Run DCE after a LoopFlatten test to reduce spurious output [nfc]
[llvm-project.git] / llvm / lib / DWARFLinker / DWARFLinker.cpp
blob80a4e2adefa6cb6180c7a427e6e47a515d7e88e9
1 //=== DWARFLinker.cpp -----------------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
9 #include "llvm/DWARFLinker/DWARFLinker.h"
10 #include "llvm/ADT/ArrayRef.h"
11 #include "llvm/ADT/BitVector.h"
12 #include "llvm/ADT/STLExtras.h"
13 #include "llvm/ADT/StringExtras.h"
14 #include "llvm/CodeGen/NonRelocatableStringpool.h"
15 #include "llvm/DWARFLinker/DWARFLinkerDeclContext.h"
16 #include "llvm/DWARFLinker/DWARFStreamer.h"
17 #include "llvm/DebugInfo/DWARF/DWARFAbbreviationDeclaration.h"
18 #include "llvm/DebugInfo/DWARF/DWARFAcceleratorTable.h"
19 #include "llvm/DebugInfo/DWARF/DWARFContext.h"
20 #include "llvm/DebugInfo/DWARF/DWARFDataExtractor.h"
21 #include "llvm/DebugInfo/DWARF/DWARFDebugLine.h"
22 #include "llvm/DebugInfo/DWARF/DWARFDebugMacro.h"
23 #include "llvm/DebugInfo/DWARF/DWARFDebugRangeList.h"
24 #include "llvm/DebugInfo/DWARF/DWARFDie.h"
25 #include "llvm/DebugInfo/DWARF/DWARFExpression.h"
26 #include "llvm/DebugInfo/DWARF/DWARFFormValue.h"
27 #include "llvm/DebugInfo/DWARF/DWARFSection.h"
28 #include "llvm/DebugInfo/DWARF/DWARFUnit.h"
29 #include "llvm/MC/MCDwarf.h"
30 #include "llvm/Support/DataExtractor.h"
31 #include "llvm/Support/Error.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/ErrorOr.h"
34 #include "llvm/Support/FormatVariadic.h"
35 #include "llvm/Support/LEB128.h"
36 #include "llvm/Support/Path.h"
37 #include "llvm/Support/ThreadPool.h"
38 #include <vector>
40 namespace llvm {
42 /// Hold the input and output of the debug info size in bytes.
43 struct DebugInfoSize {
44 uint64_t Input;
45 uint64_t Output;
48 /// Compute the total size of the debug info.
49 static uint64_t getDebugInfoSize(DWARFContext &Dwarf) {
50 uint64_t Size = 0;
51 for (auto &Unit : Dwarf.compile_units()) {
52 Size += Unit->getLength();
54 return Size;
57 /// Similar to DWARFUnitSection::getUnitForOffset(), but returning our
58 /// CompileUnit object instead.
59 static CompileUnit *getUnitForOffset(const UnitListTy &Units, uint64_t Offset) {
60 auto CU = llvm::upper_bound(
61 Units, Offset, [](uint64_t LHS, const std::unique_ptr<CompileUnit> &RHS) {
62 return LHS < RHS->getOrigUnit().getNextUnitOffset();
63 });
64 return CU != Units.end() ? CU->get() : nullptr;
67 /// Resolve the DIE attribute reference that has been extracted in \p RefValue.
68 /// The resulting DIE might be in another CompileUnit which is stored into \p
69 /// ReferencedCU. \returns null if resolving fails for any reason.
70 DWARFDie DWARFLinker::resolveDIEReference(const DWARFFile &File,
71 const UnitListTy &Units,
72 const DWARFFormValue &RefValue,
73 const DWARFDie &DIE,
74 CompileUnit *&RefCU) {
75 assert(RefValue.isFormClass(DWARFFormValue::FC_Reference));
76 uint64_t RefOffset = *RefValue.getAsReference();
77 if ((RefCU = getUnitForOffset(Units, RefOffset)))
78 if (const auto RefDie = RefCU->getOrigUnit().getDIEForOffset(RefOffset)) {
79 // In a file with broken references, an attribute might point to a NULL
80 // DIE.
81 if (!RefDie.isNULL())
82 return RefDie;
85 reportWarning("could not find referenced DIE", File, &DIE);
86 return DWARFDie();
89 /// \returns whether the passed \a Attr type might contain a DIE reference
90 /// suitable for ODR uniquing.
91 static bool isODRAttribute(uint16_t Attr) {
92 switch (Attr) {
93 default:
94 return false;
95 case dwarf::DW_AT_type:
96 case dwarf::DW_AT_containing_type:
97 case dwarf::DW_AT_specification:
98 case dwarf::DW_AT_abstract_origin:
99 case dwarf::DW_AT_import:
100 return true;
102 llvm_unreachable("Improper attribute.");
105 static bool isTypeTag(uint16_t Tag) {
106 switch (Tag) {
107 case dwarf::DW_TAG_array_type:
108 case dwarf::DW_TAG_class_type:
109 case dwarf::DW_TAG_enumeration_type:
110 case dwarf::DW_TAG_pointer_type:
111 case dwarf::DW_TAG_reference_type:
112 case dwarf::DW_TAG_string_type:
113 case dwarf::DW_TAG_structure_type:
114 case dwarf::DW_TAG_subroutine_type:
115 case dwarf::DW_TAG_typedef:
116 case dwarf::DW_TAG_union_type:
117 case dwarf::DW_TAG_ptr_to_member_type:
118 case dwarf::DW_TAG_set_type:
119 case dwarf::DW_TAG_subrange_type:
120 case dwarf::DW_TAG_base_type:
121 case dwarf::DW_TAG_const_type:
122 case dwarf::DW_TAG_constant:
123 case dwarf::DW_TAG_file_type:
124 case dwarf::DW_TAG_namelist:
125 case dwarf::DW_TAG_packed_type:
126 case dwarf::DW_TAG_volatile_type:
127 case dwarf::DW_TAG_restrict_type:
128 case dwarf::DW_TAG_atomic_type:
129 case dwarf::DW_TAG_interface_type:
130 case dwarf::DW_TAG_unspecified_type:
131 case dwarf::DW_TAG_shared_type:
132 case dwarf::DW_TAG_immutable_type:
133 return true;
134 default:
135 break;
137 return false;
140 AddressesMap::~AddressesMap() = default;
142 DwarfEmitter::~DwarfEmitter() = default;
144 bool DWARFLinker::DIECloner::getDIENames(const DWARFDie &Die,
145 AttributesInfo &Info,
146 OffsetsStringPool &StringPool,
147 bool StripTemplate) {
148 // This function will be called on DIEs having low_pcs and
149 // ranges. As getting the name might be more expansive, filter out
150 // blocks directly.
151 if (Die.getTag() == dwarf::DW_TAG_lexical_block)
152 return false;
154 if (!Info.MangledName)
155 if (const char *MangledName = Die.getLinkageName())
156 Info.MangledName = StringPool.getEntry(MangledName);
158 if (!Info.Name)
159 if (const char *Name = Die.getShortName())
160 Info.Name = StringPool.getEntry(Name);
162 if (!Info.MangledName)
163 Info.MangledName = Info.Name;
165 if (StripTemplate && Info.Name && Info.MangledName != Info.Name) {
166 StringRef Name = Info.Name.getString();
167 if (std::optional<StringRef> StrippedName = StripTemplateParameters(Name))
168 Info.NameWithoutTemplate = StringPool.getEntry(*StrippedName);
171 return Info.Name || Info.MangledName;
174 /// Resolve the relative path to a build artifact referenced by DWARF by
175 /// applying DW_AT_comp_dir.
176 static void resolveRelativeObjectPath(SmallVectorImpl<char> &Buf, DWARFDie CU) {
177 sys::path::append(Buf, dwarf::toString(CU.find(dwarf::DW_AT_comp_dir), ""));
180 /// Collect references to parseable Swift interfaces in imported
181 /// DW_TAG_module blocks.
182 static void analyzeImportedModule(
183 const DWARFDie &DIE, CompileUnit &CU,
184 swiftInterfacesMap *ParseableSwiftInterfaces,
185 std::function<void(const Twine &, const DWARFDie &)> ReportWarning) {
186 if (CU.getLanguage() != dwarf::DW_LANG_Swift)
187 return;
189 if (!ParseableSwiftInterfaces)
190 return;
192 StringRef Path = dwarf::toStringRef(DIE.find(dwarf::DW_AT_LLVM_include_path));
193 if (!Path.endswith(".swiftinterface"))
194 return;
195 // Don't track interfaces that are part of the SDK.
196 StringRef SysRoot = dwarf::toStringRef(DIE.find(dwarf::DW_AT_LLVM_sysroot));
197 if (SysRoot.empty())
198 SysRoot = CU.getSysRoot();
199 if (!SysRoot.empty() && Path.startswith(SysRoot))
200 return;
201 std::optional<const char *> Name =
202 dwarf::toString(DIE.find(dwarf::DW_AT_name));
203 if (!Name)
204 return;
205 auto &Entry = (*ParseableSwiftInterfaces)[*Name];
206 // The prepend path is applied later when copying.
207 DWARFDie CUDie = CU.getOrigUnit().getUnitDIE();
208 SmallString<128> ResolvedPath;
209 if (sys::path::is_relative(Path))
210 resolveRelativeObjectPath(ResolvedPath, CUDie);
211 sys::path::append(ResolvedPath, Path);
212 if (!Entry.empty() && Entry != ResolvedPath)
213 ReportWarning(Twine("Conflicting parseable interfaces for Swift Module ") +
214 *Name + ": " + Entry + " and " + Path,
215 DIE);
216 Entry = std::string(ResolvedPath.str());
219 /// The distinct types of work performed by the work loop in
220 /// analyzeContextInfo.
221 enum class ContextWorklistItemType : uint8_t {
222 AnalyzeContextInfo,
223 UpdateChildPruning,
224 UpdatePruning,
227 /// This class represents an item in the work list. The type defines what kind
228 /// of work needs to be performed when processing the current item. Everything
229 /// but the Type and Die fields are optional based on the type.
230 struct ContextWorklistItem {
231 DWARFDie Die;
232 unsigned ParentIdx;
233 union {
234 CompileUnit::DIEInfo *OtherInfo;
235 DeclContext *Context;
237 ContextWorklistItemType Type;
238 bool InImportedModule;
240 ContextWorklistItem(DWARFDie Die, ContextWorklistItemType T,
241 CompileUnit::DIEInfo *OtherInfo = nullptr)
242 : Die(Die), ParentIdx(0), OtherInfo(OtherInfo), Type(T),
243 InImportedModule(false) {}
245 ContextWorklistItem(DWARFDie Die, DeclContext *Context, unsigned ParentIdx,
246 bool InImportedModule)
247 : Die(Die), ParentIdx(ParentIdx), Context(Context),
248 Type(ContextWorklistItemType::AnalyzeContextInfo),
249 InImportedModule(InImportedModule) {}
252 static bool updatePruning(const DWARFDie &Die, CompileUnit &CU,
253 uint64_t ModulesEndOffset) {
254 CompileUnit::DIEInfo &Info = CU.getInfo(Die);
256 // Prune this DIE if it is either a forward declaration inside a
257 // DW_TAG_module or a DW_TAG_module that contains nothing but
258 // forward declarations.
259 Info.Prune &= (Die.getTag() == dwarf::DW_TAG_module) ||
260 (isTypeTag(Die.getTag()) &&
261 dwarf::toUnsigned(Die.find(dwarf::DW_AT_declaration), 0));
263 // Only prune forward declarations inside a DW_TAG_module for which a
264 // definition exists elsewhere.
265 if (ModulesEndOffset == 0)
266 Info.Prune &= Info.Ctxt && Info.Ctxt->getCanonicalDIEOffset();
267 else
268 Info.Prune &= Info.Ctxt && Info.Ctxt->getCanonicalDIEOffset() > 0 &&
269 Info.Ctxt->getCanonicalDIEOffset() <= ModulesEndOffset;
271 return Info.Prune;
274 static void updateChildPruning(const DWARFDie &Die, CompileUnit &CU,
275 CompileUnit::DIEInfo &ChildInfo) {
276 CompileUnit::DIEInfo &Info = CU.getInfo(Die);
277 Info.Prune &= ChildInfo.Prune;
280 /// Recursive helper to build the global DeclContext information and
281 /// gather the child->parent relationships in the original compile unit.
283 /// This function uses the same work list approach as lookForDIEsToKeep.
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 void analyzeContextInfo(
289 const DWARFDie &DIE, unsigned ParentIdx, CompileUnit &CU,
290 DeclContext *CurrentDeclContext, DeclContextTree &Contexts,
291 uint64_t ModulesEndOffset, swiftInterfacesMap *ParseableSwiftInterfaces,
292 std::function<void(const Twine &, const DWARFDie &)> ReportWarning) {
293 // LIFO work list.
294 std::vector<ContextWorklistItem> Worklist;
295 Worklist.emplace_back(DIE, CurrentDeclContext, ParentIdx, false);
297 while (!Worklist.empty()) {
298 ContextWorklistItem Current = Worklist.back();
299 Worklist.pop_back();
301 switch (Current.Type) {
302 case ContextWorklistItemType::UpdatePruning:
303 updatePruning(Current.Die, CU, ModulesEndOffset);
304 continue;
305 case ContextWorklistItemType::UpdateChildPruning:
306 updateChildPruning(Current.Die, CU, *Current.OtherInfo);
307 continue;
308 case ContextWorklistItemType::AnalyzeContextInfo:
309 break;
312 unsigned Idx = CU.getOrigUnit().getDIEIndex(Current.Die);
313 CompileUnit::DIEInfo &Info = CU.getInfo(Idx);
315 // Clang imposes an ODR on modules(!) regardless of the language:
316 // "The module-id should consist of only a single identifier,
317 // which provides the name of the module being defined. Each
318 // module shall have a single definition."
320 // This does not extend to the types inside the modules:
321 // "[I]n C, this implies that if two structs are defined in
322 // different submodules with the same name, those two types are
323 // distinct types (but may be compatible types if their
324 // definitions match)."
326 // We treat non-C++ modules like namespaces for this reason.
327 if (Current.Die.getTag() == dwarf::DW_TAG_module &&
328 Current.ParentIdx == 0 &&
329 dwarf::toString(Current.Die.find(dwarf::DW_AT_name), "") !=
330 CU.getClangModuleName()) {
331 Current.InImportedModule = true;
332 analyzeImportedModule(Current.Die, CU, ParseableSwiftInterfaces,
333 ReportWarning);
336 Info.ParentIdx = Current.ParentIdx;
337 Info.InModuleScope = CU.isClangModule() || Current.InImportedModule;
338 if (CU.hasODR() || Info.InModuleScope) {
339 if (Current.Context) {
340 auto PtrInvalidPair = Contexts.getChildDeclContext(
341 *Current.Context, Current.Die, CU, Info.InModuleScope);
342 Current.Context = PtrInvalidPair.getPointer();
343 Info.Ctxt =
344 PtrInvalidPair.getInt() ? nullptr : PtrInvalidPair.getPointer();
345 if (Info.Ctxt)
346 Info.Ctxt->setDefinedInClangModule(Info.InModuleScope);
347 } else
348 Info.Ctxt = Current.Context = nullptr;
351 Info.Prune = Current.InImportedModule;
352 // Add children in reverse order to the worklist to effectively process
353 // them in order.
354 Worklist.emplace_back(Current.Die, ContextWorklistItemType::UpdatePruning);
355 for (auto Child : reverse(Current.Die.children())) {
356 CompileUnit::DIEInfo &ChildInfo = CU.getInfo(Child);
357 Worklist.emplace_back(
358 Current.Die, ContextWorklistItemType::UpdateChildPruning, &ChildInfo);
359 Worklist.emplace_back(Child, Current.Context, Idx,
360 Current.InImportedModule);
365 static bool dieNeedsChildrenToBeMeaningful(uint32_t Tag) {
366 switch (Tag) {
367 default:
368 return false;
369 case dwarf::DW_TAG_class_type:
370 case dwarf::DW_TAG_common_block:
371 case dwarf::DW_TAG_lexical_block:
372 case dwarf::DW_TAG_structure_type:
373 case dwarf::DW_TAG_subprogram:
374 case dwarf::DW_TAG_subroutine_type:
375 case dwarf::DW_TAG_union_type:
376 return true;
378 llvm_unreachable("Invalid Tag");
381 void DWARFLinker::cleanupAuxiliarryData(LinkContext &Context) {
382 Context.clear();
384 for (DIEBlock *I : DIEBlocks)
385 I->~DIEBlock();
386 for (DIELoc *I : DIELocs)
387 I->~DIELoc();
389 DIEBlocks.clear();
390 DIELocs.clear();
391 DIEAlloc.Reset();
394 static bool isTlsAddressCode(uint8_t DW_OP_Code) {
395 return DW_OP_Code == dwarf::DW_OP_form_tls_address ||
396 DW_OP_Code == dwarf::DW_OP_GNU_push_tls_address;
399 std::pair<bool, std::optional<int64_t>>
400 DWARFLinker::getVariableRelocAdjustment(AddressesMap &RelocMgr,
401 const DWARFDie &DIE) {
402 assert((DIE.getTag() == dwarf::DW_TAG_variable ||
403 DIE.getTag() == dwarf::DW_TAG_constant) &&
404 "Wrong type of input die");
406 const auto *Abbrev = DIE.getAbbreviationDeclarationPtr();
408 // Check if DIE has DW_AT_location attribute.
409 DWARFUnit *U = DIE.getDwarfUnit();
410 std::optional<uint32_t> LocationIdx =
411 Abbrev->findAttributeIndex(dwarf::DW_AT_location);
412 if (!LocationIdx)
413 return std::make_pair(false, std::nullopt);
415 // Get offset to the DW_AT_location attribute.
416 uint64_t AttrOffset =
417 Abbrev->getAttributeOffsetFromIndex(*LocationIdx, DIE.getOffset(), *U);
419 // Get value of the DW_AT_location attribute.
420 std::optional<DWARFFormValue> LocationValue =
421 Abbrev->getAttributeValueFromOffset(*LocationIdx, AttrOffset, *U);
422 if (!LocationValue)
423 return std::make_pair(false, std::nullopt);
425 // Check that DW_AT_location attribute is of 'exprloc' class.
426 // Handling value of location expressions for attributes of 'loclist'
427 // class is not implemented yet.
428 std::optional<ArrayRef<uint8_t>> Expr = LocationValue->getAsBlock();
429 if (!Expr)
430 return std::make_pair(false, std::nullopt);
432 // Parse 'exprloc' expression.
433 DataExtractor Data(toStringRef(*Expr), U->getContext().isLittleEndian(),
434 U->getAddressByteSize());
435 DWARFExpression Expression(Data, U->getAddressByteSize(),
436 U->getFormParams().Format);
438 bool HasLocationAddress = false;
439 uint64_t CurExprOffset = 0;
440 for (DWARFExpression::iterator It = Expression.begin();
441 It != Expression.end(); ++It) {
442 DWARFExpression::iterator NextIt = It;
443 ++NextIt;
445 const DWARFExpression::Operation &Op = *It;
446 switch (Op.getCode()) {
447 case dwarf::DW_OP_const2u:
448 case dwarf::DW_OP_const4u:
449 case dwarf::DW_OP_const8u:
450 case dwarf::DW_OP_const2s:
451 case dwarf::DW_OP_const4s:
452 case dwarf::DW_OP_const8s:
453 if (NextIt == Expression.end() || !isTlsAddressCode(NextIt->getCode()))
454 break;
455 [[fallthrough]];
456 case dwarf::DW_OP_addr: {
457 HasLocationAddress = true;
458 // Check relocation for the address.
459 if (std::optional<int64_t> RelocAdjustment =
460 RelocMgr.getExprOpAddressRelocAdjustment(
461 *U, Op, AttrOffset + CurExprOffset,
462 AttrOffset + Op.getEndOffset()))
463 return std::make_pair(HasLocationAddress, *RelocAdjustment);
464 } break;
465 case dwarf::DW_OP_constx:
466 case dwarf::DW_OP_addrx: {
467 HasLocationAddress = true;
468 if (std::optional<uint64_t> AddressOffset =
469 DIE.getDwarfUnit()->getIndexedAddressOffset(
470 Op.getRawOperand(0))) {
471 // Check relocation for the address.
472 if (std::optional<int64_t> RelocAdjustment =
473 RelocMgr.getExprOpAddressRelocAdjustment(
474 *U, Op, *AddressOffset,
475 *AddressOffset + DIE.getDwarfUnit()->getAddressByteSize()))
476 return std::make_pair(HasLocationAddress, *RelocAdjustment);
478 } break;
479 default: {
480 // Nothing to do.
481 } break;
483 CurExprOffset = Op.getEndOffset();
486 return std::make_pair(HasLocationAddress, std::nullopt);
489 /// Check if a variable describing DIE should be kept.
490 /// \returns updated TraversalFlags.
491 unsigned DWARFLinker::shouldKeepVariableDIE(AddressesMap &RelocMgr,
492 const DWARFDie &DIE,
493 CompileUnit::DIEInfo &MyInfo,
494 unsigned Flags) {
495 const auto *Abbrev = DIE.getAbbreviationDeclarationPtr();
497 // Global variables with constant value can always be kept.
498 if (!(Flags & TF_InFunctionScope) &&
499 Abbrev->findAttributeIndex(dwarf::DW_AT_const_value)) {
500 MyInfo.InDebugMap = true;
501 return Flags | TF_Keep;
504 // See if there is a relocation to a valid debug map entry inside this
505 // variable's location. The order is important here. We want to always check
506 // if the variable has a valid relocation, so that the DIEInfo is filled.
507 // However, we don't want a static variable in a function to force us to keep
508 // the enclosing function, unless requested explicitly.
509 std::pair<bool, std::optional<int64_t>> LocExprAddrAndRelocAdjustment =
510 getVariableRelocAdjustment(RelocMgr, DIE);
512 if (LocExprAddrAndRelocAdjustment.first)
513 MyInfo.HasLocationExpressionAddr = true;
515 if (!LocExprAddrAndRelocAdjustment.second)
516 return Flags;
518 MyInfo.AddrAdjust = *LocExprAddrAndRelocAdjustment.second;
519 MyInfo.InDebugMap = true;
521 if (((Flags & TF_InFunctionScope) &&
522 !LLVM_UNLIKELY(Options.KeepFunctionForStatic)))
523 return Flags;
525 if (Options.Verbose) {
526 outs() << "Keeping variable DIE:";
527 DIDumpOptions DumpOpts;
528 DumpOpts.ChildRecurseDepth = 0;
529 DumpOpts.Verbose = Options.Verbose;
530 DIE.dump(outs(), 8 /* Indent */, DumpOpts);
533 return Flags | TF_Keep;
536 /// Check if a function describing DIE should be kept.
537 /// \returns updated TraversalFlags.
538 unsigned DWARFLinker::shouldKeepSubprogramDIE(
539 AddressesMap &RelocMgr, const DWARFDie &DIE, const DWARFFile &File,
540 CompileUnit &Unit, CompileUnit::DIEInfo &MyInfo, unsigned Flags) {
541 Flags |= TF_InFunctionScope;
543 auto LowPc = dwarf::toAddress(DIE.find(dwarf::DW_AT_low_pc));
544 if (!LowPc)
545 return Flags;
547 assert(LowPc && "low_pc attribute is not an address.");
548 std::optional<int64_t> RelocAdjustment =
549 RelocMgr.getSubprogramRelocAdjustment(DIE);
550 if (!RelocAdjustment)
551 return Flags;
553 MyInfo.AddrAdjust = *RelocAdjustment;
554 MyInfo.InDebugMap = true;
556 if (Options.Verbose) {
557 outs() << "Keeping subprogram DIE:";
558 DIDumpOptions DumpOpts;
559 DumpOpts.ChildRecurseDepth = 0;
560 DumpOpts.Verbose = Options.Verbose;
561 DIE.dump(outs(), 8 /* Indent */, DumpOpts);
564 if (DIE.getTag() == dwarf::DW_TAG_label) {
565 if (Unit.hasLabelAt(*LowPc))
566 return Flags;
568 DWARFUnit &OrigUnit = Unit.getOrigUnit();
569 // FIXME: dsymutil-classic compat. dsymutil-classic doesn't consider labels
570 // that don't fall into the CU's aranges. This is wrong IMO. Debug info
571 // generation bugs aside, this is really wrong in the case of labels, where
572 // a label marking the end of a function will have a PC == CU's high_pc.
573 if (dwarf::toAddress(OrigUnit.getUnitDIE().find(dwarf::DW_AT_high_pc))
574 .value_or(UINT64_MAX) <= LowPc)
575 return Flags;
576 Unit.addLabelLowPc(*LowPc, MyInfo.AddrAdjust);
577 return Flags | TF_Keep;
580 Flags |= TF_Keep;
582 std::optional<uint64_t> HighPc = DIE.getHighPC(*LowPc);
583 if (!HighPc) {
584 reportWarning("Function without high_pc. Range will be discarded.\n", File,
585 &DIE);
586 return Flags;
588 if (*LowPc > *HighPc) {
589 reportWarning("low_pc greater than high_pc. Range will be discarded.\n",
590 File, &DIE);
591 return Flags;
594 // Replace the debug map range with a more accurate one.
595 Unit.addFunctionRange(*LowPc, *HighPc, MyInfo.AddrAdjust);
596 return Flags;
599 /// Check if a DIE should be kept.
600 /// \returns updated TraversalFlags.
601 unsigned DWARFLinker::shouldKeepDIE(AddressesMap &RelocMgr, const DWARFDie &DIE,
602 const DWARFFile &File, CompileUnit &Unit,
603 CompileUnit::DIEInfo &MyInfo,
604 unsigned Flags) {
605 switch (DIE.getTag()) {
606 case dwarf::DW_TAG_constant:
607 case dwarf::DW_TAG_variable:
608 return shouldKeepVariableDIE(RelocMgr, DIE, MyInfo, Flags);
609 case dwarf::DW_TAG_subprogram:
610 case dwarf::DW_TAG_label:
611 return shouldKeepSubprogramDIE(RelocMgr, DIE, File, Unit, MyInfo, Flags);
612 case dwarf::DW_TAG_base_type:
613 // DWARF Expressions may reference basic types, but scanning them
614 // is expensive. Basic types are tiny, so just keep all of them.
615 case dwarf::DW_TAG_imported_module:
616 case dwarf::DW_TAG_imported_declaration:
617 case dwarf::DW_TAG_imported_unit:
618 // We always want to keep these.
619 return Flags | TF_Keep;
620 default:
621 break;
624 return Flags;
627 /// Helper that updates the completeness of the current DIE based on the
628 /// completeness of one of its children. It depends on the incompleteness of
629 /// the children already being computed.
630 static void updateChildIncompleteness(const DWARFDie &Die, CompileUnit &CU,
631 CompileUnit::DIEInfo &ChildInfo) {
632 switch (Die.getTag()) {
633 case dwarf::DW_TAG_structure_type:
634 case dwarf::DW_TAG_class_type:
635 case dwarf::DW_TAG_union_type:
636 break;
637 default:
638 return;
641 CompileUnit::DIEInfo &MyInfo = CU.getInfo(Die);
643 if (ChildInfo.Incomplete || ChildInfo.Prune)
644 MyInfo.Incomplete = true;
647 /// Helper that updates the completeness of the current DIE based on the
648 /// completeness of the DIEs it references. It depends on the incompleteness of
649 /// the referenced DIE already being computed.
650 static void updateRefIncompleteness(const DWARFDie &Die, CompileUnit &CU,
651 CompileUnit::DIEInfo &RefInfo) {
652 switch (Die.getTag()) {
653 case dwarf::DW_TAG_typedef:
654 case dwarf::DW_TAG_member:
655 case dwarf::DW_TAG_reference_type:
656 case dwarf::DW_TAG_ptr_to_member_type:
657 case dwarf::DW_TAG_pointer_type:
658 break;
659 default:
660 return;
663 CompileUnit::DIEInfo &MyInfo = CU.getInfo(Die);
665 if (MyInfo.Incomplete)
666 return;
668 if (RefInfo.Incomplete)
669 MyInfo.Incomplete = true;
672 /// Look at the children of the given DIE and decide whether they should be
673 /// kept.
674 void DWARFLinker::lookForChildDIEsToKeep(
675 const DWARFDie &Die, CompileUnit &CU, unsigned Flags,
676 SmallVectorImpl<WorklistItem> &Worklist) {
677 // The TF_ParentWalk flag tells us that we are currently walking up the
678 // parent chain of a required DIE, and we don't want to mark all the children
679 // of the parents as kept (consider for example a DW_TAG_namespace node in
680 // the parent chain). There are however a set of DIE types for which we want
681 // to ignore that directive and still walk their children.
682 if (dieNeedsChildrenToBeMeaningful(Die.getTag()))
683 Flags &= ~DWARFLinker::TF_ParentWalk;
685 // We're finished if this DIE has no children or we're walking the parent
686 // chain.
687 if (!Die.hasChildren() || (Flags & DWARFLinker::TF_ParentWalk))
688 return;
690 // Add children in reverse order to the worklist to effectively process them
691 // in order.
692 for (auto Child : reverse(Die.children())) {
693 // Add a worklist item before every child to calculate incompleteness right
694 // after the current child is processed.
695 CompileUnit::DIEInfo &ChildInfo = CU.getInfo(Child);
696 Worklist.emplace_back(Die, CU, WorklistItemType::UpdateChildIncompleteness,
697 &ChildInfo);
698 Worklist.emplace_back(Child, CU, Flags);
702 static bool isODRCanonicalCandidate(const DWARFDie &Die, CompileUnit &CU) {
703 CompileUnit::DIEInfo &Info = CU.getInfo(Die);
705 if (!Info.Ctxt || (Die.getTag() == dwarf::DW_TAG_namespace))
706 return false;
708 if (!CU.hasODR() && !Info.InModuleScope)
709 return false;
711 return !Info.Incomplete && Info.Ctxt != CU.getInfo(Info.ParentIdx).Ctxt;
714 void DWARFLinker::markODRCanonicalDie(const DWARFDie &Die, CompileUnit &CU) {
715 CompileUnit::DIEInfo &Info = CU.getInfo(Die);
717 Info.ODRMarkingDone = true;
718 if (Info.Keep && isODRCanonicalCandidate(Die, CU) &&
719 !Info.Ctxt->hasCanonicalDIE())
720 Info.Ctxt->setHasCanonicalDIE();
723 /// Look at DIEs referenced by the given DIE and decide whether they should be
724 /// kept. All DIEs referenced though attributes should be kept.
725 void DWARFLinker::lookForRefDIEsToKeep(
726 const DWARFDie &Die, CompileUnit &CU, unsigned Flags,
727 const UnitListTy &Units, const DWARFFile &File,
728 SmallVectorImpl<WorklistItem> &Worklist) {
729 bool UseOdr = (Flags & DWARFLinker::TF_DependencyWalk)
730 ? (Flags & DWARFLinker::TF_ODR)
731 : CU.hasODR();
732 DWARFUnit &Unit = CU.getOrigUnit();
733 DWARFDataExtractor Data = Unit.getDebugInfoExtractor();
734 const auto *Abbrev = Die.getAbbreviationDeclarationPtr();
735 uint64_t Offset = Die.getOffset() + getULEB128Size(Abbrev->getCode());
737 SmallVector<std::pair<DWARFDie, CompileUnit &>, 4> ReferencedDIEs;
738 for (const auto &AttrSpec : Abbrev->attributes()) {
739 DWARFFormValue Val(AttrSpec.Form);
740 if (!Val.isFormClass(DWARFFormValue::FC_Reference) ||
741 AttrSpec.Attr == dwarf::DW_AT_sibling) {
742 DWARFFormValue::skipValue(AttrSpec.Form, Data, &Offset,
743 Unit.getFormParams());
744 continue;
747 Val.extractValue(Data, &Offset, Unit.getFormParams(), &Unit);
748 CompileUnit *ReferencedCU;
749 if (auto RefDie =
750 resolveDIEReference(File, Units, Val, Die, ReferencedCU)) {
751 CompileUnit::DIEInfo &Info = ReferencedCU->getInfo(RefDie);
752 // If the referenced DIE has a DeclContext that has already been
753 // emitted, then do not keep the one in this CU. We'll link to
754 // the canonical DIE in cloneDieReferenceAttribute.
756 // FIXME: compatibility with dsymutil-classic. UseODR shouldn't
757 // be necessary and could be advantageously replaced by
758 // ReferencedCU->hasODR() && CU.hasODR().
760 // FIXME: compatibility with dsymutil-classic. There is no
761 // reason not to unique ref_addr references.
762 if (AttrSpec.Form != dwarf::DW_FORM_ref_addr &&
763 isODRAttribute(AttrSpec.Attr) && Info.Ctxt &&
764 Info.Ctxt->hasCanonicalDIE())
765 continue;
767 // Keep a module forward declaration if there is no definition.
768 if (!(isODRAttribute(AttrSpec.Attr) && Info.Ctxt &&
769 Info.Ctxt->hasCanonicalDIE()))
770 Info.Prune = false;
771 ReferencedDIEs.emplace_back(RefDie, *ReferencedCU);
775 unsigned ODRFlag = UseOdr ? DWARFLinker::TF_ODR : 0;
777 // Add referenced DIEs in reverse order to the worklist to effectively
778 // process them in order.
779 for (auto &P : reverse(ReferencedDIEs)) {
780 // Add a worklist item before every child to calculate incompleteness right
781 // after the current child is processed.
782 CompileUnit::DIEInfo &Info = P.second.getInfo(P.first);
783 Worklist.emplace_back(Die, CU, WorklistItemType::UpdateRefIncompleteness,
784 &Info);
785 Worklist.emplace_back(P.first, P.second,
786 DWARFLinker::TF_Keep |
787 DWARFLinker::TF_DependencyWalk | ODRFlag);
791 /// Look at the parent of the given DIE and decide whether they should be kept.
792 void DWARFLinker::lookForParentDIEsToKeep(
793 unsigned AncestorIdx, CompileUnit &CU, unsigned Flags,
794 SmallVectorImpl<WorklistItem> &Worklist) {
795 // Stop if we encounter an ancestor that's already marked as kept.
796 if (CU.getInfo(AncestorIdx).Keep)
797 return;
799 DWARFUnit &Unit = CU.getOrigUnit();
800 DWARFDie ParentDIE = Unit.getDIEAtIndex(AncestorIdx);
801 Worklist.emplace_back(CU.getInfo(AncestorIdx).ParentIdx, CU, Flags);
802 Worklist.emplace_back(ParentDIE, CU, Flags);
805 /// Recursively walk the \p DIE tree and look for DIEs to keep. Store that
806 /// information in \p CU's DIEInfo.
808 /// This function is the entry point of the DIE selection algorithm. It is
809 /// expected to walk the DIE tree in file order and (though the mediation of
810 /// its helper) call hasValidRelocation() on each DIE that might be a 'root
811 /// DIE' (See DwarfLinker class comment).
813 /// While walking the dependencies of root DIEs, this function is also called,
814 /// but during these dependency walks the file order is not respected. The
815 /// TF_DependencyWalk flag tells us which kind of traversal we are currently
816 /// doing.
818 /// The recursive algorithm is implemented iteratively as a work list because
819 /// very deep recursion could exhaust the stack for large projects. The work
820 /// list acts as a scheduler for different types of work that need to be
821 /// performed.
823 /// The recursive nature of the algorithm is simulated by running the "main"
824 /// algorithm (LookForDIEsToKeep) followed by either looking at more DIEs
825 /// (LookForChildDIEsToKeep, LookForRefDIEsToKeep, LookForParentDIEsToKeep) or
826 /// fixing up a computed property (UpdateChildIncompleteness,
827 /// UpdateRefIncompleteness).
829 /// The return value indicates whether the DIE is incomplete.
830 void DWARFLinker::lookForDIEsToKeep(AddressesMap &AddressesMap,
831 const UnitListTy &Units,
832 const DWARFDie &Die, const DWARFFile &File,
833 CompileUnit &Cu, unsigned Flags) {
834 // LIFO work list.
835 SmallVector<WorklistItem, 4> Worklist;
836 Worklist.emplace_back(Die, Cu, Flags);
838 while (!Worklist.empty()) {
839 WorklistItem Current = Worklist.pop_back_val();
841 // Look at the worklist type to decide what kind of work to perform.
842 switch (Current.Type) {
843 case WorklistItemType::UpdateChildIncompleteness:
844 updateChildIncompleteness(Current.Die, Current.CU, *Current.OtherInfo);
845 continue;
846 case WorklistItemType::UpdateRefIncompleteness:
847 updateRefIncompleteness(Current.Die, Current.CU, *Current.OtherInfo);
848 continue;
849 case WorklistItemType::LookForChildDIEsToKeep:
850 lookForChildDIEsToKeep(Current.Die, Current.CU, Current.Flags, Worklist);
851 continue;
852 case WorklistItemType::LookForRefDIEsToKeep:
853 lookForRefDIEsToKeep(Current.Die, Current.CU, Current.Flags, Units, File,
854 Worklist);
855 continue;
856 case WorklistItemType::LookForParentDIEsToKeep:
857 lookForParentDIEsToKeep(Current.AncestorIdx, Current.CU, Current.Flags,
858 Worklist);
859 continue;
860 case WorklistItemType::MarkODRCanonicalDie:
861 markODRCanonicalDie(Current.Die, Current.CU);
862 continue;
863 case WorklistItemType::LookForDIEsToKeep:
864 break;
867 unsigned Idx = Current.CU.getOrigUnit().getDIEIndex(Current.Die);
868 CompileUnit::DIEInfo &MyInfo = Current.CU.getInfo(Idx);
870 if (MyInfo.Prune) {
871 // We're walking the dependencies of a module forward declaration that was
872 // kept because there is no definition.
873 if (Current.Flags & TF_DependencyWalk)
874 MyInfo.Prune = false;
875 else
876 continue;
879 // If the Keep flag is set, we are marking a required DIE's dependencies.
880 // If our target is already marked as kept, we're all set.
881 bool AlreadyKept = MyInfo.Keep;
882 if ((Current.Flags & TF_DependencyWalk) && AlreadyKept)
883 continue;
885 if (!(Current.Flags & TF_DependencyWalk))
886 Current.Flags = shouldKeepDIE(AddressesMap, Current.Die, File, Current.CU,
887 MyInfo, Current.Flags);
889 // We need to mark context for the canonical die in the end of normal
890 // traversing(not TF_DependencyWalk) or after normal traversing if die
891 // was not marked as kept.
892 if (!(Current.Flags & TF_DependencyWalk) ||
893 (MyInfo.ODRMarkingDone && !MyInfo.Keep)) {
894 if (Current.CU.hasODR() || MyInfo.InModuleScope)
895 Worklist.emplace_back(Current.Die, Current.CU,
896 WorklistItemType::MarkODRCanonicalDie);
899 // Finish by looking for child DIEs. Because of the LIFO worklist we need
900 // to schedule that work before any subsequent items are added to the
901 // worklist.
902 Worklist.emplace_back(Current.Die, Current.CU, Current.Flags,
903 WorklistItemType::LookForChildDIEsToKeep);
905 if (AlreadyKept || !(Current.Flags & TF_Keep))
906 continue;
908 // If it is a newly kept DIE mark it as well as all its dependencies as
909 // kept.
910 MyInfo.Keep = true;
912 // We're looking for incomplete types.
913 MyInfo.Incomplete =
914 Current.Die.getTag() != dwarf::DW_TAG_subprogram &&
915 Current.Die.getTag() != dwarf::DW_TAG_member &&
916 dwarf::toUnsigned(Current.Die.find(dwarf::DW_AT_declaration), 0);
918 // After looking at the parent chain, look for referenced DIEs. Because of
919 // the LIFO worklist we need to schedule that work before any subsequent
920 // items are added to the worklist.
921 Worklist.emplace_back(Current.Die, Current.CU, Current.Flags,
922 WorklistItemType::LookForRefDIEsToKeep);
924 bool UseOdr = (Current.Flags & TF_DependencyWalk) ? (Current.Flags & TF_ODR)
925 : Current.CU.hasODR();
926 unsigned ODRFlag = UseOdr ? TF_ODR : 0;
927 unsigned ParFlags = TF_ParentWalk | TF_Keep | TF_DependencyWalk | ODRFlag;
929 // Now schedule the parent walk.
930 Worklist.emplace_back(MyInfo.ParentIdx, Current.CU, ParFlags);
934 #ifndef NDEBUG
935 /// A broken link in the keep chain. By recording both the parent and the child
936 /// we can show only broken links for DIEs with multiple children.
937 struct BrokenLink {
938 BrokenLink(DWARFDie Parent, DWARFDie Child) : Parent(Parent), Child(Child) {}
939 DWARFDie Parent;
940 DWARFDie Child;
943 /// Verify the keep chain by looking for DIEs that are kept but who's parent
944 /// isn't.
945 static void verifyKeepChain(CompileUnit &CU) {
946 std::vector<DWARFDie> Worklist;
947 Worklist.push_back(CU.getOrigUnit().getUnitDIE());
949 // List of broken links.
950 std::vector<BrokenLink> BrokenLinks;
952 while (!Worklist.empty()) {
953 const DWARFDie Current = Worklist.back();
954 Worklist.pop_back();
956 const bool CurrentDieIsKept = CU.getInfo(Current).Keep;
958 for (DWARFDie Child : reverse(Current.children())) {
959 Worklist.push_back(Child);
961 const bool ChildDieIsKept = CU.getInfo(Child).Keep;
962 if (!CurrentDieIsKept && ChildDieIsKept)
963 BrokenLinks.emplace_back(Current, Child);
967 if (!BrokenLinks.empty()) {
968 for (BrokenLink Link : BrokenLinks) {
969 WithColor::error() << formatv(
970 "Found invalid link in keep chain between {0:x} and {1:x}\n",
971 Link.Parent.getOffset(), Link.Child.getOffset());
973 errs() << "Parent:";
974 Link.Parent.dump(errs(), 0, {});
975 CU.getInfo(Link.Parent).dump();
977 errs() << "Child:";
978 Link.Child.dump(errs(), 2, {});
979 CU.getInfo(Link.Child).dump();
981 report_fatal_error("invalid keep chain");
984 #endif
986 /// Assign an abbreviation number to \p Abbrev.
988 /// Our DIEs get freed after every DebugMapObject has been processed,
989 /// thus the FoldingSet we use to unique DIEAbbrevs cannot refer to
990 /// the instances hold by the DIEs. When we encounter an abbreviation
991 /// that we don't know, we create a permanent copy of it.
992 void DWARFLinker::assignAbbrev(DIEAbbrev &Abbrev) {
993 // Check the set for priors.
994 FoldingSetNodeID ID;
995 Abbrev.Profile(ID);
996 void *InsertToken;
997 DIEAbbrev *InSet = AbbreviationsSet.FindNodeOrInsertPos(ID, InsertToken);
999 // If it's newly added.
1000 if (InSet) {
1001 // Assign existing abbreviation number.
1002 Abbrev.setNumber(InSet->getNumber());
1003 } else {
1004 // Add to abbreviation list.
1005 Abbreviations.push_back(
1006 std::make_unique<DIEAbbrev>(Abbrev.getTag(), Abbrev.hasChildren()));
1007 for (const auto &Attr : Abbrev.getData())
1008 Abbreviations.back()->AddAttribute(Attr);
1009 AbbreviationsSet.InsertNode(Abbreviations.back().get(), InsertToken);
1010 // Assign the unique abbreviation number.
1011 Abbrev.setNumber(Abbreviations.size());
1012 Abbreviations.back()->setNumber(Abbreviations.size());
1016 unsigned DWARFLinker::DIECloner::cloneStringAttribute(DIE &Die,
1017 AttributeSpec AttrSpec,
1018 const DWARFFormValue &Val,
1019 const DWARFUnit &U,
1020 AttributesInfo &Info) {
1021 std::optional<const char *> String = dwarf::toString(Val);
1022 if (!String)
1023 return 0;
1024 DwarfStringPoolEntryRef StringEntry;
1025 if (AttrSpec.Form == dwarf::DW_FORM_line_strp) {
1026 StringEntry = DebugLineStrPool.getEntry(*String);
1027 } else {
1028 StringEntry = DebugStrPool.getEntry(*String);
1030 if (AttrSpec.Attr == dwarf::DW_AT_APPLE_origin) {
1031 Info.HasAppleOrigin = true;
1032 if (std::optional<StringRef> FileName =
1033 ObjFile.Addresses->getLibraryInstallName()) {
1034 StringEntry = DebugStrPool.getEntry(*FileName);
1038 // Update attributes info.
1039 if (AttrSpec.Attr == dwarf::DW_AT_name)
1040 Info.Name = StringEntry;
1041 else if (AttrSpec.Attr == dwarf::DW_AT_MIPS_linkage_name ||
1042 AttrSpec.Attr == dwarf::DW_AT_linkage_name)
1043 Info.MangledName = StringEntry;
1044 if (U.getVersion() >= 5) {
1045 // Switch everything to DW_FORM_strx strings.
1046 auto StringOffsetIndex =
1047 StringOffsetPool.getValueIndex(StringEntry.getOffset());
1048 return Die
1049 .addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1050 dwarf::DW_FORM_strx, DIEInteger(StringOffsetIndex))
1051 ->sizeOf(U.getFormParams());
1053 // Switch everything to out of line strings.
1054 AttrSpec.Form = dwarf::DW_FORM_strp;
1056 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr), AttrSpec.Form,
1057 DIEInteger(StringEntry.getOffset()));
1058 return 4;
1061 unsigned DWARFLinker::DIECloner::cloneDieReferenceAttribute(
1062 DIE &Die, const DWARFDie &InputDIE, AttributeSpec AttrSpec,
1063 unsigned AttrSize, const DWARFFormValue &Val, const DWARFFile &File,
1064 CompileUnit &Unit) {
1065 const DWARFUnit &U = Unit.getOrigUnit();
1066 uint64_t Ref = *Val.getAsReference();
1068 DIE *NewRefDie = nullptr;
1069 CompileUnit *RefUnit = nullptr;
1071 DWARFDie RefDie =
1072 Linker.resolveDIEReference(File, CompileUnits, Val, InputDIE, RefUnit);
1074 // If the referenced DIE is not found, drop the attribute.
1075 if (!RefDie || AttrSpec.Attr == dwarf::DW_AT_sibling)
1076 return 0;
1078 CompileUnit::DIEInfo &RefInfo = RefUnit->getInfo(RefDie);
1080 // If we already have emitted an equivalent DeclContext, just point
1081 // at it.
1082 if (isODRAttribute(AttrSpec.Attr) && RefInfo.Ctxt &&
1083 RefInfo.Ctxt->getCanonicalDIEOffset()) {
1084 assert(RefInfo.Ctxt->hasCanonicalDIE() &&
1085 "Offset to canonical die is set, but context is not marked");
1086 DIEInteger Attr(RefInfo.Ctxt->getCanonicalDIEOffset());
1087 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1088 dwarf::DW_FORM_ref_addr, Attr);
1089 return U.getRefAddrByteSize();
1092 if (!RefInfo.Clone) {
1093 // We haven't cloned this DIE yet. Just create an empty one and
1094 // store it. It'll get really cloned when we process it.
1095 RefInfo.UnclonedReference = true;
1096 RefInfo.Clone = DIE::get(DIEAlloc, dwarf::Tag(RefDie.getTag()));
1098 NewRefDie = RefInfo.Clone;
1100 if (AttrSpec.Form == dwarf::DW_FORM_ref_addr ||
1101 (Unit.hasODR() && isODRAttribute(AttrSpec.Attr))) {
1102 // We cannot currently rely on a DIEEntry to emit ref_addr
1103 // references, because the implementation calls back to DwarfDebug
1104 // to find the unit offset. (We don't have a DwarfDebug)
1105 // FIXME: we should be able to design DIEEntry reliance on
1106 // DwarfDebug away.
1107 uint64_t Attr;
1108 if (Ref < InputDIE.getOffset() && !RefInfo.UnclonedReference) {
1109 // We have already cloned that DIE.
1110 uint32_t NewRefOffset =
1111 RefUnit->getStartOffset() + NewRefDie->getOffset();
1112 Attr = NewRefOffset;
1113 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1114 dwarf::DW_FORM_ref_addr, DIEInteger(Attr));
1115 } else {
1116 // A forward reference. Note and fixup later.
1117 Attr = 0xBADDEF;
1118 Unit.noteForwardReference(
1119 NewRefDie, RefUnit, RefInfo.Ctxt,
1120 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1121 dwarf::DW_FORM_ref_addr, DIEInteger(Attr)));
1123 return U.getRefAddrByteSize();
1126 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1127 dwarf::Form(AttrSpec.Form), DIEEntry(*NewRefDie));
1129 return AttrSize;
1132 void DWARFLinker::DIECloner::cloneExpression(
1133 DataExtractor &Data, DWARFExpression Expression, const DWARFFile &File,
1134 CompileUnit &Unit, SmallVectorImpl<uint8_t> &OutputBuffer,
1135 int64_t AddrRelocAdjustment, bool IsLittleEndian) {
1136 using Encoding = DWARFExpression::Operation::Encoding;
1138 uint8_t OrigAddressByteSize = Unit.getOrigUnit().getAddressByteSize();
1140 uint64_t OpOffset = 0;
1141 for (auto &Op : Expression) {
1142 auto Desc = Op.getDescription();
1143 // DW_OP_const_type is variable-length and has 3
1144 // operands. Thus far we only support 2.
1145 if ((Desc.Op.size() == 2 && Desc.Op[0] == Encoding::BaseTypeRef) ||
1146 (Desc.Op.size() == 2 && Desc.Op[1] == Encoding::BaseTypeRef &&
1147 Desc.Op[0] != Encoding::Size1))
1148 Linker.reportWarning("Unsupported DW_OP encoding.", File);
1150 if ((Desc.Op.size() == 1 && Desc.Op[0] == Encoding::BaseTypeRef) ||
1151 (Desc.Op.size() == 2 && Desc.Op[1] == Encoding::BaseTypeRef &&
1152 Desc.Op[0] == Encoding::Size1)) {
1153 // This code assumes that the other non-typeref operand fits into 1 byte.
1154 assert(OpOffset < Op.getEndOffset());
1155 uint32_t ULEBsize = Op.getEndOffset() - OpOffset - 1;
1156 assert(ULEBsize <= 16);
1158 // Copy over the operation.
1159 assert(!Op.getSubCode() && "SubOps not yet supported");
1160 OutputBuffer.push_back(Op.getCode());
1161 uint64_t RefOffset;
1162 if (Desc.Op.size() == 1) {
1163 RefOffset = Op.getRawOperand(0);
1164 } else {
1165 OutputBuffer.push_back(Op.getRawOperand(0));
1166 RefOffset = Op.getRawOperand(1);
1168 uint32_t Offset = 0;
1169 // Look up the base type. For DW_OP_convert, the operand may be 0 to
1170 // instead indicate the generic type. The same holds for
1171 // DW_OP_reinterpret, which is currently not supported.
1172 if (RefOffset > 0 || Op.getCode() != dwarf::DW_OP_convert) {
1173 RefOffset += Unit.getOrigUnit().getOffset();
1174 auto RefDie = Unit.getOrigUnit().getDIEForOffset(RefOffset);
1175 CompileUnit::DIEInfo &Info = Unit.getInfo(RefDie);
1176 if (DIE *Clone = Info.Clone)
1177 Offset = Clone->getOffset();
1178 else
1179 Linker.reportWarning(
1180 "base type ref doesn't point to DW_TAG_base_type.", File);
1182 uint8_t ULEB[16];
1183 unsigned RealSize = encodeULEB128(Offset, ULEB, ULEBsize);
1184 if (RealSize > ULEBsize) {
1185 // Emit the generic type as a fallback.
1186 RealSize = encodeULEB128(0, ULEB, ULEBsize);
1187 Linker.reportWarning("base type ref doesn't fit.", File);
1189 assert(RealSize == ULEBsize && "padding failed");
1190 ArrayRef<uint8_t> ULEBbytes(ULEB, ULEBsize);
1191 OutputBuffer.append(ULEBbytes.begin(), ULEBbytes.end());
1192 } else if (!Linker.Options.Update && Op.getCode() == dwarf::DW_OP_addrx) {
1193 if (std::optional<object::SectionedAddress> SA =
1194 Unit.getOrigUnit().getAddrOffsetSectionItem(
1195 Op.getRawOperand(0))) {
1196 // DWARFLinker does not use addrx forms since it generates relocated
1197 // addresses. Replace DW_OP_addrx with DW_OP_addr here.
1198 // Argument of DW_OP_addrx should be relocated here as it is not
1199 // processed by applyValidRelocs.
1200 OutputBuffer.push_back(dwarf::DW_OP_addr);
1201 uint64_t LinkedAddress = SA->Address + AddrRelocAdjustment;
1202 if (IsLittleEndian != sys::IsLittleEndianHost)
1203 sys::swapByteOrder(LinkedAddress);
1204 ArrayRef<uint8_t> AddressBytes(
1205 reinterpret_cast<const uint8_t *>(&LinkedAddress),
1206 OrigAddressByteSize);
1207 OutputBuffer.append(AddressBytes.begin(), AddressBytes.end());
1208 } else
1209 Linker.reportWarning("cannot read DW_OP_addrx operand.", File);
1210 } else if (!Linker.Options.Update && Op.getCode() == dwarf::DW_OP_constx) {
1211 if (std::optional<object::SectionedAddress> SA =
1212 Unit.getOrigUnit().getAddrOffsetSectionItem(
1213 Op.getRawOperand(0))) {
1214 // DWARFLinker does not use constx forms since it generates relocated
1215 // addresses. Replace DW_OP_constx with DW_OP_const[*]u here.
1216 // Argument of DW_OP_constx should be relocated here as it is not
1217 // processed by applyValidRelocs.
1218 std::optional<uint8_t> OutOperandKind;
1219 switch (OrigAddressByteSize) {
1220 case 4:
1221 OutOperandKind = dwarf::DW_OP_const4u;
1222 break;
1223 case 8:
1224 OutOperandKind = dwarf::DW_OP_const8u;
1225 break;
1226 default:
1227 Linker.reportWarning(
1228 formatv(("unsupported address size: {0}."), OrigAddressByteSize),
1229 File);
1230 break;
1233 if (OutOperandKind) {
1234 OutputBuffer.push_back(*OutOperandKind);
1235 uint64_t LinkedAddress = SA->Address + AddrRelocAdjustment;
1236 if (IsLittleEndian != sys::IsLittleEndianHost)
1237 sys::swapByteOrder(LinkedAddress);
1238 ArrayRef<uint8_t> AddressBytes(
1239 reinterpret_cast<const uint8_t *>(&LinkedAddress),
1240 OrigAddressByteSize);
1241 OutputBuffer.append(AddressBytes.begin(), AddressBytes.end());
1243 } else
1244 Linker.reportWarning("cannot read DW_OP_constx operand.", File);
1245 } else {
1246 // Copy over everything else unmodified.
1247 StringRef Bytes = Data.getData().slice(OpOffset, Op.getEndOffset());
1248 OutputBuffer.append(Bytes.begin(), Bytes.end());
1250 OpOffset = Op.getEndOffset();
1254 unsigned DWARFLinker::DIECloner::cloneBlockAttribute(
1255 DIE &Die, const DWARFDie &InputDIE, const DWARFFile &File,
1256 CompileUnit &Unit, AttributeSpec AttrSpec, const DWARFFormValue &Val,
1257 bool IsLittleEndian) {
1258 DIEValueList *Attr;
1259 DIEValue Value;
1260 DIELoc *Loc = nullptr;
1261 DIEBlock *Block = nullptr;
1262 if (AttrSpec.Form == dwarf::DW_FORM_exprloc) {
1263 Loc = new (DIEAlloc) DIELoc;
1264 Linker.DIELocs.push_back(Loc);
1265 } else {
1266 Block = new (DIEAlloc) DIEBlock;
1267 Linker.DIEBlocks.push_back(Block);
1269 Attr = Loc ? static_cast<DIEValueList *>(Loc)
1270 : static_cast<DIEValueList *>(Block);
1272 DWARFUnit &OrigUnit = Unit.getOrigUnit();
1273 // If the block is a DWARF Expression, clone it into the temporary
1274 // buffer using cloneExpression(), otherwise copy the data directly.
1275 SmallVector<uint8_t, 32> Buffer;
1276 ArrayRef<uint8_t> Bytes = *Val.getAsBlock();
1277 if (DWARFAttribute::mayHaveLocationExpr(AttrSpec.Attr) &&
1278 (Val.isFormClass(DWARFFormValue::FC_Block) ||
1279 Val.isFormClass(DWARFFormValue::FC_Exprloc))) {
1280 DataExtractor Data(StringRef((const char *)Bytes.data(), Bytes.size()),
1281 IsLittleEndian, OrigUnit.getAddressByteSize());
1282 DWARFExpression Expr(Data, OrigUnit.getAddressByteSize(),
1283 OrigUnit.getFormParams().Format);
1284 cloneExpression(Data, Expr, File, Unit, Buffer,
1285 Unit.getInfo(InputDIE).AddrAdjust, IsLittleEndian);
1286 Bytes = Buffer;
1288 for (auto Byte : Bytes)
1289 Attr->addValue(DIEAlloc, static_cast<dwarf::Attribute>(0),
1290 dwarf::DW_FORM_data1, DIEInteger(Byte));
1292 // FIXME: If DIEBlock and DIELoc just reuses the Size field of
1293 // the DIE class, this "if" could be replaced by
1294 // Attr->setSize(Bytes.size()).
1295 if (Loc)
1296 Loc->setSize(Bytes.size());
1297 else
1298 Block->setSize(Bytes.size());
1300 if (Loc)
1301 Value = DIEValue(dwarf::Attribute(AttrSpec.Attr),
1302 dwarf::Form(AttrSpec.Form), Loc);
1303 else {
1304 // The expression location data might be updated and exceed the original
1305 // size. Check whether the new data fits into the original form.
1306 if ((AttrSpec.Form == dwarf::DW_FORM_block1 &&
1307 (Bytes.size() > UINT8_MAX)) ||
1308 (AttrSpec.Form == dwarf::DW_FORM_block2 &&
1309 (Bytes.size() > UINT16_MAX)) ||
1310 (AttrSpec.Form == dwarf::DW_FORM_block4 && (Bytes.size() > UINT32_MAX)))
1311 AttrSpec.Form = dwarf::DW_FORM_block;
1313 Value = DIEValue(dwarf::Attribute(AttrSpec.Attr),
1314 dwarf::Form(AttrSpec.Form), Block);
1317 return Die.addValue(DIEAlloc, Value)->sizeOf(OrigUnit.getFormParams());
1320 unsigned DWARFLinker::DIECloner::cloneAddressAttribute(
1321 DIE &Die, const DWARFDie &InputDIE, AttributeSpec AttrSpec,
1322 unsigned AttrSize, const DWARFFormValue &Val, const CompileUnit &Unit,
1323 AttributesInfo &Info) {
1324 if (AttrSpec.Attr == dwarf::DW_AT_low_pc)
1325 Info.HasLowPc = true;
1327 if (LLVM_UNLIKELY(Linker.Options.Update)) {
1328 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1329 dwarf::Form(AttrSpec.Form), DIEInteger(Val.getRawUValue()));
1330 return AttrSize;
1333 // Cloned Die may have address attributes relocated to a
1334 // totally unrelated value. This can happen:
1335 // - If high_pc is an address (Dwarf version == 2), then it might have been
1336 // relocated to a totally unrelated value (because the end address in the
1337 // object file might be start address of another function which got moved
1338 // independently by the linker).
1339 // - If address relocated in an inline_subprogram that happens at the
1340 // beginning of its inlining function.
1341 // To avoid above cases and to not apply relocation twice (in applyValidRelocs
1342 // and here), read address attribute from InputDIE and apply Info.PCOffset
1343 // here.
1345 std::optional<DWARFFormValue> AddrAttribute = InputDIE.find(AttrSpec.Attr);
1346 if (!AddrAttribute)
1347 llvm_unreachable("Cann't find attribute.");
1349 std::optional<uint64_t> Addr = AddrAttribute->getAsAddress();
1350 if (!Addr) {
1351 Linker.reportWarning("Cann't read address attribute value.", ObjFile);
1352 return 0;
1355 if (InputDIE.getTag() == dwarf::DW_TAG_compile_unit &&
1356 AttrSpec.Attr == dwarf::DW_AT_low_pc) {
1357 if (std::optional<uint64_t> LowPC = Unit.getLowPc())
1358 Addr = *LowPC;
1359 else
1360 return 0;
1361 } else if (InputDIE.getTag() == dwarf::DW_TAG_compile_unit &&
1362 AttrSpec.Attr == dwarf::DW_AT_high_pc) {
1363 if (uint64_t HighPc = Unit.getHighPc())
1364 Addr = HighPc;
1365 else
1366 return 0;
1367 } else {
1368 *Addr += Info.PCOffset;
1371 if (AttrSpec.Form == dwarf::DW_FORM_addr) {
1372 Die.addValue(DIEAlloc, static_cast<dwarf::Attribute>(AttrSpec.Attr),
1373 AttrSpec.Form, DIEInteger(*Addr));
1374 return Unit.getOrigUnit().getAddressByteSize();
1377 auto AddrIndex = AddrPool.getValueIndex(*Addr);
1379 return Die
1380 .addValue(DIEAlloc, static_cast<dwarf::Attribute>(AttrSpec.Attr),
1381 dwarf::Form::DW_FORM_addrx, DIEInteger(AddrIndex))
1382 ->sizeOf(Unit.getOrigUnit().getFormParams());
1385 unsigned DWARFLinker::DIECloner::cloneScalarAttribute(
1386 DIE &Die, const DWARFDie &InputDIE, const DWARFFile &File,
1387 CompileUnit &Unit, AttributeSpec AttrSpec, const DWARFFormValue &Val,
1388 unsigned AttrSize, AttributesInfo &Info) {
1389 uint64_t Value;
1391 // Check for the offset to the macro table. If offset is incorrect then we
1392 // need to remove the attribute.
1393 if (AttrSpec.Attr == dwarf::DW_AT_macro_info) {
1394 if (std::optional<uint64_t> Offset = Val.getAsSectionOffset()) {
1395 const DWARFDebugMacro *Macro = File.Dwarf->getDebugMacinfo();
1396 if (Macro == nullptr || !Macro->hasEntryForOffset(*Offset))
1397 return 0;
1401 if (AttrSpec.Attr == dwarf::DW_AT_macros) {
1402 if (std::optional<uint64_t> Offset = Val.getAsSectionOffset()) {
1403 const DWARFDebugMacro *Macro = File.Dwarf->getDebugMacro();
1404 if (Macro == nullptr || !Macro->hasEntryForOffset(*Offset))
1405 return 0;
1409 if (AttrSpec.Attr == dwarf::DW_AT_str_offsets_base) {
1410 // DWARFLinker generates common .debug_str_offsets table used for all
1411 // compile units. The offset to the common .debug_str_offsets table is 8 on
1412 // DWARF32.
1413 Info.AttrStrOffsetBaseSeen = true;
1414 return Die
1415 .addValue(DIEAlloc, dwarf::DW_AT_str_offsets_base,
1416 dwarf::DW_FORM_sec_offset, DIEInteger(8))
1417 ->sizeOf(Unit.getOrigUnit().getFormParams());
1420 if (LLVM_UNLIKELY(Linker.Options.Update)) {
1421 if (auto OptionalValue = Val.getAsUnsignedConstant())
1422 Value = *OptionalValue;
1423 else if (auto OptionalValue = Val.getAsSignedConstant())
1424 Value = *OptionalValue;
1425 else if (auto OptionalValue = Val.getAsSectionOffset())
1426 Value = *OptionalValue;
1427 else {
1428 Linker.reportWarning(
1429 "Unsupported scalar attribute form. Dropping attribute.", File,
1430 &InputDIE);
1431 return 0;
1433 if (AttrSpec.Attr == dwarf::DW_AT_declaration && Value)
1434 Info.IsDeclaration = true;
1436 if (AttrSpec.Form == dwarf::DW_FORM_loclistx)
1437 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1438 dwarf::Form(AttrSpec.Form), DIELocList(Value));
1439 else
1440 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1441 dwarf::Form(AttrSpec.Form), DIEInteger(Value));
1442 return AttrSize;
1445 [[maybe_unused]] dwarf::Form OriginalForm = AttrSpec.Form;
1446 if (AttrSpec.Form == dwarf::DW_FORM_rnglistx) {
1447 // DWARFLinker does not generate .debug_addr table. Thus we need to change
1448 // all "addrx" related forms to "addr" version. Change DW_FORM_rnglistx
1449 // to DW_FORM_sec_offset here.
1450 std::optional<uint64_t> Index = Val.getAsSectionOffset();
1451 if (!Index) {
1452 Linker.reportWarning("Cannot read the attribute. Dropping.", File,
1453 &InputDIE);
1454 return 0;
1456 std::optional<uint64_t> Offset =
1457 Unit.getOrigUnit().getRnglistOffset(*Index);
1458 if (!Offset) {
1459 Linker.reportWarning("Cannot read the attribute. Dropping.", File,
1460 &InputDIE);
1461 return 0;
1464 Value = *Offset;
1465 AttrSpec.Form = dwarf::DW_FORM_sec_offset;
1466 AttrSize = Unit.getOrigUnit().getFormParams().getDwarfOffsetByteSize();
1467 } else if (AttrSpec.Form == dwarf::DW_FORM_loclistx) {
1468 // DWARFLinker does not generate .debug_addr table. Thus we need to change
1469 // all "addrx" related forms to "addr" version. Change DW_FORM_loclistx
1470 // to DW_FORM_sec_offset here.
1471 std::optional<uint64_t> Index = Val.getAsSectionOffset();
1472 if (!Index) {
1473 Linker.reportWarning("Cannot read the attribute. Dropping.", File,
1474 &InputDIE);
1475 return 0;
1477 std::optional<uint64_t> Offset =
1478 Unit.getOrigUnit().getLoclistOffset(*Index);
1479 if (!Offset) {
1480 Linker.reportWarning("Cannot read the attribute. Dropping.", File,
1481 &InputDIE);
1482 return 0;
1485 Value = *Offset;
1486 AttrSpec.Form = dwarf::DW_FORM_sec_offset;
1487 AttrSize = Unit.getOrigUnit().getFormParams().getDwarfOffsetByteSize();
1488 } else if (AttrSpec.Attr == dwarf::DW_AT_high_pc &&
1489 Die.getTag() == dwarf::DW_TAG_compile_unit) {
1490 std::optional<uint64_t> LowPC = Unit.getLowPc();
1491 if (!LowPC)
1492 return 0;
1493 // Dwarf >= 4 high_pc is an size, not an address.
1494 Value = Unit.getHighPc() - *LowPC;
1495 } else if (AttrSpec.Form == dwarf::DW_FORM_sec_offset)
1496 Value = *Val.getAsSectionOffset();
1497 else if (AttrSpec.Form == dwarf::DW_FORM_sdata)
1498 Value = *Val.getAsSignedConstant();
1499 else if (auto OptionalValue = Val.getAsUnsignedConstant())
1500 Value = *OptionalValue;
1501 else {
1502 Linker.reportWarning(
1503 "Unsupported scalar attribute form. Dropping attribute.", File,
1504 &InputDIE);
1505 return 0;
1508 DIE::value_iterator Patch =
1509 Die.addValue(DIEAlloc, dwarf::Attribute(AttrSpec.Attr),
1510 dwarf::Form(AttrSpec.Form), DIEInteger(Value));
1511 if (AttrSpec.Attr == dwarf::DW_AT_ranges ||
1512 AttrSpec.Attr == dwarf::DW_AT_start_scope) {
1513 Unit.noteRangeAttribute(Die, Patch);
1514 Info.HasRanges = true;
1515 } else if (DWARFAttribute::mayHaveLocationList(AttrSpec.Attr) &&
1516 dwarf::doesFormBelongToClass(AttrSpec.Form,
1517 DWARFFormValue::FC_SectionOffset,
1518 Unit.getOrigUnit().getVersion())) {
1520 CompileUnit::DIEInfo &LocationDieInfo = Unit.getInfo(InputDIE);
1521 Unit.noteLocationAttribute({Patch, LocationDieInfo.InDebugMap
1522 ? LocationDieInfo.AddrAdjust
1523 : Info.PCOffset});
1524 } else if (AttrSpec.Attr == dwarf::DW_AT_declaration && Value)
1525 Info.IsDeclaration = true;
1527 // check that all dwarf::DW_FORM_rnglistx are handled previously.
1528 assert((Info.HasRanges || (OriginalForm != dwarf::DW_FORM_rnglistx)) &&
1529 "Unhandled DW_FORM_rnglistx attribute");
1531 return AttrSize;
1534 /// Clone \p InputDIE's attribute described by \p AttrSpec with
1535 /// value \p Val, and add it to \p Die.
1536 /// \returns the size of the cloned attribute.
1537 unsigned DWARFLinker::DIECloner::cloneAttribute(
1538 DIE &Die, const DWARFDie &InputDIE, const DWARFFile &File,
1539 CompileUnit &Unit, const DWARFFormValue &Val, const AttributeSpec AttrSpec,
1540 unsigned AttrSize, AttributesInfo &Info, bool IsLittleEndian) {
1541 const DWARFUnit &U = Unit.getOrigUnit();
1543 switch (AttrSpec.Form) {
1544 case dwarf::DW_FORM_strp:
1545 case dwarf::DW_FORM_line_strp:
1546 case dwarf::DW_FORM_string:
1547 case dwarf::DW_FORM_strx:
1548 case dwarf::DW_FORM_strx1:
1549 case dwarf::DW_FORM_strx2:
1550 case dwarf::DW_FORM_strx3:
1551 case dwarf::DW_FORM_strx4:
1552 return cloneStringAttribute(Die, AttrSpec, Val, U, Info);
1553 case dwarf::DW_FORM_ref_addr:
1554 case dwarf::DW_FORM_ref1:
1555 case dwarf::DW_FORM_ref2:
1556 case dwarf::DW_FORM_ref4:
1557 case dwarf::DW_FORM_ref8:
1558 return cloneDieReferenceAttribute(Die, InputDIE, AttrSpec, AttrSize, Val,
1559 File, Unit);
1560 case dwarf::DW_FORM_block:
1561 case dwarf::DW_FORM_block1:
1562 case dwarf::DW_FORM_block2:
1563 case dwarf::DW_FORM_block4:
1564 case dwarf::DW_FORM_exprloc:
1565 return cloneBlockAttribute(Die, InputDIE, File, Unit, AttrSpec, Val,
1566 IsLittleEndian);
1567 case dwarf::DW_FORM_addr:
1568 case dwarf::DW_FORM_addrx:
1569 case dwarf::DW_FORM_addrx1:
1570 case dwarf::DW_FORM_addrx2:
1571 case dwarf::DW_FORM_addrx3:
1572 case dwarf::DW_FORM_addrx4:
1573 return cloneAddressAttribute(Die, InputDIE, AttrSpec, AttrSize, Val, Unit,
1574 Info);
1575 case dwarf::DW_FORM_data1:
1576 case dwarf::DW_FORM_data2:
1577 case dwarf::DW_FORM_data4:
1578 case dwarf::DW_FORM_data8:
1579 case dwarf::DW_FORM_udata:
1580 case dwarf::DW_FORM_sdata:
1581 case dwarf::DW_FORM_sec_offset:
1582 case dwarf::DW_FORM_flag:
1583 case dwarf::DW_FORM_flag_present:
1584 case dwarf::DW_FORM_rnglistx:
1585 case dwarf::DW_FORM_loclistx:
1586 case dwarf::DW_FORM_implicit_const:
1587 return cloneScalarAttribute(Die, InputDIE, File, Unit, AttrSpec, Val,
1588 AttrSize, Info);
1589 default:
1590 Linker.reportWarning("Unsupported attribute form " +
1591 dwarf::FormEncodingString(AttrSpec.Form) +
1592 " in cloneAttribute. Dropping.",
1593 File, &InputDIE);
1596 return 0;
1599 void DWARFLinker::DIECloner::addObjCAccelerator(CompileUnit &Unit,
1600 const DIE *Die,
1601 DwarfStringPoolEntryRef Name,
1602 OffsetsStringPool &StringPool,
1603 bool SkipPubSection) {
1604 std::optional<ObjCSelectorNames> Names =
1605 getObjCNamesIfSelector(Name.getString());
1606 if (!Names)
1607 return;
1608 Unit.addNameAccelerator(Die, StringPool.getEntry(Names->Selector),
1609 SkipPubSection);
1610 Unit.addObjCAccelerator(Die, StringPool.getEntry(Names->ClassName),
1611 SkipPubSection);
1612 if (Names->ClassNameNoCategory)
1613 Unit.addObjCAccelerator(
1614 Die, StringPool.getEntry(*Names->ClassNameNoCategory), SkipPubSection);
1615 if (Names->MethodNameNoCategory)
1616 Unit.addNameAccelerator(
1617 Die, StringPool.getEntry(*Names->MethodNameNoCategory), SkipPubSection);
1620 static bool
1621 shouldSkipAttribute(bool Update,
1622 DWARFAbbreviationDeclaration::AttributeSpec AttrSpec,
1623 bool SkipPC) {
1624 switch (AttrSpec.Attr) {
1625 default:
1626 return false;
1627 case dwarf::DW_AT_low_pc:
1628 case dwarf::DW_AT_high_pc:
1629 case dwarf::DW_AT_ranges:
1630 return !Update && SkipPC;
1631 case dwarf::DW_AT_rnglists_base:
1632 // In case !Update the .debug_addr table is not generated/preserved.
1633 // Thus instead of DW_FORM_rnglistx the DW_FORM_sec_offset is used.
1634 // Since DW_AT_rnglists_base is used for only DW_FORM_rnglistx the
1635 // DW_AT_rnglists_base is removed.
1636 return !Update;
1637 case dwarf::DW_AT_loclists_base:
1638 // In case !Update the .debug_addr table is not generated/preserved.
1639 // Thus instead of DW_FORM_loclistx the DW_FORM_sec_offset is used.
1640 // Since DW_AT_loclists_base is used for only DW_FORM_loclistx the
1641 // DW_AT_loclists_base is removed.
1642 return !Update;
1643 case dwarf::DW_AT_location:
1644 case dwarf::DW_AT_frame_base:
1645 return !Update && SkipPC;
1649 struct AttributeLinkedOffsetFixup {
1650 int64_t LinkedOffsetFixupVal;
1651 uint64_t InputAttrStartOffset;
1652 uint64_t InputAttrEndOffset;
1655 DIE *DWARFLinker::DIECloner::cloneDIE(const DWARFDie &InputDIE,
1656 const DWARFFile &File, CompileUnit &Unit,
1657 int64_t PCOffset, uint32_t OutOffset,
1658 unsigned Flags, bool IsLittleEndian,
1659 DIE *Die) {
1660 DWARFUnit &U = Unit.getOrigUnit();
1661 unsigned Idx = U.getDIEIndex(InputDIE);
1662 CompileUnit::DIEInfo &Info = Unit.getInfo(Idx);
1664 // Should the DIE appear in the output?
1665 if (!Unit.getInfo(Idx).Keep)
1666 return nullptr;
1668 uint64_t Offset = InputDIE.getOffset();
1669 assert(!(Die && Info.Clone) && "Can't supply a DIE and a cloned DIE");
1670 if (!Die) {
1671 // The DIE might have been already created by a forward reference
1672 // (see cloneDieReferenceAttribute()).
1673 if (!Info.Clone)
1674 Info.Clone = DIE::get(DIEAlloc, dwarf::Tag(InputDIE.getTag()));
1675 Die = Info.Clone;
1678 assert(Die->getTag() == InputDIE.getTag());
1679 Die->setOffset(OutOffset);
1680 if (isODRCanonicalCandidate(InputDIE, Unit) && Info.Ctxt &&
1681 (Info.Ctxt->getCanonicalDIEOffset() == 0)) {
1682 if (!Info.Ctxt->hasCanonicalDIE())
1683 Info.Ctxt->setHasCanonicalDIE();
1684 // We are about to emit a DIE that is the root of its own valid
1685 // DeclContext tree. Make the current offset the canonical offset
1686 // for this context.
1687 Info.Ctxt->setCanonicalDIEOffset(OutOffset + Unit.getStartOffset());
1690 // Extract and clone every attribute.
1691 DWARFDataExtractor Data = U.getDebugInfoExtractor();
1692 // Point to the next DIE (generally there is always at least a NULL
1693 // entry after the current one). If this is a lone
1694 // DW_TAG_compile_unit without any children, point to the next unit.
1695 uint64_t NextOffset = (Idx + 1 < U.getNumDIEs())
1696 ? U.getDIEAtIndex(Idx + 1).getOffset()
1697 : U.getNextUnitOffset();
1698 AttributesInfo AttrInfo;
1700 // We could copy the data only if we need to apply a relocation to it. After
1701 // testing, it seems there is no performance downside to doing the copy
1702 // unconditionally, and it makes the code simpler.
1703 SmallString<40> DIECopy(Data.getData().substr(Offset, NextOffset - Offset));
1704 Data =
1705 DWARFDataExtractor(DIECopy, Data.isLittleEndian(), Data.getAddressSize());
1707 // Modify the copy with relocated addresses.
1708 ObjFile.Addresses->applyValidRelocs(DIECopy, Offset, Data.isLittleEndian());
1710 // Reset the Offset to 0 as we will be working on the local copy of
1711 // the data.
1712 Offset = 0;
1714 const auto *Abbrev = InputDIE.getAbbreviationDeclarationPtr();
1715 Offset += getULEB128Size(Abbrev->getCode());
1717 // We are entering a subprogram. Get and propagate the PCOffset.
1718 if (Die->getTag() == dwarf::DW_TAG_subprogram)
1719 PCOffset = Info.AddrAdjust;
1720 AttrInfo.PCOffset = PCOffset;
1722 if (Abbrev->getTag() == dwarf::DW_TAG_subprogram) {
1723 Flags |= TF_InFunctionScope;
1724 if (!Info.InDebugMap && LLVM_LIKELY(!Update))
1725 Flags |= TF_SkipPC;
1726 } else if (Abbrev->getTag() == dwarf::DW_TAG_variable) {
1727 // Function-local globals could be in the debug map even when the function
1728 // is not, e.g., inlined functions.
1729 if ((Flags & TF_InFunctionScope) && Info.InDebugMap)
1730 Flags &= ~TF_SkipPC;
1731 // Location expressions referencing an address which is not in debug map
1732 // should be deleted.
1733 else if (!Info.InDebugMap && Info.HasLocationExpressionAddr &&
1734 LLVM_LIKELY(!Update))
1735 Flags |= TF_SkipPC;
1738 std::optional<StringRef> LibraryInstallName =
1739 ObjFile.Addresses->getLibraryInstallName();
1740 SmallVector<AttributeLinkedOffsetFixup> AttributesFixups;
1741 for (const auto &AttrSpec : Abbrev->attributes()) {
1742 if (shouldSkipAttribute(Update, AttrSpec, Flags & TF_SkipPC)) {
1743 DWARFFormValue::skipValue(AttrSpec.Form, Data, &Offset,
1744 U.getFormParams());
1745 continue;
1748 AttributeLinkedOffsetFixup CurAttrFixup;
1749 CurAttrFixup.InputAttrStartOffset = InputDIE.getOffset() + Offset;
1750 CurAttrFixup.LinkedOffsetFixupVal =
1751 Unit.getStartOffset() + OutOffset - CurAttrFixup.InputAttrStartOffset;
1753 DWARFFormValue Val = AttrSpec.getFormValue();
1754 uint64_t AttrSize = Offset;
1755 Val.extractValue(Data, &Offset, U.getFormParams(), &U);
1756 CurAttrFixup.InputAttrEndOffset = InputDIE.getOffset() + Offset;
1757 AttrSize = Offset - AttrSize;
1759 uint64_t FinalAttrSize =
1760 cloneAttribute(*Die, InputDIE, File, Unit, Val, AttrSpec, AttrSize,
1761 AttrInfo, IsLittleEndian);
1762 if (FinalAttrSize != 0 && ObjFile.Addresses->needToSaveValidRelocs())
1763 AttributesFixups.push_back(CurAttrFixup);
1765 OutOffset += FinalAttrSize;
1768 uint16_t Tag = InputDIE.getTag();
1769 // Add the DW_AT_APPLE_origin attribute to Compile Unit die if we have
1770 // an install name and the DWARF doesn't have the attribute yet.
1771 const bool NeedsAppleOrigin = (Tag == dwarf::DW_TAG_compile_unit) &&
1772 LibraryInstallName.has_value() &&
1773 !AttrInfo.HasAppleOrigin;
1774 if (NeedsAppleOrigin) {
1775 auto StringEntry = DebugStrPool.getEntry(LibraryInstallName.value());
1776 Die->addValue(DIEAlloc, dwarf::Attribute(dwarf::DW_AT_APPLE_origin),
1777 dwarf::DW_FORM_strp, DIEInteger(StringEntry.getOffset()));
1778 AttrInfo.Name = StringEntry;
1779 OutOffset += 4;
1782 // Look for accelerator entries.
1783 // FIXME: This is slightly wrong. An inline_subroutine without a
1784 // low_pc, but with AT_ranges might be interesting to get into the
1785 // accelerator tables too. For now stick with dsymutil's behavior.
1786 if ((Info.InDebugMap || AttrInfo.HasLowPc || AttrInfo.HasRanges) &&
1787 Tag != dwarf::DW_TAG_compile_unit &&
1788 getDIENames(InputDIE, AttrInfo, DebugStrPool,
1789 Tag != dwarf::DW_TAG_inlined_subroutine)) {
1790 if (AttrInfo.MangledName && AttrInfo.MangledName != AttrInfo.Name)
1791 Unit.addNameAccelerator(Die, AttrInfo.MangledName,
1792 Tag == dwarf::DW_TAG_inlined_subroutine);
1793 if (AttrInfo.Name) {
1794 if (AttrInfo.NameWithoutTemplate)
1795 Unit.addNameAccelerator(Die, AttrInfo.NameWithoutTemplate,
1796 /* SkipPubSection */ true);
1797 Unit.addNameAccelerator(Die, AttrInfo.Name,
1798 Tag == dwarf::DW_TAG_inlined_subroutine);
1800 if (AttrInfo.Name)
1801 addObjCAccelerator(Unit, Die, AttrInfo.Name, DebugStrPool,
1802 /* SkipPubSection =*/true);
1804 } else if (Tag == dwarf::DW_TAG_namespace) {
1805 if (!AttrInfo.Name)
1806 AttrInfo.Name = DebugStrPool.getEntry("(anonymous namespace)");
1807 Unit.addNamespaceAccelerator(Die, AttrInfo.Name);
1808 } else if (Tag == dwarf::DW_TAG_imported_declaration && AttrInfo.Name) {
1809 Unit.addNamespaceAccelerator(Die, AttrInfo.Name);
1810 } else if (isTypeTag(Tag) && !AttrInfo.IsDeclaration &&
1811 getDIENames(InputDIE, AttrInfo, DebugStrPool) && AttrInfo.Name &&
1812 AttrInfo.Name.getString()[0]) {
1813 uint32_t Hash = hashFullyQualifiedName(InputDIE, Unit, File);
1814 uint64_t RuntimeLang =
1815 dwarf::toUnsigned(InputDIE.find(dwarf::DW_AT_APPLE_runtime_class))
1816 .value_or(0);
1817 bool ObjCClassIsImplementation =
1818 (RuntimeLang == dwarf::DW_LANG_ObjC ||
1819 RuntimeLang == dwarf::DW_LANG_ObjC_plus_plus) &&
1820 dwarf::toUnsigned(InputDIE.find(dwarf::DW_AT_APPLE_objc_complete_type))
1821 .value_or(0);
1822 Unit.addTypeAccelerator(Die, AttrInfo.Name, ObjCClassIsImplementation,
1823 Hash);
1826 // Determine whether there are any children that we want to keep.
1827 bool HasChildren = false;
1828 for (auto Child : InputDIE.children()) {
1829 unsigned Idx = U.getDIEIndex(Child);
1830 if (Unit.getInfo(Idx).Keep) {
1831 HasChildren = true;
1832 break;
1836 if (Unit.getOrigUnit().getVersion() >= 5 && !AttrInfo.AttrStrOffsetBaseSeen &&
1837 Die->getTag() == dwarf::DW_TAG_compile_unit) {
1838 // No DW_AT_str_offsets_base seen, add it to the DIE.
1839 Die->addValue(DIEAlloc, dwarf::DW_AT_str_offsets_base,
1840 dwarf::DW_FORM_sec_offset, DIEInteger(8));
1841 OutOffset += 4;
1844 DIEAbbrev NewAbbrev = Die->generateAbbrev();
1845 if (HasChildren)
1846 NewAbbrev.setChildrenFlag(dwarf::DW_CHILDREN_yes);
1847 // Assign a permanent abbrev number
1848 Linker.assignAbbrev(NewAbbrev);
1849 Die->setAbbrevNumber(NewAbbrev.getNumber());
1851 uint64_t AbbrevNumberSize = getULEB128Size(Die->getAbbrevNumber());
1853 // Add the size of the abbreviation number to the output offset.
1854 OutOffset += AbbrevNumberSize;
1856 // Update fixups with the size of the abbreviation number
1857 for (AttributeLinkedOffsetFixup &F : AttributesFixups)
1858 F.LinkedOffsetFixupVal += AbbrevNumberSize;
1860 for (AttributeLinkedOffsetFixup &F : AttributesFixups)
1861 ObjFile.Addresses->updateAndSaveValidRelocs(
1862 Unit.getOrigUnit().getVersion() >= 5, Unit.getOrigUnit().getOffset(),
1863 F.LinkedOffsetFixupVal, F.InputAttrStartOffset, F.InputAttrEndOffset);
1865 if (!HasChildren) {
1866 // Update our size.
1867 Die->setSize(OutOffset - Die->getOffset());
1868 return Die;
1871 // Recursively clone children.
1872 for (auto Child : InputDIE.children()) {
1873 if (DIE *Clone = cloneDIE(Child, File, Unit, PCOffset, OutOffset, Flags,
1874 IsLittleEndian)) {
1875 Die->addChild(Clone);
1876 OutOffset = Clone->getOffset() + Clone->getSize();
1880 // Account for the end of children marker.
1881 OutOffset += sizeof(int8_t);
1882 // Update our size.
1883 Die->setSize(OutOffset - Die->getOffset());
1884 return Die;
1887 /// Patch the input object file relevant debug_ranges or debug_rnglists
1888 /// entries and emit them in the output file. Update the relevant attributes
1889 /// to point at the new entries.
1890 void DWARFLinker::generateUnitRanges(CompileUnit &Unit, const DWARFFile &File,
1891 DebugDieValuePool &AddrPool) const {
1892 if (LLVM_UNLIKELY(Options.Update))
1893 return;
1895 const auto &FunctionRanges = Unit.getFunctionRanges();
1897 // Build set of linked address ranges for unit function ranges.
1898 AddressRanges LinkedFunctionRanges;
1899 for (const AddressRangeValuePair &Range : FunctionRanges)
1900 LinkedFunctionRanges.insert(
1901 {Range.Range.start() + Range.Value, Range.Range.end() + Range.Value});
1903 // Emit LinkedFunctionRanges into .debug_aranges
1904 if (!LinkedFunctionRanges.empty())
1905 TheDwarfEmitter->emitDwarfDebugArangesTable(Unit, LinkedFunctionRanges);
1907 RngListAttributesTy AllRngListAttributes = Unit.getRangesAttributes();
1908 std::optional<PatchLocation> UnitRngListAttribute =
1909 Unit.getUnitRangesAttribute();
1911 if (!AllRngListAttributes.empty() || UnitRngListAttribute) {
1912 std::optional<AddressRangeValuePair> CachedRange;
1913 MCSymbol *EndLabel = TheDwarfEmitter->emitDwarfDebugRangeListHeader(Unit);
1915 // Read original address ranges, apply relocation value, emit linked address
1916 // ranges.
1917 for (PatchLocation &AttributePatch : AllRngListAttributes) {
1918 // Get ranges from the source DWARF corresponding to the current
1919 // attribute.
1920 AddressRanges LinkedRanges;
1921 if (Expected<DWARFAddressRangesVector> OriginalRanges =
1922 Unit.getOrigUnit().findRnglistFromOffset(AttributePatch.get())) {
1923 // Apply relocation adjustment.
1924 for (const auto &Range : *OriginalRanges) {
1925 if (!CachedRange || !CachedRange->Range.contains(Range.LowPC))
1926 CachedRange = FunctionRanges.getRangeThatContains(Range.LowPC);
1928 // All range entries should lie in the function range.
1929 if (!CachedRange) {
1930 reportWarning("inconsistent range data.", File);
1931 continue;
1934 // Store range for emiting.
1935 LinkedRanges.insert({Range.LowPC + CachedRange->Value,
1936 Range.HighPC + CachedRange->Value});
1938 } else {
1939 llvm::consumeError(OriginalRanges.takeError());
1940 reportWarning("invalid range list ignored.", File);
1943 // Emit linked ranges.
1944 TheDwarfEmitter->emitDwarfDebugRangeListFragment(
1945 Unit, LinkedRanges, AttributePatch, AddrPool);
1948 // Emit ranges for Unit AT_ranges attribute.
1949 if (UnitRngListAttribute.has_value())
1950 TheDwarfEmitter->emitDwarfDebugRangeListFragment(
1951 Unit, LinkedFunctionRanges, *UnitRngListAttribute, AddrPool);
1953 // Emit ranges footer.
1954 TheDwarfEmitter->emitDwarfDebugRangeListFooter(Unit, EndLabel);
1958 void DWARFLinker::DIECloner::generateUnitLocations(
1959 CompileUnit &Unit, const DWARFFile &File,
1960 ExpressionHandlerRef ExprHandler) {
1961 if (LLVM_UNLIKELY(Linker.Options.Update))
1962 return;
1964 const LocListAttributesTy &AllLocListAttributes =
1965 Unit.getLocationAttributes();
1967 if (AllLocListAttributes.empty())
1968 return;
1970 // Emit locations list table header.
1971 MCSymbol *EndLabel = Emitter->emitDwarfDebugLocListHeader(Unit);
1973 for (auto &CurLocAttr : AllLocListAttributes) {
1974 // Get location expressions vector corresponding to the current attribute
1975 // from the source DWARF.
1976 Expected<DWARFLocationExpressionsVector> OriginalLocations =
1977 Unit.getOrigUnit().findLoclistFromOffset(CurLocAttr.get());
1979 if (!OriginalLocations) {
1980 llvm::consumeError(OriginalLocations.takeError());
1981 Linker.reportWarning("Invalid location attribute ignored.", File);
1982 continue;
1985 DWARFLocationExpressionsVector LinkedLocationExpressions;
1986 for (DWARFLocationExpression &CurExpression : *OriginalLocations) {
1987 DWARFLocationExpression LinkedExpression;
1989 if (CurExpression.Range) {
1990 // Relocate address range.
1991 LinkedExpression.Range = {
1992 CurExpression.Range->LowPC + CurLocAttr.RelocAdjustment,
1993 CurExpression.Range->HighPC + CurLocAttr.RelocAdjustment};
1996 // Clone expression.
1997 LinkedExpression.Expr.reserve(CurExpression.Expr.size());
1998 ExprHandler(CurExpression.Expr, LinkedExpression.Expr,
1999 CurLocAttr.RelocAdjustment);
2001 LinkedLocationExpressions.push_back(LinkedExpression);
2004 // Emit locations list table fragment corresponding to the CurLocAttr.
2005 Emitter->emitDwarfDebugLocListFragment(Unit, LinkedLocationExpressions,
2006 CurLocAttr, AddrPool);
2009 // Emit locations list table footer.
2010 Emitter->emitDwarfDebugLocListFooter(Unit, EndLabel);
2013 static void patchAddrBase(DIE &Die, DIEInteger Offset) {
2014 for (auto &V : Die.values())
2015 if (V.getAttribute() == dwarf::DW_AT_addr_base) {
2016 V = DIEValue(V.getAttribute(), V.getForm(), Offset);
2017 return;
2020 llvm_unreachable("Didn't find a DW_AT_addr_base in cloned DIE!");
2023 void DWARFLinker::DIECloner::emitDebugAddrSection(
2024 CompileUnit &Unit,
2025 const uint16_t DwarfVersion) const {
2027 if (LLVM_UNLIKELY(Linker.Options.Update))
2028 return;
2030 if (DwarfVersion < 5)
2031 return;
2033 if (AddrPool.DieValues.empty())
2034 return;
2036 MCSymbol *EndLabel = Emitter->emitDwarfDebugAddrsHeader(Unit);
2037 patchAddrBase(*Unit.getOutputUnitDIE(),
2038 DIEInteger(Emitter->getDebugAddrSectionSize()));
2039 Emitter->emitDwarfDebugAddrs(AddrPool.DieValues,
2040 Unit.getOrigUnit().getAddressByteSize());
2041 Emitter->emitDwarfDebugAddrsFooter(Unit, EndLabel);
2044 /// Insert the new line info sequence \p Seq into the current
2045 /// set of already linked line info \p Rows.
2046 static void insertLineSequence(std::vector<DWARFDebugLine::Row> &Seq,
2047 std::vector<DWARFDebugLine::Row> &Rows) {
2048 if (Seq.empty())
2049 return;
2051 if (!Rows.empty() && Rows.back().Address < Seq.front().Address) {
2052 llvm::append_range(Rows, Seq);
2053 Seq.clear();
2054 return;
2057 object::SectionedAddress Front = Seq.front().Address;
2058 auto InsertPoint = partition_point(
2059 Rows, [=](const DWARFDebugLine::Row &O) { return O.Address < Front; });
2061 // FIXME: this only removes the unneeded end_sequence if the
2062 // sequences have been inserted in order. Using a global sort like
2063 // described in generateLineTableForUnit() and delaying the end_sequene
2064 // elimination to emitLineTableForUnit() we can get rid of all of them.
2065 if (InsertPoint != Rows.end() && InsertPoint->Address == Front &&
2066 InsertPoint->EndSequence) {
2067 *InsertPoint = Seq.front();
2068 Rows.insert(InsertPoint + 1, Seq.begin() + 1, Seq.end());
2069 } else {
2070 Rows.insert(InsertPoint, Seq.begin(), Seq.end());
2073 Seq.clear();
2076 static void patchStmtList(DIE &Die, DIEInteger Offset) {
2077 for (auto &V : Die.values())
2078 if (V.getAttribute() == dwarf::DW_AT_stmt_list) {
2079 V = DIEValue(V.getAttribute(), V.getForm(), Offset);
2080 return;
2083 llvm_unreachable("Didn't find DW_AT_stmt_list in cloned DIE!");
2086 void DWARFLinker::DIECloner::rememberUnitForMacroOffset(CompileUnit &Unit) {
2087 DWARFUnit &OrigUnit = Unit.getOrigUnit();
2088 DWARFDie OrigUnitDie = OrigUnit.getUnitDIE();
2090 if (std::optional<uint64_t> MacroAttr =
2091 dwarf::toSectionOffset(OrigUnitDie.find(dwarf::DW_AT_macros))) {
2092 UnitMacroMap.insert(std::make_pair(*MacroAttr, &Unit));
2093 return;
2096 if (std::optional<uint64_t> MacroAttr =
2097 dwarf::toSectionOffset(OrigUnitDie.find(dwarf::DW_AT_macro_info))) {
2098 UnitMacroMap.insert(std::make_pair(*MacroAttr, &Unit));
2099 return;
2103 void DWARFLinker::DIECloner::generateLineTableForUnit(CompileUnit &Unit) {
2104 if (LLVM_UNLIKELY(Emitter == nullptr))
2105 return;
2107 // Check whether DW_AT_stmt_list attribute is presented.
2108 DWARFDie CUDie = Unit.getOrigUnit().getUnitDIE();
2109 auto StmtList = dwarf::toSectionOffset(CUDie.find(dwarf::DW_AT_stmt_list));
2110 if (!StmtList)
2111 return;
2113 // Update the cloned DW_AT_stmt_list with the correct debug_line offset.
2114 if (auto *OutputDIE = Unit.getOutputUnitDIE())
2115 patchStmtList(*OutputDIE, DIEInteger(Emitter->getLineSectionSize()));
2117 if (const DWARFDebugLine::LineTable *LT =
2118 ObjFile.Dwarf->getLineTableForUnit(&Unit.getOrigUnit())) {
2120 DWARFDebugLine::LineTable LineTable;
2122 // Set Line Table header.
2123 LineTable.Prologue = LT->Prologue;
2125 // Set Line Table Rows.
2126 if (Linker.Options.Update) {
2127 LineTable.Rows = LT->Rows;
2128 // If all the line table contains is a DW_LNE_end_sequence, clear the line
2129 // table rows, it will be inserted again in the DWARFStreamer.
2130 if (LineTable.Rows.size() == 1 && LineTable.Rows[0].EndSequence)
2131 LineTable.Rows.clear();
2133 LineTable.Sequences = LT->Sequences;
2134 } else {
2135 // This vector is the output line table.
2136 std::vector<DWARFDebugLine::Row> NewRows;
2137 NewRows.reserve(LT->Rows.size());
2139 // Current sequence of rows being extracted, before being inserted
2140 // in NewRows.
2141 std::vector<DWARFDebugLine::Row> Seq;
2143 const auto &FunctionRanges = Unit.getFunctionRanges();
2144 std::optional<AddressRangeValuePair> CurrRange;
2146 // FIXME: This logic is meant to generate exactly the same output as
2147 // Darwin's classic dsymutil. There is a nicer way to implement this
2148 // by simply putting all the relocated line info in NewRows and simply
2149 // sorting NewRows before passing it to emitLineTableForUnit. This
2150 // should be correct as sequences for a function should stay
2151 // together in the sorted output. There are a few corner cases that
2152 // look suspicious though, and that required to implement the logic
2153 // this way. Revisit that once initial validation is finished.
2155 // Iterate over the object file line info and extract the sequences
2156 // that correspond to linked functions.
2157 for (DWARFDebugLine::Row Row : LT->Rows) {
2158 // Check whether we stepped out of the range. The range is
2159 // half-open, but consider accept the end address of the range if
2160 // it is marked as end_sequence in the input (because in that
2161 // case, the relocation offset is accurate and that entry won't
2162 // serve as the start of another function).
2163 if (!CurrRange || !CurrRange->Range.contains(Row.Address.Address)) {
2164 // We just stepped out of a known range. Insert a end_sequence
2165 // corresponding to the end of the range.
2166 uint64_t StopAddress =
2167 CurrRange ? CurrRange->Range.end() + CurrRange->Value : -1ULL;
2168 CurrRange = FunctionRanges.getRangeThatContains(Row.Address.Address);
2169 if (StopAddress != -1ULL && !Seq.empty()) {
2170 // Insert end sequence row with the computed end address, but
2171 // the same line as the previous one.
2172 auto NextLine = Seq.back();
2173 NextLine.Address.Address = StopAddress;
2174 NextLine.EndSequence = 1;
2175 NextLine.PrologueEnd = 0;
2176 NextLine.BasicBlock = 0;
2177 NextLine.EpilogueBegin = 0;
2178 Seq.push_back(NextLine);
2179 insertLineSequence(Seq, NewRows);
2182 if (!CurrRange)
2183 continue;
2186 // Ignore empty sequences.
2187 if (Row.EndSequence && Seq.empty())
2188 continue;
2190 // Relocate row address and add it to the current sequence.
2191 Row.Address.Address += CurrRange->Value;
2192 Seq.emplace_back(Row);
2194 if (Row.EndSequence)
2195 insertLineSequence(Seq, NewRows);
2198 LineTable.Rows = std::move(NewRows);
2201 Emitter->emitLineTableForUnit(LineTable, Unit, DebugStrPool,
2202 DebugLineStrPool);
2203 } else
2204 Linker.reportWarning("Cann't load line table.", ObjFile);
2207 void DWARFLinker::emitAcceleratorEntriesForUnit(CompileUnit &Unit) {
2208 for (AccelTableKind AccelTableKind : Options.AccelTables) {
2209 switch (AccelTableKind) {
2210 case AccelTableKind::Apple: {
2211 // Add namespaces.
2212 for (const auto &Namespace : Unit.getNamespaces())
2213 AppleNamespaces.addName(Namespace.Name, Namespace.Die->getOffset() +
2214 Unit.getStartOffset());
2215 // Add names.
2216 for (const auto &Pubname : Unit.getPubnames())
2217 AppleNames.addName(Pubname.Name,
2218 Pubname.Die->getOffset() + Unit.getStartOffset());
2219 // Add types.
2220 for (const auto &Pubtype : Unit.getPubtypes())
2221 AppleTypes.addName(
2222 Pubtype.Name, Pubtype.Die->getOffset() + Unit.getStartOffset(),
2223 Pubtype.Die->getTag(),
2224 Pubtype.ObjcClassImplementation ? dwarf::DW_FLAG_type_implementation
2225 : 0,
2226 Pubtype.QualifiedNameHash);
2227 // Add ObjC names.
2228 for (const auto &ObjC : Unit.getObjC())
2229 AppleObjc.addName(ObjC.Name,
2230 ObjC.Die->getOffset() + Unit.getStartOffset());
2231 } break;
2232 case AccelTableKind::Pub: {
2233 TheDwarfEmitter->emitPubNamesForUnit(Unit);
2234 TheDwarfEmitter->emitPubTypesForUnit(Unit);
2235 } break;
2236 case AccelTableKind::DebugNames: {
2237 for (const auto &Namespace : Unit.getNamespaces())
2238 DebugNames.addName(Namespace.Name, Namespace.Die->getOffset(),
2239 Namespace.Die->getTag(), Unit.getUniqueID());
2240 for (const auto &Pubname : Unit.getPubnames())
2241 DebugNames.addName(Pubname.Name, Pubname.Die->getOffset(),
2242 Pubname.Die->getTag(), Unit.getUniqueID());
2243 for (const auto &Pubtype : Unit.getPubtypes())
2244 DebugNames.addName(Pubtype.Name, Pubtype.Die->getOffset(),
2245 Pubtype.Die->getTag(), Unit.getUniqueID());
2246 } break;
2251 /// Read the frame info stored in the object, and emit the
2252 /// patched frame descriptions for the resulting file.
2254 /// This is actually pretty easy as the data of the CIEs and FDEs can
2255 /// be considered as black boxes and moved as is. The only thing to do
2256 /// is to patch the addresses in the headers.
2257 void DWARFLinker::patchFrameInfoForObject(LinkContext &Context) {
2258 DWARFContext &OrigDwarf = *Context.File.Dwarf;
2259 unsigned SrcAddrSize = OrigDwarf.getDWARFObj().getAddressSize();
2261 StringRef FrameData = OrigDwarf.getDWARFObj().getFrameSection().Data;
2262 if (FrameData.empty())
2263 return;
2265 RangesTy AllUnitsRanges;
2266 for (std::unique_ptr<CompileUnit> &Unit : Context.CompileUnits) {
2267 for (auto CurRange : Unit->getFunctionRanges())
2268 AllUnitsRanges.insert(CurRange.Range, CurRange.Value);
2271 DataExtractor Data(FrameData, OrigDwarf.isLittleEndian(), 0);
2272 uint64_t InputOffset = 0;
2274 // Store the data of the CIEs defined in this object, keyed by their
2275 // offsets.
2276 DenseMap<uint64_t, StringRef> LocalCIES;
2278 while (Data.isValidOffset(InputOffset)) {
2279 uint64_t EntryOffset = InputOffset;
2280 uint32_t InitialLength = Data.getU32(&InputOffset);
2281 if (InitialLength == 0xFFFFFFFF)
2282 return reportWarning("Dwarf64 bits no supported", Context.File);
2284 uint32_t CIEId = Data.getU32(&InputOffset);
2285 if (CIEId == 0xFFFFFFFF) {
2286 // This is a CIE, store it.
2287 StringRef CIEData = FrameData.substr(EntryOffset, InitialLength + 4);
2288 LocalCIES[EntryOffset] = CIEData;
2289 // The -4 is to account for the CIEId we just read.
2290 InputOffset += InitialLength - 4;
2291 continue;
2294 uint64_t Loc = Data.getUnsigned(&InputOffset, SrcAddrSize);
2296 // Some compilers seem to emit frame info that doesn't start at
2297 // the function entry point, thus we can't just lookup the address
2298 // in the debug map. Use the AddressInfo's range map to see if the FDE
2299 // describes something that we can relocate.
2300 std::optional<AddressRangeValuePair> Range =
2301 AllUnitsRanges.getRangeThatContains(Loc);
2302 if (!Range) {
2303 // The +4 is to account for the size of the InitialLength field itself.
2304 InputOffset = EntryOffset + InitialLength + 4;
2305 continue;
2308 // This is an FDE, and we have a mapping.
2309 // Have we already emitted a corresponding CIE?
2310 StringRef CIEData = LocalCIES[CIEId];
2311 if (CIEData.empty())
2312 return reportWarning("Inconsistent debug_frame content. Dropping.",
2313 Context.File);
2315 // Look if we already emitted a CIE that corresponds to the
2316 // referenced one (the CIE data is the key of that lookup).
2317 auto IteratorInserted = EmittedCIEs.insert(
2318 std::make_pair(CIEData, TheDwarfEmitter->getFrameSectionSize()));
2319 // If there is no CIE yet for this ID, emit it.
2320 if (IteratorInserted.second) {
2321 LastCIEOffset = TheDwarfEmitter->getFrameSectionSize();
2322 IteratorInserted.first->getValue() = LastCIEOffset;
2323 TheDwarfEmitter->emitCIE(CIEData);
2326 // Emit the FDE with updated address and CIE pointer.
2327 // (4 + AddrSize) is the size of the CIEId + initial_location
2328 // fields that will get reconstructed by emitFDE().
2329 unsigned FDERemainingBytes = InitialLength - (4 + SrcAddrSize);
2330 TheDwarfEmitter->emitFDE(IteratorInserted.first->getValue(), SrcAddrSize,
2331 Loc + Range->Value,
2332 FrameData.substr(InputOffset, FDERemainingBytes));
2333 InputOffset += FDERemainingBytes;
2337 uint32_t DWARFLinker::DIECloner::hashFullyQualifiedName(DWARFDie DIE,
2338 CompileUnit &U,
2339 const DWARFFile &File,
2340 int ChildRecurseDepth) {
2341 const char *Name = nullptr;
2342 DWARFUnit *OrigUnit = &U.getOrigUnit();
2343 CompileUnit *CU = &U;
2344 std::optional<DWARFFormValue> Ref;
2346 while (true) {
2347 if (const char *CurrentName = DIE.getName(DINameKind::ShortName))
2348 Name = CurrentName;
2350 if (!(Ref = DIE.find(dwarf::DW_AT_specification)) &&
2351 !(Ref = DIE.find(dwarf::DW_AT_abstract_origin)))
2352 break;
2354 if (!Ref->isFormClass(DWARFFormValue::FC_Reference))
2355 break;
2357 CompileUnit *RefCU;
2358 if (auto RefDIE =
2359 Linker.resolveDIEReference(File, CompileUnits, *Ref, DIE, RefCU)) {
2360 CU = RefCU;
2361 OrigUnit = &RefCU->getOrigUnit();
2362 DIE = RefDIE;
2366 unsigned Idx = OrigUnit->getDIEIndex(DIE);
2367 if (!Name && DIE.getTag() == dwarf::DW_TAG_namespace)
2368 Name = "(anonymous namespace)";
2370 if (CU->getInfo(Idx).ParentIdx == 0 ||
2371 // FIXME: dsymutil-classic compatibility. Ignore modules.
2372 CU->getOrigUnit().getDIEAtIndex(CU->getInfo(Idx).ParentIdx).getTag() ==
2373 dwarf::DW_TAG_module)
2374 return djbHash(Name ? Name : "", djbHash(ChildRecurseDepth ? "" : "::"));
2376 DWARFDie Die = OrigUnit->getDIEAtIndex(CU->getInfo(Idx).ParentIdx);
2377 return djbHash(
2378 (Name ? Name : ""),
2379 djbHash((Name ? "::" : ""),
2380 hashFullyQualifiedName(Die, *CU, File, ++ChildRecurseDepth)));
2383 static uint64_t getDwoId(const DWARFDie &CUDie) {
2384 auto DwoId = dwarf::toUnsigned(
2385 CUDie.find({dwarf::DW_AT_dwo_id, dwarf::DW_AT_GNU_dwo_id}));
2386 if (DwoId)
2387 return *DwoId;
2388 return 0;
2391 static std::string remapPath(StringRef Path,
2392 const objectPrefixMap &ObjectPrefixMap) {
2393 if (ObjectPrefixMap.empty())
2394 return Path.str();
2396 SmallString<256> p = Path;
2397 for (const auto &Entry : ObjectPrefixMap)
2398 if (llvm::sys::path::replace_path_prefix(p, Entry.first, Entry.second))
2399 break;
2400 return p.str().str();
2403 static std::string getPCMFile(const DWARFDie &CUDie,
2404 objectPrefixMap *ObjectPrefixMap) {
2405 std::string PCMFile = dwarf::toString(
2406 CUDie.find({dwarf::DW_AT_dwo_name, dwarf::DW_AT_GNU_dwo_name}), "");
2408 if (PCMFile.empty())
2409 return PCMFile;
2411 if (ObjectPrefixMap)
2412 PCMFile = remapPath(PCMFile, *ObjectPrefixMap);
2414 return PCMFile;
2417 std::pair<bool, bool> DWARFLinker::isClangModuleRef(const DWARFDie &CUDie,
2418 std::string &PCMFile,
2419 LinkContext &Context,
2420 unsigned Indent,
2421 bool Quiet) {
2422 if (PCMFile.empty())
2423 return std::make_pair(false, false);
2425 // Clang module DWARF skeleton CUs abuse this for the path to the module.
2426 uint64_t DwoId = getDwoId(CUDie);
2428 std::string Name = dwarf::toString(CUDie.find(dwarf::DW_AT_name), "");
2429 if (Name.empty()) {
2430 if (!Quiet)
2431 reportWarning("Anonymous module skeleton CU for " + PCMFile,
2432 Context.File);
2433 return std::make_pair(true, true);
2436 if (!Quiet && Options.Verbose) {
2437 outs().indent(Indent);
2438 outs() << "Found clang module reference " << PCMFile;
2441 auto Cached = ClangModules.find(PCMFile);
2442 if (Cached != ClangModules.end()) {
2443 // FIXME: Until PR27449 (https://llvm.org/bugs/show_bug.cgi?id=27449) is
2444 // fixed in clang, only warn about DWO_id mismatches in verbose mode.
2445 // ASTFileSignatures will change randomly when a module is rebuilt.
2446 if (!Quiet && Options.Verbose && (Cached->second != DwoId))
2447 reportWarning(Twine("hash mismatch: this object file was built against a "
2448 "different version of the module ") +
2449 PCMFile,
2450 Context.File);
2451 if (!Quiet && Options.Verbose)
2452 outs() << " [cached].\n";
2453 return std::make_pair(true, true);
2456 return std::make_pair(true, false);
2459 bool DWARFLinker::registerModuleReference(const DWARFDie &CUDie,
2460 LinkContext &Context,
2461 objFileLoader Loader,
2462 CompileUnitHandler OnCUDieLoaded,
2463 unsigned Indent) {
2464 std::string PCMFile = getPCMFile(CUDie, Options.ObjectPrefixMap);
2465 std::pair<bool, bool> IsClangModuleRef =
2466 isClangModuleRef(CUDie, PCMFile, Context, Indent, false);
2468 if (!IsClangModuleRef.first)
2469 return false;
2471 if (IsClangModuleRef.second)
2472 return true;
2474 if (Options.Verbose)
2475 outs() << " ...\n";
2477 // Cyclic dependencies are disallowed by Clang, but we still
2478 // shouldn't run into an infinite loop, so mark it as processed now.
2479 ClangModules.insert({PCMFile, getDwoId(CUDie)});
2481 if (Error E = loadClangModule(Loader, CUDie, PCMFile, Context, OnCUDieLoaded,
2482 Indent + 2)) {
2483 consumeError(std::move(E));
2484 return false;
2486 return true;
2489 Error DWARFLinker::loadClangModule(objFileLoader Loader, const DWARFDie &CUDie,
2490 const std::string &PCMFile,
2491 LinkContext &Context,
2492 CompileUnitHandler OnCUDieLoaded,
2493 unsigned Indent) {
2495 uint64_t DwoId = getDwoId(CUDie);
2496 std::string ModuleName = dwarf::toString(CUDie.find(dwarf::DW_AT_name), "");
2498 /// Using a SmallString<0> because loadClangModule() is recursive.
2499 SmallString<0> Path(Options.PrependPath);
2500 if (sys::path::is_relative(PCMFile))
2501 resolveRelativeObjectPath(Path, CUDie);
2502 sys::path::append(Path, PCMFile);
2503 // Don't use the cached binary holder because we have no thread-safety
2504 // guarantee and the lifetime is limited.
2506 if (Loader == nullptr) {
2507 reportError("Could not load clang module: loader is not specified.\n",
2508 Context.File);
2509 return Error::success();
2512 auto ErrOrObj = Loader(Context.File.FileName, Path);
2513 if (!ErrOrObj)
2514 return Error::success();
2516 std::unique_ptr<CompileUnit> Unit;
2517 for (const auto &CU : ErrOrObj->Dwarf->compile_units()) {
2518 OnCUDieLoaded(*CU);
2519 // Recursively get all modules imported by this one.
2520 auto ChildCUDie = CU->getUnitDIE();
2521 if (!ChildCUDie)
2522 continue;
2523 if (!registerModuleReference(ChildCUDie, Context, Loader, OnCUDieLoaded,
2524 Indent)) {
2525 if (Unit) {
2526 std::string Err =
2527 (PCMFile +
2528 ": Clang modules are expected to have exactly 1 compile unit.\n");
2529 reportError(Err, Context.File);
2530 return make_error<StringError>(Err, inconvertibleErrorCode());
2532 // FIXME: Until PR27449 (https://llvm.org/bugs/show_bug.cgi?id=27449) is
2533 // fixed in clang, only warn about DWO_id mismatches in verbose mode.
2534 // ASTFileSignatures will change randomly when a module is rebuilt.
2535 uint64_t PCMDwoId = getDwoId(ChildCUDie);
2536 if (PCMDwoId != DwoId) {
2537 if (Options.Verbose)
2538 reportWarning(
2539 Twine("hash mismatch: this object file was built against a "
2540 "different version of the module ") +
2541 PCMFile,
2542 Context.File);
2543 // Update the cache entry with the DwoId of the module loaded from disk.
2544 ClangModules[PCMFile] = PCMDwoId;
2547 // Add this module.
2548 Unit = std::make_unique<CompileUnit>(*CU, UniqueUnitID++, !Options.NoODR,
2549 ModuleName);
2553 if (Unit)
2554 Context.ModuleUnits.emplace_back(RefModuleUnit{*ErrOrObj, std::move(Unit)});
2556 return Error::success();
2559 uint64_t DWARFLinker::DIECloner::cloneAllCompileUnits(
2560 DWARFContext &DwarfContext, const DWARFFile &File, bool IsLittleEndian) {
2561 uint64_t OutputDebugInfoSize =
2562 (Emitter == nullptr) ? 0 : Emitter->getDebugInfoSectionSize();
2563 const uint64_t StartOutputDebugInfoSize = OutputDebugInfoSize;
2565 for (auto &CurrentUnit : CompileUnits) {
2566 const uint16_t DwarfVersion = CurrentUnit->getOrigUnit().getVersion();
2567 const uint32_t UnitHeaderSize = DwarfVersion >= 5 ? 12 : 11;
2568 auto InputDIE = CurrentUnit->getOrigUnit().getUnitDIE();
2569 CurrentUnit->setStartOffset(OutputDebugInfoSize);
2570 if (!InputDIE) {
2571 OutputDebugInfoSize = CurrentUnit->computeNextUnitOffset(DwarfVersion);
2572 continue;
2574 if (CurrentUnit->getInfo(0).Keep) {
2575 // Clone the InputDIE into your Unit DIE in our compile unit since it
2576 // already has a DIE inside of it.
2577 CurrentUnit->createOutputDIE();
2578 rememberUnitForMacroOffset(*CurrentUnit);
2579 cloneDIE(InputDIE, File, *CurrentUnit, 0 /* PC offset */, UnitHeaderSize,
2580 0, IsLittleEndian, CurrentUnit->getOutputUnitDIE());
2583 OutputDebugInfoSize = CurrentUnit->computeNextUnitOffset(DwarfVersion);
2585 if (Emitter != nullptr) {
2587 generateLineTableForUnit(*CurrentUnit);
2589 Linker.emitAcceleratorEntriesForUnit(*CurrentUnit);
2591 if (LLVM_UNLIKELY(Linker.Options.Update))
2592 continue;
2594 Linker.generateUnitRanges(*CurrentUnit, File, AddrPool);
2596 auto ProcessExpr = [&](SmallVectorImpl<uint8_t> &SrcBytes,
2597 SmallVectorImpl<uint8_t> &OutBytes,
2598 int64_t RelocAdjustment) {
2599 DWARFUnit &OrigUnit = CurrentUnit->getOrigUnit();
2600 DataExtractor Data(SrcBytes, IsLittleEndian,
2601 OrigUnit.getAddressByteSize());
2602 cloneExpression(Data,
2603 DWARFExpression(Data, OrigUnit.getAddressByteSize(),
2604 OrigUnit.getFormParams().Format),
2605 File, *CurrentUnit, OutBytes, RelocAdjustment,
2606 IsLittleEndian);
2608 generateUnitLocations(*CurrentUnit, File, ProcessExpr);
2609 emitDebugAddrSection(*CurrentUnit, DwarfVersion);
2611 AddrPool.clear();
2614 if (Emitter != nullptr) {
2615 assert(Emitter);
2616 // Emit macro tables.
2617 Emitter->emitMacroTables(File.Dwarf.get(), UnitMacroMap, DebugStrPool);
2619 // Emit all the compile unit's debug information.
2620 for (auto &CurrentUnit : CompileUnits) {
2621 CurrentUnit->fixupForwardReferences();
2623 if (!CurrentUnit->getOutputUnitDIE())
2624 continue;
2626 unsigned DwarfVersion = CurrentUnit->getOrigUnit().getVersion();
2628 assert(Emitter->getDebugInfoSectionSize() ==
2629 CurrentUnit->getStartOffset());
2630 Emitter->emitCompileUnitHeader(*CurrentUnit, DwarfVersion);
2631 Emitter->emitDIE(*CurrentUnit->getOutputUnitDIE());
2632 assert(Emitter->getDebugInfoSectionSize() ==
2633 CurrentUnit->computeNextUnitOffset(DwarfVersion));
2637 return OutputDebugInfoSize - StartOutputDebugInfoSize;
2640 void DWARFLinker::copyInvariantDebugSection(DWARFContext &Dwarf) {
2641 TheDwarfEmitter->emitSectionContents(Dwarf.getDWARFObj().getLocSection().Data,
2642 "debug_loc");
2643 TheDwarfEmitter->emitSectionContents(
2644 Dwarf.getDWARFObj().getRangesSection().Data, "debug_ranges");
2645 TheDwarfEmitter->emitSectionContents(
2646 Dwarf.getDWARFObj().getFrameSection().Data, "debug_frame");
2647 TheDwarfEmitter->emitSectionContents(Dwarf.getDWARFObj().getArangesSection(),
2648 "debug_aranges");
2649 TheDwarfEmitter->emitSectionContents(
2650 Dwarf.getDWARFObj().getAddrSection().Data, "debug_addr");
2651 TheDwarfEmitter->emitSectionContents(
2652 Dwarf.getDWARFObj().getRnglistsSection().Data, "debug_rnglists");
2653 TheDwarfEmitter->emitSectionContents(
2654 Dwarf.getDWARFObj().getLoclistsSection().Data, "debug_loclists");
2657 void DWARFLinker::addObjectFile(DWARFFile &File, objFileLoader Loader,
2658 CompileUnitHandler OnCUDieLoaded) {
2659 ObjectContexts.emplace_back(LinkContext(File));
2661 if (ObjectContexts.back().File.Dwarf) {
2662 for (const std::unique_ptr<DWARFUnit> &CU :
2663 ObjectContexts.back().File.Dwarf->compile_units()) {
2664 DWARFDie CUDie = CU->getUnitDIE();
2666 if (!CUDie)
2667 continue;
2669 OnCUDieLoaded(*CU);
2671 if (!LLVM_UNLIKELY(Options.Update))
2672 registerModuleReference(CUDie, ObjectContexts.back(), Loader,
2673 OnCUDieLoaded);
2678 Error DWARFLinker::link() {
2679 assert((Options.TargetDWARFVersion != 0) &&
2680 "TargetDWARFVersion should be set");
2682 // First populate the data structure we need for each iteration of the
2683 // parallel loop.
2684 unsigned NumObjects = ObjectContexts.size();
2686 // This Dwarf string pool which is used for emission. It must be used
2687 // serially as the order of calling getStringOffset matters for
2688 // reproducibility.
2689 OffsetsStringPool DebugStrPool(StringsTranslator, true);
2690 OffsetsStringPool DebugLineStrPool(StringsTranslator, false);
2691 DebugDieValuePool StringOffsetPool;
2693 // ODR Contexts for the optimize.
2694 DeclContextTree ODRContexts;
2696 for (LinkContext &OptContext : ObjectContexts) {
2697 if (Options.Verbose) {
2698 if (DwarfLinkerClientID == DwarfLinkerClient::Dsymutil)
2699 outs() << "DEBUG MAP OBJECT: " << OptContext.File.FileName << "\n";
2700 else
2701 outs() << "OBJECT FILE: " << OptContext.File.FileName << "\n";
2704 if (!OptContext.File.Dwarf)
2705 continue;
2707 if (Options.VerifyInputDWARF)
2708 verifyInput(OptContext.File);
2710 // Look for relocations that correspond to address map entries.
2712 // there was findvalidrelocations previously ... probably we need to gather
2713 // info here
2714 if (LLVM_LIKELY(!Options.Update) &&
2715 !OptContext.File.Addresses->hasValidRelocs()) {
2716 if (Options.Verbose)
2717 outs() << "No valid relocations found. Skipping.\n";
2719 // Set "Skip" flag as a signal to other loops that we should not
2720 // process this iteration.
2721 OptContext.Skip = true;
2722 continue;
2725 // Setup access to the debug info.
2726 if (!OptContext.File.Dwarf)
2727 continue;
2729 // Check whether type units are presented.
2730 if (!OptContext.File.Dwarf->types_section_units().empty()) {
2731 reportWarning("type units are not currently supported: file will "
2732 "be skipped",
2733 OptContext.File);
2734 OptContext.Skip = true;
2735 continue;
2738 // Clone all the clang modules with requires extracting the DIE units. We
2739 // don't need the full debug info until the Analyze phase.
2740 OptContext.CompileUnits.reserve(
2741 OptContext.File.Dwarf->getNumCompileUnits());
2742 for (const auto &CU : OptContext.File.Dwarf->compile_units()) {
2743 auto CUDie = CU->getUnitDIE(/*ExtractUnitDIEOnly=*/true);
2744 if (Options.Verbose) {
2745 outs() << "Input compilation unit:";
2746 DIDumpOptions DumpOpts;
2747 DumpOpts.ChildRecurseDepth = 0;
2748 DumpOpts.Verbose = Options.Verbose;
2749 CUDie.dump(outs(), 0, DumpOpts);
2753 for (auto &CU : OptContext.ModuleUnits) {
2754 if (Error Err = cloneModuleUnit(OptContext, CU, ODRContexts, DebugStrPool,
2755 DebugLineStrPool, StringOffsetPool))
2756 reportWarning(toString(std::move(Err)), CU.File);
2760 // At this point we know how much data we have emitted. We use this value to
2761 // compare canonical DIE offsets in analyzeContextInfo to see if a definition
2762 // is already emitted, without being affected by canonical die offsets set
2763 // later. This prevents undeterminism when analyze and clone execute
2764 // concurrently, as clone set the canonical DIE offset and analyze reads it.
2765 const uint64_t ModulesEndOffset =
2766 (TheDwarfEmitter == nullptr) ? 0
2767 : TheDwarfEmitter->getDebugInfoSectionSize();
2769 // These variables manage the list of processed object files.
2770 // The mutex and condition variable are to ensure that this is thread safe.
2771 std::mutex ProcessedFilesMutex;
2772 std::condition_variable ProcessedFilesConditionVariable;
2773 BitVector ProcessedFiles(NumObjects, false);
2775 // Analyzing the context info is particularly expensive so it is executed in
2776 // parallel with emitting the previous compile unit.
2777 auto AnalyzeLambda = [&](size_t I) {
2778 auto &Context = ObjectContexts[I];
2780 if (Context.Skip || !Context.File.Dwarf)
2781 return;
2783 for (const auto &CU : Context.File.Dwarf->compile_units()) {
2784 // Previously we only extracted the unit DIEs. We need the full debug info
2785 // now.
2786 auto CUDie = CU->getUnitDIE(/*ExtractUnitDIEOnly=*/false);
2787 std::string PCMFile = getPCMFile(CUDie, Options.ObjectPrefixMap);
2789 if (!CUDie || LLVM_UNLIKELY(Options.Update) ||
2790 !isClangModuleRef(CUDie, PCMFile, Context, 0, true).first) {
2791 Context.CompileUnits.push_back(std::make_unique<CompileUnit>(
2792 *CU, UniqueUnitID++, !Options.NoODR && !Options.Update, ""));
2796 // Now build the DIE parent links that we will use during the next phase.
2797 for (auto &CurrentUnit : Context.CompileUnits) {
2798 auto CUDie = CurrentUnit->getOrigUnit().getUnitDIE();
2799 if (!CUDie)
2800 continue;
2801 analyzeContextInfo(CurrentUnit->getOrigUnit().getUnitDIE(), 0,
2802 *CurrentUnit, &ODRContexts.getRoot(), ODRContexts,
2803 ModulesEndOffset, Options.ParseableSwiftInterfaces,
2804 [&](const Twine &Warning, const DWARFDie &DIE) {
2805 reportWarning(Warning, Context.File, &DIE);
2810 // For each object file map how many bytes were emitted.
2811 StringMap<DebugInfoSize> SizeByObject;
2813 // And then the remaining work in serial again.
2814 // Note, although this loop runs in serial, it can run in parallel with
2815 // the analyzeContextInfo loop so long as we process files with indices >=
2816 // than those processed by analyzeContextInfo.
2817 auto CloneLambda = [&](size_t I) {
2818 auto &OptContext = ObjectContexts[I];
2819 if (OptContext.Skip || !OptContext.File.Dwarf)
2820 return;
2822 // Then mark all the DIEs that need to be present in the generated output
2823 // and collect some information about them.
2824 // Note that this loop can not be merged with the previous one because
2825 // cross-cu references require the ParentIdx to be setup for every CU in
2826 // the object file before calling this.
2827 if (LLVM_UNLIKELY(Options.Update)) {
2828 for (auto &CurrentUnit : OptContext.CompileUnits)
2829 CurrentUnit->markEverythingAsKept();
2830 copyInvariantDebugSection(*OptContext.File.Dwarf);
2831 } else {
2832 for (auto &CurrentUnit : OptContext.CompileUnits) {
2833 lookForDIEsToKeep(*OptContext.File.Addresses, OptContext.CompileUnits,
2834 CurrentUnit->getOrigUnit().getUnitDIE(),
2835 OptContext.File, *CurrentUnit, 0);
2836 #ifndef NDEBUG
2837 verifyKeepChain(*CurrentUnit);
2838 #endif
2842 // The calls to applyValidRelocs inside cloneDIE will walk the reloc
2843 // array again (in the same way findValidRelocsInDebugInfo() did). We
2844 // need to reset the NextValidReloc index to the beginning.
2845 if (OptContext.File.Addresses->hasValidRelocs() ||
2846 LLVM_UNLIKELY(Options.Update)) {
2847 SizeByObject[OptContext.File.FileName].Input =
2848 getDebugInfoSize(*OptContext.File.Dwarf);
2849 SizeByObject[OptContext.File.FileName].Output =
2850 DIECloner(*this, TheDwarfEmitter.get(), OptContext.File, DIEAlloc,
2851 OptContext.CompileUnits, Options.Update, DebugStrPool,
2852 DebugLineStrPool, StringOffsetPool)
2853 .cloneAllCompileUnits(*OptContext.File.Dwarf, OptContext.File,
2854 OptContext.File.Dwarf->isLittleEndian());
2856 if ((TheDwarfEmitter != nullptr) && !OptContext.CompileUnits.empty() &&
2857 LLVM_LIKELY(!Options.Update))
2858 patchFrameInfoForObject(OptContext);
2860 // Clean-up before starting working on the next object.
2861 cleanupAuxiliarryData(OptContext);
2864 auto EmitLambda = [&]() {
2865 // Emit everything that's global.
2866 if (TheDwarfEmitter != nullptr) {
2867 TheDwarfEmitter->emitAbbrevs(Abbreviations, Options.TargetDWARFVersion);
2868 TheDwarfEmitter->emitStrings(DebugStrPool);
2869 TheDwarfEmitter->emitStringOffsets(StringOffsetPool.DieValues,
2870 Options.TargetDWARFVersion);
2871 TheDwarfEmitter->emitLineStrings(DebugLineStrPool);
2872 for (AccelTableKind TableKind : Options.AccelTables) {
2873 switch (TableKind) {
2874 case AccelTableKind::Apple:
2875 TheDwarfEmitter->emitAppleNamespaces(AppleNamespaces);
2876 TheDwarfEmitter->emitAppleNames(AppleNames);
2877 TheDwarfEmitter->emitAppleTypes(AppleTypes);
2878 TheDwarfEmitter->emitAppleObjc(AppleObjc);
2879 break;
2880 case AccelTableKind::Pub:
2881 // Already emitted by emitAcceleratorEntriesForUnit.
2882 // Already emitted by emitAcceleratorEntriesForUnit.
2883 break;
2884 case AccelTableKind::DebugNames:
2885 TheDwarfEmitter->emitDebugNames(DebugNames);
2886 break;
2892 auto AnalyzeAll = [&]() {
2893 for (unsigned I = 0, E = NumObjects; I != E; ++I) {
2894 AnalyzeLambda(I);
2896 std::unique_lock<std::mutex> LockGuard(ProcessedFilesMutex);
2897 ProcessedFiles.set(I);
2898 ProcessedFilesConditionVariable.notify_one();
2902 auto CloneAll = [&]() {
2903 for (unsigned I = 0, E = NumObjects; I != E; ++I) {
2905 std::unique_lock<std::mutex> LockGuard(ProcessedFilesMutex);
2906 if (!ProcessedFiles[I]) {
2907 ProcessedFilesConditionVariable.wait(
2908 LockGuard, [&]() { return ProcessedFiles[I]; });
2912 CloneLambda(I);
2914 EmitLambda();
2917 // To limit memory usage in the single threaded case, analyze and clone are
2918 // run sequentially so the OptContext is freed after processing each object
2919 // in endDebugObject.
2920 if (Options.Threads == 1) {
2921 for (unsigned I = 0, E = NumObjects; I != E; ++I) {
2922 AnalyzeLambda(I);
2923 CloneLambda(I);
2925 EmitLambda();
2926 } else {
2927 ThreadPool Pool(hardware_concurrency(2));
2928 Pool.async(AnalyzeAll);
2929 Pool.async(CloneAll);
2930 Pool.wait();
2933 if (Options.Statistics) {
2934 // Create a vector sorted in descending order by output size.
2935 std::vector<std::pair<StringRef, DebugInfoSize>> Sorted;
2936 for (auto &E : SizeByObject)
2937 Sorted.emplace_back(E.first(), E.second);
2938 llvm::sort(Sorted, [](auto &LHS, auto &RHS) {
2939 return LHS.second.Output > RHS.second.Output;
2942 auto ComputePercentange = [](int64_t Input, int64_t Output) -> float {
2943 const float Difference = Output - Input;
2944 const float Sum = Input + Output;
2945 if (Sum == 0)
2946 return 0;
2947 return (Difference / (Sum / 2));
2950 int64_t InputTotal = 0;
2951 int64_t OutputTotal = 0;
2952 const char *FormatStr = "{0,-45} {1,10}b {2,10}b {3,8:P}\n";
2954 // Print header.
2955 outs() << ".debug_info section size (in bytes)\n";
2956 outs() << "----------------------------------------------------------------"
2957 "---------------\n";
2958 outs() << "Filename Object "
2959 " dSYM Change\n";
2960 outs() << "----------------------------------------------------------------"
2961 "---------------\n";
2963 // Print body.
2964 for (auto &E : Sorted) {
2965 InputTotal += E.second.Input;
2966 OutputTotal += E.second.Output;
2967 llvm::outs() << formatv(
2968 FormatStr, sys::path::filename(E.first).take_back(45), E.second.Input,
2969 E.second.Output, ComputePercentange(E.second.Input, E.second.Output));
2971 // Print total and footer.
2972 outs() << "----------------------------------------------------------------"
2973 "---------------\n";
2974 llvm::outs() << formatv(FormatStr, "Total", InputTotal, OutputTotal,
2975 ComputePercentange(InputTotal, OutputTotal));
2976 outs() << "----------------------------------------------------------------"
2977 "---------------\n\n";
2980 return Error::success();
2983 Error DWARFLinker::cloneModuleUnit(LinkContext &Context, RefModuleUnit &Unit,
2984 DeclContextTree &ODRContexts,
2985 OffsetsStringPool &DebugStrPool,
2986 OffsetsStringPool &DebugLineStrPool,
2987 DebugDieValuePool &StringOffsetPool,
2988 unsigned Indent) {
2989 assert(Unit.Unit.get() != nullptr);
2991 if (!Unit.Unit->getOrigUnit().getUnitDIE().hasChildren())
2992 return Error::success();
2994 if (Options.Verbose) {
2995 outs().indent(Indent);
2996 outs() << "cloning .debug_info from " << Unit.File.FileName << "\n";
2999 // Analyze context for the module.
3000 analyzeContextInfo(Unit.Unit->getOrigUnit().getUnitDIE(), 0, *(Unit.Unit),
3001 &ODRContexts.getRoot(), ODRContexts, 0,
3002 Options.ParseableSwiftInterfaces,
3003 [&](const Twine &Warning, const DWARFDie &DIE) {
3004 reportWarning(Warning, Context.File, &DIE);
3006 // Keep everything.
3007 Unit.Unit->markEverythingAsKept();
3009 // Clone unit.
3010 UnitListTy CompileUnits;
3011 CompileUnits.emplace_back(std::move(Unit.Unit));
3012 assert(TheDwarfEmitter);
3013 DIECloner(*this, TheDwarfEmitter.get(), Unit.File, DIEAlloc, CompileUnits,
3014 Options.Update, DebugStrPool, DebugLineStrPool, StringOffsetPool)
3015 .cloneAllCompileUnits(*Unit.File.Dwarf, Unit.File,
3016 Unit.File.Dwarf->isLittleEndian());
3017 return Error::success();
3020 void DWARFLinker::verifyInput(const DWARFFile &File) {
3021 assert(File.Dwarf);
3024 std::string Buffer;
3025 raw_string_ostream OS(Buffer);
3026 DIDumpOptions DumpOpts;
3027 if (!File.Dwarf->verify(OS, DumpOpts.noImplicitRecursion())) {
3028 if (Options.InputVerificationHandler)
3029 Options.InputVerificationHandler(File, OS.str());
3033 Error DWARFLinker::createEmitter(const Triple &TheTriple,
3034 OutputFileType FileType,
3035 raw_pwrite_stream &OutFile) {
3037 TheDwarfEmitter = std::make_unique<DwarfStreamer>(
3038 FileType, OutFile, StringsTranslator, WarningHandler);
3040 return TheDwarfEmitter->init(TheTriple, "__DWARF");
3043 DwarfEmitter *DWARFLinker::getEmitter() { return TheDwarfEmitter.get(); }
3045 } // namespace llvm