[RFC / RFH] Re-opened C++/51213 (access control under SFINAE)

[Paolo Carlini]

On 08/01/2012 07:35 PM, Jason Merrill wrote:
> On 08/01/2012 11:49 AM, Paolo Carlini wrote:
>> static_assert(sizeof(g<C>(0)) == 2, "Ouch");          // line 11
>>
>> typedef int testg[sizeof(g<C>(0)) == 2 ? 1 : -1];      // line 13
>>
>> what happens is that line 13 is mishandled:
>>
>> sfinae37_red.C:13:48: error: size of array ‘testg’ is negative
>>
>> However, *if I comment out line 11*, things work for line 13! If I swap
>> line 11 and line 13 then the declaration of testg is accepted and the
>> static_assert triggers.
>
> Curious.  I guess that the second time we see the call the compiler 
> thinks it already has the candidate it needs, but I don't know why 
> that would be.  Are we not getting to type_unification_real from 
> add_template_candidate the second time?
So, I'm in the middle of this (got distracted earlier today). I can tell 
you what I have.

For the second evaluation, the second time we call 
instantiate_template_1, thus for the interesting g(int) overload, here:

   spec = retrieve_specialization (gen_tmpl, targ_ptr, 0);

   gcc_assert (tmpl == gen_tmpl
           || ((fndecl = retrieve_specialization (tmpl, orig_args, 0))
           == spec)
           || fndecl == NULL_TREE);

   if (spec != NULL_TREE)
     {
       if (FNDECL_RECHECK_ACCESS_P (spec) && (complain & tf_error))
     recheck_decl_substitution (spec, gen_tmpl, targ_ptr);
       return spec;
     }

things are completely different, because spec != NULL_TREE and, more 
importantly, complain is tf_none, thus recheck_decl_substitution is not 
called, we just return immediately.

Compare to the first evaluation: in that case we call enforce_access 
*way* below, with the perform_deferred_access_checks call near the end 
of instantiate_template_1.

Thus, looks like the recheck_decl_substitution mechanism is not working 
by design because of complain == tf_none?!?

Note that while I'm debugging this, I see instantiate_template_1 always 
getting complain == tf_none, something seems weird about the && 
(complain & tf_error) above...

Paolo.