RFR: 8295486: Inconsistent constant field values observed during compilation

[Jatin Bhateja]

On Thu, 5 Jan 2023 13:26:11 GMT, Tobias Hartmann <thartmann at openjdk.org> wrote:

> We hit a "not monotonic" assert because the new type of a load from a stable final field is more narrow than the old type which contradicts the assumption that types should only go from TOP to BOTTOM during CCP:
> 
> old: `narrowoop: java/lang/Integer:BotPTR:exact *`
> new: `narrowoop: java/lang/Integer java.lang.Integer {0x000000062c41e548} ...`
> 
> or 
> 
> old: `narrowoop: java/lang/Integer java.lang.Integer {0x000000062c41e538} ...`
> new: `narrowoop: java/lang/Integer java.lang.Integer {0x000000062c41e548} ...`
> 
> The problem is that a stable field can be (re-)initialized during compilation and since the value is not cached, contradicting types can be observed. In `LoadNode::Value`, we re-read the field value each time: 
> 
> https://github.com/openjdk/jdk/blob/872384707e89d03ede655aad16f360dc94f10152/src/hotspot/share/opto/memnode.cpp#L1994-L1997
> 
> https://github.com/openjdk/jdk/blob/872384707e89d03ede655aad16f360dc94f10152/src/hotspot/share/opto/type.cpp#L332-L337
> 
> The same problem exists for loads from stable arrays:
> https://github.com/openjdk/jdk/blob/872384707e89d03ede655aad16f360dc94f10152/src/hotspot/share/opto/memnode.cpp#L1923
> 
> Caching the field value is not feasible as it would require a cache per ciInstance for all the fields and per ciArray for all the elements. Alternatively, we could keep track of the lookup and only do it once but that would also be lots of additional complexity for a benign issue.
> 
> Instead, I propose to skip verification during CCP when folding loads from stable fields. Non-stable, constant fields are not affected as `null` is a valid value for them and they would already be folded before CCP.
> 
> Thanks,
> Tobias

> > Stable field value should be added to method dependencies such that compilation becomes invalid if stable value changes later on.
> 
> It's too much for @stable. It's allowed to observe stale values when concurrent modifications happen. It would require dependency checking on every `@Stable` field or array element update which is infeasible for arrays (all array stores have to be intercepted).
> 
> On the other hand, caching a single value which is used across the whole compilation has only performance impact (whether non-null value is observed or not).

Agree, but there should be a way for method to deoptimize if stable field value changes past method registration. Only full proof solution is to enforce that stable field value does not change after single initialization apart from null to non-null transitions.

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

PR: https://git.openjdk.org/jdk/pull/11861