[clang] [clang][C23] N3006 Underspecified object declarations (PR #79845)

[Aaron Ballman via cfe-commits]

================
@@ -7639,6 +7639,8 @@ def err_attribute_arm_mve_polymorphism : Error<
   "'__clang_arm_mve_strict_polymorphism' attribute can only be applied to an MVE/NEON vector type">;
 def err_attribute_webassembly_funcref : Error<
   "'__funcref' attribute can only be applied to a function pointer type">;
+def err_c23_underspecified_object_declaration: Error<
+  "'%select{struct|<ERROR>|union|<ERROR>|enum}0 %1' is defined as an underspecified object initializer">;
----------------
AaronBallman wrote:

I think the terminology used in the standard is really hard for mortals to understand, but I also think the definition of "underspecified" makes it hard to capture this in a single diagnostic and still be precise. But I think existing diagnostics cover the other cases, mostly.

C23 6.7.1p12: A declaration such that the declaration specifiers contain no type specifier or that is declared with constexpr is said to be underspecified. If such a declaration is not a definition, if it declares no or more than one ordinary identifier, if the declared identifier already has a declaration in the same scope, if the declared entity is not an object, or if anywhere within the sequence of tokens making up the declaration identifiers that are not ordinary are declared, the behavior is implementation-defined.

So I think the best I can come up with is:

`%select{declaration using a deduced type|constexpr declaration}0 %1 also declares %2 which is not an ordinary identifier`

I think "if such a declaration is not a definition" is already covered by `constexpr variable 'i' must be initialized by a constant expression`: https://godbolt.org/z/nvKEzv975
I think "if it declares no or more than one ordinary identifier" is already covered by `expected identifier or '('` or is supported by Clang: https://godbolt.org/z/Yas4od5Ev (Note, we should add documentation for the implementation-defined extension accepting multiple declarations in a group.)
I think "if the declared identifier already has a declaration in the same scope" is not correctly supported: https://godbolt.org/z/Kab7GvW85 but this should be handled by an existing redeclaration diagnostic when we fix it.
I think "if the declared entity is not an object" is already covered by "'auto' not allowed in function return type": https://godbolt.org/z/KzeTeczY7
I think "if anywhere within the sequence of tokens making up the declaration identifiers that are not ordinary are declared" is the only thing we're missing a diagnostic for, and the above suggested wording would handle it.

So I think we can get away without inflicting "underspecified" on users.

The other half of this paper has to do with scope of identifiers, we may end up needing to get creative with diagnostic wording when lookup fails to find an identifier from an underspecified declaration.

https://github.com/llvm/llvm-project/pull/79845