RFR: 8015831: Add lint check for calling overridable methods from a constructor [v5]

[Maurizio Cimadamore]

On Wed, 11 Jan 2023 14:12:46 GMT, Maurizio Cimadamore <mcimadamore at openjdk.org> wrote:

>> Actually I think it would take 1 + 1 + 3 iterations.
>> 
>> During the first two iterations of the outer loop, nothing changes after the first go round of the inner loop - i.e., the total set of possible references in existence does not change, because all of the assignments in the inner loop won't involve any 'this' references.
>> 
>> It's only the during the third iteration of the outer loop that any 'this' references seep into any of the variables seen by the inner loop. Then it will take 3 cycles for the reference set to converge again.
>> 
>> However, this is not to say that there aren't some pathological examples out there. I guess the question is could they exist in "normal" code.
>
> True - probably 3 * 3 can be achieved if this:
> 
> 
>     ThisEscapeLoop ref21 = ref14;
> 
> Is replaced with   
> 
> 
>     ThisEscapeLoop ref21 = this;
> 
> 
> In which case the inner loop won't converge immediately (as it will have to propagate from ref21 to ref22 to ref23 to ref24).
> 
> I guess what I'm uncomfortable with is that we have effectively unbounded computation here (especially when we also consider the fact that the analysis "follows" method bodies as well, if they are found in the same compilation unit).
> 
> I suggest one experiment where you:
> 1. downgrade the warnings to notes (so that they won't make the JDK build fail)
> 2. enable this Lint everywhere
> 3. compare JDK `clean images` time w/ and w/o the Lint

Also, looking at the loop test more closely, it seems to me that what the compiler needs to do is to prove that there can be possible paths by which a `this` can land into ref4.

If we build a graph of all the assignments, we get:

ref4 <- ref3 <- ref2 <- ref1 <- this

So, if we ask "can ref4 possibly contain `this`?" we could "walk" the variable dependencies backwards and discover that, yes, there exist a possible path in which `this` would get there.

Now, of course without a loop this can never be a real issue (since you can never send a `this` fully down the chain) - but again, this is a question of how much effort should be spend to handle false negatives in what look like a pathological case.

-------------

PR: https://git.openjdk.org/jdk/pull/11874