[PATCH] Start adding support for dllimport/dllexport on classes (PR11170)

[Hans Wennborg]

On Thu, May 29, 2014 at 4:23 PM, Reid Kleckner <rnk at google.com> wrote:
> On Thu, May 29, 2014 at 4:01 PM, Hans Wennborg <hans at chromium.org> wrote:
>>
>> I tested my patch with Nico's tests from
>> https://github.com/gix/clang/commit/ea6cb1d71c6ef8dac2afd5f40219de9460a92b3f,
>> and fixed the two asserts we were hitting when exporting defaulted trivial
>> constructors and destructors.
>>
>> There were two remaining test failures: 1) we emit RTTI type descriptor,
>> etc. as weak_odr instead of linkonce_odr when exporting the class. Not a big
>> deal since we're exporting the vtable anyway, but I'm not sure why it's
>> happening.
>
>
> It's pretty simple: I told Warren to use the vftable linkage for RTTI data.

That makes sense :) I'll see if I can add the dllexport case to that logic.

> Can RTTI data be imported? What happens when you use typeid?

MSVC doesn't export the RTTI data, so it can't be imported. When I use
typeid it seems to materialize the things it needs, it doesn't try to
import anything.

> I seem to
> recall they don't let you import the type descriptor and you have to use
> string comparison on the mangled name to compare types for equality.  In
> that case, linkonce_odr is correct.

Yes, I think linkonce_odr would be the correct one, but I don't think
emitting it as weak_odr until we fix it is breaking anything. Since
we're exporting the vtable, which references these things, we can't
drop them anyway.

Does the patch look good to you otherwise?