RFR [10]: Move (Java)Thread::_gc_state to lower offset to optimize barrier fast-path encoding

[Roman Kennke]

Am 13.06.2018 um 13:05 schrieb Aleksey Shipilev:
> http://cr.openjdk.java.net/~shade/shenandoah/gcstate-low/webrev.01/
> 
> We get this automatically with sh/jdk, because there we reuse the "GC thread locals" area. We can do
> the same trick in our backports, by moving the _gc_state field closer to Thread header, where GC
> thread locals would be. This saves us 3 bytes per barrier fastpath check on x86.
> 
> WB check before:
> 
>      0: testb  $0x14,0x390(%r15)
>     +8: jne    0x00007ff85bbb8900
>    +10: ...
> 
> WB check after:
> 
>      0: testb  $0x14,0x1c(%r15)
>     +5: jne    0x00007f8a57bb887a
>     +7: ...
> 
> Testing: tier3_gc_shenandoah, eyeballing assembly
> 
> Thanks,
> -Aleksey
> 

Good idea! Do it!

Thanks,
Roman