependencies of root DIEs, this function is
900 /// also called, but during these dependency walks the file order is
901 /// not respected. The TF_DependencyWalk flag tells us which kind of
902 /// traversal we are currently doing.
904 /// The return value indicates whether the DIE is incomplete.
905 void DwarfLinker::lookForDIEsToKeep(RelocationManager &RelocMgr,
906 RangesTy &Ranges, const UnitListTy &Units,
907 const DWARFDie &Die,
908 const DebugMapObject &DMO, CompileUnit &CU,
909 unsigned Flags) {
910 // LIFO work list.
911 SmallVector<WorklistItem, 4> Worklist;
912 Worklist.emplace_back(Die, Flags);
914 while (!Worklist.empty()) {
915 WorklistItem Current = Worklist.back();
916 Worklist.pop_back();
918 if (Current.IsContinuation) {
919 updateIncompleteness(Current.Die, *Current.ChildInfo, CU);
920 continue;
923 unsigned Idx = CU.getOrigUnit().getDIEIndex(Current.Die);
924 CompileUnit::DIEInfo &MyInfo = CU.getInfo(Idx);
926 // At this point we are guaranteed to have a continuation marker before us
927 // in the worklist, except for the last DIE.
928 if (!Worklist.empty())
929 Worklist.back().ChildInfo = &MyInfo;
931 if (MyInfo.Prune)
932 continue;
934 // If the Keep flag is set, we are marking a required DIE's dependencies.
935 // If our target is already marked as kept, we're all set.
936 bool AlreadyKept = MyInfo.Keep;
937 if ((Current.Flags & TF_DependencyWalk) && AlreadyKept)
938 continue;
940 // We must not call shouldKeepDIE while called from keepDIEAndDependencies,
941 // because it would screw up the relocation finding logic.
942 if (!(Current.Flags & TF_DependencyWalk))
943 Current.Flags = shouldKeepDIE(RelocMgr, Ranges, Current.Die, DMO, CU,
944 MyInfo, Current.Flags);
946 // If it is a newly kept DIE mark it as well as all its dependencies as
947 // kept.
948 if (!AlreadyKept && (Current.Flags & TF_Keep)) {
949 bool UseOdr = (Current.Flags & TF_DependencyWalk)
950 ? (Current.Flags & TF_ODR)
951 : CU.hasODR();
952 keepDIEAndDependencies(RelocMgr, Ranges, Units, Current.Die, MyInfo, DMO,
953 CU, UseOdr);
956 // The TF_ParentWalk flag tells us that we are currently walking up
957 // the parent chain of a required DIE, and we don't want to mark all
958 // the children of the parents as kept (consider for example a
959 // DW_TAG_namespace node in the parent chain). There are however a
960 // set of DIE types for which we want to ignore that directive and still
961 // walk their children.
962 if (dieNeedsChildrenToBeMeaningful(Current.Die.getTag()))
963 Current.Flags &= ~TF_ParentWalk;
965 if (!Current.Die.hasChildren() || (Current.Flags & TF_ParentWalk))
966 continue;
968 // Add children in reverse order to the worklist to effectively process
969 // them in order.
970 for (auto Child : reverse(Current.Die.children())) {
971 // Add continuation marker before every child to calculate incompleteness
972 // after the last child is processed. We can't store this information in
973 // the same item because we might have to process other continuations
974 // first.
975 Worklist.emplace_back(Current.Die);
976 Worklist.emplace_back(Child, Current.Flags);
981 /// Assign an abbreviation number to \p Abbrev.
983 /// Our DIEs get freed after every DebugMapObject has been processed,
984 /// thus the FoldingSet we use to unique DIEAbbrevs cannot refer to
985 /// the instances hold by the DIEs. When we encounter an abbreviation
986 /// that we don't know, we create a permanent copy of it.
987 void DwarfLinker::AssignAbbrev(DIEAbbrev &Abbrev) {
988 // Check the set for priors.
989 FoldingSetNodeID ID;
990 Abbrev.Profile(ID);
991 void *InsertToken;
992 DIEAbbrev *InSet = AbbreviationsSet.FindNodeOrInsertPos(ID, InsertToken);
994 // If it's newly added.
995 if (InSet) {
996 // Assign existing abbreviation number.
997 Abbrev.setNumber(InSet->getNumber());
998 } else {
999 // Add to abbreviation list.
1000 Abbreviations.push_back(
1001 std::make_unique<DIEAbbrev>(Abbrev.getTag(), Abbrev.hasChildren()));
1002 for (const auto &Attr : Abbrev.getData())
1003 Abbreviations.back()->AddAttribute(Attr.getAttribute(), Attr.getForm());
1004 AbbreviationsSet.InsertNode(Abbreviations.back().get(), InsertToken);
1005 // Assign the unique abbreviation number.
1006 Abbrev.setNumber(Abbreviations.size());
1007 Abbreviations.back()->setNumber(Abbreviations.size());
1011 unsigned DwarfLinker::DIECloner::cloneStringAttribute(
1012 DIE &Die, AttributeSpec AttrSpec, const DWARFFormValue &Val,
1013 const DWARFUnit &U, OffsetsStringPool &StringPool, AttributesInfo &Info) {
1014 // Switch everything to out of line strings.
1015 const char *String = *Val.getAsCString();
1016 auto StringEntry = StringPool.getEntry(String);
1018 // Update attributes info.
1019 if (AttrSpec.Attr == dwarf::DW_AT_name)
1020 Info.Name = StringEntry;
1021 else if (AttrSpec.Attr == dwarf::DW_AT_MIPS_linkage_name ||
1022 AttrSpec.Attr == dwarf::DW_AT_linkage_name)
1023 Info.MangledName = StringEntry;
1025 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr), dwarf::DW_FORM_strp,
1026 DIEInteger(StringEntry.getOffset()));
1028 return 4;
1031 unsigned DwarfLinker::DIECloner::cloneDieReferenceAttribute(
1032 DIE &Die, const DWARFDie &InputDIE, AttributeSpec AttrSpec,
1033 unsigned AttrSize, const DWARFFormValue &Val, const DebugMapObject &DMO,
1034 CompileUnit &Unit) {
1035 const DWARFUnit &U = Unit.getOrigUnit();
1036 uint64_t Ref = *Val.getAsReference();
1037 DIE *NewRefDie = nullptr;
1038 CompileUnit *RefUnit = nullptr;
1039 DeclContext *Ctxt = nullptr;
1041 DWARFDie RefDie =
1042 resolveDIEReference(Linker, DMO, CompileUnits, Val, InputDIE, RefUnit);
1044 // If the referenced DIE is not found, drop the attribute.
1045 if (!RefDie || AttrSpec.Attr == dwarf::DW_AT_sibling)
1046 return 0;
1048 unsigned Idx = RefUnit->getOrigUnit().getDIEIndex(RefDie);
1049 CompileUnit::DIEInfo &RefInfo = RefUnit->getInfo(Idx);
1051 // If we already have emitted an equivalent DeclContext, just point
1052 // at it.
1053 if (isODRAttribute(AttrSpec.Attr)) {
1054 Ctxt = RefInfo.Ctxt;
1055 if (Ctxt && Ctxt->getCanonicalDIEOffset()) {
1056 DIEInteger Attr(Ctxt->getCanonicalDIEOffset());
1057 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1058 dwarf::DW_FORM_ref_addr, Attr);
1059 return U.getRefAddrByteSize();
1063 if (!RefInfo.Clone) {
1064 assert(Ref > InputDIE.getOffset());
1065 // We haven't cloned this DIE yet. Just create an empty one and
1066 // store it. It'll get really cloned when we process it.
1067 RefInfo.Clone = DIE::get(DIEAlloc, dwarf::Tag(RefDie.getTag()));
1069 NewRefDie = RefInfo.Clone;
1071 if (AttrSpec.Form == dwarf::DW_FORM_ref_addr ||
1072 (Unit.hasODR() && isODRAttribute(AttrSpec.Attr))) {
1073 // We cannot currently rely on a DIEEntry to emit ref_addr
1074 // references, because the implementation calls back to DwarfDebug
1075 // to find the unit offset. (We don't have a DwarfDebug)
1076 // FIXME: we should be able to design DIEEntry reliance on
1077 // DwarfDebug away.
1078 uint64_t Attr;
1079 if (Ref < InputDIE.getOffset()) {
1080 // We must have already cloned that DIE.
1081 uint32_t NewRefOffset =
1082 RefUnit->getStartOffset() + NewRefDie->getOffset();
1083 Attr = NewRefOffset;
1084 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1085 dwarf::DW_FORM_ref_addr, DIEInteger(Attr));
1086 } else {
1087 // A forward reference. Note and fixup later.
1088 Attr = 0xBADDEF;
1089 Unit.noteForwardReference(
1090 NewRefDie, RefUnit, Ctxt,
1091 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1092 dwarf::DW_FORM_ref_addr, DIEInteger(Attr)));
1094 return U.getRefAddrByteSize();
1097 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1098 dwarf::Form(AttrSpec.Form), DIEEntry(*NewRefDie));
1099 return AttrSize;
1102 void DwarfLinker::DIECloner::cloneExpression(
1103 DataExtractor &Data, DWARFExpression Expression, const DebugMapObject &DMO,
1104 CompileUnit &Unit, SmallVectorImpl<uint8_t> &OutputBuffer) {
1105 using Encoding = DWARFExpression::Operation::Encoding;
1107 uint64_t OpOffset = 0;
1108 for (auto &Op : Expression) {
1109 auto Description = Op.getDescription();
1110 // DW_OP_const_type is variable-length and has 3
1111 // operands. DWARFExpression thus far only supports 2.
1112 auto Op0 = Description.Op[0];
1113 auto Op1 = Description.Op[1];
1114 if ((Op0 == Encoding::BaseTypeRef && Op1 != Encoding::SizeNA) ||
1115 (Op1 == Encoding::BaseTypeRef && Op0 != Encoding::Size1))
1116 Linker.reportWarning("Unsupported DW_OP encoding.", DMO);
1118 if ((Op0 == Encoding::BaseTypeRef && Op1 == Encoding::SizeNA) ||
1119 (Op1 == Encoding::BaseTypeRef && Op0 == Encoding::Size1)) {
1120 // This code assumes that the other non-typeref operand fits into 1 byte.
1121 assert(OpOffset < Op.getEndOffset());
1122 uint32_t ULEBsize = Op.getEndOffset() - OpOffset - 1;
1123 assert(ULEBsize <= 16);
1125 // Copy over the operation.
1126 OutputBuffer.push_back(Op.getCode());
1127 uint64_t RefOffset;
1128 if (Op1 == Encoding::SizeNA) {
1129 RefOffset = Op.getRawOperand(0);
1130 } else {
1131 OutputBuffer.push_back(Op.getRawOperand(0));
1132 RefOffset = Op.getRawOperand(1);
1134 auto RefDie = Unit.getOrigUnit().getDIEForOffset(RefOffset);
1135 uint32_t RefIdx = Unit.getOrigUnit().getDIEIndex(RefDie);
1136 CompileUnit::DIEInfo &Info = Unit.getInfo(RefIdx);
1137 uint32_t Offset = 0;
1138 if (DIE *Clone = Info.Clone)
1139 Offset = Clone->getOffset();
1140 else
1141 Linker.reportWarning("base type ref doesn't point to DW_TAG_base_type.",
1142 DMO);
1143 uint8_t ULEB[16];
1144 unsigned RealSize = encodeULEB128(Offset, ULEB, ULEBsize);
1145 if (RealSize > ULEBsize) {
1146 // Emit the generic type as a fallback.
1147 RealSize = encodeULEB128(0, ULEB, ULEBsize);
1148 Linker.reportWarning("base type ref doesn't fit.", DMO);
1150 assert(RealSize == ULEBsize && "padding failed");
1151 ArrayRef<uint8_t> ULEBbytes(ULEB, ULEBsize);
1152 OutputBuffer.append(ULEBbytes.begin(), ULEBbytes.end());
1153 } else {
1154 // Copy over everything else unmodified.
1155 StringRef Bytes = Data.getData().slice(OpOffset, Op.getEndOffset());
1156 OutputBuffer.append(Bytes.begin(), Bytes.end());
1158 OpOffset = Op.getEndOffset();
1162 unsigned DwarfLinker::DIECloner::cloneBlockAttribute(
1163 DIE &Die, const DebugMapObject &DMO, CompileUnit &Unit,
1164 AttributeSpec AttrSpec, const DWARFFormValue &Val, unsigned AttrSize,
1165 bool IsLittleEndian) {
1166 DIEValueList *Attr;
1167 DIEValue Value;
1168 DIELoc *Loc = nullptr;
1169 DIEBlock *Block = nullptr;
1170 if (AttrSpec.Form == dwarf::DW_FORM_exprloc) {
1171 Loc = new (DIEAlloc) DIELoc;
1172 Linker.DIELocs.push_back(Loc);
1173 } else {
1174 Block = new (DIEAlloc) DIEBlock;
1175 Linker.DIEBlocks.push_back(Block);
1177 Attr = Loc ? static_cast<DIEValueList *>(Loc)
1178 : static_cast<DIEValueList *>(Block);
1180 if (Loc)
1181 Value = DIEValue(dwarf::Attribute(AttrSpec.Attr),
1182 dwarf::Form(AttrSpec.Form), Loc);
1183 else
1184 Value = DIEValue(dwarf::Attribute(AttrSpec.Attr),
1185 dwarf::Form(AttrSpec.Form), Block);
1187 // If the block is a DWARF Expression, clone it into the temporary
1188 // buffer using cloneExpression(), otherwise copy the data directly.
1189 SmallVector<uint8_t, 32> Buffer;
1190 ArrayRef<uint8_t> Bytes = *Val.getAsBlock();
1191 if (DWARFAttribute::mayHaveLocationDescription(AttrSpec.Attr) &&
1192 (Val.isFormClass(DWARFFormValue::FC_Block) ||
1193 Val.isFormClass(DWARFFormValue::FC_Exprloc))) {
1194 DWARFUnit &OrigUnit = Unit.getOrigUnit();
1195 DataExtractor Data(StringRef((const char *)Bytes.data(), Bytes.size()),
1196 IsLittleEndian, OrigUnit.getAddressByteSize());
1197 DWARFExpression Expr(Data, OrigUnit.getVersion(),
1198 OrigUnit.getAddressByteSize());
1199 cloneExpression(Data, Expr, DMO, Unit, Buffer);
1200 Bytes = Buffer;
1202 for (auto Byte : Bytes)
1203 Attr->addValue(DIEAlloc, static_cast<dwarf::Attribute>(0),
1204 dwarf::DW_FORM_data1, DIEInteger(Byte));
1206 // FIXME: If DIEBlock and DIELoc just reuses the Size field of
1207 // the DIE class, this if could be replaced by
1208 // Attr->setSize(Bytes.size()).
1209 if (Linker.Streamer) {
1210 auto *AsmPrinter = &Linker.Streamer->getAsmPrinter();
1211 if (Loc)
1212 Loc->ComputeSize(AsmPrinter);
1213 else
1214 Block->ComputeSize(AsmPrinter);
1216 Die.addValue(DIEAlloc, Value);
1217 return AttrSize;
1220 unsigned DwarfLinker::DIECloner::cloneAddressAttribute(
1221 DIE &Die, AttributeSpec AttrSpec, const DWARFFormValue &Val,
1222 const CompileUnit &Unit, AttributesInfo &Info) {
1223 uint64_t Addr = *Val.getAsAddress();
1225 if (LLVM_UNLIKELY(Linker.Options.Update)) {
1226 if (AttrSpec.Attr == dwarf::DW_AT_low_pc)
1227 Info.HasLowPc = true;
1228 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1229 dwarf::Form(AttrSpec.Form), DIEInteger(Addr));
1230 return Unit.getOrigUnit().getAddressByteSize();
1233 if (AttrSpec.Attr == dwarf::DW_AT_low_pc) {
1234 if (Die.getTag() == dwarf::DW_TAG_inlined_subroutine ||
1235 Die.getTag() == dwarf::DW_TAG_lexical_block)
1236 // The low_pc of a block or inline subroutine might get
1237 // relocated because it happens to match the low_pc of the
1238 // enclosing subprogram. To prevent issues with that, always use
1239 // the low_pc from the input DIE if relocations have been applied.
1240 Addr = (Info.OrigLowPc != std::numeric_limits<uint64_t>::max()
1241 ? Info.OrigLowPc
1242 : Addr) +
1243 Info.PCOffset;
1244 else if (Die.getTag() == dwarf::DW_TAG_compile_unit) {
1245 Addr = Unit.getLowPc();
1246 if (Addr == std::numeric_limits<uint64_t>::max())
1247 return 0;
1249 Info.HasLowPc = true;
1250 } else if (AttrSpec.Attr == dwarf::DW_AT_high_pc) {
1251 if (Die.getTag() == dwarf::DW_TAG_compile_unit) {
1252 if (uint64_t HighPc = Unit.getHighPc())
1253 Addr = HighPc;
1254 else
1255 return 0;
1256 } else
1257 // If we have a high_pc recorded for the input DIE, use
1258 // it. Otherwise (when no relocations where applied) just use the
1259 // one we just decoded.
1260 Addr = (Info.OrigHighPc ? Info.OrigHighPc : Addr) + Info.PCOffset;
1263 Die.addValue(DIEAlloc, static_cast<dwarf::Attribute>(AttrSpec.Attr),
1264 static_cast<dwarf::Form>(AttrSpec.Form), DIEInteger(Addr));
1265 return Unit.getOrigUnit().getAddressByteSize();
1268 unsigned DwarfLinker::DIECloner::cloneScalarAttribute(
1269 DIE &Die, const DWARFDie &InputDIE, const DebugMapObject &DMO,
1270 CompileUnit &Unit, AttributeSpec AttrSpec, const DWARFFormValue &Val,
1271 unsigned AttrSize, AttributesInfo &Info) {
1272 uint64_t Value;
1274 if (LLVM_UNLIKELY(Linker.Options.Update)) {
1275 if (auto OptionalValue = Val.getAsUnsignedConstant())
1276 Value = *OptionalValue;
1277 else if (auto OptionalValue = Val.getAsSignedConstant())
1278 Value = *OptionalValue;
1279 else if (auto OptionalValue = Val.getAsSectionOffset())
1280 Value = *OptionalValue;
1281 else {
1282 Linker.reportWarning(
1283 "Unsupported scalar attribute form. Dropping attribute.", DMO,
1284 &InputDIE);
1285 return 0;
1287 if (AttrSpec.Attr == dwarf::DW_AT_declaration && Value)
1288 Info.IsDeclaration = true;
1289 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1290 dwarf::Form(AttrSpec.Form), DIEInteger(Value));
1291 return AttrSize;
1294 if (AttrSpec.Attr == dwarf::DW_AT_high_pc &&
1295 Die.getTag() == dwarf::DW_TAG_compile_unit) {
1296 if (Unit.getLowPc() == -1ULL)
1297 return 0;
1298 // Dwarf >= 4 high_pc is an size, not an address.
1299 Value = Unit.getHighPc() - Unit.getLowPc();
1300 } else if (AttrSpec.Form == dwarf::DW_FORM_sec_offset)
1301 Value = *Val.getAsSectionOffset();
1302 else if (AttrSpec.Form == dwarf::DW_FORM_sdata)
1303 Value = *Val.getAsSignedConstant();
1304 else if (auto OptionalValue = Val.getAsUnsignedConstant())
1305 Value = *OptionalValue;
1306 else {
1307 Linker.reportWarning(
1308 "Unsupported scalar attribute form. Dropping attribute.", DMO,
1309 &InputDIE);
1310 return 0;
1312 PatchLocation Patch =
1313 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1314 dwarf::Form(AttrSpec.Form), DIEInteger(Value));
1315 if (AttrSpec.Attr == dwarf::DW_AT_ranges) {
1316 Unit.noteRangeAttribute(Die, Patch);
1317 Info.HasRanges = true;
1320 // A more generic way to check for location attributes would be
1321 // nice, but it's very unlikely that any other attribute needs a
1322 // location list.
1323 // FIXME: use DWARFAttribute::mayHaveLocationDescription().
1324 else if (AttrSpec.Attr == dwarf::DW_AT_location ||
1325 AttrSpec.Attr == dwarf::DW_AT_frame_base)
1326 Unit.noteLocationAttribute(Patch, Info.PCOffset);
1327 else if (AttrSpec.Attr == dwarf::DW_AT_declaration && Value)
1328 Info.IsDeclaration = true;
1330 return AttrSize;
1333 /// Clone \p InputDIE's attribute described by \p AttrSpec with
1334 /// value \p Val, and add it to \p Die.
1335 /// \returns the size of the cloned attribute.
1336 unsigned DwarfLinker::DIECloner::cloneAttribute(
1337 DIE &Die, const DWARFDie &InputDIE, const DebugMapObject &DMO,
1338 CompileUnit &Unit, OffsetsStringPool &StringPool, const DWARFFormValue &Val,
1339 const AttributeSpec AttrSpec, unsigned AttrSize, AttributesInfo &Info,
1340 bool IsLittleEndian) {
1341 const DWARFUnit &U = Unit.getOrigUnit();
1343 switch (AttrSpec.Form) {
1344 case dwarf::DW_FORM_strp:
1345 case dwarf::DW_FORM_string:
1346 return cloneStringAttribute(Die, AttrSpec, Val, U, StringPool, Info);
1347 case dwarf::DW_FORM_ref_addr:
1348 case dwarf::DW_FORM_ref1:
1349 case dwarf::DW_FORM_ref2:
1350 case dwarf::DW_FORM_ref4:
1351 case dwarf::DW_FORM_ref8:
1352 return cloneDieReferenceAttribute(Die, InputDIE, AttrSpec, AttrSize, Val,
1353 DMO, Unit);
1354 case dwarf::DW_FORM_block:
1355 case dwarf::DW_FORM_block1:
1356 case dwarf::DW_FORM_block2:
1357 case dwarf::DW_FORM_block4:
1358 case dwarf::DW_FORM_exprloc:
1359 return cloneBlockAttribute(Die, DMO, Unit, AttrSpec, Val, AttrSize,
1360 IsLittleEndian);
1361 case dwarf::DW_FORM_addr:
1362 return cloneAddressAttribute(Die, AttrSpec, Val, Unit, Info);
1363 case dwarf::DW_FORM_data1:
1364 case dwarf::DW_FORM_data2:
1365 case dwarf::DW_FORM_data4:
1366 case dwarf::DW_FORM_data8:
1367 case dwarf::DW_FORM_udata:
1368 case dwarf::DW_FORM_sdata:
1369 case dwarf::DW_FORM_sec_offset:
1370 case dwarf::DW_FORM_flag:
1371 case dwarf::DW_FORM_flag_present:
1372 return cloneScalarAttribute(Die, InputDIE, DMO, Unit, AttrSpec, Val,
1373 AttrSize, Info);
1374 default:
1375 Linker.reportWarning(
1376 "Unsupported attribute form in cloneAttribute. Dropping.", DMO,
1377 &InputDIE);
1380 return 0;
1383 /// Apply the valid relocations found by findValidRelocs() to
1384 /// the buffer \p Data, taking into account that Data is at \p BaseOffset
1385 /// in the debug_info section.
1387 /// Like for findValidRelocs(), this function must be called with
1388 /// monotonic \p BaseOffset values.
1390 /// \returns whether any reloc has been applied.
1391 bool DwarfLinker::RelocationManager::applyValidRelocs(
1392 MutableArrayRef<char> Data, uint64_t BaseOffset, bool IsLittleEndian) {
1393 assert((NextValidReloc == 0 ||
1394 BaseOffset > ValidRelocs[NextValidReloc - 1].Offset) &&
1395 "BaseOffset should only be increasing.");
1396 if (NextValidReloc >= ValidRelocs.size())
1397 return false;
1399 // Skip relocs that haven't been applied.
1400 while (NextValidReloc < ValidRelocs.size() &&
1401 ValidRelocs[NextValidReloc].Offset < BaseOffset)
1402 ++NextValidReloc;
1404 bool Applied = false;
1405 uint64_t EndOffset = BaseOffset + Data.size();
1406 while (NextValidReloc < ValidRelocs.size() &&
1407 ValidRelocs[NextValidReloc].Offset >= BaseOffset &&
1408 ValidRelocs[NextValidReloc].Offset < EndOffset) {
1409 const auto &ValidReloc = ValidRelocs[NextValidReloc++];
1410 assert(ValidReloc.Offset - BaseOffset < Data.size());
1411 assert(ValidReloc.Offset - BaseOffset + ValidReloc.Size <= Data.size());
1412 char Buf[8];
1413 uint64_t Value = ValidReloc.Mapping->getValue().BinaryAddress;
1414 Value += ValidReloc.Addend;
1415 for (unsigned i = 0; i != ValidReloc.Size; ++i) {
1416 unsigned Index = IsLittleEndian ? i : (ValidReloc.Size - i - 1);
1417 Buf[i] = uint8_t(Value >> (Index * 8));
1419 assert(ValidReloc.Size <= sizeof(Buf));
1420 memcpy(&Data[ValidReloc.Offset - BaseOffset], Buf, ValidReloc.Size);
1421 Applied = true;
1424 return Applied;
1427 static bool isObjCSelector(StringRef Name) {
1428 return Name.size() > 2 && (Name[0] == '-' || Name[0] == '+') &&
1429 (Name[1] == '[');
1432 void DwarfLinker::DIECloner::addObjCAccelerator(CompileUnit &Unit,
1433 const DIE *Die,
1434 DwarfStringPoolEntryRef Name,
1435 OffsetsStringPool &StringPool,
1436 bool SkipPubSection) {
1437 assert(isObjCSelector(Name.getString()) && "not an objc selector");
1438 // Objective C method or class function.
1439 // "- [Class(Category) selector :withArg ...]"
1440 StringRef ClassNameStart(Name.getString().drop_front(2));
1441 size_t FirstSpace = ClassNameStart.find(' ');
1442 if (FirstSpace == StringRef::npos)
1443 return;
1445 StringRef SelectorStart(ClassNameStart.data() + FirstSpace + 1);
1446 if (!SelectorStart.size())
1447 return;
1449 StringRef Selector(SelectorStart.data(), SelectorStart.size() - 1);
1450 Unit.addNameAccelerator(Die, StringPool.getEntry(Selector), SkipPubSection);
1452 // Add an entry for the class name that points to this
1453 // method/class function.
1454 StringRef ClassName(ClassNameStart.data(), FirstSpace);
1455 Unit.addObjCAccelerator(Die, StringPool.getEntry(ClassName), SkipPubSection);
1457 if (ClassName[ClassName.size() - 1] == ')') {
1458 size_t OpenParens = ClassName.find('(');
1459 if (OpenParens != StringRef::npos) {
1460 StringRef ClassNameNoCategory(ClassName.data(), OpenParens);
1461 Unit.addObjCAccelerator(Die, StringPool.getEntry(ClassNameNoCategory),
1462 SkipPubSection);
1464 std::string MethodNameNoCategory(Name.getString().data(), OpenParens + 2);
1465 // FIXME: The missing space here may be a bug, but
1466 // dsymutil-classic also does it this way.
1467 MethodNameNoCategory.append(SelectorStart);
1468 Unit.addNameAccelerator(Die, StringPool.getEntry(MethodNameNoCategory),
1469 SkipPubSection);
1474 static bool
1475 shouldSkipAttribute(DWARFAbbreviationDeclaration::AttributeSpec AttrSpec,
1476 uint16_t Tag, bool InDebugMap, bool SkipPC,
1477 bool InFunctionScope) {
1478 switch (AttrSpec.Attr) {
1479 default:
1480 return false;
1481 case dwarf::DW_AT_low_pc:
1482 case dwarf::DW_AT_high_pc:
1483 case dwarf::DW_AT_ranges:
1484 return SkipPC;
1485 case dwarf::DW_AT_location:
1486 case dwarf::DW_AT_frame_base:
1487 // FIXME: for some reason dsymutil-classic keeps the location attributes
1488 // when they are of block type (i.e. not location lists). This is totally
1489 // wrong for globals where we will keep a wrong address. It is mostly
1490 // harmless for locals, but there is no point in keeping these anyway when
1491 // the function wasn't linked.
1492 return (SkipPC || (!InFunctionScope && Tag == dwarf::DW_TAG_variable &&
1493 !InDebugMap)) &&
1494 !DWARFFormValue(AttrSpec.Form).isFormClass(DWARFFormValue::FC_Block);
1498 DIE *DwarfLinker::DIECloner::cloneDIE(
1499 const DWARFDie &InputDIE, const DebugMapObject &DMO, CompileUnit &Unit,
1500 OffsetsStringPool &StringPool, int64_t PCOffset, uint32_t OutOffset,
1501 unsigned Flags, bool IsLittleEndian, DIE *Die) {
1502 DWARFUnit &U = Unit.getOrigUnit();
1503 unsigned Idx = U.getDIEIndex(InputDIE);
1504 CompileUnit::DIEInfo &Info = Unit.getInfo(Idx);
1506 // Should the DIE appear in the output?
1507 if (!Unit.getInfo(Idx).Keep)
1508 return nullptr;
1510 uint64_t Offset = InputDIE.getOffset();
1511 assert(!(Die && Info.Clone) && "Can't supply a DIE and a cloned DIE");
1512 if (!Die) {
1513 // The DIE might have been already created by a forward reference
1514 // (see cloneDieReferenceAttribute()).
1515 if (!Info.Clone)
1516 Info.Clone = DIE::get(DIEAlloc, dwarf::Tag(InputDIE.getTag()));
1517 Die = Info.Clone;
1520 assert(Die->getTag() == InputDIE.getTag());
1521 Die->setOffset(OutOffset);
1522 if ((Unit.hasODR() || Unit.isClangModule()) && !Info.Incomplete &&
1523 Die->getTag() != dwarf::DW_TAG_namespace && Info.Ctxt &&
1524 Info.Ctxt != Unit.getInfo(Info.ParentIdx).Ctxt &&
1525 !Info.Ctxt->getCanonicalDIEOffset()) {
1526 // We are about to emit a DIE that is the root of its own valid
1527 // DeclContext tree. Make the current offset the canonical offset
1528 // for this context.
1529 Info.Ctxt->setCanonicalDIEOffset(OutOffset + Unit.getStartOffset());
1532 // Extract and clone every attribute.
1533 DWARFDataExtractor Data = U.getDebugInfoExtractor();
1534 // Point to the next DIE (generally there is always at least a NULL
1535 // entry after the current one). If this is a lone
1536 // DW_TAG_compile_unit without any children, point to the next unit.
1537 uint64_t NextOffset = (Idx + 1 < U.getNumDIEs())
1538 ? U.getDIEAtIndex(Idx + 1).getOffset()
1539 : U.getNextUnitOffset();
1540 AttributesInfo AttrInfo;
1542 // We could copy the data only if we need to apply a relocation to it. After
1543 // testing, it seems there is no performance downside to doing the copy
1544 // unconditionally, and it makes the code simpler.
1545 SmallString<40> DIECopy(Data.getData().substr(Offset, NextOffset - Offset));
1546 Data =
1547 DWARFDataExtractor(DIECopy, Data.isLittleEndian(), Data.getAddressSize());
1548 // Modify the copy with relocated addresses.
1549 if (RelocMgr.applyValidRelocs(DIECopy, Offset, Data.isLittleEndian())) {
1550 // If we applied relocations, we store the value of high_pc that was
1551 // potentially stored in the input DIE. If high_pc is an address
1552 // (Dwarf version == 2), then it might have been relocated to a
1553 // totally unrelated value (because the end address in the object
1554 // file might be start address of another function which got moved
1555 // independently by the linker). The computation of the actual
1556 // high_pc value is done in cloneAddressAttribute().
1557 AttrInfo.OrigHighPc =
1558 dwarf::toAddress(InputDIE.find(dwarf::DW_AT_high_pc), 0);
1559 // Also store the low_pc. It might get relocated in an
1560 // inline_subprogram that happens at the beginning of its
1561 // inlining function.
1562 AttrInfo.OrigLowPc = dwarf::toAddress(InputDIE.find(dwarf::DW_AT_low_pc),
1563 std::numeric_limits<uint64_t>::max());
1566 // Reset the Offset to 0 as we will be working on the local copy of
1567 // the data.
1568 Offset = 0;
1570 const auto *Abbrev = InputDIE.getAbbreviationDeclarationPtr();
1571 Offset += getULEB128Size(Abbrev->getCode());
1573 // We are entering a subprogram. Get and propagate the PCOffset.
1574 if (Die->getTag() == dwarf::DW_TAG_subprogram)
1575 PCOffset = Info.AddrAdjust;
1576 AttrInfo.PCOffset = PCOffset;
1578 if (Abbrev->getTag() == dwarf::DW_TAG_subprogram) {
1579 Flags |= TF_InFunctionScope;
1580 if (!Info.InDebugMap && LLVM_LIKELY(!Options.Update))
1581 Flags |= TF_SkipPC;
1584 bool Copied = false;
1585 for (const auto &AttrSpec : Abbrev->attributes()) {
1586 if (LLVM_LIKELY(!Options.Update) &&
1587 shouldSkipAttribute(AttrSpec, Die->getTag(), Info.InDebugMap,
1588 Flags & TF_SkipPC, Flags & TF_InFunctionScope)) {
1589 DWARFFormValue::skipValue(AttrSpec.Form, Data, &Offset,
1590 U.getFormParams());
1591 // FIXME: dsymutil-classic keeps the old abbreviation around
1592 // even if it's not used. We can remove this (and the copyAbbrev
1593 // helper) as soon as bit-for-bit compatibility is not a goal anymore.
1594 if (!Copied) {
1595 copyAbbrev(*InputDIE.getAbbreviationDeclarationPtr(), Unit.hasODR());
1596 Copied = true;
1598 continue;
1601 DWARFFormValue Val(AttrSpec.Form);
1602 uint64_t AttrSize = Offset;
1603 Val.extractValue(Data, &Offset, U.getFormParams(), &U);
1604 AttrSize = Offset - AttrSize;
1606 OutOffset += cloneAttribute(*Die, InputDIE, DMO, Unit, StringPool, Val,
1607 AttrSpec, AttrSize, AttrInfo, IsLittleEndian);
1610 // Look for accelerator entries.
1611 uint16_t Tag = InputDIE.getTag();
1612 // FIXME: This is slightly wrong. An inline_subroutine without a
1613 // low_pc, but with AT_ranges might be interesting to get into the
1614 // accelerator tables too. For now stick with dsymutil's behavior.
1615 if ((Info.InDebugMap || AttrInfo.HasLowPc || AttrInfo.HasRanges) &&
1616 Tag != dwarf::DW_TAG_compile_unit &&
1617 getDIENames(InputDIE, AttrInfo, StringPool,
1618 Tag != dwarf::DW_TAG_inlined_subroutine)) {
1619 if (AttrInfo.MangledName && AttrInfo.MangledName != AttrInfo.Name)
1620 Unit.addNameAccelerator(Die, AttrInfo.MangledName,
1621 Tag == dwarf::DW_TAG_inlined_subroutine);
1622 if (AttrInfo.Name) {
1623 if (AttrInfo.NameWithoutTemplate)
1624 Unit.addNameAccelerator(Die, AttrInfo.NameWithoutTemplate,
1625 /* SkipPubSection */ true);
1626 Unit.addNameAccelerator(Die, AttrInfo.Name,
1627 Tag == dwarf::DW_TAG_inlined_subroutine);
1629 if (AttrInfo.Name && isObjCSelector(AttrInfo.Name.getString()))
1630 addObjCAccelerator(Unit, Die, AttrInfo.Name, StringPool,
1631 /* SkipPubSection =*/true);
1633 } else if (Tag == dwarf::DW_TAG_namespace) {
1634 if (!AttrInfo.Name)
1635 AttrInfo.Name = StringPool.getEntry("(anonymous namespace)");
1636 Unit.addNamespaceAccelerator(Die, AttrInfo.Name);
1637 } else if (isTypeTag(Tag) && !AttrInfo.IsDeclaration &&
1638 getDIENames(InputDIE, AttrInfo, StringPool) && AttrInfo.Name &&
1639 AttrInfo.Name.getString()[0]) {
1640 uint32_t Hash = hashFullyQualifiedName(InputDIE, Unit, DMO);
1641 uint64_t RuntimeLang =
1642 dwarf::toUnsigned(InputDIE.find(dwarf::DW_AT_APPLE_runtime_class))
1643 .getValueOr(0);
1644 bool ObjCClassIsImplementation =
1645 (RuntimeLang == dwarf::DW_LANG_ObjC ||
1646 RuntimeLang == dwarf::DW_LANG_ObjC_plus_plus) &&
1647 dwarf::toUnsigned(InputDIE.find(dwarf::DW_AT_APPLE_objc_complete_type))
1648 .getValueOr(0);
1649 Unit.addTypeAccelerator(Die, AttrInfo.Name, ObjCClassIsImplementation,
1650 Hash);
1653 // Determine whether there are any children that we want to keep.
1654 bool HasChildren = false;
1655 for (auto Child : InputDIE.children()) {
1656 unsigned Idx = U.getDIEIndex(Child);
1657 if (Unit.getInfo(Idx).Keep) {
1658 HasChildren = true;
1659 break;
1663 DIEAbbrev NewAbbrev = Die->generateAbbrev();
1664 if (HasChildren)
1665 NewAbbrev.setChildrenFlag(dwarf::DW_CHILDREN_yes);
1666 // Assign a permanent abbrev number
1667 Linker.AssignAbbrev(NewAbbrev);
1668 Die->setAbbrevNumber(NewAbbrev.getNumber());
1670 // Add the size of the abbreviation number to the output offset.
1671 OutOffset += getULEB128Size(Die->getAbbrevNumber());
1673 if (!HasChildren) {
1674 // Update our size.
1675 Die->setSize(OutOffset - Die->getOffset());
1676 return Die;
1679 // Recursively clone children.
1680 for (auto Child : InputDIE.children()) {
1681 if (DIE *Clone = cloneDIE(Child, DMO, Unit, StringPool, PCOffset, OutOffset,
1682 Flags, IsLittleEndian)) {
1683 Die->addChild(Clone);
1684 OutOffset = Clone->getOffset() + Clone->getSize();
1688 // Account for the end of children marker.
1689 OutOffset += sizeof(int8_t);
1690 // Update our size.
1691 Die->setSize(OutOffset - Die->getOffset());
1692 return Die;
1695 /// Patch the input object file relevant debug_ranges entries
1696 /// and emit them in the output file. Update the relevant attributes
1697 /// to point at the new entries.
1698 void DwarfLinker::patchRangesForUnit(const CompileUnit &Unit,
1699 DWARFContext &OrigDwarf,
1700 const DebugMapObject &DMO) const {
1701 DWARFDebugRangeList RangeList;
1702 const auto &FunctionRanges = Unit.getFunctionRanges();
1703 unsigned AddressSize = Unit.getOrigUnit().getAddressByteSize();
1704 DWARFDataExtractor RangeExtractor(OrigDwarf.getDWARFObj(),
1705 OrigDwarf.getDWARFObj().getRangesSection(),
1706 OrigDwarf.isLittleEndian(), AddressSize);
1707 auto InvalidRange = FunctionRanges.end(), CurrRange = InvalidRange;
1708 DWARFUnit &OrigUnit = Unit.getOrigUnit();
1709 auto OrigUnitDie = OrigUnit.getUnitDIE(false);
1710 uint64_t OrigLowPc =
1711 dwarf::toAddress(OrigUnitDie.find(dwarf::DW_AT_low_pc), -1ULL);
1712 // Ranges addresses are based on the unit's low_pc. Compute the
1713 // offset we need to apply to adapt to the new unit's low_pc.
1714 int64_t UnitPcOffset = 0;
1715 if (OrigLowPc != -1ULL)
1716 UnitPcOffset = int64_t(OrigLowPc) - Unit.getLowPc();
1718 for (const auto &RangeAttribute : Unit.getRangesAttributes()) {
1719 uint64_t Offset = RangeAttribute.get();
1720 RangeAttribute.set(Streamer->getRangesSectionSize());
1721 if (Error E = RangeList.extract(RangeExtractor, &Offset)) {
1722 llvm::consumeError(std::move(E));
1723 reportWarning("invalid range list ignored.", DMO);
1724 RangeList.clear();
1726 const auto &Entries = RangeList.getEntries();
1727 if (!Entries.empty()) {
1728 const DWARFDebugRangeList::RangeListEntry &First = Entries.front();
1730 if (CurrRange == InvalidRange ||
1731 First.StartAddress + OrigLowPc < CurrRange.start() ||
1732 First.StartAddress + OrigLowPc >= CurrRange.stop()) {
1733 CurrRange = FunctionRanges.find(First.StartAddress + OrigLowPc);
1734 if (CurrRange == InvalidRange ||
1735 CurrRange.start() > First.StartAddress + OrigLowPc) {
1736 reportWarning("no mapping for range.", DMO);
1737 continue;
1742 Streamer->emitRangesEntries(UnitPcOffset, OrigLowPc, CurrRange, Entries,
1743 AddressSize);
1747 /// Generate the debug_aranges entries for \p Unit and if the
1748 /// unit has a DW_AT_ranges attribute, also emit the debug_ranges
1749 /// contribution for this attribute.
1750 /// FIXME: this could actually be done right in patchRangesForUnit,
1751 /// but for the sake of initial bit-for-bit compatibility with legacy
1752 /// dsymutil, we have to do it in a delayed pass.
1753 void DwarfLinker::generateUnitRanges(CompileUnit &Unit) const {
1754 auto Attr = Unit.getUnitRangesAttribute();
1755 if (Attr)
1756 Attr->set(Streamer->getRangesSectionSize());
1757 Streamer->emitUnitRangesEntries(Unit, static_cast<bool>(Attr));
1760 /// Insert the new line info sequence \p Seq into the current
1761 /// set of already linked line info \p Rows.
1762 static void insertLineSequence(std::vector<DWARFDebugLine::Row> &Seq,
1763 std::vector<DWARFDebugLine::Row> &Rows) {
1764 if (Seq.empty())
1765 return;
1767 if (!Rows.empty() && Rows.back().Address < Seq.front().Address) {
1768 Rows.insert(Rows.end(), Seq.begin(), Seq.end());
1769 Seq.clear();
1770 return;
1773 object::SectionedAddress Front = Seq.front().Address;
1774 auto InsertPoint = partition_point(
1775 Rows, [=](const DWARFDebugLine::Row &O) { return O.Address < Front; });
1777 // FIXME: this only removes the unneeded end_sequence if the
1778 // sequences have been inserted in order. Using a global sort like
1779 // described in patchLineTableForUnit() and delaying the end_sequene
1780 // elimination to emitLineTableForUnit() we can get rid of all of them.
1781 if (InsertPoint != Rows.end() && InsertPoint->Address == Front &&
1782 InsertPoint->EndSequence) {
1783 *InsertPoint = Seq.front();
1784 Rows.insert(InsertPoint + 1, Seq.begin() + 1, Seq.end());
1785 } else {
1786 Rows.insert(InsertPoint, Seq.begin(), Seq.end());
1789 Seq.clear();
1792 static void patchStmtList(DIE &Die, DIEInteger Offset) {
1793 for (auto &V : Die.values())
1794 if (V.getAttribute() == dwarf::DW_AT_stmt_list) {
1795 V = DIEValue(V.getAttribute(), V.getForm(), Offset);
1796 return;
1799 llvm_unreachable("Didn't find DW_AT_stmt_list in cloned DIE!");
1802 /// Extract the line table for \p Unit from \p OrigDwarf, and
1803 /// recreate a relocated version of these for the address ranges that
1804 /// are present in the binary.
1805 void DwarfLinker::patchLineTableForUnit(CompileUnit &Unit,
1806 DWARFContext &OrigDwarf,
1807 RangesTy &Ranges,
1808 const DebugMapObject &DMO) {
1809 DWARFDie CUDie = Unit.getOrigUnit().getUnitDIE();
1810 auto StmtList = dwarf::toSectionOffset(CUDie.find(dwarf::DW_AT_stmt_list));
1811 if (!StmtList)
1812 return;
1814 // Update the cloned DW_AT_stmt_list with the correct debug_line offset.
1815 if (auto *OutputDIE = Unit.getOutputUnitDIE())
1816 patchStmtList(*OutputDIE, DIEInteger(Streamer->getLineSectionSize()));
1818 // Parse the original line info for the unit.
1819 DWARFDebugLine::LineTable LineTable;
1820 uint64_t StmtOffset = *StmtList;
1821 DWARFDataExtractor LineExtractor(
1822 OrigDwarf.getDWARFObj(), OrigDwarf.getDWARFObj().getLineSection(),
1823 OrigDwarf.isLittleEndian(), Unit.getOrigUnit().getAddressByteSize());
1824 if (Options.Translator)
1825 return Streamer->translateLineTable(LineExtractor, StmtOffset);
1827 Error Err = LineTable.parse(LineExtractor, &StmtOffset, OrigDwarf,
1828 &Unit.getOrigUnit(), DWARFContext::dumpWarning);
1829 DWARFContext::dumpWarning(std::move(Err));
1831 // This vector is the output line table.
1832 std::vector<DWARFDebugLine::Row> NewRows;
1833 NewRows.reserve(LineTable.Rows.size());
1835 // Current sequence of rows being extracted, before being inserted
1836 // in NewRows.
1837 std::vector<DWARFDebugLine::Row> Seq;
1838 const auto &FunctionRanges = Unit.getFunctionRanges();
1839 auto InvalidRange = FunctionRanges.end(), CurrRange = InvalidRange;
1841 // FIXME: This logic is meant to generate exactly the same output as
1842 // Darwin's classic dsymutil. There is a nicer way to implement this
1843 // by simply putting all the relocated line info in NewRows and simply
1844 // sorting NewRows before passing it to emitLineTableForUnit. This
1845 // should be correct as sequences for a function should stay
1846 // together in the sorted output. There are a few corner cases that
1847 // look suspicious though, and that required to implement the logic
1848 // this way. Revisit that once initial validation is finished.
1850 // Iterate over the object file line info and extract the sequences
1851 // that correspond to linked functions.
1852 for (auto &Row : LineTable.Rows) {
1853 // Check whether we stepped out of the range. The range is
1854 // half-open, but consider accept the end address of the range if
1855 // it is marked as end_sequence in the input (because in that
1856 // case, the relocation offset is accurate and that entry won't
1857 // serve as the start of another function).
1858 if (CurrRange == InvalidRange || Row.Address.Address < CurrRange.start() ||
1859 Row.Address.Address > CurrRange.stop() ||
1860 (Row.Address.Address == CurrRange.stop() && !Row.EndSequence)) {
1861 // We just stepped out of a known range. Insert a end_sequence
1862 // corresponding to the end of the range.
1863 uint64_t StopAddress = CurrRange != InvalidRange
1864 ? CurrRange.stop() + CurrRange.value()
1865 : -1ULL;
1866 CurrRange = FunctionRanges.find(Row.Address.Address);
1867 bool CurrRangeValid =
1868 CurrRange != InvalidRange && CurrRange.start() <= Row.Address.Address;
1869 if (!CurrRangeValid) {
1870 CurrRange = InvalidRange;
1871 if (StopAddress != -1ULL) {
1872 // Try harder by looking in the DebugMapObject function
1873 // ranges map. There are corner cases where this finds a
1874 // valid entry. It's unclear if this is right or wrong, but
1875 // for now do as dsymutil.
1876 // FIXME: Understand exactly what cases this addresses and
1877 // potentially remove it along with the Ranges map.
1878 auto Range = Ranges.lower_bound(Row.Address.Address);
1879 if (Range != Ranges.begin() && Range != Ranges.end())
1880 --Range;
1882 if (Range != Ranges.end() && Range->first <= Row.Address.Address &&
1883 Range->second.HighPC >= Row.Address.Address) {
1884 StopAddress = Row.Address.Address + Range->second.Offset;
1888 if (StopAddress != -1ULL && !Seq.empty()) {
1889 // Insert end sequence row with the computed end address, but
1890 // the same line as the previous one.
1891 auto NextLine = Seq.back();
1892 NextLine.Address.Address = StopAddress;
1893 NextLine.EndSequence = 1;
1894 NextLine.PrologueEnd = 0;
1895 NextLine.BasicBlock = 0;
1896 NextLine.EpilogueBegin = 0;
1897 Seq.push_back(NextLine);
1898 insertLineSequence(Seq, NewRows);
1901 if (!CurrRangeValid)
1902 continue;
1905 // Ignore empty sequences.
1906 if (Row.EndSequence && Seq.empty())
1907 continue;
1909 // Relocate row address and add it to the current sequence.
1910 Row.Address.Address += CurrRange.value();
1911 Seq.emplace_back(Row);
1913 if (Row.EndSequence)
1914 insertLineSequence(Seq, NewRows);
1917 // Finished extracting, now emit the line tables.
1918 // FIXME: LLVM hard-codes its prologue values. We just copy the
1919 // prologue over and that works because we act as both producer and
1920 // consumer. It would be nicer to have a real configurable line
1921 // table emitter.
1922 if (LineTable.Prologue.getVersion() < 2 ||
1923 LineTable.Prologue.getVersion() > 5 ||
1924 LineTable.Prologue.DefaultIsStmt != DWARF2_LINE_DEFAULT_IS_STMT ||
1925 LineTable.Prologue.OpcodeBase > 13)
1926 reportWarning("line table parameters mismatch. Cannot emit.", DMO);
1927 else {
1928 uint32_t PrologueEnd = *StmtList + 10 + LineTable.Prologue.PrologueLength;
1929 // DWARF v5 has an extra 2 bytes of information before the header_length
1930 // field.
1931 if (LineTable.Prologue.getVersion() == 5)
1932 PrologueEnd += 2;
1933 StringRef LineData = OrigDwarf.getDWARFObj().getLineSection().Data;
1934 MCDwarfLineTableParams Params;
1935 Params.DWARF2LineOpcodeBase = LineTable.Prologue.OpcodeBase;
1936 Params.DWARF2LineBase = LineTable.Prologue.LineBase;
1937 Params.DWARF2LineRange = LineTable.Prologue.LineRange;
1938 Streamer->emitLineTableForUnit(Params,
1939 LineData.slice(*StmtList + 4, PrologueEnd),
1940 LineTable.Prologue.MinInstLength, NewRows,
1941 Unit.getOrigUnit().getAddressByteSize());
1945 void DwarfLinker::emitAcceleratorEntriesForUnit(CompileUnit &Unit) {
1946 switch (Options.TheAccelTableKind) {
1947 case AccelTableKind::Apple:
1948 emitAppleAcceleratorEntriesForUnit(Unit);
1949 break;
1950 case AccelTableKind::Dwarf:
1951 emitDwarfAcceleratorEntriesForUnit(Unit);
1952 break;
1953 case AccelTableKind::Default:
1954 llvm_unreachable("The default must be updated to a concrete value.");
1955 break;
1959 void DwarfLinker::emitAppleAcceleratorEntriesForUnit(CompileUnit &Unit) {
1960 // Add namespaces.
1961 for (const auto &Namespace : Unit.getNamespaces())
1962 AppleNamespaces.addName(Namespace.Name,
1963 Namespace.Die->getOffset() + Unit.getStartOffset());
1965 /// Add names.
1966 if (!Options.Minimize)
1967 Streamer->emitPubNamesForUnit(Unit);
1968 for (const auto &Pubname : Unit.getPubnames())
1969 AppleNames.addName(Pubname.Name,
1970 Pubname.Die->getOffset() + Unit.getStartOffset());
1972 /// Add types.
1973 if (!Options.Minimize)
1974 Streamer->emitPubTypesForUnit(Unit);
1975 for (const auto &Pubtype : Unit.getPubtypes())
1976 AppleTypes.addName(
1977 Pubtype.Name, Pubtype.Die->getOffset() + Unit.getStartOffset(),
1978 Pubtype.Die->getTag(),
1979 Pubtype.ObjcClassImplementation ? dwarf::DW_FLAG_type_implementation
1980 : 0,
1981 Pubtype.QualifiedNameHash);
1983 /// Add ObjC names.
1984 for (const auto &ObjC : Unit.getObjC())
1985 AppleObjc.addName(ObjC.Name, ObjC.Die->getOffset() + Unit.getStartOffset());
1988 void DwarfLinker::emitDwarfAcceleratorEntriesForUnit(CompileUnit &Unit) {
1989 for (const auto &Namespace : Unit.getNamespaces())
1990 DebugNames.addName(Namespace.Name, Namespace.Die->getOffset(),
1991 Namespace.Die->getTag(), Unit.getUniqueID());
1992 for (const auto &Pubname : Unit.getPubnames())
1993 DebugNames.addName(Pubname.Name, Pubname.Die->getOffset(),
1994 Pubname.Die->getTag(), Unit.getUniqueID());
1995 for (const auto &Pubtype : Unit.getPubtypes())
1996 DebugNames.addName(Pubtype.Name, Pubtype.Die->getOffset(),
1997 Pubtype.Die->getTag(), Unit.getUniqueID());
2000 /// Read the frame info stored in the object, and emit the
2001 /// patched frame descriptions for the linked binary.
2003 /// This is actually pretty easy as the data of the CIEs and FDEs can
2004 /// be considered as black boxes and moved as is. The only thing to do
2005 /// is to patch the addresses in the headers.
2006 void DwarfLinker::patchFrameInfoForObject(const DebugMapObject &DMO,
2007 RangesTy &Ranges,
2008 DWARFContext &OrigDwarf,
2009 unsigned AddrSize) {
2010 StringRef FrameData = OrigDwarf.getDWARFObj().getFrameSection().Data;
2011 if (FrameData.empty())
2012 return;
2014 DataExtractor Data(FrameData, OrigDwarf.isLittleEndian(), 0);
2015 uint64_t InputOffset = 0;
2017 // Store the data of the CIEs defined in this object, keyed by their
2018 // offsets.
2019 DenseMap<uint64_t, StringRef> LocalCIES;
2021 while (Data.isValidOffset(InputOffset)) {
2022 uint64_t EntryOffset = InputOffset;
2023 uint32_t InitialLength = Data.getU32(&InputOffset);
2024 if (InitialLength == 0xFFFFFFFF)
2025 return reportWarning("Dwarf64 bits no supported", DMO);
2027 uint32_t CIEId = Data.getU32(&InputOffset);
2028 if (CIEId == 0xFFFFFFFF) {
2029 // This is a CIE, store it.
2030 StringRef CIEData = FrameData.substr(EntryOffset, InitialLength + 4);
2031 LocalCIES[EntryOffset] = CIEData;
2032 // The -4 is to account for the CIEId we just read.
2033 InputOffset += InitialLength - 4;
2034 continue;
2037 uint32_t Loc = Data.getUnsigned(&InputOffset, AddrSize);
2039 // Some compilers seem to emit frame info that doesn't start at
2040 // the function entry point, thus we can't just lookup the address
2041 // in the debug map. Use the linker's range map to see if the FDE
2042 // describes something that we can relocate.
2043 auto Range = Ranges.upper_bound(Loc);
2044 if (Range != Ranges.begin())
2045 --Range;
2046 if (Range == Ranges.end() || Range->first > Loc ||
2047 Range->second.HighPC <= Loc) {
2048 // The +4 is to account for the size of the InitialLength field itself.
2049 InputOffset = EntryOffset + InitialLength + 4;
2050 continue;
2053 // This is an FDE, and we have a mapping.
2054 // Have we already emitted a corresponding CIE?
2055 StringRef CIEData = LocalCIES[CIEId];
2056 if (CIEData.empty())
2057 return reportWarning("Inconsistent debug_frame content. Dropping.", DMO);
2059 // Look if we already emitted a CIE that corresponds to the
2060 // referenced one (the CIE data is the key of that lookup).
2061 auto IteratorInserted = EmittedCIEs.insert(
2062 std::make_pair(CIEData, Streamer->getFrameSectionSize()));
2063 // If there is no CIE yet for this ID, emit it.
2064 if (IteratorInserted.second ||
2065 // FIXME: dsymutil-classic only caches the last used CIE for
2066 // reuse. Mimic that behavior for now. Just removing that
2067 // second half of the condition and the LastCIEOffset variable
2068 // makes the code DTRT.
2069 LastCIEOffset != IteratorInserted.first->getValue()) {
2070 LastCIEOffset = Streamer->getFrameSectionSize();
2071 IteratorInserted.first->getValue() = LastCIEOffset;
2072 Streamer->emitCIE(CIEData);
2075 // Emit the FDE with updated address and CIE pointer.
2076 // (4 + AddrSize) is the size of the CIEId + initial_location
2077 // fields that will get reconstructed by emitFDE().
2078 unsigned FDERemainingBytes = InitialLength - (4 + AddrSize);
2079 Streamer->emitFDE(IteratorInserted.first->getValue(), AddrSize,
2080 Loc + Range->second.Offset,
2081 FrameData.substr(InputOffset, FDERemainingBytes));
2082 InputOffset += FDERemainingBytes;
2086 void DwarfLinker::DIECloner::copyAbbrev(
2087 const DWARFAbbreviationDeclaration &Abbrev, bool hasODR) {
2088 DIEAbbrev Copy(dwarf::Tag(Abbrev.getTag()),
2089 dwarf::Form(Abbrev.hasChildren()));
2091 for (const auto &Attr : Abbrev.attributes()) {
2092 uint16_t Form = Attr.Form;
2093 if (hasODR && isODRAttribute(Attr.Attr))
2094 Form = dwarf::DW_FORM_ref_addr;
2095 Copy.AddAttribute(dwarf::Attribute(Attr.Attr), dwarf::Form(Form));
2098 Linker.AssignAbbrev(Copy);
2101 uint32_t
2102 DwarfLinker::DIECloner::hashFullyQualifiedName(DWARFDie DIE, CompileUnit &U,
2103 const DebugMapObject &DMO,
2104 int ChildRecurseDepth) {
2105 const char *Name = nullptr;
2106 DWARFUnit *OrigUnit = &U.getOrigUnit();
2107 CompileUnit *CU = &U;
2108 Optional<DWARFFormValue> Ref;
2110 while (1) {
2111 if (const char *CurrentName = DIE.getName(DINameKind::ShortName))
2112 Name = CurrentName;
2114 if (!(Ref = DIE.find(dwarf::DW_AT_specification)) &&
2115 !(Ref = DIE.find(dwarf::DW_AT_abstract_origin)))
2116 break;
2118 if (!Ref->isFormClass(DWARFFormValue::FC_Reference))
2119 break;
2121 CompileUnit *RefCU;
2122 if (auto RefDIE =
2123 resolveDIEReference(Linker, DMO, CompileUnits, *Ref, DIE, RefCU)) {
2124 CU = RefCU;
2125 OrigUnit = &RefCU->getOrigUnit();
2126 DIE = RefDIE;
2130 unsigned Idx = OrigUnit->getDIEIndex(DIE);
2131 if (!Name && DIE.getTag() == dwarf::DW_TAG_namespace)
2132 Name = "(anonymous namespace)";
2134 if (CU->getInfo(Idx).ParentIdx == 0 ||
2135 // FIXME: dsymutil-classic compatibility. Ignore modules.
2136 CU->getOrigUnit().getDIEAtIndex(CU->getInfo(Idx).ParentIdx).getTag() ==
2137 dwarf::DW_TAG_module)
2138 return djbHash(Name ? Name : "", djbHash(ChildRecurseDepth ? "" : "::"));
2140 DWARFDie Die = OrigUnit->getDIEAtIndex(CU->getInfo(Idx).ParentIdx);
2141 return djbHash(
2142 (Name ? Name : ""),
2143 djbHash((Name ? "::" : ""),
2144 hashFullyQualifiedName(Die, *CU, DMO, ++ChildRecurseDepth)));
2147 static uint64_t getDwoId(const DWARFDie &CUDie, const DWARFUnit &Unit) {
2148 auto DwoId = dwarf::toUnsigned(
2149 CUDie.find({dwarf::DW_AT_dwo_id, dwarf::DW_AT_GNU_dwo_id}));
2150 if (DwoId)
2151 return *DwoId;
2152 return 0;
2155 bool DwarfLinker::registerModuleReference(
2156 DWARFDie CUDie, const DWARFUnit &Unit, DebugMap &ModuleMap,
2157 const DebugMapObject &DMO, RangesTy &Ranges, OffsetsStringPool &StringPool,
2158 UniquingStringPool &UniquingStringPool, DeclContextTree &ODRContexts,
2159 uint64_t ModulesEndOffset, unsigned &UnitID, bool IsLittleEndian,
2160 unsigned Indent, bool Quiet) {
2161 std::string PCMfile = dwarf::toString(
2162 CUDie.find({dwarf::DW_AT_dwo_name, dwarf::DW_AT_GNU_dwo_name}), "");
2163 if (PCMfile.empty())
2164 return false;
2166 // Clang module DWARF skeleton CUs abuse this for the path to the module.
2167 uint64_t DwoId = getDwoId(CUDie, Unit);
2169 std::string Name = dwarf::toString(CUDie.find(dwarf::DW_AT_name), "");
2170 if (Name.empty()) {
2171 if (!Quiet)
2172 reportWarning("Anonymous module skeleton CU for " + PCMfile, DMO);
2173 return true;
2176 if (!Quiet && Options.Verbose) {
2177 outs().indent(Indent);
2178 outs() << "Found clang module reference " << PCMfile;
2181 auto Cached = ClangModules.find(PCMfile);
2182 if (Cached != ClangModules.end()) {
2183 // FIXME: Until PR27449 (https://llvm.org/bugs/show_bug.cgi?id=27449) is
2184 // fixed in clang, only warn about DWO_id mismatches in verbose mode.
2185 // ASTFileSignatures will change randomly when a module is rebuilt.
2186 if (!Quiet && Options.Verbose && (Cached->second != DwoId))
2187 reportWarning(Twine("hash mismatch: this object file was built against a "
2188 "different version of the module ") +
2189 PCMfile,
2190 DMO);
2191 if (!Quiet && Options.Verbose)
2192 outs() << " [cached].\n";
2193 return true;
2195 if (!Quiet && Options.Verbose)
2196 outs() << " ...\n";
2198 // Cyclic dependencies are disallowed by Clang, but we still
2199 // shouldn't run into an infinite loop, so mark it as processed now.
2200 ClangModules.insert({PCMfile, DwoId});
2202 if (Error E = loadClangModule(CUDie, PCMfile, Name, DwoId, ModuleMap, DMO,
2203 Ranges, StringPool, UniquingStringPool,
2204 ODRContexts, ModulesEndOffset, UnitID,
2205 IsLittleEndian, Indent + 2, Quiet)) {
2206 consumeError(std::move(E));
2207 return false;
2209 return true;
2212 ErrorOr<const object::ObjectFile &>
2213 DwarfLinker::loadObject(const DebugMapObject &Obj, const DebugMap &Map) {
2214 auto ObjectEntry =
2215 BinHolder.getObjectEntry(Obj.getObjectFilename(), Obj.getTimestamp());
2216 if (!ObjectEntry) {
2217 auto Err = ObjectEntry.takeError();
2218 reportWarning(
2219 Twine(Obj.getObjectFilename()) + ": " + toString(std::move(Err)), Obj);
2220 return errorToErrorCode(std::move(Err));
2223 auto Object = ObjectEntry->getObject(Map.getTriple());
2224 if (!Object) {
2225 auto Err = Object.takeError();
2226 reportWarning(
2227 Twine(Obj.getObjectFilename()) + ": " + toString(std::move(Err)), Obj);
2228 return errorToErrorCode(std::move(Err));
2231 return *Object;
2234 Error DwarfLinker::loadClangModule(
2235 DWARFDie CUDie, StringRef Filename, StringRef ModuleName, uint64_t DwoId,
2236 DebugMap &ModuleMap, const DebugMapObject &DMO, RangesTy &Ranges,
2237 OffsetsStringPool &StringPool, UniquingStringPool &UniquingStringPool,
2238 DeclContextTree &ODRContexts, uint64_t ModulesEndOffset, unsigned &UnitID,
2239 bool IsLittleEndian, unsigned Indent, bool Quiet) {
2240 /// Using a SmallString<0> because loadClangModule() is recursive.
2241 SmallString<0> Path(Options.PrependPath);
2242 if (sys::path::is_relative(Filename))
2243 resolveRelativeObjectPath(Path, CUDie);
2244 sys::path::append(Path, Filename);
2245 // Don't use the cached binary holder because we have no thread-safety
2246 // guarantee and the lifetime is limited.
2247 auto &Obj = ModuleMap.addDebugMapObject(
2248 Path, sys::TimePoint<std::chrono::seconds>(), MachO::N_OSO);
2249 auto ErrOrObj = loadObject(Obj, ModuleMap);
2250 if (!ErrOrObj) {
2251 // Try and emit more helpful warnings by applying some heuristics.
2252 StringRef ObjFile = DMO.getObjectFilename();
2253 bool isClangModule = sys::path::extension(Filename).equals(".pcm");
2254 bool isArchive = ObjFile.endswith(")");
2255 if (isClangModule) {
2256 StringRef ModuleCacheDir = sys::path::parent_path(Path);
2257 if (sys::fs::exists(ModuleCacheDir)) {
2258 // If the module's parent directory exists, we assume that the module
2259 // cache has expired and was pruned by clang. A more adventurous
2260 // dsymutil would invoke clang to rebuild the module now.
2261 if (!ModuleCacheHintDisplayed) {
2262 WithColor::note() << "The clang module cache may have expired since "
2263 "this object file was built. Rebuilding the "
2264 "object file will rebuild the module cache.\n";
2265 ModuleCacheHintDisplayed = true;
2267 } else if (isArchive) {
2268 // If the module cache directory doesn't exist at all and the object
2269 // file is inside a static library, we assume that the static library
2270 // was built on a different machine. We don't want to discourage module
2271 // debugging for convenience libraries within a project though.
2272 if (!ArchiveHintDisplayed) {
2273 WithColor::note()
2274 << "Linking a static library that was built with "
2275 "-gmodules, but the module cache was not found. "
2276 "Redistributable static libraries should never be "
2277 "built with module debugging enabled. The debug "
2278 "experience will be degraded due to incomplete "
2279 "debug information.\n";
2280 ArchiveHintDisplayed = true;
2284 return Error::success();
2287 std::unique_ptr<CompileUnit> Unit;
2289 // Setup access to the debug info.
2290 auto DwarfContext = DWARFContext::create(*ErrOrObj);
2291 RelocationManager RelocMgr(*this);
2293 for (const auto &CU : DwarfContext->compile_units()) {
2294 updateDwarfVersion(CU->getVersion());
2295 // Recursively get all modules imported by this one.
2296 auto CUDie = CU->getUnitDIE(false);
2297 if (!CUDie)
2298 continue;
2299 if (!registerModuleReference(CUDie, *CU, ModuleMap, DMO, Ranges, StringPool,
2300 UniquingStringPool, ODRContexts,
2301 ModulesEndOffset, UnitID, IsLittleEndian,
2302 Indent, Quiet)) {
2303 if (Unit) {
2304 std::string Err =
2305 (Filename +
2306 ": Clang modules are expected to have exactly 1 compile unit.\n")
2307 .str();
2308 error(Err);
2309 return make_error<StringError>(Err, inconvertibleErrorCode());
2311 // FIXME: Until PR27449 (https://llvm.org/bugs/show_bug.cgi?id=27449) is
2312 // fixed in clang, only warn about DWO_id mismatches in verbose mode.
2313 // ASTFileSignatures will change randomly when a module is rebuilt.
2314 uint64_t PCMDwoId = getDwoId(CUDie, *CU);
2315 if (PCMDwoId != DwoId) {
2316 if (!Quiet && Options.Verbose)
2317 reportWarning(
2318 Twine("hash mismatch: this object file was built against a "
2319 "different version of the module ") +
2320 Filename,
2321 DMO);
2322 // Update the cache entry with the DwoId of the module loaded from disk.
2323 ClangModules[Filename] = PCMDwoId;
2326 // Add this module.
2327 Unit = std::make_unique<CompileUnit>(*CU, UnitID++, !Options.NoODR,
2328 ModuleName);
2329 Unit->setHasInterestingContent();
2330 analyzeContextInfo(CUDie, 0, *Unit, &ODRContexts.getRoot(),
2331 UniquingStringPool, ODRContexts, ModulesEndOffset,
2332 ParseableSwiftInterfaces,
2333 [&](const Twine &Warning, const DWARFDie &DIE) {
2334 reportWarning(Warning, DMO, &DIE);
2336 // Keep everything.
2337 Unit->markEverythingAsKept();
2340 if (!Unit->getOrigUnit().getUnitDIE().hasChildren())
2341 return Error::success();
2342 if (!Quiet && Options.Verbose) {
2343 outs().indent(Indent);
2344 outs() << "cloning .debug_info from " << Filename << "\n";
2347 UnitListTy CompileUnits;
2348 CompileUnits.push_back(std::move(Unit));
2349 DIECloner(*this, RelocMgr, DIEAlloc, CompileUnits, Options)
2350 .cloneAllCompileUnits(*DwarfContext, DMO, Ranges, StringPool,
2351 IsLittleEndian);
2352 return Error::success();
2355 void DwarfLinker::DIECloner::cloneAllCompileUnits(
2356 DWARFContext &DwarfContext, const DebugMapObject &DMO, RangesTy &Ranges,
2357 OffsetsStringPool &StringPool, bool IsLittleEndian) {
2358 if (!Linker.Streamer)
2359 return;
2361 for (auto &CurrentUnit : CompileUnits) {
2362 auto InputDIE = CurrentUnit->getOrigUnit().getUnitDIE();
2363 CurrentUnit->setStartOffset(Linker.OutputDebugInfoSize);
2364 if (!InputDIE) {
2365 Linker.OutputDebugInfoSize = CurrentUnit->computeNextUnitOffset();
2366 continue;
2368 if (CurrentUnit->getInfo(0).Keep) {
2369 // Clone the InputDIE into your Unit DIE in our compile unit since it
2370 // already has a DIE inside of it.
2371 CurrentUnit->createOutputDIE();
2372 cloneDIE(InputDIE, DMO, *CurrentUnit, StringPool, 0 /* PC offset */,
2373 11 /* Unit Header size */, 0, IsLittleEndian,
2374 CurrentUnit->getOutputUnitDIE());
2377 Linker.OutputDebugInfoSize = CurrentUnit->computeNextUnitOffset();
2379 if (Linker.Options.NoOutput)
2380 continue;
2382 // FIXME: for compatibility with the classic dsymutil, we emit
2383 // an empty line table for the unit, even if the unit doesn't
2384 // actually exist in the DIE tree.
2385 if (LLVM_LIKELY(!Linker.Options.Update) || Linker.Options.Translator)
2386 Linker.patchLineTableForUnit(*CurrentUnit, DwarfContext, Ranges, DMO);
2388 Linker.emitAcceleratorEntriesForUnit(*CurrentUnit);
2390 if (LLVM_UNLIKELY(Linker.Options.Update))
2391 continue;
2393 Linker.patchRangesForUnit(*CurrentUnit, DwarfContext, DMO);
2394 auto ProcessExpr = [&](StringRef Bytes, SmallVectorImpl<uint8_t> &Buffer) {
2395 DWARFUnit &OrigUnit = CurrentUnit->getOrigUnit();
2396 DataExtractor Data(Bytes, IsLittleEndian, OrigUnit.getAddressByteSize());
2397 cloneExpression(Data,
2398 DWARFExpression(Data, OrigUnit.getVersion(),
2399 OrigUnit.getAddressByteSize()),
2400 DMO, *CurrentUnit, Buffer);
2402 Linker.Streamer->emitLocationsForUnit(*CurrentUnit, DwarfContext,
2403 ProcessExpr);
2406 if (Linker.Options.NoOutput)
2407 return;
2409 // Emit all the compile unit's debug information.
2410 for (auto &CurrentUnit : CompileUnits) {
2411 if (LLVM_LIKELY(!Linker.Options.Update))
2412 Linker.generateUnitRanges(*CurrentUnit);
2414 CurrentUnit->fixupForwardReferences();
2416 if (!CurrentUnit->getOutputUnitDIE())
2417 continue;
2419 Linker.Streamer->emitCompileUnitHeader(*CurrentUnit);
2420 Linker.Streamer->emitDIE(*CurrentUnit->getOutputUnitDIE());
2424 void DwarfLinker::updateAccelKind(DWARFContext &Dwarf) {
2425 if (Options.TheAccelTableKind != AccelTableKind::Default)
2426 return;
2428 auto &DwarfObj = Dwarf.getDWARFObj();
2430 if (!AtLeastOneDwarfAccelTable &&
2431 (!DwarfObj.getAppleNamesSection().Data.empty() ||
2432 !DwarfObj.getAppleTypesSection().Data.empty() ||
2433 !DwarfObj.getAppleNamespacesSection().Data.empty() ||
2434 !DwarfObj.getAppleObjCSection().Data.empty())) {
2435 AtLeastOneAppleAccelTable = true;
2438 if (!AtLeastOneDwarfAccelTable &&
2439 !DwarfObj.getNamesSection().Data.empty()) {
2440 AtLeastOneDwarfAccelTable = true;
2444 bool DwarfLinker::emitPaperTrailWarnings(const DebugMapObject &DMO,
2445 const DebugMap &Map,
2446 OffsetsStringPool &StringPool) {
2447 if (DMO.getWarnings().empty() || !DMO.empty())
2448 return false;
2450 Streamer->switchToDebugInfoSection(/* Version */ 2);
2451 DIE *CUDie = DIE::get(DIEAlloc, dwarf::DW_TAG_compile_unit);
2452 CUDie->setOffset(11);
2453 StringRef Producer = StringPool.internString("dsymutil");
2454 StringRef File = StringPool.internString(DMO.getObjectFilename());
2455 CUDie->addValue(DIEAlloc, dwarf::DW_AT_producer, dwarf::DW_FORM_strp,
2456 DIEInteger(StringPool.getStringOffset(Producer)));
2457 DIEBlock *String = new (DIEAlloc) DIEBlock();
2458 DIEBlocks.push_back(String);
2459 for (auto &C : File)
2460 String->addValue(DIEAlloc, dwarf::Attribute(0), dwarf::DW_FORM_data1,
2461 DIEInteger(C));
2462 String->addValue(DIEAlloc, dwarf::Attribute(0), dwarf::DW_FORM_data1,
2463 DIEInteger(0));
2465 CUDie->addValue(DIEAlloc, dwarf::DW_AT_name, dwarf::DW_FORM_string, String);
2466 for (const auto &Warning : DMO.getWarnings()) {
2467 DIE &ConstDie = CUDie->addChild(DIE::get(DIEAlloc, dwarf::DW_TAG_constant));
2468 ConstDie.addValue(
2469 DIEAlloc, dwarf::DW_AT_name, dwarf::DW_FORM_strp,
2470 DIEInteger(StringPool.getStringOffset("dsymutil_warning")));
2471 ConstDie.addValue(DIEAlloc, dwarf::DW_AT_artificial, dwarf::DW_FORM_flag,
2472 DIEInteger(1));
2473 ConstDie.addValue(DIEAlloc, dwarf::DW_AT_const_value, dwarf::DW_FORM_strp,
2474 DIEInteger(StringPool.getStringOffset(Warning)));
2476 unsigned Size = 4 /* FORM_strp */ + File.size() + 1 +
2477 DMO.getWarnings().size() * (4 + 1 + 4) +
2478 1 /* End of children */;
2479 DIEAbbrev Abbrev = CUDie->generateAbbrev();
2480 AssignAbbrev(Abbrev);
2481 CUDie->setAbbrevNumber(Abbrev.getNumber());
2482 Size += getULEB128Size(Abbrev.getNumber());
2483 // Abbreviation ordering needed for classic compatibility.
2484 for (auto &Child : CUDie->children()) {
2485 Abbrev = Child.generateAbbrev();
2486 AssignAbbrev(Abbrev);
2487 Child.setAbbrevNumber(Abbrev.getNumber());
2488 Size += getULEB128Size(Abbrev.getNumber());
2490 CUDie->setSize(Size);
2491 auto &Asm = Streamer->getAsmPrinter();
2492 Asm.emitInt32(11 + CUDie->getSize() - 4);
2493 Asm.emitInt16(2);
2494 Asm.emitInt32(0);
2495 Asm.emitInt8(Map.getTriple().isArch64Bit() ? 8 : 4);
2496 Streamer->emitDIE(*CUDie);
2497 OutputDebugInfoSize += 11 /* Header */ + Size;
2499 return true;
2502 static Error copySwiftInterfaces(
2503 const std::map<std::string, std::string> &ParseableSwiftInterfaces,
2504 StringRef Architecture, const LinkOptions &Options) {
2505 std::error_code EC;
2506 SmallString<128> InputPath;
2507 SmallString<128> Path;
2508 sys::path::append(Path, *Options.ResourceDir, "Swift", Architecture);
2509 if ((EC = sys::fs::create_directories(Path.str(), true,
2510 sys::fs::perms::all_all)))
2511 return make_error<StringError>(
2512 "cannot create directory: " + toString(errorCodeToError(EC)), EC);
2513 unsigned BaseLength = Path.size();
2515 for (auto &I : ParseableSwiftInterfaces) {
2516 StringRef ModuleName = I.first;
2517 StringRef InterfaceFile = I.second;
2518 if (!Options.PrependPath.empty()) {
2519 InputPath.clear();
2520 sys::path::append(InputPath, Options.PrependPath, InterfaceFile);
2521 InterfaceFile = InputPath;
2523 sys::path::append(Path, ModuleName);
2524 Path.append(".swiftinterface");
2525 if (Options.Verbose)
2526 outs() << "copy parseable Swift interface " << InterfaceFile << " -> "
2527 << Path.str() << '\n';
2529 // copy_file attempts an APFS clone first, so this should be cheap.
2530 if ((EC = sys::fs::copy_file(InterfaceFile, Path.str())))
2531 warn(Twine("cannot copy parseable Swift interface ") +
2532 InterfaceFile + ": " +
2533 toString(errorCodeToError(EC)));
2534 Path.resize(BaseLength);
2536 return Error::success();
2539 bool DwarfLinker::link(const DebugMap &Map) {
2540 if (!createStreamer(Map.getTriple(), OutFile))
2541 return false;
2543 // Size of the DIEs (and headers) generated for the linked output.
2544 OutputDebugInfoSize = 0;
2545 // A unique ID that identifies each compile unit.
2546 unsigned UnitID = 0;
2547 DebugMap ModuleMap(Map.getTriple(), Map.getBinaryPath());
2549 // First populate the data structure we need for each iteration of the
2550 // parallel loop.
2551 unsigned NumObjects = Map.getNumberOfObjects();
2552 std::vector<LinkContext> ObjectContexts;
2553 ObjectContexts.reserve(NumObjects);
2554 for (const auto &Obj : Map.objects()) {
2555 ObjectContexts.emplace_back(Map, *this, *Obj.get());
2556 LinkContext &LC = ObjectContexts.back();
2557 if (LC.ObjectFile)
2558 updateAccelKind(*LC.DwarfContext);
2561 // This Dwarf string pool which is only used for uniquing. This one should
2562 // never be used for offsets as its not thread-safe or predictable.
2563 UniquingStringPool UniquingStringPool;
2565 // This Dwarf string pool which is used for emission. It must be used
2566 // serially as the order of calling getStringOffset matters for
2567 // reproducibility.
2568 OffsetsStringPool OffsetsStringPool(Options.Translator);
2570 // ODR Contexts for the link.
2571 DeclContextTree ODRContexts;
2573 // If we haven't decided on an accelerator table kind yet, we base ourselves
2574 // on the DWARF we have seen so far. At this point we haven't pulled in debug
2575 // information from modules yet, so it is technically possible that they
2576 // would affect the decision. However, as they're built with the same
2577 // compiler and flags, it is safe to assume that they will follow the
2578 // decision made here.
2579 if (Options.TheAccelTableKind == AccelTableKind::Default) {
2580 if (AtLeastOneDwarfAccelTable && !AtLeastOneAppleAccelTable)
2581 Options.TheAccelTableKind = AccelTableKind::Dwarf;
2582 else
2583 Options.TheAccelTableKind = AccelTableKind::Apple;
2586 for (LinkContext &LinkContext : ObjectContexts) {
2587 if (Options.Verbose)
2588 outs() << "DEBUG MAP OBJECT: " << LinkContext.DMO.getObjectFilename()
2589 << "\n";
2591 // N_AST objects (swiftmodule files) should get dumped directly into the
2592 // appropriate DWARF section.
2593 if (LinkContext.DMO.getType() == MachO::N_AST) {
2594 StringRef File = LinkContext.DMO.getObjectFilename();
2595 auto ErrorOrMem = MemoryBuffer::getFile(File);
2596 if (!ErrorOrMem) {
2597 warn("Could not open '" + File + "'\n");
2598 continue;
2600 sys::fs::file_status Stat;
2601 if (auto Err = sys::fs::status(File, Stat)) {
2602 warn(Err.message());
2603 continue;
2605 if (!Options.NoTimestamp) {
2606 // The modification can have sub-second precision so we need to cast
2607 // away the extra precision that's not present in the debug map.
2608 auto ModificationTime =
2609 std::chrono::time_point_cast<std::chrono::seconds>(
2610 Stat.getLastModificationTime());
2611 if (ModificationTime != LinkContext.DMO.getTimestamp()) {
2612 // Not using the helper here as we can easily stream TimePoint<>.
2613 WithColor::warning()
2614 << "Timestamp mismatch for " << File << ": "
2615 << Stat.getLastModificationTime() << " and "
2616 << sys::TimePoint<>(LinkContext.DMO.getTimestamp()) << "\n";
2617 continue;
2621 // Copy the module into the .swift_ast section.
2622 if (!Options.NoOutput)
2623 Streamer->emitSwiftAST((*ErrorOrMem)->getBuffer());
2624 continue;
2627 if (emitPaperTrailWarnings(LinkContext.DMO, Map, OffsetsStringPool))
2628 continue;
2630 if (!LinkContext.ObjectFile)
2631 continue;
2633 // Look for relocations that correspond to debug map entries.
2635 if (LLVM_LIKELY(!Options.Update) &&
2636 !LinkContext.RelocMgr.findValidRelocsInDebugInfo(
2637 *LinkContext.ObjectFile, LinkContext.DMO)) {
2638 if (Options.Verbose)
2639 outs() << "No valid relocations found. Skipping.\n";
2641 // Clear this ObjFile entry as a signal to other loops that we should not
2642 // process this iteration.
2643 LinkContext.ObjectFile = nullptr;
2644 continue;
2647 // Setup access to the debug info.
2648 if (!LinkContext.DwarfContext)
2649 continue;
2651 startDebugObject(LinkContext);
2653 // In a first phase, just read in the debug info and load all clang modules.
2654 LinkContext.CompileUnits.reserve(
2655 LinkContext.DwarfContext->getNumCompileUnits());
2657 for (const auto &CU : LinkContext.DwarfContext->compile_units()) {
2658 updateDwarfVersion(CU->getVersion());
2659 auto CUDie = CU->getUnitDIE(false);
2660 if (Options.Verbose) {
2661 outs() << "Input compilation unit:";
2662 DIDumpOptions DumpOpts;
2663 DumpOpts.ChildRecurseDepth = 0;
2664 DumpOpts.Verbose = Options.Verbose;
2665 CUDie.dump(outs(), 0, DumpOpts);
2667 if (CUDie && !LLVM_UNLIKELY(Options.Update))
2668 registerModuleReference(CUDie, *CU, ModuleMap, LinkContext.DMO,
2669 LinkContext.Ranges, OffsetsStringPool,
2670 UniquingStringPool, ODRContexts, 0, UnitID,
2671 LinkContext.DwarfContext->isLittleEndian());
2675 // If we haven't seen any CUs, pick an arbitrary valid Dwarf version anyway.
2676 if (MaxDwarfVersion == 0)
2677 MaxDwarfVersion = 3;
2679 // At this point we know how much data we have emitted. We use this value to
2680 // compare canonical DIE offsets in analyzeContextInfo to see if a definition
2681 // is already emitted, without being affected by canonical die offsets set
2682 // later. This prevents undeterminism when analyze and clone execute
2683 // concurrently, as clone set the canonical DIE offset and analyze reads it.
2684 const uint64_t ModulesEndOffset = OutputDebugInfoSize;
2686 // These variables manage the list of processed object files.
2687 // The mutex and condition variable are to ensure that this is thread safe.
2688 std::mutex ProcessedFilesMutex;
2689 std::condition_variable ProcessedFilesConditionVariable;
2690 BitVector ProcessedFiles(NumObjects, false);
2692 // Analyzing the context info is particularly expensive so it is executed in
2693 // parallel with emitting the previous compile unit.
2694 auto AnalyzeLambda = [&](size_t i) {
2695 auto &LinkContext = ObjectContexts[i];
2697 if (!LinkContext.ObjectFile || !LinkContext.DwarfContext)
2698 return;
2700 for (const auto &CU : LinkContext.DwarfContext->compile_units()) {
2701 updateDwarfVersion(CU->getVersion());
2702 // The !registerModuleReference() condition effectively skips
2703 // over fully resolved skeleton units. This second pass of
2704 // registerModuleReferences doesn't do any new work, but it
2705 // will collect top-level errors, which are suppressed. Module
2706 // warnings were already displayed in the first iteration.
2707 bool Quiet = true;
2708 auto CUDie = CU->getUnitDIE(false);
2709 if (!CUDie || LLVM_UNLIKELY(Options.Update) ||
2710 !registerModuleReference(CUDie, *CU, ModuleMap, LinkContext.DMO,
2711 LinkContext.Ranges, OffsetsStringPool,
2712 UniquingStringPool, ODRContexts,
2713 ModulesEndOffset, UnitID, Quiet)) {
2714 LinkContext.CompileUnits.push_back(std::make_unique<CompileUnit>(
2715 *CU, UnitID++, !Options.NoODR && !Options.Update, ""));
2719 // Now build the DIE parent links that we will use during the next phase.
2720 for (auto &CurrentUnit : LinkContext.CompileUnits) {
2721 auto CUDie = CurrentUnit->getOrigUnit().getUnitDIE();
2722 if (!CUDie)
2723 continue;
2724 analyzeContextInfo(CurrentUnit->getOrigUnit().getUnitDIE(), 0,
2725 *CurrentUnit, &ODRContexts.getRoot(),
2726 UniquingStringPool, ODRContexts, ModulesEndOffset,
2727 ParseableSwiftInterfaces,
2728 [&](const Twine &Warning, const DWARFDie &DIE) {
2729 reportWarning(Warning, LinkContext.DMO, &DIE);
2734 // And then the remaining work in serial again.
2735 // Note, although this loop runs in serial, it can run in parallel with
2736 // the analyzeContextInfo loop so long as we process files with indices >=
2737 // than those processed by analyzeContextInfo.
2738 auto CloneLambda = [&](size_t i) {
2739 auto &LinkContext = ObjectContexts[i];
2740 if (!LinkContext.ObjectFile)
2741 return;
2743 // Then mark all the DIEs that need to be present in the linked output
2744 // and collect some information about them.
2745 // Note that this loop can not be merged with the previous one because
2746 // cross-cu references require the ParentIdx to be setup for every CU in
2747 // the object file before calling this.
2748 if (LLVM_UNLIKELY(Options.Update)) {
2749 for (auto &CurrentUnit : LinkContext.CompileUnits)
2750 CurrentUnit->markEverythingAsKept();
2751 Streamer->copyInvariantDebugSection(*LinkContext.ObjectFile);
2752 } else {
2753 for (auto &CurrentUnit : LinkContext.CompileUnits)
2754 lookForDIEsToKeep(LinkContext.RelocMgr, LinkContext.Ranges,
2755 LinkContext.CompileUnits,
2756 CurrentUnit->getOrigUnit().getUnitDIE(),
2757 LinkContext.DMO, *CurrentUnit, 0);
2760 // The calls to applyValidRelocs inside cloneDIE will walk the reloc
2761 // array again (in the same way findValidRelocsInDebugInfo() did). We
2762 // need to reset the NextValidReloc index to the beginning.
2763 LinkContext.RelocMgr.resetValidRelocs();
2764 if (LinkContext.RelocMgr.hasValidRelocs() || LLVM_UNLIKELY(Options.Update))
2765 DIECloner(*this, LinkContext.RelocMgr, DIEAlloc, LinkContext.CompileUnits,
2766 Options)
2767 .cloneAllCompileUnits(*LinkContext.DwarfContext, LinkContext.DMO,
2768 LinkContext.Ranges, OffsetsStringPool,
2769 LinkContext.DwarfContext->isLittleEndian());
2770 if (!Options.NoOutput && !LinkContext.CompileUnits.empty() &&
2771 LLVM_LIKELY(!Options.Update))
2772 patchFrameInfoForObject(
2773 LinkContext.DMO, LinkContext.Ranges, *LinkContext.DwarfContext,
2774 LinkContext.CompileUnits[0]->getOrigUnit().getAddressByteSize());
2776 // Clean-up before starting working on the next object.
2777 endDebugObject(LinkContext);
2780 auto EmitLambda = [&]() {
2781 // Emit everything that's global.
2782 if (!Options.NoOutput) {
2783 Streamer->emitAbbrevs(Abbreviations, MaxDwarfVersion);
2784 Streamer->emitStrings(OffsetsStringPool);
2785 switch (Options.TheAccelTableKind) {
2786 case AccelTableKind::Apple:
2787 Streamer->emitAppleNames(AppleNames);
2788 Streamer->emitAppleNamespaces(AppleNamespaces);
2789 Streamer->emitAppleTypes(AppleTypes);
2790 Streamer->emitAppleObjc(AppleObjc);
2791 break;
2792 case AccelTableKind::Dwarf:
2793 Streamer->emitDebugNames(DebugNames);
2794 break;
2795 case AccelTableKind::Default:
2796 llvm_unreachable("Default should have already been resolved.");
2797 break;
2802 auto AnalyzeAll = [&]() {
2803 for (unsigned i = 0, e = NumObjects; i != e; ++i) {
2804 AnalyzeLambda(i);
2806 std::unique_lock<std::mutex> LockGuard(ProcessedFilesMutex);
2807 ProcessedFiles.set(i);
2808 ProcessedFilesConditionVariable.notify_one();
2812 auto CloneAll = [&]() {
2813 for (unsigned i = 0, e = NumObjects; i != e; ++i) {
2815 std::unique_lock<std::mutex> LockGuard(ProcessedFilesMutex);
2816 if (!ProcessedFiles[i]) {
2817 ProcessedFilesConditionVariable.wait(
2818 LockGuard, [&]() { return ProcessedFiles[i]; });
2822 CloneLambda(i);
2824 EmitLambda();
2827 // To limit memory usage in the single threaded case, analyze and clone are
2828 // run sequentially so the LinkContext is freed after processing each object
2829 // in endDebugObject.
2830 if (Options.Threads == 1) {
2831 for (unsigned i = 0, e = NumObjects; i != e; ++i) {
2832 AnalyzeLambda(i);
2833 CloneLambda(i);
2835 EmitLambda();
2836 } else {
2837 ThreadPool pool(2);
2838 pool.async(AnalyzeAll);
2839 pool.async(CloneAll);
2840 pool.wait();
2843 if (Options.NoOutput)
2844 return true;
2846 if (Options.ResourceDir && !ParseableSwiftInterfaces.empty()) {
2847 StringRef ArchName = Triple::getArchTypeName(Map.getTriple().getArch());
2848 if (auto E =
2849 copySwiftInterfaces(ParseableSwiftInterfaces, ArchName, Options))
2850 return error(toString(std::move(E)));
2853 return Streamer->finish(Map, Options.Translator);
2854 } // namespace dsymutil
2856 bool linkDwarf(raw_fd_ostream &OutFile, BinaryHolder &BinHolder,
2857 const DebugMap &DM, LinkOptions Options) {
2858 DwarfLinker Linker(OutFile, BinHolder, std::move(Options));
2859 return Linker.link(DM);
2862 } // namespace dsymutil
2863 } // namespace llvm