[Modules] Delay deserialization of preferred_name attribute at r…

clang/include/clang/AST/Attr.h

@@ -60,6 +60,8 @@ class Attr : public AttributeCommonInfo {
   unsigned IsLateParsed : 1;
   LLVM_PREFERRED_TYPE(bool)
   unsigned InheritEvenIfAlreadyPresent : 1;
+  LLVM_PREFERRED_TYPE(bool)
+  unsigned DeferDeserialization : 1;
 
   void *operator new(size_t bytes) noexcept {
     llvm_unreachable("Attrs cannot be allocated with regular 'new'.");
@@ -80,10 +82,11 @@ class Attr : public AttributeCommonInfo {
 
 protected:
   Attr(ASTContext &Context, const AttributeCommonInfo &CommonInfo,
-       attr::Kind AK, bool IsLateParsed)
+       attr::Kind AK, bool IsLateParsed, bool DeferDeserialization = false)
       : AttributeCommonInfo(CommonInfo), AttrKind(AK), Inherited(false),
         IsPackExpansion(false), Implicit(false), IsLateParsed(IsLateParsed),
-        InheritEvenIfAlreadyPresent(false) {}
+        InheritEvenIfAlreadyPresent(false),
+        DeferDeserialization(DeferDeserialization) {}
 
 public:
   attr::Kind getKind() const { return static_cast<attr::Kind>(AttrKind); }
@@ -105,6 +108,8 @@ class Attr : public AttributeCommonInfo {
   void setPackExpansion(bool PE) { IsPackExpansion = PE; }
   bool isPackExpansion() const { return IsPackExpansion; }
 
+  bool shouldDeferDeserialization() const { return DeferDeserialization; }
+
   // Clone this attribute.
   Attr *clone(ASTContext &C) const;
 
@@ -146,8 +151,9 @@ class InheritableAttr : public Attr {
 protected:
   InheritableAttr(ASTContext &Context, const AttributeCommonInfo &CommonInfo,
                   attr::Kind AK, bool IsLateParsed,
-                  bool InheritEvenIfAlreadyPresent)
-      : Attr(Context, CommonInfo, AK, IsLateParsed) {
+                  bool InheritEvenIfAlreadyPresent,
+                  bool DeferDeserialization = false)
+      : Attr(Context, CommonInfo, AK, IsLateParsed, DeferDeserialization) {
     this->InheritEvenIfAlreadyPresent = InheritEvenIfAlreadyPresent;
   }
 

clang/include/clang/Basic/Attr.td

@@ -713,6 +713,12 @@ class Attr {
   // attribute may be documented under multiple categories, more than one
   // Documentation entry may be listed.
   list<Documentation> Documentation;
+  // Set to true if deserialization of this attribute must be deferred until 
+  // the parent Decl is fully deserialized (during header module file
+  // deserialization). E.g., this is the case for the preferred_name attribute,
+  // since its type deserialization depends on its target Decl type.
+  // (See https://github.com/llvm/llvm-project/issues/56490 for details).
+  bit DeferDeserialization = 0;
 }
 
 /// Used to define a set of mutually exclusive attributes.
@@ -3240,6 +3246,11 @@ def PreferredName : InheritableAttr {
   let InheritEvenIfAlreadyPresent = 1;
   let MeaningfulToClassTemplateDefinition = 1;
   let TemplateDependent = 1;
+  // Type of this attribute depends on the target Decl type.
+  // Therefore, its deserialization must be deferred until
+  // deserialization of the target Decl is complete
+  // (for header modules).
+  let DeferDeserialization = 1;
 }
 
 def PreserveMost : DeclOrTypeAttr {

clang/include/clang/Serialization/ASTReader.h

@@ -1205,6 +1205,24 @@ class ASTReader
   /// been completed.
   std::deque<PendingDeclContextInfo> PendingDeclContextInfos;
 
+  /// Deserialization of some attributes must be deferred since they refer
+  /// to themselves in their type (e.g., preferred_name attribute refers to the
+  /// typedef that refers back to the template specialization of the template
+  /// that the attribute is attached to).
+  /// More attributes that store TypeSourceInfo might be potentially affected,
+  /// see https://github.com/llvm/llvm-project/issues/56490 for details.
+  struct DeferredAttribute {
+    // Index of the deferred attribute in the Record of the TargetedDecl.
+    uint64_t RecordIdx;
+    // Decl to attach a deferred attribute to.
+    Decl *TargetedDecl;
+  };
+
+  /// The collection of Decls that have been loaded but some of their attributes
+  /// have been deferred, paired with the index inside the record pointing
+  /// at the skipped attribute.
+  SmallVector<DeferredAttribute> PendingDeferredAttributes;
+
   template <typename DeclTy>
   using DuplicateObjCDecls = std::pair<DeclTy *, DeclTy *>;
 
@@ -1551,6 +1569,7 @@ class ASTReader
   void loadPendingDeclChain(Decl *D, uint64_t LocalOffset);
   void loadObjCCategories(GlobalDeclID ID, ObjCInterfaceDecl *D,
                           unsigned PreviousGeneration = 0);
+  void loadDeferredAttribute(const DeferredAttribute &DA);
 
   RecordLocation getLocalBitOffset(uint64_t GlobalOffset);
   uint64_t getGlobalBitOffset(ModuleFile &M, uint64_t LocalOffset);

clang/include/clang/Serialization/ASTRecordReader.h

@@ -83,6 +83,12 @@ class ASTRecordReader
   /// Returns the current value in this record, without advancing.
   uint64_t peekInt() { return Record[Idx]; }
 
+  /// Returns the next N values in this record, without advancing.
+  uint64_t peekInts(unsigned N) { return Record[Idx + N]; }
+
+  /// Skips the current value.
+  void skipInt() { Idx += 1; }
+
   /// Skips the specified number of values.
   void skipInts(unsigned N) { Idx += N; }
 
@@ -335,7 +341,12 @@ class ASTRecordReader
   Attr *readAttr();
 
   /// Reads attributes from the current stream position, advancing Idx.
-  void readAttributes(AttrVec &Attrs);
+  /// For some attributes (where type depends on itself recursively), defer
+  /// reading the attribute until the type has been read.
+  void readAttributes(AttrVec &Attrs, Decl *D = nullptr);
+
+  /// Reads one attribute from the current stream position, advancing Idx.
+  Attr *readOrDeferAttrFor(Decl *D);
 
   /// Read an BTFTypeTagAttr object.
   BTFTypeTagAttr *readBTFTypeTagAttr() {

clang/lib/Serialization/ASTReader.cpp

@@ -10079,6 +10079,11 @@ void ASTReader::finishPendingActions() {
     }
     PendingDeducedVarTypes.clear();
 
+    // Load the delayed preferred name attributes.
+    for (unsigned I = 0; I != PendingDeferredAttributes.size(); ++I)
+      loadDeferredAttribute(PendingDeferredAttributes[I]);
+    PendingDeferredAttributes.clear();
+
     // For each decl chain that we wanted to complete while deserializing, mark
     // it as "still needs to be completed".
     for (unsigned I = 0; I != PendingIncompleteDeclChains.size(); ++I) {

clang/lib/Serialization/ASTReaderDecl.cpp

@@ -612,7 +612,7 @@ void ASTDeclReader::VisitDecl(Decl *D) {
 
   if (HasAttrs) {
     AttrVec Attrs;
-    Record.readAttributes(Attrs);
+    Record.readAttributes(Attrs, D);
     // Avoid calling setAttrs() directly because it uses Decl::getASTContext()
     // internally which is unsafe during derialization.
     D->setAttrsImpl(Attrs, Reader.getContext());
@@ -3088,6 +3088,8 @@ class AttrReader {
     return Reader.readInt();
   }
 
+  uint64_t peekInts(unsigned N) { return Reader.peekInts(N); }
+
   bool readBool() { return Reader.readBool(); }
 
   SourceRange readSourceRange() {
@@ -3118,18 +3120,29 @@ class AttrReader {
     return Reader.readVersionTuple();
   }
 
+  void skipInt() { Reader.skipInts(1); }
+
+  void skipInts(unsigned N) { Reader.skipInts(N); }
+
+  unsigned getCurrentIdx() { return Reader.getIdx(); }
+
   OMPTraitInfo *readOMPTraitInfo() { return Reader.readOMPTraitInfo(); }
 
   template <typename T> T *readDeclAs() { return Reader.readDeclAs<T>(); }
 };
 }
 
+/// Reads one attribute from the current stream position, advancing Idx.
 Attr *ASTRecordReader::readAttr() {
   AttrReader Record(*this);
   auto V = Record.readInt();
   if (!V)
     return nullptr;
 
+  // Read and ignore the skip count, since attribute deserialization is not
+  // deferred on this pass.
+  Record.skipInt();
+
   Attr *New = nullptr;
   // Kind is stored as a 1-based integer because 0 is used to indicate a null
   // Attr pointer.
@@ -3159,13 +3172,28 @@ Attr *ASTRecordReader::readAttr() {
   return New;
 }
 
-/// Reads attributes from the current stream position.
-void ASTRecordReader::readAttributes(AttrVec &Attrs) {
+/// Reads attributes from the current stream position, advancing Idx.
+/// For some attributes (where type depends on itself recursively), defer
+/// reading the attribute until the type has been read.
+void ASTRecordReader::readAttributes(AttrVec &Attrs, Decl *D) {
   for (unsigned I = 0, E = readInt(); I != E; ++I)
-    if (auto *A = readAttr())
+    if (auto *A = readOrDeferAttrFor(D))
       Attrs.push_back(A);
 }
 
+/// Reads one attribute from the current stream position, advancing Idx.
+/// For some attributes (where type depends on itself recursively), defer
+/// reading the attribute until the type has been read.
+Attr *ASTRecordReader::readOrDeferAttrFor(Decl *D) {
+  AttrReader Record(*this);
+  unsigned SkipCount = Record.peekInts(1);
+  if (!SkipCount)
+    return readAttr();
+  Reader->PendingDeferredAttributes.push_back({Record.getCurrentIdx(), D});
+  Record.skipInts(SkipCount);
+  return nullptr;
+}
+
 //===----------------------------------------------------------------------===//
 // ASTReader Implementation
 //===----------------------------------------------------------------------===//
@@ -4424,6 +4452,49 @@ void ASTReader::loadPendingDeclChain(Decl *FirstLocal, uint64_t LocalOffset) {
   ASTDeclReader::attachLatestDecl(CanonDecl, MostRecent);
 }
 
+void ASTReader::loadDeferredAttribute(const DeferredAttribute &DA) {
+  Decl *D = DA.TargetedDecl;
+  ModuleFile *M = getOwningModuleFile(D);
+
+  unsigned LocalDeclIndex = D->getGlobalID().getLocalDeclIndex();
+  const DeclOffset &DOffs = M->DeclOffsets[LocalDeclIndex];
+  RecordLocation Loc(M, DOffs.getBitOffset(M->DeclsBlockStartOffset));
+
+  llvm::BitstreamCursor &Cursor = Loc.F->DeclsCursor;
+  SavedStreamPosition SavedPosition(Cursor);
+  if (llvm::Error Err = Cursor.JumpToBit(Loc.Offset)) {
+    Error(std::move(Err));
+  }
+
+  Expected<unsigned> MaybeCode = Cursor.ReadCode();
+  if (!MaybeCode) {
+    llvm::report_fatal_error(
+        Twine("ASTReader::loadPreferredNameAttribute failed reading code: ") +
+        toString(MaybeCode.takeError()));
+  }
+  unsigned Code = MaybeCode.get();
+
+  ASTRecordReader Record(*this, *Loc.F);
+  Expected<unsigned> MaybeRecCode = Record.readRecord(Cursor, Code);
+  if (!MaybeRecCode) {
+    llvm::report_fatal_error(
+        Twine(
+            "ASTReader::loadPreferredNameAttribute failed reading rec code: ") +
+        toString(MaybeCode.takeError()));
+  }
+  unsigned RecCode = MaybeRecCode.get();
+  if (RecCode < DECL_TYPEDEF || RecCode > DECL_LAST) {
+    llvm::report_fatal_error(
+        Twine("ASTReader::loadPreferredNameAttribute failed reading rec code: "
+              "expected valid DeclCode") +
+        toString(MaybeCode.takeError()));
+  }
+
+  Record.skipInts(DA.RecordIdx);
+  Attr *A = Record.readAttr();
+  getContext().getDeclAttrs(D).push_back(A);
+}
+
 namespace {
 
   /// Given an ObjC interface, goes through the modules and links to the

clang/lib/Serialization/ASTWriter.cpp

@@ -37,6 +37,7 @@
 #include "clang/AST/Type.h"
 #include "clang/AST/TypeLoc.h"
 #include "clang/AST/TypeLocVisitor.h"
+#include "clang/Basic/AttrKinds.h"
 #include "clang/Basic/Diagnostic.h"
 #include "clang/Basic/DiagnosticOptions.h"
 #include "clang/Basic/FileEntry.h"
@@ -4906,15 +4907,14 @@ void ASTWriter::WriteModuleFileExtension(Sema &SemaRef,
 
 void ASTRecordWriter::AddAttr(const Attr *A) {
   auto &Record = *this;
-  // FIXME: Clang can't handle the serialization/deserialization of
-  // preferred_name properly now. See
-  // https://github.com/llvm/llvm-project/issues/56490 for example.
-  if (!A || (isa<PreferredNameAttr>(A) &&
-             Writer->isWritingStdCXXNamedModules()))
+  if (!A)
     return Record.push_back(0);
 
   Record.push_back(A->getKind() + 1); // FIXME: stable encoding, target attrs
 
+  auto SkipIdx = Record.size();
+  // Add placeholder for the size of deferred attribute.
+  Record.push_back(0);
   Record.AddIdentifierRef(A->getAttrName());
   Record.AddIdentifierRef(A->getScopeName());
   Record.AddSourceRange(A->getRange());
@@ -4925,6 +4925,12 @@ void ASTRecordWriter::AddAttr(const Attr *A) {
   Record.push_back(A->isRegularKeywordAttribute());
 
 #include "clang/Serialization/AttrPCHWrite.inc"
+
+  if (A->shouldDeferDeserialization()) {
+    // Record the actual size of deferred attribute (+ 1 to count the attribute
+    // kind).
+    Record[SkipIdx] = Record.size() - SkipIdx + 1;
+  }
 }
 
 /// Emit the list of attributes to the specified record.

clang/test/Modules/preferred_name.cppm

@@ -53,10 +53,16 @@ import A;
 export using ::foo_templ;
 
 //--- Use1.cpp
-import A;         // expected-[email protected]:8 {{attribute declaration must precede definition}}
-#include "foo.h"  // [email protected]:9 {{previous definition is here}}
-
+// expected-no-diagnostics
+import A;
+#include "foo.h"
 //--- Use2.cpp
 // expected-no-diagnostics
 #include "foo.h"
 import A;
+
+//--- Use3.cpp
+#include "foo.h"
+import A;
+foo test;
+int size = test.size(); // expected-error {{no member named 'size' in 'foo'}}

clang/utils/TableGen/ClangAttrEmitter.cpp

@@ -3043,6 +3043,10 @@ static void emitAttributes(const RecordKeeper &Records, raw_ostream &OS,
            << (R.getValueAsBit("InheritEvenIfAlreadyPresent") ? "true"
                                                               : "false");
       }
+      if (R.getValueAsBit("DeferDeserialization")) {
+        OS << ", "
+           << "/*DeferDeserialization=*/true";
+      }
       OS << ")\n";
 
       for (auto const &ai : Args) {