[13] RFR(XS):8226287: Make process_users_of_allocation handle gc barriers

[Nils Eliasson]

New version:

The replace is unconditional like you suggested, and then the rest 
happens in an assert block.

I also improved it to handle multiple barriers. That required a new 
barrierset method that you will have to implement in Shenandoah, but the 
base implementation should do unless you have hit this problem too. I 
also added a check that the barrier traversal makes progress to catch 
any instances of missing implementations.

http://cr.openjdk.java.net/~neliasso/8226287/webrev.02

Regards,

Nils



On 2019-06-26 16:09, Roland Westrelin wrote:
>> _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control)) should always
>> work.
>>
>> I added the checks to be able to keep the
>> "assert(init->in(TypeFunc::Control) == _fallthroughcatchproj,
>> "allocation control projection");"
>>
>> Mostly because it could be nice to catch any other malformed control
>> flow here. And from the beginning I was a bit unsure if there where
>> other code paths that assumes the
>> alloc->proj->catch->catchproj->initialize pattern.
>>
>> I can remove it if you want to. It definitely gets easier to read.
> It makes sense now. I think it would be better to not make:
>
> _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
>
> conditional but modify the assert:
>
> assert(init->in(TypeFunc::Control) == _fallthroughcatchproj ||
> (bs->is_gc_barrier_node(init->in(TypeFunc::Control ...)
>
> Otherwise it seems that sometimes we don't remove the init barrier but
> it's unclear why. Also a comment would be nice.
>
> Roland.