Re: [C++ Patch] Prefer error + inform to two errors in check_template_shadow

[Paolo Carlini <paolo dot carlini at oracle dot com>]

Hi again,

On 07/14/2015 05:43 PM, Paolo Carlini wrote:
> Hi,
>
> On 07/14/2015 05:07 PM, Jason Merrill wrote:
> > On 07/13/2015 09:41 AM, Paolo Carlini wrote:
> > > +++ testsuite/g++.dg/template/crash81.C (working copy)
> > > @@ -3,6 +3,6 @@
> > >   struct A
> > >   {
> > >     template<T::X> struct X; // { dg-error "'T' has not been 
> > > declared" "T" }
> > > -  // { dg-error "declaration of 'template<int X> struct A::X'" 
> > > "A::X" { target *-*-* } 5 }
> > > -  // { dg-error "shadows template parm 'int X'" "shadow" { target 
> > > *-*-* } 5 }
> > > +  // { dg-error "declaration of 'template<int X> struct A::X' 
> > > shadows" "A::X" { target *-*-* } 5 }
> > > +  // { dg-message "template parameter 'X'" "" { target *-*-* } 5 }
> >
> > I don't see any reason to check for specific diagnostics here; the 
> > latter two messages are poor error-recovery, not something to test for.
> Indeed, I noticed the error recovery issue, which I didn't know, and 
> considered looking into it, when I'm done with some other things. In 
> the meanwhile I'm shortening the expected error/inform.
I had a look and tried various things... Ultimately the issue is due to 
the type_was_error_mark_node becomes integer_type_node trick in 
grokdeclarator (which I don't like that much ;) Thus for this testcase 
the latter simply returns a PARM_DECL of that type and no trace remains 
of the error, then the following bad error recovery. Now, 
cp_parser_template_parameter knows how to handle error_mark_node thus in 
principle we should be able to just return it and be done. The only 
problem with that, testsuite-wise, is the very broken cpp/pr64127.C 
which ends up regressing, since we emit a spurious additional error 
about "default argument for template parameter for class enclosing" from 
check_default_tmpl_args. Note that for this specific testcase 
grokdeclarator gets also a null declarator thus for it isn't the same if 
grokdeclarator returns error_mark_node immediately or just sets type = 
error_mark_node at the beginning (which results in grokdeclarator not 
returning an error_mark_node, instead a PARM_DECL with error_mark_node 
as type).

Thus I have two different proposals: 1- More conservative, just set type 
= error_mark_node instead of integer_type_node when we have a template 
parameter; 2- Return error_mark_node immediately and adjust 
check_default_tmpl_args to cope correctly with pr64127.C.

Thanks!
Paolo.

Attachment: patch_error_recovery
Description: Text document

Attachment: patch_error_recovery_2
Description: Text document