[VM] Iterative drop of vm values

[fEst1ck]

Description

Implements a custom drop for values to prevent overflow when dropping deeply nested values.

How Has This Been Tested?

All existing vm tests.

Type of Change

• Bug fix

Which Components or Systems Does This Change Impact?

• Move/Aptos Virtual Machine

[musitdev]

It seems fine. My only remarks is that there can have some memory leak because ValueImpls can contain a ValueImpls that point to the initial ValueImpls, creating a reference cycle. So the ValueImpls are never dropped.
I don't know if it's possible in the usage but if yes the case can be avoid by using a new container enum and a struct with Weak ref like:

pub(crate) struct WeakValueImpls(Weak<RefCell<Vec<ValueImpl>>>);

enum Container {
  ...
   Parent(WeakValueImpls),
}

[fEst1ck]

@musitdev Thanks for pointing out. Cyclic values shouldn't be constructed given the Move type system. I'll see if it's worth the extra effort to prevent against this.

[fEst1ck]

I'll also investigate if closure dropping can be unbounded.

[seanyoung]

This code is tricky, it really needs a big fat comment and a test case.

Ideally, I would like to see a different fix. I do realise this might not be simple or might cost too much time. If we allow types to be nested 40,000 deep, then surely we will encounter all sorts of problems.

I think there should be some limit on the type depth rather than working around the problems of the same types in rust (e.g. recursion in Drop). I suspect there will be other problems For example, validator CPU cycles, 40,000 malloc's and free's are very costly too.

[musitdev]

Yes, if the depth can be limited, it will be better. In fact, how big trees can be constructed in this context?