[C++ Patch] PR 34892

[Paolo Carlini]

Hi,

On 10/24/2012 07:30 PM, Jason Merrill wrote:
> On 10/24/2012 01:20 PM, Paolo Carlini wrote:
>> +      if (parm == error_mark_node
>> +      || TREE_PURPOSE (parm) == error_mark_node)
>
> It seems odd to bail out early if the default argument is bad even if 
> we aren't trying to use it.  Doesn't it work to check this further 
> down where we actually look at the default argument?
The problem is that the first time we go through the loop, when parm_idx 
== 0 and TREE_PURPOSE is error_mark_node, the condition:

       if (template_parameter_pack_p (TREE_VALUE (parm))
       && !(arg && ARGUMENT_PACK_P (arg)))

near the beginning of the loop is true, thus we call 
coerce_template_parameter_pack, no error and we continue to the next 
iteration. The next iteration, arg == NULL_TREE and require_all_args is 
true, we actually use TREE_PURPOSE (parm) but for the next parm, when 
things are fine, then we get to the:

TREE_VEC_ELT (new_inner_args, arg_idx) = arg;

at the bottom of the loop and we are in big troubles because 
new_inner_args has only 1 element and arg_idx is 2.

Thus, in summary, it really seems to me that we should do *something* 
the *first* time through the loop, when currently we only use TREE_VALUE 
per the above and we don't notice that TREE_PURPOSE is error_mark_node. 
Did I explain myself well enough?

Thanks,
Paolo.