RFR: JDK-8203157: Object equals abstraction for BarrierSetAssembler

[Roman Kennke]

Am 11.06.2018 um 19:11 schrieb Andrew Haley:
> On 06/11/2018 04:56 PM, Andrew Haley wrote:
>> On 06/08/2018 09:17 PM, Roman Kennke wrote:
>>> Why is it better? And how would I do that? It sounds like a fairly
>>> complex undertaking for a special case. Notice that if the oop doesn't
>>> qualify as immediate operand (quite likely for an oop?) it used to be
>>> moved into rscratch1 anyway a few lines below.
>>
>> Sorry for the slow reply.  I'm looking now.
> 
> OK.  The problem is that this is a very bad code smell:
> 
>       case T_ARRAY:
>         jobject2reg(opr2->as_constant_ptr()->as_jobject(), rscratch1);
>         __ cmpoop(reg1, rscratch1);
> 
> I can't tell that this is correct.  rscratch1 is used by assembler
> macros, and I don't know if some other GC (e.g. ZGC) might need to use
> rscratch1 inside cmpoop.  The risk here is obvious.  The Right Thing
> to do IMO is to generate a scratch register for pointer comparisons.
> 
> Unless, I guess, we know that cmpoop never ever needs a scratch
> register for any forseeable garbage collector.
> 

I do know that Shenandoah does not require a tmp reg. I also do know
that no other collector currently needs equals-barriers at all. I cannot
see into the future. I prefer to be pragmatic and solve existing
problems. How about I add a comment to the obj_equals() API that says
'don't use tmp reg X, and if you really need to, push/pop it or let the
compiler generate one for you' ?

Roman