[C++ PATCH] Diagnose invalid uses of auto type specifier (PR c++/37962)

[Jason Merrill]

Jakub Jelinek wrote:
> (the type is error_mark_node, but due to lots of tentative parsing no error
> is ever reported).

Yes, there doesn't seem to be an easy way to defer an error to the point 
  where we accept the parse; cp_parser_simulate_error simulates a parse 
error, which this isn't, it's a semantic violation.

I would expect most of these cases to be caught by the preexisting code 
that requires a complete type in these different situations.  If they 
aren't, that should be fixed rather than adding in new special cases for 
auto everywhere.  For instance, add_exception_specifier doesn't complain 
because it unconditionally allows TEMPLATE_TYPE_PARM; that seems 
unnecessary, and removing that if ought to fix that case.

Jason