[SYCL] Do not crash on attempt to visit invalid decl

[Fznamznon]

No description provided.

[Fznamznon]

This is the case when patch itself is very simple, but it is almost impossible to cover all test cases. I was able to figure some test cases, but I would appreciate suggestions.

[erichkeane]

Now seeing how simple the patch is, I think you've covered most of the cases I can think of. I didn't realize an invalid field/base made the parent invalid.

It might be a different bug, but I seem to remember seeing a problem with a member function passing something of its own type to a kernel (since it is incomplete!) and causing trouble.

Obviously the same problem exists for in the definition of the base class:

struct Base {
  struct Derived : Base { void use() const{}
  };
  void foo() {
    Derived d;
    kernel<class Whatever>([=](){d.use();});
  }
}; // End of Base

If that is sufficiently different of a bug, feel free to do that somewhere else.

[Fznamznon]

Now seeing how simple the patch is, I think you've covered most of the cases I can think of. I didn't realize an invalid field/base made the parent invalid.

It might be a different bug, but I seem to remember seeing a problem with a member function passing something of its own type to a kernel (since it is incomplete!) and causing trouble.

Obviously the same problem exists for in the definition of the base class:

struct Base {
  struct Derived : Base { void use() const{}
  };
  void foo() {
    Derived d;
    kernel<class Whatever>([=](){d.use();});
  }
}; // End of Base

If that is sufficiently different of a bug, feel free to do that somewhere else.

I've just tried this test case and the output seems valid for me:

test.cpp:3:20: error: base class has incomplete type
  struct Derived : Base { void use() const{}
                   ^~~~
test.cpp:2:8: note: definition of 'Base' is not complete until the closing '}'
struct Base {
       ^
1 error generated.

Looking into AST dump, it seems it is also the case when kernel object becomes invalid:
-LambdaExpr 0x561b64cdeec8 <col:50, col:64> '(lambda at test.cpp:7:50)
|-CXXRecordDecl 0x561b64cdea60 col:50 col:50 implicit invalid class definition
... etc

[erichkeane]

Now seeing how simple the patch is, I think you've covered most of the cases I can think of. I didn't realize an invalid field/base made the parent invalid.
It might be a different bug, but I seem to remember seeing a problem with a member function passing something of its own type to a kernel (since it is incomplete!) and causing trouble.
Obviously the same problem exists for in the definition of the base class:

struct Base {
  struct Derived : Base { void use() const{}
  };
  void foo() {
    Derived d;
    kernel<class Whatever>([=](){d.use();});
  }
}; // End of Base

If that is sufficiently different of a bug, feel free to do that somewhere else.

I've just tried this test case and the output seems valid for me:

test.cpp:3:20: error: base class has incomplete type
  struct Derived : Base { void use() const{}
                   ^~~~
test.cpp:2:8: note: definition of 'Base' is not complete until the closing '}'
struct Base {
       ^
1 error generated.

Looking into AST dump, it seems it is also the case when kernel object becomes invalid:
-LambdaExpr 0x561b64cdeec8 <col:50, col:64> '(lambda at test.cpp:7:50)
|-CXXRecordDecl 0x561b64cdea60 col:50 col:50 implicit invalid class definition
... etc

Perfect, thanks!