pointer check in unit tests

[Miklos Vajna]

Hi Stephan,

On Tue, Nov 03, 2020 at 10:21:34AM +0100, Stephan Bergmann <sbergman at redhat.com> wrote:
> On 03/11/2020 09:37, Miklos Vajna wrote:
> > (If you see a case where a pointer is returned and it can't be ever
> > nullptr, then we should fix the return type to a be reference. Caolan
> > did lots of fixes like that recently.)
> 
> That's up for debate.  For example, if a sufficiently large fraction of call
> sites wants a pointer, it can be awkward to change a function's return type
> from pointer to reference.  (And doing so without carefully auditing all
> call sites can silently introduce regressions.)
> 
> T& is something rather different than T* plus "cannot be null".  Just as the
> C++ type system isn't capable of expressing the type "int, but never 42", it
> isn't capable of expressing the type "T*, but never nullptr".

Somehow I assumed that we always try to return std::unique_ptr<> for
owning pointers and references for non-owning pointers (at least in new
code), but perhaps that was too naive. :-)

If a never-nullptr pointer is returned, it would be really useful to
somehow state that in the function's signature. Are you aware of any way
to do that? There is C's __attribute__((nonnull)), but it only works on
function parameters, I think.

Regards,

Miklos