[PATCH] D57540: [C++17] Don't crash while diagnosing different access specifier of a deduction guide.

[Richard Smith - zygoloid via Phabricator via cfe-commits]

rsmith added inline comments.


================
Comment at: test/SemaCXX/cxx1z-class-template-argument-deduction.cpp:360-361
+  struct Type {
+    Typo(); // expected-error{{deduction guide must be declared in the same scope}}
+    // expected-error at -1{{deduction guide declaration without trailing return type}}
+  };
----------------
aaron.ballman wrote:
> Rakete1111 wrote:
> > aaron.ballman wrote:
> > > These errors are pretty unfortunate -- they don't really help the user to understand what's gone wrong here. They're an improvement over the crash, but I think we should try to make the errors more useful if we can.
> > > 
> > > Why is `Typo()` being treated as a deduction guide? Perhaps we could look to see if there is a `->` before making that determination?
> > > Perhaps we could look to see if there is a -> before making that determination?
> > 
> > Yes that would be possible. The diagnostic would change for the following code:
> > 
> > ```
> > template <typename>
> > struct Foo {};
> > 
> > Foo();// -> Foo<int>; 
> > // currently: deduction guide missing ->
> > // after: C++ requires type specifier for every declaration
> > ```
> > 
> > Is that acceptable? Or I guess I could restrict this to partial deduction guides in classes.
> I think the original diagnostic is better in that case. If you restrict to partial deduction guides, do we get all the good diagnostics?
I think the most likely intent in that case was to declare a constructor of `Foo`, and either it was accidentally written after the end of the class, or the in-class declaration got copy-pasted and the programmer forgot to fix it up. And moreso for a case such as

```
template <typename>
struct Foo { Foo(); };

template<typename T>
Foo() {
    // ...
}
```

... where we currently give three errors about deduction guides and no hint that a qualified name is required to define a constructor.

I think it's comparatively unlikely that someone would forget the `->` in a deduction guide, given how central it is to the purpose of the declaration. So always disambiguating as a failed constructor rather than a failed deduction guide seems reasonable to me.


Repository:
  rC Clang

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D57540/new/

https://reviews.llvm.org/D57540