enable-checking: error_mark accessed as type

[Martin v. Loewis]

> I especially have a problem with creating ERROR_MARK nodes that
> are not error_mark_node.
> 
> Basically, I think we need a cleaner solution to this problem.  How does
> the C frontend deal with this issue?

I assume they don't have such problems. The specific problem with
build_dummy_object is that it is passed to integer_zerop. That could
be solved by extending STRIP_NOP to expect error_mark_node, or by not
putting ERROR_MARKs into your dummy objects.

I don't understand why the dummy object *has* to have an
error_mark_node inside, it seems any recognizable pattern would do.

Other than that, my patch does not produce new ERROR_MARKs. As for
checking for both error_mark and error_type: Based on the current test
suite, this only rarely has to be done. If we put an error_mark into a
type, we *already* printed an error message. The code just has to get
through somehow, for error stabilization.

Regards,
Martin