[patch] Fix PR c++/27572: ICE on typedef in parameter declaration

[Volker Reichelt]

On 14 May, Mark Mitchell wrote:
> Volker Reichelt wrote:
> 
>> The patch is mostly mechanical, as it replaces all uses of
>> "declspecs->specs[(int)ds_typedef]" by "typedefp". The only interesting
>> part is the third hunk where typedefp is reset for parameters.
> 
> Why not just return error_mark_node at that point?

Because then the following will cause an ICE:

template<typedef> struct A {};
A<int> a;

bug.cc:1: error: ISO C++ forbids declaration of 'parameter' with no type
bug.cc:1: error: typedef declaration invalid in parameter declaration
bug.cc:1: internal compiler error: tree check: accessed elt 1 of tree_vec with 0 elts in get_innermost_template_args, at cp/pt.c:490
Please submit a full bug report, [etc.]

Btw, there are more PRs (PR26938, PR27397, PR27493, PR27494) with
invalid accesses of elts of tree_vecs. Maybe one can get away returning
error_mark_node once these get fixed, but I haven't looked into them.

Do you want to have a look?

>> Btw, "typedefp" is now the sixth variable to track the status of
>> declspecs locally in grokdeclarator. Wouldn't it make more sense in
>> the long run to skip the "const" in "declspecs" and modify the
>> original specs direcly?
> 
> IIRC, the point is that the same specifiers are reused for multiple
> declarators, in something like "static const int i, *j;" and we want to
> avoid modifying the behavior on the second declarator.

Makes sense. (I only thought of function decls.)

> You could indeed
> make a copy of the specs in grokdeclarator, and use that.

If we have to make a copy anyway, we might be better off with the
local variables. Maybe we can convert the predicates from int to bool
sometimes to better reflect their intention.

Regards,
Volker