[cfe-commits] PATCH: Improve diagnostics for conversion failures during initialization

[John McCall]

On Feb 27, 2010, at 3:43 AM, Chandler Carruth wrote:
> I've not finished updating all the tests to match the new format, but
> I've fixed the crashers. I really want feedback on whether these
> diagnostics are useful, do too much or too little, and wording or
> phrasing suggestions. In particular, Jeff already raised concerns
> about the incomplete type diagnostic (which motivated this cleanup).
> I've toned it down considerably based on his feedback, but he also
> mentioned that we might issue different diagnostics for a pointer type
> where we *know* that it must be an inheritance relationship if not
> exactly which one... Not sure how much i like this, but it's an
> interesting idea.

I haven't looked at the patch yet, but some comments about the diagnostics.  In general, that's a lot of text.  The overload-resolution diagnostics have to bend over backwards to establish context because they have to put the caret on the candidate's declaration.  That's not true here;  the caret and source ranges should make it perfectly clear that we're talking about a return value or a variable initializer.  That means you can cut a *lot* of unnecessary verbiage out of the diagnostics.

> t.cc:3:34: error: cannot initialize the return object with an lvalue
> pointing to an incomplete type 'class Derived': cannot check
> conversion to a pointer to 'class Base' without a complete type
> Base* test1(Derived* d) { return d; }

"can't convert pointer to incomplete type 'class Derived' to 'class Base*'"

I think this is clear enough without calling out the fact that it has to be an inheritance relationship.

> t.cc:5:9: error: cannot initialize the variable 'b' with an lvalue
> pointing to an incomplete type 'class Derived': cannot check
> conversion to a pointer to 'class Base' without a complete type
>  Base* b = d;
>        ^   ~

Since you can let the caret establish context, you can re-use the same diagnostic here:  "can't convert pointer to incomplete type 'class Derived' to 'class Base*'"

> % ./bin/clang -fsyntax-only .../test/SemaCXX/addr-of-overloaded-function.cpp
> .../test/SemaCXX/addr-of-overloaded-function.cpp:12:7: error: cannot
> initialize the variable 'pfe': no overload of the function 'f' matches
> 'int (*)(...)'

This is pretty good.

John.