object.classinfo: Return a TypeInfo_Class instead of a TypeInfo_Interface

[kinke]

This may be enough to fix #1471 (at least there's no crash anymore for @rainers' minimal testcase).

[JohanEngelen]

Could you also add a testcase for this?

Perhaps something like:

// RUN: %ldc -c -output-ll -of=%t.ll %s && FileCheck %s < %t.ll

class Node
{
    final void _clone()
    {
// CHECK-NOT: load %object.TypeInfo_Interface*
// CHECK: load %object.TypeInfo_Class*
        this.classinfo.create();
    }
}

[kinke]

Thanks for the ready-to-paste test, but I'm not convinced this deserves a test on its own; it rather seems like a bug due to lack of focus or a typo to me, with an extremely low probability of being reintroduced. ;)
Let's see if it passes the tests and whether it also fixes Rainer's problems first.

Edit: But if there's really not a single test accessing a ClassInfo this way, then we definitely need one (preferably upstream).

[JohanEngelen]

OK. Pretty much the first thing I think of now is: "does it have tests?" More tests, more better ;-)
I am not sure the IR has to be checked, but a simple "does it compile" check would be nice, no? It would check for many other things besides your typo ;)

[kinke]

I'm pretty sure this typo isn't mine. ;)
Yeah that's what I meant (not checking the IR, just checking whether it compiles and works), but I'd rather see such tests go upstream so that all compilers benefit (I don't know whether GDC uses dmd-testsuite too, but I guess and hope so) from improved test coverage. DMD's test coverage is pretty bad, especially wrt. basic compiler and language stuff - #1327 wouldn't exist if there was a good test coverage for the evaluation order of expressions, which is sadly a constant source of errors.

[JohanEngelen]

Opps, sorry misread :(
Upstreaming is good.

[rainers]

LGTM, but I haven't tried it yet (I replaced all occurences with typeid). I wonder why this isn't lowered to TypeidExp in the front end, though.
A test should verify runtime behaviour, too.

[kinke]

This may not be correct. TypeidExp (thanks for the hint, Rainer) returns a TypeInfo_Interface for interfaces, not a TypeInfo_Class. Don't know if that is supposed to happen for .classinfo too.

[rainers]

According to the doc and Kenji, typeid() and .classinfo are supposed to do the same. There have been some discussions around what should be returned and that strange things happen with dmd, too: dlang/dmd#4711, https://issues.dlang.org/show_bug.cgi?id=14612 and a number of linked reports.

[kinke]

Updated so that .classinfo for interfaces returns a TypeInfo_Interface, as before and compatible to TypeidExp codegen.

[rainers]

LGTM. It's compatible with dmd with respect to not returning the actual class for interfaces.