[clang] [clang] Implement P2582R1: CTAD from inherited constructors (PR #98788)

[Haojian Wu via cfe-commits]

================
@@ -936,19 +940,21 @@ Expr *buildIsDeducibleConstraint(Sema &SemaRef,
         Context.DeclarationNames.getCXXDeductionGuideName(AliasTemplate));
   };
 
+  TemplateDecl *TD = DeducingTemplate ? DeducingTemplate : AliasTemplate;
+
   SmallVector<TypeSourceInfo *> IsDeducibleTypeTraitArgs = {
       Context.getTrivialTypeSourceInfo(
           Context.getDeducedTemplateSpecializationType(
-              TemplateName(AliasTemplate), /*DeducedType=*/QualType(),
+              TemplateName(TD), /*DeducedType=*/QualType(),
----------------
hokein wrote:

Thanks for the detailed analysis.

> I think you're right that the `is_deducible` constraint is technically unnecessary-- as far as I can tell, any case where this constraint would fail would cause the `CC<R>::type` instantiation to fail before the constraint is evaluated. There are some other caveats with removing it, however. Consider this example:
> 
> ```c++
> template<typename T>
> struct Base1 { };
> 
> template<typename T>
> struct Base2 { };
> 
> template<typename T = int>
> struct Derived : public Base1<T>, Base2<T> {
>     using Base1<T>::Base1;
>     using Base2<T>::Base2;
> };
> 
> Derived d;
> ```
> 
> https://godbolt.org/z/4YcdvTMov
> 
> Under [[over.match.best.general]p2.5](https://eel.is/c++draft/over.match.best.general#2.5) (and more specifically [[temp.func.order]p6.4](https://eel.is/c++draft/temp.func.order#6.4)), the bases' deduction guides are more specialized than any from `Derived` (due to the `is_deducible` constraint), and so the implementation as it is sees this as ambiguous.

Hmm, this is an interesting case. I think this is valid code, and it is accept by all compilers in C++ 20 (or before) mode. Now we reject it in C++23, this seems like a regression, https://godbolt.org/z/zxx4W36zq.

Removing the constraint seems to result in a reasonable behavior, at least this case will be compilable in C++23.

We should probably ask CWG for clarifications. @cor3ntin what's the best way to do that? 






https://github.com/llvm/llvm-project/pull/98788