[Clang][Sema] Reject undeduced static data members without initializers (#173349)

[HamzaHassanain]

#173349

The Problem

Clang crashes when encountering a static data member declared with an alias template but lacking template arguments (e.g., static E u;).

Because of C++ CTAD rules, Clang's parser assigns this a placeholder AutoType. If this member is later referenced, the backend (CodeGen) attempts to emit the variable. Since the type remains undeduced and no initializer exists, the compiler hits an llvm_unreachable in CodeGenTypes.cpp.

The Fix

The fix is implemented in Sema::ActOnUninitializedDecl. We now explicitly check if a static data member still possesses an undeduced placeholder type after all deduction attempts have failed.

• Logic: If the type contains an undeduced type, and the variable is a static member without an initializer, we emit err_auto_var_requires_init.

Changes

• SemaDecl.cpp: Added validation logic to ActOnUninitializedDecl to catch undeduced static members.
• alias-template-static-member.cpp: Added a regression test to ensure this scenario results in a diagnostic rather than a crash.

Notes on formatting

When I clang-format the SemaDecl.cpp after applying my changes, I found the whole file got changed, so I undid the formatting.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Hamza Hassanain (HamzaHassanain)

Changes

The Problem

Clang crashes when encountering a static data member declared with an alias template but lacking template arguments (e.g., static E u;).

Because of C++ CTAD rules, Clang's parser assigns this a placeholder AutoType. If this member is later referenced, the backend (CodeGen) attempts to emit the variable. Since the type remains undeduced and no initializer exists, the compiler hits an llvm_unreachable in CodeGenTypes.cpp.

The Fix

The fix is implemented in Sema::ActOnUninitializedDecl. We now explicitly check if a static data member still possesses an undeduced placeholder type after all deduction attempts have failed.

• Logic: If the type contains an undeduced type, and the variable is a static member without an initializer, we emit err_auto_var_requires_init.

Changes

• SemaDecl.cpp: Added validation logic to ActOnUninitializedDecl to catch undeduced static members.
• alias-template-static-member.cpp: Added a regression test to ensure this scenario results in a diagnostic rather than a crash.

Notes on formatting

When I clang-format the SemaDecl.cpp after applying my changes, I found the whole file got changed, so I undid the formatting.

---

Full diff: https://github.com/llvm/llvm-project/pull/174281.diff

2 Files Affected:

• (modified) clang/lib/Sema/SemaDecl.cpp (+10)
• (added) clang/test/SemaCXX/alias-template-static-member.cpp (+11)

diff --git a/clang/lib/Sema/SemaDecl.cpp b/clang/lib/Sema/SemaDecl.cpp
index 11323803e1910..50a6a4176d1ec 100644
--- a/clang/lib/Sema/SemaDecl.cpp
+++ b/clang/lib/Sema/SemaDecl.cpp
@@ -14337,8 +14337,18 @@ void Sema::ActOnUninitializedDecl(Decl *RealDecl) {
         DeduceVariableDeclarationType(Var, false, nullptr))
       return;
 
+    Type = Var->getType();
     this->CheckAttributesOnDeducedType(RealDecl);
 
+    if (auto *Deduced = Type->getContainedDeducedType()) {
+      if (Var->isStaticDataMember() && Deduced->getDeducedType().isNull()) {
+        Diag(Var->getLocation(), diag::err_auto_var_requires_init)
+            << Var->getDeclName() << Type;
+        Var->setInvalidDecl();
+        return;
+      }
+    }
+
     // C++11 [class.static.data]p3: A static data member can be declared with
     // the constexpr specifier; if so, its declaration shall specify
     // a brace-or-equal-initializer.
diff --git a/clang/test/SemaCXX/alias-template-static-member.cpp b/clang/test/SemaCXX/alias-template-static-member.cpp
new file mode 100644
index 0000000000000..0089aea33cc80
--- /dev/null
+++ b/clang/test/SemaCXX/alias-template-static-member.cpp
@@ -0,0 +1,11 @@
+// RUN: %clang_cc1 -fsyntax-only -verify -std=c++17 %s
+
+template <class T>
+struct A {
+  template <class U>
+  using E = U;
+
+  static E u; // expected-error {{declaration of variable 'u' with deduced type 'E' requires an initializer}}
+};
+
+decltype(A<int>::u) a;
\ No newline at end of file

[zwuis]

Please add a release note entry to "clang/docs/ReleaseNotes.rst".

---

We already diagnose correctly if we move E and u out of templates. Can you merge these logic?

[HamzaHassanain]

@zwuis

Switched from checking Type->isUndeducedType() to Type->getContainedDeducedType()
.

As far as I understand:

isUndeducedType() looks at E, sees it's an alias, and returns false because E itself isn't technically a "deduced type" node until you unwrap it.

So, should I leave it as Type->isUndeducedType() and find another condition to force calling Sema::DeduceVariableDeclarationType (so as to call deduceVarTypeFromInitializer to hit the place where we handle E, U)

[cor3ntin]

Missing EOF - also, can we put that test in an existing file ? test/SemaCXX/alias-template.cpp would be a good candidate

[HamzaHassanain]

Ok, I added that!

[zwuis]

While we are here, could you handle the case that u has initializer?

---

So, should I leave it as Type->isUndeducedType() and find another condition to force calling Sema::DeduceVariableDeclarationType (so as to call deduceVarTypeFromInitializer to hit the place where we handle E, U)

I'm not familiar enough with this part of code. Please wait for other reviewers.

[zwuis]

It would be nice to emit same diagnostic messages as the case that E and u are out of the definition of A.

template <class U>
using E = U;

static E u;

[HamzaHassanain]

While we are here, could you handle the case that u has an initializer?

It is already handled

[HamzaHassanain]

@cor3ntin

About the current diagnostic, expected-error {{declaration of variable 'u' with deduced type 'E' requires an initializer}} is this the correct diagnostic? (err_auto_var_requires_init) Or should it be:

expected-error {{alias template 'E' requires template arguments; argument deduction only allowed for class templates or alias templates}} (err_deduced_non_class_or_alias_template_specialization_type) ?

I believe the first one is fine. If not, could you please tell me, should I try to handle it inside

SemaDecl/Sema::deduceVarTypeFromInitializer

Or would it be inside

SemaInit.cpp/Sema::DeduceTemplateSpecializationFromInitializer

Also, about moving to a separate file, the file you mentioned uses c++14, while the whole issue is with C++17, C++ 20, C++23., So I believe we should keep it in a separate file?

I am very confused also with @zwuis comments, could you please explain a bit more? Or at least give me a little clue?

[zwuis]

While we are here, could you handle the case that u has an initializer?

It is already handled

We can refer to how it is handled.

template <typename> struct A {
  template <typename U>
  using E = U;

  // both of them should be rejected
  static E u1;
  static E u2 = 0;
};

Also, about moving to a separate file, the file you mentioned uses c++14, while the whole issue is with C++17, C++ 20, C++23., So I believe we should keep it in a separate file?

We can write multiple // RUN in one test file, each // RUN specifies different C++ versions.

[zwuis]

While we are here, could you handle the case that u has an initializer?

It is already handled

Oh, I see. We already emit diagnostics for static E u2 = 0; during template instantiation. With this patch, we emit diagnostics for static E u; before template instantiation. It would be better to handle the former case before template instantiation.

[zwuis]

Can we remove this #if?

[HamzaHassanain]

How then can I make some of the file be tested only with c++ 14, and my tests only run with c++ 20 ?