Hi,

in this relatively serious ICE on invalid regression (we don't emit any 
message before ICEing) make_typename_type, called by strip_typedefs, 
returns error_mark_node which then isn't handled by typedef_variant_p. 
Indeed, Jakub noticed that the same in principle can happen when 
strip_typedefs calls finish_decltype_type. make_typename_type is passed 
tf_none therefore cannot emit the diagnostic we want before the ICE 
anyway, thus it seems to me that we also have one of those well known 
issues about not propagating a tusbst_flags_t flag: 
canonicalize_type_argument (and canonicalize_expr_argument) get a 
tusbst_flags_t but doesn't propagate it to strip_typedefs (and 
strip_typedefs_expr).

Implementing the straightforward set of changes, consistently for 
strip_typedefs, strip_typedefs_expr and the related functions plus 
checking the return values of make_typename_type and 
finish_decltype_type for error_mark_node (only make_typename_type would 
be enough for the testcase at issue) appears to solve the problem 
without introducing regressions. Does that make sense?

Thanks,
Paolo.

/////////////////////////////