https://github.com/pdschbrt created 
https://github.com/llvm/llvm-project/pull/226942

CodeGenModule::DeferredAnnotations was keyed by StringRef, but not every 
mangled name passed to GetOrCreateLLVMFunction outlives the call. 
CodeGenVTables::maybeEmitThunk mangles the thunk name into a stack-local 
SmallString and hands it to GetAddrOfThunk, so for an annotated virtual 
function the map retained a reference into a frame that was gone by the time 
EmitGlobalAnnotations looked the key up.

ASan reports this as a stack-use-after-return in EmitGlobalAnnotations, with 
the freed frame being maybeEmitThunk's 'Name'. The user-visible effect is that 
annotations silently disappear from the this-adjusting thunks once the dead 
frame has been reused.

Make the key own its storage, using StringMap<unsigned> as the MapVector map 
type so lookups still hash a StringRef without allocating.

>From 4d3952b9313cec51867de92f32b354ab5c8f6eec Mon Sep 17 00:00:00 2001
From: Philipp Schubert <[email protected]>
Date: Mon, 28 Sep 2026 12:04:53 +0200
Subject: [PATCH] [clang][CodeGen] Fix stack-use-after-return in deferred
 annotations

CodeGenModule::DeferredAnnotations was keyed by StringRef, but not every
mangled name passed to GetOrCreateLLVMFunction outlives the call.
CodeGenVTables::maybeEmitThunk mangles the thunk name into a stack-local
SmallString and hands it to GetAddrOfThunk, so for an annotated virtual
function the map retained a reference into a frame that was gone by the
time EmitGlobalAnnotations looked the key up.

ASan reports this as a stack-use-after-return in EmitGlobalAnnotations,
with the freed frame being maybeEmitThunk's 'Name'. The user-visible
effect is that annotations silently disappear from the this-adjusting
thunks once the dead frame has been reused.

Make the key own its storage, using StringMap<unsigned> as the MapVector
map type so lookups still hash a StringRef without allocating.
---
 clang/docs/ReleaseNotes.md                    |  4 ++
 clang/lib/CodeGen/CodeGenModule.cpp           |  4 +-
 clang/lib/CodeGen/CodeGenModule.h             |  5 ++-
 clang/test/CodeGenCXX/attr-annotate-thunk.cpp | 40 +++++++++++++++++++
 4 files changed, 50 insertions(+), 3 deletions(-)
 create mode 100644 clang/test/CodeGenCXX/attr-annotate-thunk.cpp

diff --git a/clang/docs/ReleaseNotes.md b/clang/docs/ReleaseNotes.md
index 3c6acf353f93f..adc941e287607 100644
--- a/clang/docs/ReleaseNotes.md
+++ b/clang/docs/ReleaseNotes.md
@@ -802,6 +802,10 @@ features cannot lower the translation-unit ABI level;
 - Fixed an assertion failure when a method or function definition follows an
   Objective-C `@implementation` that was ended by a nested `@interface`,
   `@protocol` or `@implementation` before its `@end`. (#GH209503)
+- Fixed `annotate` attributes being dropped from the `this`-adjusting thunks of
+  annotated virtual functions. The deferred-annotation map kept a `StringRef`
+  into the caller's stack buffer, so the lookup at the end of the translation
+  unit read a dangling key.
 
 ### OpenACC Specific Changes
 
diff --git a/clang/lib/CodeGen/CodeGenModule.cpp 
b/clang/lib/CodeGen/CodeGenModule.cpp
index 7274a8588670f..277764fa300c3 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -4863,7 +4863,7 @@ void CodeGenModule::EmitGlobal(GlobalDecl GD) {
     if (FD->hasAttr<AnnotateAttr>()) {
       StringRef MangledName = getMangledName(GD);
       if (GetGlobalValue(MangledName))
-        DeferredAnnotations[MangledName] = FD;
+        DeferredAnnotations[MangledName.str()] = FD;
     }
 
     // Forward declarations are emitted lazily on first use.
@@ -5784,7 +5784,7 @@ llvm::Constant *CodeGenModule::GetOrCreateLLVMFunction(
   // Store the declaration associated with this function so it is potentially
   // updated by further declarations or definitions and emitted at the end.
   if (D && D->hasAttr<AnnotateAttr>())
-    DeferredAnnotations[MangledName] = cast<ValueDecl>(D);
+    DeferredAnnotations[MangledName.str()] = cast<ValueDecl>(D);
 
   // If we already created a function with the same mangled name (but different
   // type) before, take its name and add it to the list of functions to be
diff --git a/clang/lib/CodeGen/CodeGenModule.h 
b/clang/lib/CodeGen/CodeGenModule.h
index 9e3f073c20f4a..28e9bcf9e866b 100644
--- a/clang/lib/CodeGen/CodeGenModule.h
+++ b/clang/lib/CodeGen/CodeGenModule.h
@@ -506,7 +506,10 @@ class CodeGenModule : public CodeGenTypeCache {
 
   // Store deferred function annotations so they can be emitted at the end with
   // most up to date ValueDecl that will have all the inherited annotations.
-  llvm::MapVector<StringRef, const ValueDecl *> DeferredAnnotations;
+  // The key owns its storage: not every mangled name handed to
+  // GetOrCreateLLVMFunction outlives the call.
+  llvm::MapVector<std::string, const ValueDecl *, llvm::StringMap<unsigned>>
+      DeferredAnnotations;
 
   /// Map used to get unique annotation strings.
   llvm::StringMap<llvm::Constant*> AnnotationStrings;
diff --git a/clang/test/CodeGenCXX/attr-annotate-thunk.cpp 
b/clang/test/CodeGenCXX/attr-annotate-thunk.cpp
new file mode 100644
index 0000000000000..5aa57877c1719
--- /dev/null
+++ b/clang/test/CodeGenCXX/attr-annotate-thunk.cpp
@@ -0,0 +1,40 @@
+// RUN: %clang_cc1 %s -emit-llvm -triple x86_64-unknown-linux-gnu -o - | 
FileCheck %s
+
+// Annotations on a virtual function are deferred to the end of the TU, keyed
+// by mangled name. For a this-adjusting thunk that name is mangled into a
+// stack buffer in CodeGenVTables::maybeEmitThunk, so the deferred-annotation
+// map must own a copy of the key instead of referencing the caller's storage.
+
+struct A {
+  virtual void f();
+  virtual ~A();
+};
+
+struct B {
+  virtual void g();
+  virtual ~B();
+};
+
+struct C : A, B {
+  void f() override;
+  __attribute__((annotate("annotated_method"))) void g() override;
+  __attribute__((annotate("annotated_dtor"))) ~C() override;
+};
+
+void C::f() {}
+void C::g() {}
+C::~C() {}
+
+// Each annotation is recorded for the function itself and for the thunk that
+// adjusts `this` to the B subobject.
+
+// CHECK: @[[METHOD:.*]] = private unnamed_addr constant [17 x i8] 
c"annotated_method\00", section "llvm.metadata"
+// CHECK: @[[DTOR:.*]] = private unnamed_addr constant [15 x i8] 
c"annotated_dtor\00", section "llvm.metadata"
+// CHECK: @llvm.global.annotations = appending global [7 x { ptr, ptr, ptr, 
i32, ptr }] [
+// CHECK-SAME: { ptr @_ZN1C1gEv, ptr @[[METHOD]],
+// CHECK-SAME: { ptr @_ZThn8_N1C1gEv, ptr @[[METHOD]],
+// CHECK-SAME: { ptr @_ZN1CD2Ev, ptr @[[DTOR]],
+// CHECK-SAME: { ptr @_ZN1CD1Ev, ptr @[[DTOR]],
+// CHECK-SAME: { ptr @_ZThn8_N1CD1Ev, ptr @[[DTOR]],
+// CHECK-SAME: { ptr @_ZN1CD0Ev, ptr @[[DTOR]],
+// CHECK-SAME: { ptr @_ZThn8_N1CD0Ev, ptr @[[DTOR]],

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

Reply via email to