Author: Dmitry Polukhin
Date: 2026-06-01T10:02:26+01:00
New Revision: 63c29df8ecebb57bdcc04de652afc8d2d98650d7

URL: 
https://github.com/llvm/llvm-project/commit/63c29df8ecebb57bdcc04de652afc8d2d98650d7
DIFF: 
https://github.com/llvm/llvm-project/commit/63c29df8ecebb57bdcc04de652afc8d2d98650d7.diff

LOG: [Serialization] Fix assertion on re-deserialized friend template spec… 
(#200566)

…ialization in PCH (#198133)

A friend function-template specialization declared inside a class
template is serialized into a PCH. When the class template is later
instantiated while loading the PCH, the friend specialization can be
deserialized re-entrantly (VisitFriendDecl -> VisitFunctionDecl -> ...
-> VisitFunctionDecl for the same specialization) at the same time as
the canonical copy, producing two redeclarations of the same
specialization in the template's specialization set.

ASTDeclReader::VisitFunctionDecl asserted that this collision could only
happen when merging declarations from different modules. Since
38b3d87bd384, friend functions defined inside dependent class templates
are loaded eagerly, so the collision can now also occur within a single
PCH/AST file (non-modules build), tripping the assertion:

  Assertion failed: (Reader.getContext().getLangOpts().Modules &&
  "already deserialized this template specialization"), function
  VisitFunctionDecl

The merge that follows (mergeRedeclarable) already links the two
redeclarations correctly regardless of whether modules are enabled, so
the fix is to drop the modules-only assumption and let the merge run.

Fixes https://github.com/llvm/llvm-project/issues/198133

Added: 
    clang/test/PCH/friend-template-spec-redecl.cpp

Modified: 
    clang/lib/Serialization/ASTReaderDecl.cpp

Removed: 
    


################################################################################
diff  --git a/clang/lib/Serialization/ASTReaderDecl.cpp 
b/clang/lib/Serialization/ASTReaderDecl.cpp
index 6815a27537034..fb291a4b0f2c5 100644
--- a/clang/lib/Serialization/ASTReaderDecl.cpp
+++ b/clang/lib/Serialization/ASTReaderDecl.cpp
@@ -996,8 +996,6 @@ void ASTDeclReader::VisitFunctionDecl(FunctionDecl *FD) {
       if (InsertPos)
         CommonPtr->Specializations.InsertNode(FTInfo, InsertPos);
       else {
-        assert(Reader.getContext().getLangOpts().Modules &&
-               "already deserialized this template specialization");
         Existing = ExistingInfo->getFunction();
       }
     }

diff  --git a/clang/test/PCH/friend-template-spec-redecl.cpp 
b/clang/test/PCH/friend-template-spec-redecl.cpp
new file mode 100644
index 0000000000000..5ae5bbe22aa5e
--- /dev/null
+++ b/clang/test/PCH/friend-template-spec-redecl.cpp
@@ -0,0 +1,34 @@
+// Regression test for https://github.com/llvm/llvm-project/issues/198133
+//
+// A friend function-template specialization declared inside a class template
+// is serialized into a PCH.  When the class template is later instantiated
+// while loading the PCH, the friend specialization could be deserialized
+// re-entrantly (VisitFriendDecl -> VisitFunctionDecl -> ... -> 
VisitFunctionDecl
+// for the same specialization).  This used to trip the assertion
+//   "already deserialized this template specialization"
+// in ASTReaderDecl::VisitFunctionDecl for non-modules (PCH) builds.
+
+// RUN: %clang_cc1 -std=c++17 -x c++-header -emit-pch %s -o %t.pch
+// RUN: %clang_cc1 -std=c++17 -include-pch %t.pch %s -fsyntax-only -verify
+
+#ifndef HEADER
+#define HEADER
+
+template <bool = false> int get_extents(const int &);
+template <typename> struct BoundingBoxBase {
+  BoundingBoxBase(int) {}
+  friend int get_extents<>(const int &);
+};
+template <class> struct BoundingBox3Base {
+  BoundingBox3Base();
+};
+struct BoundingBoxf : BoundingBoxBase<int> {
+  BoundingBoxf(int points) : BoundingBoxBase(points) {}
+};
+
+#else
+
+// expected-no-diagnostics
+void f() { BoundingBox3Base<int> build_volume; }
+
+#endif


        
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to