[PATCH] D157201: [Clang] Support qualified name as member designator in offsetof

[Yichi Lee via Phabricator via cfe-commits]

yichi170 added inline comments.


================
Comment at: clang/test/SemaCXX/offsetof.cpp:106
+int x3[__builtin_offsetof(struct X2, X2::static_a) == 0 ? 1 : -1]; // expected-error{{no member named 'static_a'}}
+int x4[__builtin_offsetof(struct X2, X2::X2) == 0 ? 1 : -1]; // expected-error{{no member named 'X2'}}
+
----------------
cor3ntin wrote:
> yichi170 wrote:
> > cor3ntin wrote:
> > > yichi170 wrote:
> > > > hubert.reinterpretcast wrote:
> > > > > aaron.ballman wrote:
> > > > > > yichi170 wrote:
> > > > > > > aaron.ballman wrote:
> > > > > > > > There's one more test I'd like to see:
> > > > > > > > ```
> > > > > > > > struct S {
> > > > > > > >   int Foo;
> > > > > > > > };
> > > > > > > > 
> > > > > > > > template <typename Ty>
> > > > > > > > void func() {
> > > > > > > >   static_assert(__builtin_offsetof(Ty, Ty::Foo) == 0, "");
> > > > > > > > }
> > > > > > > > 
> > > > > > > > void inst() {
> > > > > > > >   func<S>();
> > > > > > > > }
> > > > > > > > ```
> > > > > > > It would get the compile error in the current patch, but I think it should be compiled without any error, right?
> > > > > > Correct, that should be accepted: https://godbolt.org/z/1f6a9Yaxa
> > > > > Should expect this to pass too:
> > > > > ```
> > > > > template <typename T>
> > > > > struct Z {
> > > > >   static_assert(!__builtin_offsetof(T, template Q<T>::x));
> > > > > };
> > > > > 
> > > > > struct A {
> > > > >   template <typename T> using Q = T;
> > > > >   int x;
> > > > > };
> > > > > 
> > > > > Z<A> za;
> > > > > ```
> > > > Wow. Does it mean we cannot simply parse the identifier, `::`, `.` and brackets in `__builtin_offsetof`?
> > > GCC seems to support that. 
> > > 
> > > We probably want to call `ParseOptionalCXXScopeSpecifier` and store the `NestedNameSpecifierLoc` we'd get from it (and then parse the (sequence of) identifier(s) corresponding to the member) as we do now.
> > > 
> > > The documentation of https://gcc.gnu.org/onlinedocs/gcc-13.2.0/gcc/Offsetof.html#index-_005f_005fbuiltin_005foffsetof 
> > > seems inaccurate,
> > > 
> > > it seems to be
> > > 
> > > `"__builtin_offsetof" "(" typename ","  nested-name-specifier offsetof_member_designator ")"`
> > > 
> > > 
> > > Note that you will have to take care of transforming the nested name in TreeTransform and handle type dependencies. Let me know if you have further questions,
> > > it's more involved than what you signed for. Sorry for not spotting that earlier (Thanks @hubert.reinterpretcast !)
> > Thank you for all the help! I'll take a look at it!
> I was wrong, we need another approach.
> 
> I think the grammar is actually
> ```
> member-designator:
>           qualified-id
>           member-designator.qualified-id
>           member-designator.qualified-id
> ```
> IE, we should support https://godbolt.org/z/eEq8snMc8
> 
> Unfortunately, this looks like a much bigger change that we envisioned when we tagged this as a good first issue, to the extent I'm not sure what is actually the right design is would be.
> 
> For each component I imagine we want to store
> `NestedNameSpecifierLoc + DeclarationNameInfo`
> 
> The parser would have to produce a CXXScopeSpec + UnqualifiedId pair for each component.
> 
> The expression is dependent if any of the component is type dependent,
> 
> `OffsetOfNode` would have to change, but i think we can get away
> Only changing the identifier case (ie the dependent case)  
> 
Would it be better for me to close this patch and submit a new one if I find out how to implement it? I hope others won't hesitate to contribute because I'm working on this. I don't want to cause any delays in the release plan!


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D157201/new/

https://reviews.llvm.org/D157201