[PATCH] D64380: Add 'require_designated_init' and 'required' attribute to clang

[Aaron Ballman via Phabricator via cfe-commits]

aaron.ballman added inline comments.


================
Comment at: clang/include/clang/Basic/Attr.td:1948
+def RequiresDesignator : InheritableAttr {
+  let Spellings = [Clang<"requires_designator">];
+  let Subjects = SubjectList<[Record]>;
----------------
emmettneyman wrote:
> compnerd wrote:
> > aaron.ballman wrote:
> > > Hmm, after making this suggestion, I noticed that GCC seems to support a similar attribute named `designated_init` (https://gcc.gnu.org/onlinedocs/gcc/Common-Type-Attributes.html#Common-Type-Attributes). Was your goal to support the same thing GCC supported?
> > `designated_init` is a suggestion, the original patch seemed to be a stronger version.  I think that we should be supporting the GNU spelling for `designated_init` and can support a Clang spelling of `requires_designated_init` if the goal is to have the stronger guarantee that this *must* happen.
> I hadn't known about the GCC attribute until now. Yes, the `requires_designator` (originally `require_designated_init`) attribute wants to enforce the same thing I believe. I couldn't find more documentation for the GCC `designated_init` attribute so it's a little tough to tell whether the behavior is the exact same. The attribute in this patch allows a field to be default constructed (unless the other attribute is applied to that specific field) but enforces that a brace initializer must be used. So `Foo foo {};` would be valid (every field is default constructed) but `Foo foo;` would not be valid. I'm not sure if that's the same behavior the GCC attribute is trying to enforce. But on a high level, both are trying to prohibit using positional args when declaring a struct.
> 
> @compnerd I don't mind this attribute generating warnings rather than errors. It's ok for this attribute to be a "suggestion" as well. Like @aaron.ballman mentioned, "users can always use -Werror to strengthen their own requirements."
I think the correct approach then is: add this attribute under the name `designated_init` with the `GCC` spelling kind and match GCC's behavior for it (or, if we don't want to match GCC's behavior, note where we differ and why). We should put its diagnostics under a `-Wdesignated-init` warning group.

The next question is: are you interested in finding out whether GCC would be willing to implement the `requires_init` attribute? If GCC is interested in picking it up, then we can name it with the `GCC` spelling as well. If the GCC folks are not interested in picking up the attribute, then we should probably leave it with the `Clang` spelling. I'd recommend we put the diagnostics into the `-Wdesignated-init` warning group as well.

I think both of these attributes should also be available in C2x mode as well, but that can probably be done in a follow-up patch as it will involve tablegen changes and opens up other questions (i.e., does GCC handle C2x attributes yet, and if they do, are they making `designated_init` available in that mode?).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D64380/new/

https://reviews.llvm.org/D64380