RFR 8202201: All oop stores in the x64 interpreter are treated as volatile when using G1
david.holmes at oracle.com
Mon Sep 17 04:55:10 UTC 2018
Wow. Have you run any performance tests to see what we've gained by
eliding all the unnecessary "lock addl" psuedo-mfence?
On 15/09/2018 1:44 AM, coleen.phillimore at oracle.com wrote:
> Summary: ran out of registers, generated volatile and non-volatile
> open webrev at http://cr.openjdk.java.net/~coleenp/8202201.01/webrev
Generally looks good. Your webrev seems to have picked up some noise
from Mikael's unused Label changes ??
One query. Can you comment on the change to the 32-bit code where we no
longer use MO_RELAXED for the store? I'm assuming this is just a
potential optimization in the case that the store might be ordered
somehow by default. Though why this would be 32-bit specific I don't know.
> bug link https://bugs.openjdk.java.net/browse/JDK-8202201
> Tested with applications/jcstress, tier1&2 and jvmti tests (sometimes
> they filled up interpreter fixed space).
> putfield bytecode size is 2176 bytes now, 1248 bytes before
> putstatic bytecode size is 1344 bytes now, 864 before
> fast_xputfield is same 96 bytes.
More information about the hotspot-runtime-dev