RFR: JDK-8215482: check for cycles in type variables can provoke NPE

[Maurizio Cimadamore]

As per my suggestion to split into multiple stages, it was mostly a 
suggestion to take full advantage of the tiered architecture of 
TypeEnter. You are basically caching some types in a 'todo later' list, 
and then you are coming back at them - avoiding these kind of queuing is 
what the (relatively) new TypeEnter code is for.

If the cyclic check is moved to a later phase, then I think you just 
need to fetch the type variables from the symbol/type/tree and check 
them; the types will be already set, no need to stash them into a map. 
At least in theory :-)

Maurizio

On 11/01/2019 18:05, Vicente Romero wrote:
> I have realized that cloning the type variable is not necessary as the 
> Pair<JCTypeParameter, TypeVar> element in the table will keep a 
> reference to the type variable
>
> Vicente
>
>
> On 1/11/19 12:29 PM, Vicente Romero wrote:
>>
>>
>> On 1/11/19 12:21 PM, Maurizio Cimadamore wrote:
>>>
>>> It seems that another thing that your path is doing is storing all 
>>> the type vars to check keyed by outermost class, so that only when 
>>> you have finished entering the outermost symbol you actually start 
>>> checking all pending typevars for cycles. I guess this delay is 
>>> necessary otherwise you would hit a problem anyway when checking Bc 
>>> for cyclicity?
>>>
>>
>> correct
>>
>>> Have you consider moving the attribution and cyclicity check in 
>>> different type enter phases? For instance, leave attribution in 
>>> HeaderPhase, but move the cycle check in MembersPhase ?
>>>
>>
>> I didn't try, but why would this be preferable? also as annotation 
>> processing can nuke the type we would have to make sure that the 
>> types are there before doing the cycle check which is something we 
>> can guarantee now by keeping the check for cycles in the Header phase
>>
>>> Maurizio
>>>
>>
>> Vicente
>>
>>> On 11/01/2019 16:47, Vicente Romero wrote:
>>>> Please review the fix for [1] at [2]. The NPE showed up in code like:
>>>>
>>>> class Outer<A extends Outer.Inner, B> {
>>>>     class Inner<Bc extends B> {}
>>>> }
>>>>
>>>> here attribution of type variable `A` during type enter phase will 
>>>> trigger attribution of class `Inner` while type variable `B` hasn't 
>>>> been attributed yet and thus its bound is still set to null. A 
>>>> similar problem arose a while ago see [3]. The issue provoking the 
>>>> current bug is that checks for cycles in type variables are done as 
>>>> soon as the type variable is attributed but in cases like the one 
>>>> above we can't do that until the type variable for the outer class 
>>>> has been attributed too.
>>>>
>>>> My first try was to create a fixup table that maps the outer type 
>>>> to the list of type variables defined by it or its members that 
>>>> happen to be types too, and once the compiler finish entering the 
>>>> outer class it would be safe to check for cycles in all the 
>>>> concerning type variables. I had a mild success here as there were 
>>>> some trivial cases that were working before that started failing. I 
>>>> realized that it was because the annotation processing phase was 
>>>> setting all the types to null, no bueno. So I decided to clone the 
>>>> type variables to be stored in the fixup table and do the cycle 
>>>> check on the clones which is what the current patch is doing.
>>>>
>>>> Thanks,
>>>> Vicente
>>>>
>>>> [1] https://bugs.openjdk.java.net/browse/JDK-8215482
>>>> [2] http://cr.openjdk.java.net/~vromero/8215482/
>>>> [3] https://bugs.openjdk.java.net/browse/JDK-6660289
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.java.net/pipermail/compiler-dev/attachments/20190111/ea593a33/attachment.html>