RFR: Eliminate LoadAddr edge from LoadReferenceBarrier

[Roman Kennke]

>> I've had some observations about Phis though: It seems we see the case
>> fairly often that an input load gets split through phis, but the
>> corresponding LRB is not. And then we sometimes get values on one path
>> of the phi that would allow to optimize away the LRB, e.g. constants or
>> allocations. Would it be worth to maybe also split the LRB through phis
>> and enable to optimize them on paths that don't need them?
> 
> Splitting LRBs through phi makes sense to me. Then the only LRBs that
> would be left would have a load as input, right? And figuring the
> load address would be straightforward.

Ok.


I wonder if it makes sense to pursue that and the safepointing issue
(which seems important), or go for late barrier insertion to begin with,
which would solve both. I have no idea about the complexities of either
path, and also if late insertion has any new problems of its own. I'll
leave that to you ;-)

Thanks,
Roman