On 01/10/2013 07:48 PM, Aleksey Shipilev wrote:
On 01/10/2013 09:15 PM, Chris Hegarty wrote:
On 01/10/2013 05:05 PM, Aleksey Shipilev wrote:
On 01/10/2013 08:40 PM, Chris Hegarty wrote:
Doug, Aleksey,
I updated the appropriate methods in the Atomic classes to use the
instinsics defined by 7023898 , Unsafe getAndAddInt, getAndSetInt,
getAndAddLong, getAndSetLong, getAndSetObject.
http://cr.openjdk.java.net/~chegar/8006007/webrev.00/webrev/
Good. Two comments:
a) Any java-concurrency-torture [1] failures for these classes?
Can you give me a brief introduction to running these? I have run the
JDK regression tests and the appropriate JCK tests, all pass.
Build it, run it, see results/index.html. Should be 100% pass rate. If
not, drill down to exact tests.
Maven has just finished downloading the dependencies to build this
project! ;-) All tests pass.
You can probably remove the *.atomic.*V8 source and tests once these
changes are integrated.
b) Can we delegate all the suitable methods to Unsafe directly, without
calling the middleman (i.e. getAndDec() -> getAndAdd() -> unsafe), as in
[2]?
Yes, we could. The existing implementation was not consistent.
I took the view that this was not performance critical, since some
existing methods already delegate, and my preference, for simplicity, is
for the middleman ;-) Do you think there is a perf benefit to changing
this, or is this a style issue?
Yeah, that's mostly stylistic issue. If that's not in Doug's repo, you
can just disregard this. (There is a tempting desire to not to blow up
the call tree to help inliner, since the delegating method is not private).
I'll leave it as is, I find it much less error prone. We can revisit if
necessary.
Thanks,
-Chris.
-Aleksey.