On 11-4-2023 07:23, Hairy Pixels via fpc-devel wrote:
Strange question but I heard it brought up during a discussion and I was very
curious what compiler people think about it.
I hope you don't mind me replying as non compiler person :-)
The question is, instead of using a virtual method table would it be feasible
to use a case statement on the real class type and make a dispatch routine for
each method call?
For example assume “type” is the internal type of the class (where a VTable
would be) and a call to “animal.DoSomething()” is called which would internally
basically be this dispatch table:
case animal.type of
TDog: TDog(animal).DoSomething;
TCat: TCat(animal).DoSomething;
TMouse: TMouse(animal).DoSomething;
end;
This doesn't happen. There is no class that is TDog,Cat and mouse.
Usually a VMT governs the relation between parent and child, i.e. TDog
and TAnimal. The TDog nor the TAnimal implementation has no knowledge
of the other mammals.
The reason this was brought up was because virtual methods can’t be inlined and
this would allow that but I think there would be more overhead overall to have
all these case statements on every virtual method call.
Much more, keep in mind your example above doesn't even have parameters.
Keep in mind that the virtual call is fairly cheap to begin with, you
load the VMT from the instance references (first field), and then the
method pointer at an fixed offset relative to that. Two instructions on
x86 with its many addressing modes, only slightly more on non-x86, I
would guess, and no branches.
Your solution would also need to load some reference to the class type,
(the first instruction) and then a compare and branch for each case.
But the core problem is that you don't have an overview of all other
descendants in the compiler. Those can be fragmented over multiple
units, and then you touch the problem that whole program optimization
like de-virtualization tries to solve.
_______________________________________________
fpc-devel maillist - fpc-devel@lists.freepascal.org
https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
