Re: [sigc] C++11 'auto' with sigc::mem_fun doesn't track object lifetime (potential crash)

[Andrejs Hanins <andrejs hanins ubnt com>]

Hi Paul,

On 03/02/2016 03:10 PM, Paul Davis wrote:
> On Wed, Mar 2, 2016 at 7:45 AM, Andrejs Hanins <andrejs hanins ubnt com <mailto:andrejs hanins ubnt com>> 
> wrote:
>
>     Hi,
>
>         Consider simple code below which results in Foo::Call method to be actually called for _deleted 
> instance_:
>
>         class Foo : public sigc::trackable  {
>             public:
>             void Call() { printf("Foo call\n"); }
>         };
>
>         Foo* inst = new Foo();
>         auto fun = sigc::mem_fun(inst, &Foo::Call); // <-- bad
>         //sigc::slot<void> fun = sigc::mem_fun(inst, &Foo::Call); // <-- good
>         delete inst;
>         fun(); // <--- ooops!
>
>
>         The documentation for mem_fun says that "...only if the object type inherits from sigc::trackable 
> <https://developer.gnome.org/libsigc++/stable/structsigc_1_1trackable.html> is the slot automatically 
> cleared...", but the problem is that result of mem_fun is not a slot, but a bound_mem_fun which for some 
> reason doesn't track referenced object lifetime even if it inherits from trackable. If result of mem_fun is 
> put explicitly into a slot, then everything works fine and Call() method is not called after instance is 
> freed.
>
>         For me it looks like a typical "shoot in the foot" for C++11 programs and should be indeed fixed. 
> Just in case - issue is reproducible in current master too.
>
>
> your expectations are too high. the behaviour of sigc::trackable as documented applies ONLY to a slot.
>
> a sigc::mem_fun is just a closure, aka a functor aka a functor object aka an invokable object. it has not 
> semantics beyond operator(). In C terms, it is really nothing more than:
>
>      struct {
>          SomeThing* instance_of_thing;
>          ReturnType (*function_taking_something)(SomeThing*);
>      };
>
> and operator() is just a call to function_taking_something (instance_of_thing); there are no magic 
> semantics here, no lifetime tracking. this is intentional (and correct).
>
> if you want the behaviour of sigc::trackable, then you MUST use sigc::slot.  there's no "shoot in the foot" 
> issue here.
That makes sense, but isn't documentation still confusing, as it mentions 'slot' when speaking about mem_fun. 
Slot is indeed present in the example code below mem_fun description, but it is not obvious which 'slot' the 
sentence is talking about. I'm not insisting on this, just a space for a potential improvement. Maybe change 
the docs to "Note that _in the example below_ only if the object..."
>  
> the same behaviour applies to boost::bind() for example, which also returns a functor/closure.