RFR: 8257197: Add additional verification code to PhaseCCP [v4]

[Tobias Hartmann]

On Thu, 8 Dec 2022 08:46:14 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> **Targetted for JDK-21.**
>> 
>> We have had many bugs that could be tracked back to missing optimizations during PhaseCCP.
>> Often the problem is that a node `x` has an input (of input of input) `y` which is modified, but `y` does not notify `x` (does not push it to the worklist). For one this is a missed opportunity to optimize, but it can also lead to failed assumptions later: often we assume that all that can be optimize is already optimized.
>> This verification helps us debug faster, and can also help when `Value` optimizations suddenly do further-reaching traversals, which would require further-reaching notifications.
>> 
>> Sadly, the verification is not total: we have some exceptions. Especially for `LoadNode`, which perform a walk up the memory inputs, which can go arbitrarily far, to look for relates `StoreNode`s. Notification would thus have to put very many nodes on the worklist. The question is if the potential additional optimization is worth the compile-time. If we decided yes, then one we might want to implement a listener-style notification: when a node visits inputs during `Value`, it could subscribe to all (or the relevant) visited input-nodes for future updates. Currently, we just mostly do fixed 1-hop or 2-hop notification of output nodes.
>> 
>> FYI: I plan to do a similar verification, and a refactoring of `PhaseCCP::push_child_nodes_to_worklist` and `PhaseIterGVN::add_users_to_worklist` in [JDK-8298094](https://bugs.openjdk.org/browse/JDK-8298094).
>
> Emanuel Peter has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Apply suggestions from code review
>   
>   Thanks Christian
>   
>   Co-authored-by: Christian Hagedorn <christian.hagedorn at oracle.com>

Looks good!

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

Marked as reviewed by thartmann (Reviewer).

PR: https://git.openjdk.org/jdk/pull/11529