RFR (M): 8000805: JMM issue: short loads are non-atomic

[Aleksey Shipilev]

Please double-check. Reproduces perfectly with 32-bit build on my x86_64
Linux:

$ ~/Install.32/jdk7u4/bin/java -client -version
java version "1.7.0_04"
Java(TM) SE Runtime Environment (build 1.7.0_04-b20)
Java HotSpot(TM) Client VM (build 23.0-b21, mixed mode)

$ ~/Install.32/jdk7u4/bin/java -client -jar
target/concurrency-torture.jar  -t ".*ShortAtomicity.*"
Java Concurrency Torture Tests
---------------------------------------------------------------------------------
Running in forked mode...

Running
net.shipilev.concurrent.torture.tests.primitiveAtomicity.ShortAtomicityTest
Warmup .....
                     Observed state  Occurrences          Expectation
Interpretation
                            [0, 0]  (    878998)             REQUIRED
                           [-1, -1] (  10025286)             REQUIRED
                           [-1, 0]  (     15596)            FORBIDDEN

Running
net.shipilev.concurrent.torture.tests.primitiveAtomicity.VolatileShortAtomicityTest
Warmup .....
                     Observed state  Occurrences          Expectation
Interpretation
                             [0, 0] (   5322930)             REQUIRED
                           [-1, -1] (   1774531)             REQUIRED
                            [-1, 0] (     13709)            FORBIDDEN


-Aleksey.


On 10/20/2012 12:37 AM, Vladimir Ivanov wrote:
> Disregard my comment about C1.
> 
> I vaguely remember I saw it fails and CR has a comment by Vladimir K.
> about C1 being also affected, but I'm not able to reproduce the problem
> anymore.
> 
> Best regards,
> Vladimir Ivanov
> 
> On 10/19/12 10:39 PM, Vladimir Ivanov wrote:
>> http://cr.openjdk.java.net/~vlivanov/8000805/webrev.00/
>> 282 lines changed: 255 ins; 27 del; 0 mod
>>
>> Ideal transformations during parsing & IGVN may rematerialize loads in
>> order to reify useful information (like signed/unsigned load). Such
>> behavior breaks JMM - instead of a single atomic load, multiple loads
>> are performed violating consistency of loaded data.
>>
>> The fix is to disable such transformations at all, but perform relevant
>> optimizations during matching.
>>
>> It fixes only C2 part of problem. С1 has similar deficiency, but it'll
>> be addressed separately.
>>
>> Testing: failing test, test for new matching rules, JPRT, CTW (x86, x64
>> & sparc).
>>
>> Best regards,
>> Vladimir Ivanov