Re: unique_ptr's new static_assert check whether deleter can be invoked fails

[Jonathan Wakely <jwakely at redhat dot com>]

On 17/09/18 16:41 +0200, Stephan Bergmann wrote:
> Compiling LibreOffice with recent GCC trunk, I came across an error 
> like with this reproducer
>
> > $ cat test14.cc
> > #include <memory>
> > struct S1;
> > struct S2;
> > struct D { void operator ()(S1 *); };
> > class C {
> >    std::unique_ptr<S2, D> m;
> >    C();
> >    C(C const &);
> >    ~C();
> >    C & operator =(C const &);
> > };
> >
> > $ gcc/inst/bin/g++ -c test14.cc
> > In file included from /home/sbergman/gcc/inst/include/c++/9.0.0/memory:80,
> >                 from test14.cc:1:
> > /home/sbergman/gcc/inst/include/c++/9.0.0/bits/unique_ptr.h: In instantiation of ‘class std::__uniq_ptr_impl<S2, D>’:
> > /home/sbergman/gcc/inst/include/c++/9.0.0/bits/unique_ptr.h:172:33:   required from ‘class std::unique_ptr<S2, D>’
> > test14.cc:6:28:   required from here
> > /home/sbergman/gcc/inst/include/c++/9.0.0/bits/unique_ptr.h:145:22: error: static assertion failed: unique_ptr's deleter must be invocable with a pointer
> > 145 |       static_assert( __is_invocable<_Dp&, pointer&>::value,
> >    |                      ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>
> where S2 is an incomplete type, so the static_assert in unique_ptr 
> (added recently with <https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=199f729ee3a6bd564151cda4caac289fbe7105cc> 
> "Implement LWG 2905 changes to constrain unique_ptr constructors") 
> can't determine that D can be invoked with an S2* (which it can if S2 
> is derived from S1).

Oh dear, this is another example of the problem described in
https://cplusplus.github.io/LWG/lwg-active.html#3099

I think strictly speaking, your code is undefined according to
[unique.ptr.single] p1. That requires the is_invocable condition to be
true. I think that should be considered a defect in the standard, so
I've requested a new issue to fix that.

> I'm not sure whether this is a bug in LibreOffice (which would need to 
> make sure S2 is a complete type when defining C),

Or you can define D::pointer as S1* and then the assertion passes.

struct D {
 using pointer = S1*;
 void operator ()(S1 *);
};


> or a too aggressive static_assert in libstdc++?

Yes, I'll have to remove it.