On 11/16/18 9:16 PM, Alexandre Oliva wrote:
> We used to preserve deferred access check along with resolved template
> ids, but a tentative parsing firewall introduced additional layers of
> deferred access checks, so that we don't preserve the checks we
> want to any more.
> 
> This patch collapses the access check levels introduced by the
> firewall with those we wanted to preserve to begin with, saves them,
> pushes them one more layer up so that the firewall doesn't drop them,
> and then pushes new empty levels so balance out with the upcoming
> popping.
> 
> We may want to avoid the pop_to_parent and one push after the save, and
> instead leave the firewall's scope before the final pop_to_parent.
> However, this patch is what I regression-tested successfully with
> check-g++ on x86_64-linux-gnu.  Regstrapping on i686- and
> x86_64-linux-gnu now.  Ok to install?
> 
> 
> for  gcc/cp/ChangeLog
> 
> 	PR c++/86823
> 	* parser.c (cp_parser_template_id): Merge access checks from
> 	outside the tentative parsing firewall with those from inside,
> 	and save them all with the template id token.
> 
> for  gcc/testsuite/ChangeLog
> 
> 	PR c++/86823
> 	* g++.dg/pr86823.C: New.
> ---
>   gcc/cp/parser.c                |   12 ++++++++++++
>   gcc/testsuite/g++.dg/pr86823.C |   15 +++++++++++++++
>   2 files changed, 27 insertions(+)
>   create mode 100644 gcc/testsuite/g++.dg/pr86823.C
> 
> diff --git a/gcc/cp/parser.c b/gcc/cp/parser.c
> index db0f0338179e..6a08b09715a7 100644
> --- a/gcc/cp/parser.c
> +++ b/gcc/cp/parser.c
> @@ -16185,7 +16185,19 @@ cp_parser_template_id (cp_parser *parser,
>   	 so the memory will not be reclaimed during token replacing below.  */
>         token->u.tree_check_value = ggc_cleared_alloc<struct tree_check> ();
>         token->u.tree_check_value->value = template_id;
> +      /* Collapse the access check levels introduced by
> +	 tentative_firewall with the one we introduced ourselves.  */
> +      pop_to_parent_deferring_access_checks ();
> +      pop_to_parent_deferring_access_checks ();
>         token->u.tree_check_value->checks = get_deferred_access_checks ();
> +      /* Push the checks up one more level, so that the firewall
> +	 doesn't drop them on the floor when we return.  Then, push
> +	 empty levels back in place so that they are popped
> +	 properly.  */
> +      pop_to_parent_deferring_access_checks ();
> +      push_deferring_access_checks (dk_deferred);
> +      push_deferring_access_checks (dk_deferred);
> +      push_deferring_access_checks (dk_deferred);

Hmm, I'm uncomfortable with how this depends on the specific 
implementation of tentative_firewall.  What do you think of this 
alternate approach (untested other than with the testcase)?

Jason