[Serialization] Support loading template specializations lazily (llvm#119333)

Reland llvm#83237 --- (Original comments) Currently all the specializations of a template (including instantiation, specialization and partial specializations) will be loaded at once if we want to instantiate another instance for the template, or find instantiation for the template, or just want to complete the redecl chain. This means basically we need to load every specializations for the template once the template declaration got loaded. This is bad since when we load a specialization, we need to load all of its template arguments. Then we have to deserialize a lot of unnecessary declarations. For example, ``` // M.cppm export module M; export template <class T> class A {}; export class ShouldNotBeLoaded {}; export class Temp { A<ShouldNotBeLoaded> AS; }; // use.cpp import M; A<int> a; ``` We have a specialization ` A<ShouldNotBeLoaded>` in `M.cppm` and we instantiate the template `A` in `use.cpp`. Then we will deserialize `ShouldNotBeLoaded` surprisingly when compiling `use.cpp`. And this patch tries to avoid that. Given that the templates are heavily used in C++, this is a pain point for the performance. This patch adds MultiOnDiskHashTable for specializations in the ASTReader. Then we will only deserialize the specializations with the same template arguments. We made that by using ODRHash for the template arguments as the key of the hash table. To review this patch, I think `ASTReaderDecl::AddLazySpecializations` may be a good entry point.

Author: ChuanqiXu9

clang/include/clang/AST/DeclTemplate.h

@@ -735,6 +735,7 @@ class RedeclarableTemplateDecl : public TemplateDecl,
  }
 
  void anchor() override;
+
 protected:
  template <typename EntryType> struct SpecEntryTraits {
  using DeclType = EntryType;
@@ -775,13 +776,22 @@ class RedeclarableTemplateDecl : public TemplateDecl,
  return SpecIterator<EntryType>(isEnd ? Specs.end() : Specs.begin());
  }
 
- void loadLazySpecializationsImpl() const;
+ void loadLazySpecializationsImpl(bool OnlyPartial = false) const;
+
+ bool loadLazySpecializationsImpl(llvm::ArrayRef<TemplateArgument> Args,
+ TemplateParameterList *TPL = nullptr) const;
 
  template <class EntryType, typename ...ProfileArguments>
  typename SpecEntryTraits<EntryType>::DeclType*
  findSpecializationImpl(llvm::FoldingSetVector<EntryType> &Specs,
  void *&InsertPos, ProfileArguments &&...ProfileArgs);
 
+ template <class EntryType, typename... ProfileArguments>
+ typename SpecEntryTraits<EntryType>::DeclType *
+ findSpecializationLocally(llvm::FoldingSetVector<EntryType> &Specs,
+ void *&InsertPos,
+ ProfileArguments &&...ProfileArgs);
+
  template <class Derived, class EntryType>
  void addSpecializationImpl(llvm::FoldingSetVector<EntryType> &Specs,
  EntryType *Entry, void *InsertPos);
@@ -796,13 +806,6 @@ class RedeclarableTemplateDecl : public TemplateDecl,
  /// was explicitly specialized.
  llvm::PointerIntPair<RedeclarableTemplateDecl *, 1, bool>
  InstantiatedFromMember;
-
- /// If non-null, points to an array of specializations (including
- /// partial specializations) known only by their external declaration IDs.
- ///
- /// The first value in the array is the number of specializations/partial
- /// specializations that follow.
- GlobalDeclID *LazySpecializations = nullptr;
  };
 
  /// Pointer to the common data shared by all declarations of this
@@ -2283,7 +2286,7 @@ class ClassTemplateDecl : public RedeclarableTemplateDecl {
  friend class TemplateDeclInstantiator;
 
  /// Load any lazily-loaded specializations from the external source.
- void LoadLazySpecializations() const;
+ void LoadLazySpecializations(bool OnlyPartial = false) const;
 
  /// Get the underlying class declarations of the template.
  CXXRecordDecl *getTemplatedDecl() const {
@@ -3033,7 +3036,7 @@ class VarTemplateDecl : public RedeclarableTemplateDecl {
  friend class ASTDeclWriter;
 
  /// Load any lazily-loaded specializations from the external source.
- void LoadLazySpecializations() const;
+ void LoadLazySpecializations(bool OnlyPartial = false) const;
 
  /// Get the underlying variable declarations of the template.
  VarDecl *getTemplatedDecl() const {

clang/include/clang/AST/ExternalASTSource.h

@@ -152,6 +152,21 @@ class ExternalASTSource : public RefCountedBase<ExternalASTSource> {
  virtual bool
  FindExternalVisibleDeclsByName(const DeclContext *DC, DeclarationName Name);
 
+ /// Load all the external specializations for the Decl \param D if \param
+ /// OnlyPartial is false. Otherwise, load all the external **partial**
+ /// specializations for the \param D.
+ ///
+ /// Return true if any new specializations get loaded. Return false otherwise.
+ virtual bool LoadExternalSpecializations(const Decl *D, bool OnlyPartial);
+
+ /// Load all the specializations for the Decl \param D with the same template
+ /// args specified by \param TemplateArgs.
+ ///
+ /// Return true if any new specializations get loaded. Return false otherwise.
+ virtual bool
+ LoadExternalSpecializations(const Decl *D,
+ ArrayRef<TemplateArgument> TemplateArgs);
+
  /// Ensures that the table of all visible declarations inside this
  /// context is up to date.
  ///

clang/include/clang/Sema/MultiplexExternalSemaSource.h

@@ -97,6 +97,12 @@ class MultiplexExternalSemaSource : public ExternalSemaSource {
  bool FindExternalVisibleDeclsByName(const DeclContext *DC,
  DeclarationName Name) override;
 
+ bool LoadExternalSpecializations(const Decl *D, bool OnlyPartial) override;
+
+ bool
+ LoadExternalSpecializations(const Decl *D,
+ ArrayRef<TemplateArgument> TemplateArgs) override;
+
  /// Ensures that the table of all visible declarations inside this
  /// context is up to date.
  void completeVisibleDeclsMap(const DeclContext *DC) override;

clang/include/clang/Serialization/ASTBitCodes.h

@@ -733,6 +733,13 @@ enum ASTRecordTypes {
  /// Record code for Sema's vector of functions/blocks with effects to
  /// be verified.
  DECLS_WITH_EFFECTS_TO_VERIFY = 72,
+
+ /// Record code for updated specialization
+ UPDATE_SPECIALIZATION = 73,
+
+ CXX_ADDED_TEMPLATE_SPECIALIZATION = 74,
+
+ CXX_ADDED_TEMPLATE_PARTIAL_SPECIALIZATION = 75,
 };
 
 /// Record types used within a source manager block.
@@ -1502,6 +1509,12 @@ enum DeclCode {
  /// An ImplicitConceptSpecializationDecl record.
  DECL_IMPLICIT_CONCEPT_SPECIALIZATION,
 
+ // A decls specilization record.
+ DECL_SPECIALIZATIONS,
+
+ // A decls specilization record.
+ DECL_PARTIAL_SPECIALIZATIONS,
+
  DECL_LAST = DECL_IMPLICIT_CONCEPT_SPECIALIZATION
 };
 

clang/include/clang/Serialization/ASTReader.h

@@ -354,6 +354,9 @@ class ASTIdentifierLookupTrait;
 /// The on-disk hash table(s) used for DeclContext name lookup.
 struct DeclContextLookupTable;
 
+/// The on-disk hash table(s) used for specialization decls.
+struct LazySpecializationInfoLookupTable;
+
 } // namespace reader
 
 } // namespace serialization
@@ -632,20 +635,40 @@ class ASTReader
  llvm::DenseMap<const DeclContext *,
  serialization::reader::DeclContextLookupTable> Lookups;
 
+ using SpecLookupTableTy =
+ llvm::DenseMap<const Decl *,
+ serialization::reader::LazySpecializationInfoLookupTable>;
+ /// Map from decls to specialized decls.
+ SpecLookupTableTy SpecializationsLookups;
+ /// Split partial specialization from specialization to speed up lookups.
+ SpecLookupTableTy PartialSpecializationsLookups;
+
+ bool LoadExternalSpecializationsImpl(SpecLookupTableTy &SpecLookups,
+ const Decl *D);
+ bool LoadExternalSpecializationsImpl(SpecLookupTableTy &SpecLookups,
+ const Decl *D,
+ ArrayRef<TemplateArgument> TemplateArgs);
+
  // Updates for visible decls can occur for other contexts than just the
  // TU, and when we read those update records, the actual context may not
  // be available yet, so have this pending map using the ID as a key. It
- // will be realized when the context is actually loaded.
- struct PendingVisibleUpdate {
+ // will be realized when the data is actually loaded.
+ struct UpdateData {
  ModuleFile *Mod;
  const unsigned char *Data;
  };
- using DeclContextVisibleUpdates = SmallVector<PendingVisibleUpdate, 1>;
+ using DeclContextVisibleUpdates = SmallVector<UpdateData, 1>;
 
  /// Updates to the visible declarations of declaration contexts that
  /// haven't been loaded yet.
  llvm::DenseMap<GlobalDeclID, DeclContextVisibleUpdates> PendingVisibleUpdates;
 
+ using SpecializationsUpdate = SmallVector<UpdateData, 1>;
+ using SpecializationsUpdateMap =
+ llvm::DenseMap<GlobalDeclID, SpecializationsUpdate>;
+ SpecializationsUpdateMap PendingSpecializationsUpdates;
+ SpecializationsUpdateMap PendingPartialSpecializationsUpdates;
+
  /// The set of C++ or Objective-C classes that have forward
  /// declarations that have not yet been linked to their definitions.
  llvm::SmallPtrSet<Decl *, 4> PendingDefinitions;
@@ -678,6 +701,11 @@ class ASTReader
  llvm::BitstreamCursor &Cursor,
  uint64_t Offset, GlobalDeclID ID);
 
+ bool ReadSpecializations(ModuleFile &M, llvm::BitstreamCursor &Cursor,
+ uint64_t Offset, Decl *D, bool IsPartial);
+ void AddSpecializations(const Decl *D, const unsigned char *Data,
+ ModuleFile &M, bool IsPartial);
+
  /// A vector containing identifiers that have already been
  /// loaded.
  ///
@@ -1419,6 +1447,14 @@ class ASTReader
  const serialization::reader::DeclContextLookupTable *
  getLoadedLookupTables(DeclContext *Primary) const;
 
+ /// Get the loaded specializations lookup tables for \p D,
+ /// if any.
+ serialization::reader::LazySpecializationInfoLookupTable *
+ getLoadedSpecializationsLookupTables(const Decl *D, bool IsPartial);
+
+ /// If we have any unloaded specialization for \p D
+ bool haveUnloadedSpecializations(const Decl *D) const;
+
 private:
  struct ImportedModule {
  ModuleFile *Mod;
@@ -2076,6 +2112,12 @@ class ASTReader
  unsigned BlockID,
  uint64_t *StartOfBlockOffset = nullptr);
 
+ bool LoadExternalSpecializations(const Decl *D, bool OnlyPartial) override;
+
+ bool
+ LoadExternalSpecializations(const Decl *D,
+ ArrayRef<TemplateArgument> TemplateArgs) override;
+
  /// Finds all the visible declarations with a given name.
  /// The current implementation of this method just loads the entire
  /// lookup table as unmaterialized references.

clang/include/clang/Serialization/ASTWriter.h

@@ -423,6 +423,13 @@ class ASTWriter : public ASTDeserializationListener,
  /// Only meaningful for reduced BMI.
  DeclUpdateMap DeclUpdatesFromGMF;
 
+ /// Mapping from decl templates and its new specialization in the
+ /// current TU.
+ using SpecializationUpdateMap =
+ llvm::MapVector<const NamedDecl *, SmallVector<const Decl *>>;
+ SpecializationUpdateMap SpecializationsUpdates;
+ SpecializationUpdateMap PartialSpecializationsUpdates;
+
  using FirstLatestDeclMap = llvm::DenseMap<Decl *, Decl *>;
 
  /// Map of first declarations from a chained PCH that point to the
@@ -575,6 +582,12 @@ class ASTWriter : public ASTDeserializationListener,
 
  bool isLookupResultExternal(StoredDeclsList &Result, DeclContext *DC);
 
+ void GenerateSpecializationInfoLookupTable(
+ const NamedDecl *D, llvm::SmallVectorImpl<const Decl *> &Specializations,
+ llvm::SmallVectorImpl<char> &LookupTable, bool IsPartial);
+ uint64_t WriteSpecializationInfoLookupTable(
+ const NamedDecl *D, llvm::SmallVectorImpl<const Decl *> &Specializations,
+ bool IsPartial);
  void GenerateNameLookupTable(ASTContext &Context, const DeclContext *DC,
  llvm::SmallVectorImpl<char> &LookupTable);
  uint64_t WriteDeclContextLexicalBlock(ASTContext &Context,
@@ -590,6 +603,7 @@ class ASTWriter : public ASTDeserializationListener,
  void WriteDeclAndTypes(ASTContext &Context);
  void PrepareWritingSpecialDecls(Sema &SemaRef);
  void WriteSpecialDeclRecords(Sema &SemaRef);
+ void WriteSpecializationsUpdates(bool IsPartial);
  void WriteDeclUpdatesBlocks(ASTContext &Context,
  RecordDataImpl &OffsetsRecord);
  void WriteDeclContextVisibleUpdate(ASTContext &Context,
@@ -619,6 +633,9 @@ class ASTWriter : public ASTDeserializationListener,
  unsigned DeclEnumAbbrev = 0;
  unsigned DeclObjCIvarAbbrev = 0;
  unsigned DeclCXXMethodAbbrev = 0;
+ unsigned DeclSpecializationsAbbrev = 0;
+ unsigned DeclPartialSpecializationsAbbrev = 0;
+
  unsigned DeclDependentNonTemplateCXXMethodAbbrev = 0;
  unsigned DeclTemplateCXXMethodAbbrev = 0;
  unsigned DeclMemberSpecializedCXXMethodAbbrev = 0;