[fpc-devel] Nested function closures

[Sven Barth]

Am 27.04.2021 um 17:44 schrieb Ryan Joseph via fpc-devel:
> Continued from our discussion at https://bugs.freepascal.org/view.php?id=24481.
>
>> if the compiler devs will allow me as soon as this is finished I want to allow the existing nested functions functionality to work with anonymous functions, so at the very least we don't need to generate the expensive interface based object which often times is not even needed. At that point we would need to make nested functions inline-able, which they are currently not. But we're not there yet so lets not complicated anything by proposing extensions to a feature that doesn't even exist yet.
> Sven replies:
>
>> Getting rid of the interface only works in very narrow circumstances that are so seldom in real world code that it is not worth the effort.
> I'm referring to my test I did a few years ago (https://github.com/graemeg/freepascal/compare/master...genericptr:anon_funcs) where I say we can use existing nested functions as a closure when passing is not required. As you can see I already implemented this quite easily but it is not related to the new forthcoming closures feature. I did in fact try to replace the interface with a record on the old closures branch but I ran into many problems I decided it wasn't the best route.
>
> Indeed there are many times where we don't want a heap allocated interface you can pass around  but rather a simple inline function pointer like below. Consider this loop is run 60 times a second and allocating a useless class every time for no gain. This could easily be 1000*60=60,000 constructions and allocations of a class.
>
>    for i := 0 to entities.Count - 1 do
>      begin
>        value := entities[i];
>        value.SortEntities(function(a, b: TEntity): integer
>          begin
>            // do stuff
>          end
>        );
>      end;
>
>
> So anyways what I propose is if a closure is never passed outside of scope (i.e. temporary) then use anonymous nested functions instead (like in my GitHub branch). If this is an acceptable approach I will personally do what is required to get it implemented along side the real closures.
As soon as *any* function is passed to a "reference to 
procedure/function" it *must* be an interface, because that's how 
"reference to procedure/function" is internally implemented. Anything 
that's calling a "reference to procedure/function" is expecting to call 
an interface method, thus it *must* be an interface. Also if you capture 
anything then that *must* be contained in a capture context, because 
whatever you pass that function reference to might store that for later 
calling and then the stack context might be long gone. The only 
situations where the compiler might optimize this is if it's inside the 
same function, maybe inside the same implementation section of the unit 
or possibly if WPO is involved (with a dedicated WPO pass), but those 
are complex optimizations.

Also your example is wrong, cause it will not create an interface for 
each loop iteration. Instead the pseudo code essentially looks like this:

=== code begin ===

procedure Foo;
type
   ISort = interface
     function Invoke(a, b: TEntity): Integer;
   end;
   TCaptureObject = class(TInterfacedObject, ISort)
     function Invoke(a, b: TEntity): Integer;
   end;

function TCaptureObject.Invoke(a, b: TEntity): Integer;
begin
   // do stuff
end;

var
   context: TCaptureObject;
begin
   context := TCaptureObject.Create;

   for i := 0 to entities.Count - 1 do
      begin
       value := entities[i];
       value.SortEntities(ISort(context));
      end;
end;

=== code end ===

If you capture variables they'll be part of TCaptureObject instead of 
the stack.

Regards,
Sven