[clang] Correct Microsoft mangling of lifetime extended temporary objects.

[tahonermann]

Lifetime extended temporary objects that are bound to references with static storage duration may have external linkage and therefore require mangled symbol names. Clang uses an extension of the Microsoft ABI to give these symbols an implicit name of '$RT' followed by a discriminator and then mangles them similarly to the variable they are bound to. Clang's mangling scheme differs from the one used by MSVC.

Previously, the $RT<discriminator> portion of the name was not registered as a back reference candidate and this resulted in incorrect back references for enclosing class and/or namespace scopes that might be referenced in the type of the object.

This is an ABI change and has the potential to cause backward compatibility issues with previous Clang releases. Since MSVC uses a different mangling scheme, this change does not affect compatibility with MSVC.

This fixes one of the name mangling concerns reported in #85423.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Tom Honermann (tahonermann)

Changes

Lifetime extended temporary objects that are bound to references with static storage duration may have external linkage and therefore require mangled symbol names. Clang uses an extension of the Microsoft ABI to give these symbols an implicit name of '$RT' followed by a discriminator and then mangles them similarly to the variable they are bound to. Clang's mangling scheme differs from the one used by MSVC.

Previously, the $RT<discriminator> portion of the name was not registered as a back reference candidate and this resulted in incorrect back references for enclosing class and/or namespace scopes that might be referenced in the type of the object.

This is an ABI change and has the potential to cause backward compatibility issues with previous Clang releases. Since MSVC uses a different mangling scheme, this change does not affect compatibility with MSVC.

This fixes one of the name mangling concerns reported in #85423.

---

Full diff: https://github.com/llvm/llvm-project/pull/85529.diff

3 Files Affected:

• (modified) clang/docs/ReleaseNotes.rst (+6)
• (modified) clang/lib/AST/MicrosoftMangle.cpp (+2-1)
• (modified) clang/test/CodeGenCXX/mangle-ms-back-references.cpp (+13)

diff --git a/clang/docs/ReleaseNotes.rst b/clang/docs/ReleaseNotes.rst
index 1ae35e6881d2f8..f0da6ca497b575 100644
--- a/clang/docs/ReleaseNotes.rst
+++ b/clang/docs/ReleaseNotes.rst
@@ -54,6 +54,12 @@ ABI Changes in This Version
   inline member function that contains a static local variable with a dynamic
   initializer is declared with ``__declspec(dllimport)``. (#GH83616).
 
+- Fixed Microsoft name mangling of lifetime extended temporary objects. This
+  change corrects missing back reference registrations that could result in
+  incorrect back reference indexes and suprising demangled name results. Since
+  MSVC uses a different mangling for these objects, compatibility is not affected.
+  (#GH85423).
+
 AST Dumping Potentially Breaking Changes
 ----------------------------------------
 
diff --git a/clang/lib/AST/MicrosoftMangle.cpp b/clang/lib/AST/MicrosoftMangle.cpp
index aa26bb7ed46f48..cf9c2093a8f6a1 100644
--- a/clang/lib/AST/MicrosoftMangle.cpp
+++ b/clang/lib/AST/MicrosoftMangle.cpp
@@ -3911,7 +3911,8 @@ void MicrosoftMangleContextImpl::mangleReferenceTemporary(
   msvc_hashing_ostream MHO(Out);
   MicrosoftCXXNameMangler Mangler(*this, MHO);
 
-  Mangler.getStream() << "?$RT" << ManglingNumber << '@';
+  Mangler.getStream() << "?";
+  Mangler.mangleSourceName("$RT" + llvm::utostr(ManglingNumber));
   Mangler.mangle(VD, "");
 }
 
diff --git a/clang/test/CodeGenCXX/mangle-ms-back-references.cpp b/clang/test/CodeGenCXX/mangle-ms-back-references.cpp
index b27a9c5acacb77..8707bff9534070 100644
--- a/clang/test/CodeGenCXX/mangle-ms-back-references.cpp
+++ b/clang/test/CodeGenCXX/mangle-ms-back-references.cpp
@@ -1,5 +1,18 @@
 // RUN: %clang_cc1 -fms-extensions -fblocks -emit-llvm %s -o - -triple=i386-pc-win32 | FileCheck %s
 
+namespace NS {
+// The name "RT1" for the name of the class below has been specifically
+// chosen to ensure that back reference lookup does not match against the
+// implicitly generated "$RT1" name of the reference temporary symbol.
+struct RT1 {
+  static const RT1& singleton;
+  int i;
+};
+const RT1& RT1::singleton = RT1{1};
+}
+// CHECK: "?$RT1@singleton@RT1@NS@@2ABU23@B"
+// CHECK: "?singleton@RT1@NS@@2ABU12@B"
+
 void f1(const char* a, const char* b) {}
 // CHECK: "?f1@@YAXPBD0@Z"
 

[tahonermann]

Commit e77de75 claims that Microsoft doesn't have a mangling for these objects and that the $RT mangling is a Clang invention. The claim that MSVC doesn't implement a mangling for these doesn't seem right though. Consider https://godbolt.org/z/n715zPbc6. It appears that MSVC (back to at least v19.14) uses a $S<discriminator> mangling for these; like that used for other unnamed objects like for structured bindings, anonymous unions, and (non-thread-safe) guards for static local variables. Further, I think https://godbolt.org/z/x1Kh57818 demonstrates a Clang/MSVC compatibility issue. The code Clang emits for h() for the inlining of g() attempts to import __imp_?$RT1@singleton@?1??g@@YAHXZ@4ABUS@@B while MSVC's code attempts to import __imp_?$S1@?1??g@@YAHXZ@4US@@B.

Assuming agreement with the analysis above, I'll file a new issue for the divergent mangling.

[zmodem]

Commit e77de75 claims that Microsoft doesn't have a mangling for these objects and that the $RT mangling is a Clang invention. The claim that MSVC doesn't implement a mangling for these doesn't seem right though. Consider https://godbolt.org/z/n715zPbc6.

Looks like things have changed then. I tried the example code with MSVC 18.00.31101.0 (from VS2013) and there the object is just _$S1:

??__E?singleton@RT1@@2ABU1@B@@YAXXZ PROC                ; ??__E?singleton@RT1@@2ABU1@B@@YAXXZ, COMDAT
; File z:\tmp\a.cc
; Line 6
        push    ebp
        mov     ebp, esp
        push    ecx
        call    ?f@@YA?AURT1@@XZ                        ; f
        mov     DWORD PTR $T1[ebp], eax
        mov     eax, DWORD PTR $T1[ebp]
        mov     DWORD PTR _$S1, eax
        mov     DWORD PTR ?singleton@RT1@@2ABU1@B, OFFSET _$S1 ; RT1::singleton
        mov     esp, ebp
        pop     ebp
        ret     0

[majnemer]

LGTM

[tahonermann]

Thanks for the reviews. I'll land this in the next day or two and will file a new issue to track switching to the $S<discriminator> mangling. I'll look into making that change as well.

[tahonermann]

A new issue to address the inconsistent use of the $S<discriminator> mangling (as well as an additional issue with mangling of anonymous unions) has been filed as #86525.

[zmodem]

This looks reasonable to me, but maybe give rnk and/or majnemer a day to comments as well since they're the real experts.

re: the incompatibility, yes if MSVC is mangling these now, clang should probably try to match.