[clang][AST] Support AST files larger than 512M

[DmitryPolukhin]

Summary

Clang uses 32-bit integers for storing bit offsets from the beginning of
the file that results in 512M limit on AST file. This diff replaces
absolute offsets with relative offsets from the beginning of
corresponding data structure when it is possible. And uses 64-bit
offsets for DeclOffests and TypeOffssts because these coder AST
section may easily exceeds 512M alone.

This diff breaks AST file format compatibility so VERSION_MAJOR bumped.

Test Plan:
Existing clang AST serialization tests
Tested on clangd with ~700M and ~900M preamble files

[DmitryPolukhin]

clang-format for ASTBitCodes.h

[sammccall]

This looks reasonable to me, though I'm not familiar enough with the code to do a good review.

Just wanted to confirm my understanding: the problematic offsets that are now section-relative instead of file-relative are only used in two sections of the file (macro directives and sloc entries). These sections seem likely to be small relative to others that hold e.g. the AST. So this should raise the effective size limit by a lot, more likely 100x than 5?

(We've coincidentally started to see large crashing preambles this week, so thanks for this! I think we probably don't need to add any detection/recovery if this limit is high enough)

[DmitryPolukhin]

In D76594#1937289, @sammccall wrote:

This looks reasonable to me, though I'm not familiar enough with the code to do a good review.

Just wanted to confirm my understanding: the problematic offsets that are now section-relative instead of file-relative are only used in two sections of the file (macro directives and sloc entries). These sections seem likely to be small relative to others that hold e.g. the AST. So this should raise the effective size limit by a lot, more likely 100x than 5?

Yes, problematic offsets are now section relative and the sections several times smaller than the whole file. In suspect in some artificial cases these section can be pretty large. For example, you can create source file with only macro and you can exceed limit for on macro size alone. Moreover I'm not sure that files bigger than 4G can be processed correctly. Because there are other places where 32 bit numbers are used for some IDs or byte offsets.

(We've coincidentally started to see large crashing preambles this week, so thanks for this! I think we probably don't need to add any detection/recovery if this limit is high enough)

I added asserts for all new places where overflow can happen but perhaps we also need something for release builds. If you have cases when you reach size limit, it would be very helpful if you can try this patch on these cases. I tested it on couple ~700M files but my examples relatively unified and follow the same pattern.

[DmitryPolukhin]

Use 64-bit offsets for TypesOffsests and DeclOffsests

[xbolva00]

clang/lib/Serialization/ASTReader.cpp

3102

use C++ cast style

[DmitryPolukhin]

Resolve merge conflict, revert clang-format for ASTBitCodes.h to avoid further conflicts, use C++ cast

[DmitryPolukhin]

Rebase, all tests passes locally

[sammccall]

clang/include/clang/Serialization/ASTBitCodes.h

220–221

Is there one of these for every decl in the module? It seems like we're probably giving up a good fraction of the 4% increase for just using absolute 64 bit offsets everywhere :-( Is there still a significant gain from using section-relative elsewhere?

If there are indeed lots of these, giving up 4 bytes to padding (in addition to the wide offset) seems unfortunate and because we memcpy the structs into the AST file seems like a sad reason :-)
Can we align this to 4 bytes instead?
(e.g. by splitting into two fields and encapsulating the few direct accesses, though there's probably a neater way)

[DmitryPolukhin]

clang/include/clang/Serialization/ASTBitCodes.h

220–221

Yes, it seems that all referenced decls should have an entry in this table. 4% increase was counted on previous diff before I discovered DeclOffsets and TypeOffsets. This diff uses relative offsets for previous cases and increase is only 2 additional 64-bit integers per file. DeclOffsets are about 4.7% and TypeOffsets are about 2.5% of the whole file. With this diff both number will be 2x, splitting DeclOffset in two separate arrays could make the increase only 1.5x for DeclOffsets. If approach from this diff looks fine, I can split array of DeclOffset into 2 separate arrays. It could be a bit less efficient because of worse memory locality but will save some space.

[DmitryPolukhin]

@rsmith, @dexonsmith, @jdoerfert could you please take a look to this diff?
If you think that there are other reviewers who might have more context AST persistence, please add them.

[dexonsmith]

In D76594#1957452, @DmitryPolukhin wrote:

@rsmith, @dexonsmith, @jdoerfert could you please take a look to this diff?
If you think that there are other reviewers who might have more context AST persistence, please add them.

@Bigcheese, can you take a look?

[hokein]

I'm not familiar with the code, but this patch does fix the large-preamble crashes we encountered internally.

[sammccall]

We're starting to see quite a lot of production crashes with this and would very much like a fix.
At this point it seems like we're best moving ahead and if someone comes up with a better idea, we can improve it later.

clang/include/clang/Serialization/ASTBitCodes.h

220–221

Or even just 32-bit BitOffsetLow and BitOffsetHigh fields so the overall struct is 32-bit aligned... then you'd still be local?

[DmitryPolukhin]

Split BitOffset in DeclOffset in high/low parts; rebase

[DmitryPolukhin]

Split BitOffset in DeclOffset in high/low parts; rebase

[DmitryPolukhin]

Fix nit

[DmitryPolukhin]

@sammccall thank you for review!
I'll wait for one more day for additional feedback if any, and land this diff.

[martong]

I am not sure, but maybe this patch causes an undefined behavior?
http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-fast/builds/40472/steps/check-clang%20ubsan/logs/stdio

/b/sanitizer-x86_64-linux-fast/build/llvm-project/clang/lib/Serialization/ASTReader.cpp:6327:28: runtime error: load of misaligned address 0x7feac40a55bc for type 'const uint64_t' (aka 'const unsigned long'), which requires 8 byte alignment
0x7feac40a55bc: note: pointer points here
  00 00 00 00 a3 c7 01 00  00 00 00 00 0c c8 01 00  00 00 00 00 29 c8 01 00  00 00 00 00 7a cc 01 00
              ^ 
    #0 0x3be2fe4 in clang::ASTReader::TypeCursorForIndex(unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/clang/lib/Serialization/ASTReader.cpp:6327:28
    #1 0x3be30a0 in clang::ASTReader::readTypeRecord(unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/clang/lib/Serialization/ASTReader.cpp:6348:24
    #2 0x3bd3d4a in clang::ASTReader::GetType(unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/clang/lib/Serialization/ASTReader.cpp:6985:26
    ...

(I am in the blamelist too of that build, so that's why I am sniffing.)

[DmitryPolukhin]

In D76594#1986846, @martong wrote:

I am not sure, but maybe this patch causes an undefined behavior?
http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-fast/builds/40472/steps/check-clang%20ubsan/logs/stdio

/b/sanitizer-x86_64-linux-fast/build/llvm-project/clang/lib/Serialization/ASTReader.cpp:6327:28: runtime error: load of misaligned address 0x7feac40a55bc for type 'const uint64_t' (aka 'const unsigned long'), which requires 8 byte alignment
0x7feac40a55bc: note: pointer points here
  00 00 00 00 a3 c7 01 00  00 00 00 00 0c c8 01 00  00 00 00 00 29 c8 01 00  00 00 00 00 7a cc 01 00
              ^ 
    #0 0x3be2fe4 in clang::ASTReader::TypeCursorForIndex(unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/clang/lib/Serialization/ASTReader.cpp:6327:28
    #1 0x3be30a0 in clang::ASTReader::readTypeRecord(unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/clang/lib/Serialization/ASTReader.cpp:6348:24
    #2 0x3bd3d4a in clang::ASTReader::GetType(unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/clang/lib/Serialization/ASTReader.cpp:6985:26
    ...

(I am in the blamelist too of that build, so that's why I am sniffing.)

Temporary reverted in a8f85da9f538a400dfea00e4954e403bf5f3269c

[DmitryPolukhin]

Split TypeOffsets to low/high parts too

[sammccall]

A little sad it's not possible to just 64-bit align the typeoffsets array, but 32 seems to be the magic number.

clang/include/clang/Serialization/ASTBitCodes.h

225

may be worth mentioning that blobs in bitstream are specifically 32-bit aligned as part of the motivation here.

228

consider making this a class and these members private

256

Alternatively you could factor this as a class UnderalignedInt64, use that for type-offsets, and embed it in DeclOffset...

[DmitryPolukhin]

Comments resolved

clang/include/clang/Serialization/ASTBitCodes.h

228

I kept members public because they public for all other similar struct in the file. I think initial idea to keep them public was to have POD struct but it is actually not the case due to user defined constructor.

[JackKemp99]

clang/lib/Serialization/ASTWriter.cpp

1908

clangd crashes here:

[DmitryPolukhin]

clang/lib/Serialization/ASTWriter.cpp

1908

@JackKemp99 It means that you have a source file that has number of source location that exceeds 32-bit offset. Do you have a reproducer that you can share?

Is it something Boost related? There is a discussion about too many source location with Boost https://lists.llvm.org/pipermail/cfe-dev/2021-February/067599.html Your example may not exceed amount in of SourceLocation IDs but exceeds lower bound on offsets during preamble generation.

clang/include/clang/Serialization/ASTBitCodes.h

namespace serialization {
     /// Whenever the AST file format changes in a way that makes it
     /// incompatible with previous versions (such that a reader
     /// designed for the previous version could not support reading
     /// the new version), this number should be increased.
     ///
     /// Version 4 of AST files also requires that the version control branch and
     /// revision match exactly, since there is no backward compatibility of
     /// AST files at this time.
-    const unsigned VERSION_MAJOR = 9;
+    const unsigned VERSION_MAJOR = 10;
 
     /// AST file minor version number supported by this version of
     /// Clang.
     ///
     /// Whenever the AST format changes in a way that is still
     /// compatible with previous versions (such that a reader designed
     /// for the previous version could still support reading the new
     /// version by ignoring new kinds of subblocks), this number
namespace serialization {
     /// Source range/offset of a preprocessed entity.
     struct PPEntityOffset {
       /// Raw source location of beginning of range.
       unsigned Begin;
 
       /// Raw source location of end of range.
       unsigned End;
 
-      /// Offset in the AST file.
+      /// Offset in the AST file relative to ModuleFile::MacroOffsetsBase.
       uint32_t BitOffset;
 
       PPEntityOffset(SourceRange R, uint32_t BitOffset)
         : Begin(R.getBegin().getRawEncoding()),
           End(R.getEnd().getRawEncoding()), BitOffset(BitOffset) {}
 
       SourceLocation getBegin() const {
         return SourceLocation::getFromRawEncoding(Begin);
struct PPSkippedRange {
         return SourceLocation::getFromRawEncoding(Begin);
       }
       SourceLocation getEnd() const {
         return SourceLocation::getFromRawEncoding(End);
       }
     };
 
     /// Source range/offset of a preprocessed entity.
     struct DeclOffset {
       /// Raw source location.
-      unsigned Loc = 0;
+      uint32_t Loc = 0;
 
-      /// Offset in the AST file.
-      uint32_t BitOffset = 0;
+      /// Offset in the AST file. Split 64-bit integer into low/high parts
+      /// to keep structure alignment 32-bit and don't have padding gap.
+      /// This structure is serialized "as is" to the AST file and undefined
+      /// value in the padding affects AST hash.
+      uint32_t BitOffsetLow = 0;
+      uint32_t BitOffsetHigh = 0;
 
       DeclOffset() = default;
-      DeclOffset(SourceLocation Loc, uint32_t BitOffset)
-        : Loc(Loc.getRawEncoding()), BitOffset(BitOffset) {}
+      DeclOffset(SourceLocation Loc, uint64_t BitOffset) {
+        setLocation(Loc);
+        setBitOffset(BitOffset);
+      }
 
       void setLocation(SourceLocation L) {
         Loc = L.getRawEncoding();
       }
 
       SourceLocation getLocation() const {
         return SourceLocation::getFromRawEncoding(Loc);
       }
+
+      void setBitOffset(uint64_t Offset) {
+        BitOffsetLow = Offset;
+        BitOffsetHigh = Offset >> 32;
+      }
+
+      uint64_t getBitOffset() const {
+        return BitOffsetLow | (uint64_t(BitOffsetHigh) << 32);
+      }
     };
 
     /// The number of predefined preprocessed entity IDs.
     const unsigned int NUM_PREDEF_PP_ENTITY_IDS = 1;
 
     /// Describes the various kinds of blocks that occur within
     /// an AST file.
     enum BlockIDs {
       /// The AST block, which acts as a container around the
       /// full AST block.
       AST_BLOCK_ID = llvm::bitc::FIRST_APPLICATION_BLOCKID,
 

clang/include/clang/Serialization/ASTReader.h

private:
   llvm::DenseMap<Selector, unsigned> SelectorGeneration;
 
   /// Whether a selector is out of date. We mark a selector as out of date
   /// if we load another module after the method pool entry was pulled in.
   llvm::DenseMap<Selector, bool> SelectorOutOfDate;
 
   struct PendingMacroInfo {
     ModuleFile *M;
-    uint64_t MacroDirectivesOffset;
+    /// Offset relative to ModuleFile::MacroOffsetsBase.
+    uint32_t MacroDirectivesOffset;
 
-    PendingMacroInfo(ModuleFile *M, uint64_t MacroDirectivesOffset)
+    PendingMacroInfo(ModuleFile *M, uint32_t MacroDirectivesOffset)
         : M(M), MacroDirectivesOffset(MacroDirectivesOffset) {}
   };
 
   using PendingMacroIDsMap =
       llvm::MapVector<IdentifierInfo *, SmallVector<PendingMacroInfo, 2>>;
 
   /// Mapping from identifiers that have a macro history to the global
   /// IDs have not yet been deserialized to the global IDs of those macros.
public:
 
   /// Add a macro to deserialize its macro directive history.
   ///
   /// \param II The name of the macro.
   /// \param M The module file.
   /// \param MacroDirectivesOffset Offset of the serialized macro directive
   /// history.
   void addPendingMacro(IdentifierInfo *II, ModuleFile *M,
-                       uint64_t MacroDirectivesOffset);
+                       uint32_t MacroDirectivesOffset);
 
   /// Read the set of macros defined by this external macro source.
   void ReadDefinedMacros() override;
 
   /// Update an out-of-date identifier.
   void updateOutOfDateIdentifier(IdentifierInfo &II) override;
 
   /// Note that this identifier is up-to-date.

clang/include/clang/Serialization/ASTWriter.h

private:
   /// stored in the stream, the ID number is shifted by 2 bits to
   /// allow for the const/volatile qualifiers.
   ///
   /// Keys in the map never have const/volatile qualifiers.
   TypeIdxMap TypeIdxs;
 
   /// Offset of each type in the bitstream, indexed by
   /// the type's ID.
-  std::vector<uint32_t> TypeOffsets;
+  std::vector<uint64_t> TypeOffsets;
 
   /// The first ID number we can use for our own identifiers.
   serialization::IdentID FirstIdentID = serialization::NUM_PREDEF_IDENT_IDS;
 
   /// The identifier ID that will be assigned to the next new identifier.
   serialization::IdentID NextIdentID = FirstIdentID;
 
   /// Map that provides the ID numbers of each identifier in
struct MacroInfoToEmitData {
     const IdentifierInfo *Name;
     MacroInfo *MI;
     serialization::MacroID ID;
   };
 
   /// The macro infos to emit.
   std::vector<MacroInfoToEmitData> MacroInfosToEmit;
 
-  llvm::DenseMap<const IdentifierInfo *, uint64_t> IdentMacroDirectivesOffsetMap;
+  llvm::DenseMap<const IdentifierInfo *, uint32_t>
+      IdentMacroDirectivesOffsetMap;
 
   /// @name FlushStmt Caches
   /// @{
 
   /// Set of parent Stmts for the currently serializing sub-stmt.
   llvm::DenseSet<Stmt *> ParentStmts;
 
   /// Offsets of sub-stmts already serialized. The offset points
private:
   static ASTFileSignature createSignature(StringRef Bytes);
 
   void WriteInputFiles(SourceManager &SourceMgr, HeaderSearchOptions &HSOpts,
                        bool Modules);
   void WriteSourceManagerBlock(SourceManager &SourceMgr,
                                const Preprocessor &PP);
   void WritePreprocessor(const Preprocessor &PP, bool IsModule);
   void WriteHeaderSearch(const HeaderSearch &HS);
-  void WritePreprocessorDetail(PreprocessingRecord &PPRec);
+  void WritePreprocessorDetail(PreprocessingRecord &PPRec,
+                               uint64_t MacroOffsetsBase);
   void WriteSubmodules(Module *WritingModule);
 
   void WritePragmaDiagnosticMappings(const DiagnosticsEngine &Diag,
                                      bool isModule);
 
   unsigned TypeExtQualAbbrev = 0;
   unsigned TypeFunctionProtoAbbrev = 0;
   void WriteTypeAbbrevs();
public:
   serialization::IdentID getIdentifierRef(const IdentifierInfo *II);
 
   /// Get the unique number used to refer to the given macro.
   serialization::MacroID getMacroRef(MacroInfo *MI, const IdentifierInfo *Name);
 
   /// Determine the ID of an already-emitted macro.
   serialization::MacroID getMacroID(MacroInfo *MI);
 
-  uint64_t getMacroDirectivesOffset(const IdentifierInfo *Name);
+  uint32_t getMacroDirectivesOffset(const IdentifierInfo *Name);
 
   /// Emit a reference to a type.
   void AddTypeRef(QualType T, RecordDataImpl &Record);
 
   /// Force a type to be emitted and get its ID.
   serialization::TypeID GetOrCreateTypeID(QualType T);
 
   /// Determine the type ID of an already-emitted type.

clang/include/clang/Serialization/ModuleFile.h

public:
   unsigned LocalNumSLocEntries = 0;
 
   /// The base ID in the source manager's view of this module.
   int SLocEntryBaseID = 0;
 
   /// The base offset in the source manager's view of this module.
   unsigned SLocEntryBaseOffset = 0;
 
+  /// Base file offset for the offsets in SLocEntryOffsets. Real file offset
+  /// for the entry is SLocEntryOffsetsBase + SLocEntryOffsets[i].
+  uint64_t SLocEntryOffsetsBase = 0;
+
   /// Offsets for all of the source location entries in the
   /// AST file.
   const uint32_t *SLocEntryOffsets = nullptr;
 
   /// SLocEntries that we're going to preload.
   SmallVector<uint64_t, 4> PreloadSLocEntries;
 
   /// Remapping table for source locations in this module.
public:
 
   /// The cursor to the start of the preprocessor block, which stores
   /// all of the macro definitions.
   llvm::BitstreamCursor MacroCursor;
 
   /// The number of macros in this AST file.
   unsigned LocalNumMacros = 0;
 
+  /// Base file offset for the offsets in MacroOffsets. Real file offset for
+  /// the entry is MacroOffsetsBase + MacroOffsets[i].
+  uint64_t MacroOffsetsBase = 0;
+
   /// Offsets of macros in the preprocessor block.
   ///
   /// This array is indexed by the macro ID (-1), and provides
   /// the offset into the preprocessor block where macro definitions are
   /// stored.
   const uint32_t *MacroOffsets = nullptr;
 
   /// Base macro ID for macros local to this module.
public:
 
   // === Types ===
 
   /// The number of types in this AST file.
   unsigned LocalNumTypes = 0;
 
   /// Offset of each type within the bitstream, indexed by the
   /// type ID, or the representation of a Type*.
-  const uint32_t *TypeOffsets = nullptr;
+  const uint64_t *TypeOffsets = nullptr;
 
   /// Base type ID for types local to this module as represented in
   /// the global type ID space.
   serialization::TypeID BaseTypeIndex = 0;
 
   /// Remapping table for type IDs in this module.
   ContinuousRangeMap<uint32_t, int, 2> TypeRemap;
 

clang/lib/Serialization/ASTReader.cpp

auto ReadBuffer = [this](
     } else {
       Error("AST record has invalid code");
       return nullptr;
     }
   };
 
   ModuleFile *F = GlobalSLocEntryMap.find(-ID)->second;
   if (llvm::Error Err = F->SLocEntryCursor.JumpToBit(
+          F->SLocEntryOffsetsBase +
           F->SLocEntryOffsets[ID - F->SLocEntryBaseID])) {
     Error(std::move(Err));
     return true;
   }
 
   BitstreamCursor &SLocEntryCursor = F->SLocEntryCursor;
   unsigned BaseOffset = F->SLocEntryBaseOffset;
 
HeaderFileInfoTrait::ReadData(internal_key_ref key, const unsigned char *d,
   }
 
   // This HeaderFileInfo was externally loaded.
   HFI.External = true;
   HFI.IsValid = true;
   return HFI;
 }
 
-void ASTReader::addPendingMacro(IdentifierInfo *II,
-                                ModuleFile *M,
-                                uint64_t MacroDirectivesOffset) {
+void ASTReader::addPendingMacro(IdentifierInfo *II, ModuleFile *M,
+                                uint32_t MacroDirectivesOffset) {
   assert(NumCurrentElementsDeserializing > 0 &&"Missing deserialization guard");
   PendingMacroIDs[II].push_back(PendingMacroInfo(M, MacroDirectivesOffset));
 }
 
 void ASTReader::ReadDefinedMacros() {
   // Note that we are loading defined macros.
   Deserializing Macros(this);
 

 }
 
 void ASTReader::resolvePendingMacro(IdentifierInfo *II,
                                     const PendingMacroInfo &PMInfo) {
   ModuleFile &M = *PMInfo.M;
 
   BitstreamCursor &Cursor = M.MacroCursor;
   SavedStreamPosition SavedPosition(Cursor);
-  if (llvm::Error Err = Cursor.JumpToBit(PMInfo.MacroDirectivesOffset)) {
+  if (llvm::Error Err =
+          Cursor.JumpToBit(M.MacroOffsetsBase + PMInfo.MacroDirectivesOffset)) {
     Error(std::move(Err));
     return;
   }
 
   struct ModuleMacroRecord {
     SubmoduleID SubModID;
     MacroInfo *MI;
     SmallVector<SubmoduleID, 8> Overrides;
while (true) {
     default:  // Default behavior: ignore.
       break;
 
     case TYPE_OFFSET: {
       if (F.LocalNumTypes != 0) {
         Error("duplicate TYPE_OFFSET record in AST file");
         return Failure;
       }
-      F.TypeOffsets = (const uint32_t *)Blob.data();
+      F.TypeOffsets = reinterpret_cast<const uint64_t *>(Blob.data());
       F.LocalNumTypes = Record[0];
       unsigned LocalBaseTypeIndex = Record[1];
       F.BaseTypeIndex = getTotalNumTypes();
 
       if (F.LocalNumTypes > 0) {
         // Introduce the global -> local mapping for types within this module.
         GlobalTypeMap.insert(std::make_pair(getTotalNumTypes(), &F));
 
case FILE_SORTED_DECLS:
       F.FileSortedDecls = (const DeclID *)Blob.data();
       F.NumFileSortedDecls = Record[0];
       break;
 
     case SOURCE_LOCATION_OFFSETS: {
       F.SLocEntryOffsets = (const uint32_t *)Blob.data();
       F.LocalNumSLocEntries = Record[0];
       unsigned SLocSpaceSize = Record[1];
+      F.SLocEntryOffsetsBase = Record[2];
       std::tie(F.SLocEntryBaseID, F.SLocEntryBaseOffset) =
           SourceMgr.AllocateLoadedSLocEntries(F.LocalNumSLocEntries,
                                               SLocSpaceSize);
       if (!F.SLocEntryBaseID) {
         Error("ran out of source locations");
         break;
       }
       // Make our entry in the range map. BaseID is negative and growing, so
while (true) {
     case MACRO_OFFSET: {
       if (F.LocalNumMacros != 0) {
         Error("duplicate MACRO_OFFSET record in AST file");
         return Failure;
       }
       F.MacroOffsets = (const uint32_t *)Blob.data();
       F.LocalNumMacros = Record[0];
       unsigned LocalBaseMacroID = Record[1];
+      F.MacroOffsetsBase = Record[2];
       F.BaseMacroID = getTotalNumMacros();
 
       if (F.LocalNumMacros > 0) {
         // Introduce the global -> local mapping for macros within this module.
         GlobalMacroMap.insert(std::make_pair(getTotalNumMacros() + 1, &F));
 
         // Introduce the local -> global mapping for macros within this module.
         F.MacroRemap.insertOrReplace(
PreprocessedEntity *ASTReader::ReadPreprocessedEntity(unsigned Index) {
   const PPEntityOffset &PPOffs = M.PreprocessedEntityOffsets[LocalIndex];
 
   if (!PP.getPreprocessingRecord()) {
     Error("no preprocessing record");
     return nullptr;
   }
 
   SavedStreamPosition SavedPosition(M.PreprocessorDetailCursor);
-  if (llvm::Error Err =
-          M.PreprocessorDetailCursor.JumpToBit(PPOffs.BitOffset)) {
+  if (llvm::Error Err = M.PreprocessorDetailCursor.JumpToBit(
+          M.MacroOffsetsBase + PPOffs.BitOffset)) {
     Error(std::move(Err));
     return nullptr;
   }
 
   Expected<llvm::BitstreamEntry> MaybeEntry =
       M.PreprocessorDetailCursor.advance(BitstreamCursor::AF_DontPopBlockAtEnd);
   if (!MaybeEntry) {
     Error(MaybeEntry.takeError());
MacroInfo *ASTReader::getMacro(MacroID ID) {
 
   ID -= NUM_PREDEF_MACRO_IDS;
   if (!MacrosLoaded[ID]) {
     GlobalMacroMapType::iterator I
       = GlobalMacroMap.find(ID + NUM_PREDEF_MACRO_IDS);
     assert(I != GlobalMacroMap.end() && "Corrupted global macro map");
     ModuleFile *M = I->second;
     unsigned Index = ID - M->BaseMacroID;
-    MacrosLoaded[ID] = ReadMacroRecord(*M, M->MacroOffsets[Index]);
+    MacrosLoaded[ID] =
+        ReadMacroRecord(*M, M->MacroOffsetsBase + M->MacroOffsets[Index]);
 
     if (DeserializationListener)
       DeserializationListener->MacroRead(ID + NUM_PREDEF_MACRO_IDS,
                                          MacrosLoaded[ID]);
   }
 
   return MacrosLoaded[ID];
 }

clang/lib/Serialization/ASTReaderDecl.cpp

 ASTReader::RecordLocation
 ASTReader::DeclCursorForID(DeclID ID, SourceLocation &Loc) {
   GlobalDeclMapType::iterator I = GlobalDeclMap.find(ID);
   assert(I != GlobalDeclMap.end() && "Corrupted global declaration map");
   ModuleFile *M = I->second;
   const DeclOffset &DOffs =
       M->DeclOffsets[ID - M->BaseDeclID - NUM_PREDEF_DECL_IDS];
   Loc = TranslateSourceLocation(*M, DOffs.getLocation());
-  return RecordLocation(M, DOffs.BitOffset);
+  return RecordLocation(M, DOffs.getBitOffset());
 }
 
 ASTReader::RecordLocation ASTReader::getLocalBitOffset(uint64_t GlobalOffset) {
   auto I = GlobalBitOffsetsMap.find(GlobalOffset);
 
   assert(I != GlobalBitOffsetsMap.end() && "Corrupted global bit offsets map");
   return RecordLocation(I->second, GlobalOffset - I->second->GlobalBitOffset);
 }

clang/lib/Serialization/ASTWriter.cpp

void ASTWriter::WriteSourceManagerBlock(SourceManager &SourceMgr,
   unsigned SLocBufferBlobAbbrv = CreateSLocBufferBlobAbbrev(Stream, false);
   unsigned SLocBufferBlobCompressedAbbrv =
       CreateSLocBufferBlobAbbrev(Stream, true);
   unsigned SLocExpansionAbbrv = CreateSLocExpansionAbbrev(Stream);
 
   // Write out the source location entry table. We skip the first
   // entry, which is always the same dummy entry.
   std::vector<uint32_t> SLocEntryOffsets;
+  uint64_t SLocEntryOffsetsBase = Stream.GetCurrentBitNo();
   RecordData PreloadSLocs;
   SLocEntryOffsets.reserve(SourceMgr.local_sloc_entry_size() - 1);
   for (unsigned I = 1, N = SourceMgr.local_sloc_entry_size();
        I != N; ++I) {
     // Get this source location entry.
     const SrcMgr::SLocEntry *SLoc = &SourceMgr.getLocalSLocEntry(I);
     FileID FID = FileID::get(I);
     assert(&SourceMgr.getSLocEntry(FID) == SLoc);
 
     // Record the offset of this source-location entry.
-    SLocEntryOffsets.push_back(Stream.GetCurrentBitNo());
+    uint64_t Offset = Stream.GetCurrentBitNo() - SLocEntryOffsetsBase;
+    assert((Offset >> 32) == 0 && "SLocEntry offset too large");
+    SLocEntryOffsets.push_back(Offset);
 
     // Figure out which record code to use.
     unsigned Code;
     if (SLoc->isFile()) {
       const SrcMgr::ContentCache *Cache = SLoc->getFile().getContentCache();
       if (Cache->OrigEntry) {
         Code = SM_SLOC_FILE_ENTRY;
       } else
void ASTWriter::WriteSourceManagerBlock(SourceManager &SourceMgr,
   // Write the source-location offsets table into the AST block. This
   // table is used for lazily loading source-location information.
   using namespace llvm;
 
   auto Abbrev = std::make_shared<BitCodeAbbrev>();
   Abbrev->Add(BitCodeAbbrevOp(SOURCE_LOCATION_OFFSETS));
   Abbrev->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 16)); // # of slocs
   Abbrev->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 16)); // total size
+  Abbrev->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 32)); // base offset
   Abbrev->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::Blob)); // offsets
   unsigned SLocOffsetsAbbrev = Stream.EmitAbbrev(std::move(Abbrev));
   {
     RecordData::value_type Record[] = {
         SOURCE_LOCATION_OFFSETS, SLocEntryOffsets.size(),
-        SourceMgr.getNextLocalOffset() - 1 /* skip dummy */};
+        SourceMgr.getNextLocalOffset() - 1 /* skip dummy */,
+        SLocEntryOffsetsBase};
     Stream.EmitRecordWithBlob(SLocOffsetsAbbrev, Record,
                               bytes(SLocEntryOffsets));
   }
   // Write the source location entry preloads array, telling the AST
   // reader which source locations entries it should load eagerly.
   Stream.EmitRecord(SOURCE_LOCATION_PRELOADS, PreloadSLocs);
 
   // Write the line table. It depends on remapping working, so it must come
static bool shouldIgnoreMacro(MacroDirective *MD, bool IsModule,
   }
 
   return false;
 }
 
 /// Writes the block containing the serialized form of the
 /// preprocessor.
 void ASTWriter::WritePreprocessor(const Preprocessor &PP, bool IsModule) {
+  uint64_t MacroOffsetsBase = Stream.GetCurrentBitNo();
+
   PreprocessingRecord *PPRec = PP.getPreprocessingRecord();
   if (PPRec)
-    WritePreprocessorDetail(*PPRec);
+    WritePreprocessorDetail(*PPRec, MacroOffsetsBase);
 
   RecordData Record;
   RecordData ModuleMacroRecord;
 
   // If the preprocessor __COUNTER__ value has been bumped, remember it.
   if (PP.getCounterValue() != 0) {
     RecordData::value_type Record[] = {PP.getCounterValue()};
     Stream.EmitRecord(PP_COUNTER_VALUE, Record);
void ASTWriter::WritePreprocessor(const Preprocessor &PP, bool IsModule) {
   // Sort the set of macro definitions that need to be serialized by the
   // name of the macro, to provide a stable ordering.
   llvm::sort(MacroIdentifiers, llvm::deref<std::less<>>());
 
   // Emit the macro directives as a list and associate the offset with the
   // identifier they belong to.
   for (const IdentifierInfo *Name : MacroIdentifiers) {
     MacroDirective *MD = PP.getLocalMacroDirectiveHistory(Name);
-    auto StartOffset = Stream.GetCurrentBitNo();
+    uint64_t StartOffset = Stream.GetCurrentBitNo() - MacroOffsetsBase;
+    assert((StartOffset >> 32) == 0 && "Macro identifiers offset too large");
 
     // Emit the macro directives in reverse source order.
     for (; MD; MD = MD->getPrevious()) {
       // Once we hit an ignored macro, we're done: the rest of the chain
       // will all be ignored macros.
       if (shouldIgnoreMacro(MD, IsModule, PP))
         break;
 
for (unsigned I = 0, N = MacroInfosToEmit.size(); I != N; ++I) {
 
     if (ID < FirstMacroID) {
       assert(0 && "Loaded MacroInfo entered MacroInfosToEmit ?");
       continue;
     }
 
     // Record the local offset of this macro.
     unsigned Index = ID - FirstMacroID;
-    if (Index == MacroOffsets.size())
-      MacroOffsets.push_back(Stream.GetCurrentBitNo());
-    else {
-      if (Index > MacroOffsets.size())
-        MacroOffsets.resize(Index + 1);
+    if (Index >= MacroOffsets.size())
 
-      MacroOffsets[Index] = Stream.GetCurrentBitNo();
-    }
+    uint64_t Offset = Stream.GetCurrentBitNo() - MacroOffsetsBase;
+    assert((Offset >> 32) == 0 && "Macro offset too large");
+    MacroOffsets[Index] = Offset;
 
     AddIdentifierRef(Name, Record);
     AddSourceLocation(MI->getDefinitionLoc(), Record);
     AddSourceLocation(MI->getDefinitionEndLoc(), Record);
     Record.push_back(MI->isUsed());
     Record.push_back(MI->isUsedForHeaderGuard());
     unsigned Code;
     if (MI->isObjectLike()) {
void ASTWriter::WritePreprocessor(const Preprocessor &PP, bool IsModule) {
 
   // Write the offsets table for macro IDs.
   using namespace llvm;
 
   auto Abbrev = std::make_shared<BitCodeAbbrev>();
   Abbrev->Add(BitCodeAbbrevOp(MACRO_OFFSET));
   Abbrev->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::Fixed, 32)); // # of macros
   Abbrev->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::Fixed, 32)); // first ID
+  Abbrev->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::VBR, 32));   // base offset
   Abbrev->Add(BitCodeAbbrevOp(BitCodeAbbrevOp::Blob));
 
   unsigned MacroOffsetAbbrev = Stream.EmitAbbrev(std::move(Abbrev));
   {
     RecordData::value_type Record[] = {MACRO_OFFSET, MacroOffsets.size(),
-                                       FirstMacroID - NUM_PREDEF_MACRO_IDS};
+                                       FirstMacroID - NUM_PREDEF_MACRO_IDS,
+                                       MacroOffsetsBase};
     Stream.EmitRecordWithBlob(MacroOffsetAbbrev, Record, bytes(MacroOffsets));
   }
 }
 
-void ASTWriter::WritePreprocessorDetail(PreprocessingRecord &PPRec) {
+void ASTWriter::WritePreprocessorDetail(PreprocessingRecord &PPRec,
+                                        uint64_t MacroOffsetsBase) {
   if (PPRec.local_begin() == PPRec.local_end())
     return;
 
   SmallVector<PPEntityOffset, 64> PreprocessedEntityOffsets;
 
   // Enter the preprocessor block.
   Stream.EnterSubblock(PREPROCESSOR_DETAIL_BLOCK_ID, 3);
 
void ASTWriter::WritePreprocessorDetail(PreprocessingRecord &PPRec,
   unsigned NextPreprocessorEntityID = FirstPreprocessorEntityID;
   RecordData Record;
   for (PreprocessingRecord::iterator E = PPRec.local_begin(),
                                   EEnd = PPRec.local_end();
        E != EEnd;
        (void)++E, ++NumPreprocessingRecords, ++NextPreprocessorEntityID) {
     Record.clear();
 
+    uint64_t Offset = Stream.GetCurrentBitNo() - MacroOffsetsBase;
+    assert((Offset >> 32) == 0 && "Preprocessed entity offset too large");
     PreprocessedEntityOffsets.push_back(
-        PPEntityOffset((*E)->getSourceRange(), Stream.GetCurrentBitNo()));
+        PPEntityOffset((*E)->getSourceRange(), Offset));
 
     if (auto *MD = dyn_cast<MacroDefinitionRecord>(*E)) {
       // Record this macro definition's ID.
       MacroDefinitions[MD] = NextPreprocessorEntityID;
 
       AddIdentifierRef(MD->getName(), Record);
       Stream.EmitRecord(PPD_MACRO_DEFINITION, Record);
       continue;

 MacroID ASTWriter::getMacroID(MacroInfo *MI) {
   if (!MI || MI->isBuiltinMacro())
     return 0;
 
   assert(MacroIDs.find(MI) != MacroIDs.end() && "Macro not emitted!");
   return MacroIDs[MI];
 }
 
-uint64_t ASTWriter::getMacroDirectivesOffset(const IdentifierInfo *Name) {
+uint32_t ASTWriter::getMacroDirectivesOffset(const IdentifierInfo *Name) {
   return IdentMacroDirectivesOffsetMap.lookup(Name);
 }
 
 void ASTRecordWriter::AddSelectorRef(const Selector SelRef) {
   Record->push_back(Writer->getSelectorRef(SelRef));
 }
 
 SelectorID ASTWriter::getSelectorRef(Selector Sel) {

clang/lib/Serialization/ASTWriterDecl.cpp

void ASTWriter::WriteDecl(ASTContext &Context, Decl *D) {
 
   // Emit this declaration to the bitstream.
   uint64_t Offset = W.Emit(D);
 
   // Record the offset for this declaration
   SourceLocation Loc = D->getLocation();
   unsigned Index = ID - FirstDeclID;
   if (DeclOffsets.size() == Index)
-    DeclOffsets.push_back(DeclOffset(Loc, Offset));
+    DeclOffsets.emplace_back(Loc, Offset);
   else if (DeclOffsets.size() < Index) {
     // FIXME: Can/should this happen?
     DeclOffsets.resize(Index+1);
     DeclOffsets[Index].setLocation(Loc);
-    DeclOffsets[Index].BitOffset = Offset;
+    DeclOffsets[Index].setBitOffset(Offset);
   } else {
     llvm_unreachable("declarations should be emitted in ID order");
   }
 
   SourceManager &SM = Context.getSourceManager();
   if (Loc.isValid() && SM.isLocalSourceLocation(Loc))
     associateDeclWithFile(D, ID);