Re: [C++ RFC/Patch] PR 39813

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

Hi,

On 04/28/2015 02:59 PM, Jason Merrill wrote:
> On 04/28/2015 08:54 AM, Paolo Carlini wrote:
> > Hi,
> >
> > On 04/28/2015 02:45 PM, Jason Merrill wrote:
> > > On 04/28/2015 06:59 AM, Paolo Carlini wrote:
> > > >        && !DECL_FRIEND_PSEUDO_TEMPLATE_INSTANTIATION (t)
> > > > the name
> > > > of the macro seems weird to me: it mentions friends (and friends are
> > > > mentioned in its comment too) but isn't used only for DECL_FRIEND_P 
> > > > true
> > > > and indeed it is true for the non-friend 'fn' in the snippet at 
> > > > issue).
> > >
> > > It shouldn't be true for fn, that's a bug.
> > Ok, but then what? Its definition boils down to very basic macros:
> >
> > #define DECL_FRIEND_PSEUDO_TEMPLATE_INSTANTIATION(DECL) \
> >    (DECL_TEMPLATE_INFO (DECL) && !DECL_USE_TEMPLATE (DECL))
> >
> > It could be only matter of adding a DECL_FRIEND_P check at the 
> > beginning?
> Sure, that sounds likely.
The simple idea almost works. The below however has an additional 
STRIP_TEMPLATE, which I added to avoid ICEs for testcases like 
g++.dg/template/friend26.C, where DECL_FRIEND_P of plain tmpl is false, 
whereas DECL_FRIEND_P of STRIP_TEMPLATE (tmpl) is true for the use of 
DECL_FRIEND_PSEUDO_TEMPLATE_INSTANTIATION in template_for_substitution. 
Does that make sense? Then we come to the two *remaining* regressions:

    g++.old-deja/g++.pt/friend10.C
    g++.old-deja/g++.pt/link1.C

which fail at link time. Those would disappear if I replace the already 
mentioned DECL_FRIEND_PSEUDO_TEMPLATE_INSTANTIATION in 
template_for_substitution with the old implemention of it, just checking 
DECL_TEMPLATE_INFO && !DECL_USE_TEMPLATE. Interestingly, note that with 
current (and old) clang too these two tests do nor link. I'm a bit stuck 
on this...

Thanks!
Paolo.

/////////////////////

Attachment: patch_39813_draft
Description: Text document