[concurrency-interest] a volatile bug?
Aleksey Shipilev
aleksey.shipilev at gmail.com
Wed May 16 14:22:17 PDT 2012
Well, I do not want to sound alarming, but... if I understand the C1
code correctly, then C1 GVN does not account prior volatile reads at
all. I can not find any code in C1 GVN code which actually prevents
killing second non-volatile read after volatile one, which is required
by JMM semantics.
I think I'll stop here. The impact of this issue is limited, given
most of the guys run -server (even by default on most machines), so
there is always the workaround for running with -server. Also, I would
*speculate* turning off GVN with -XX:-UseGlobalValueNumbering when
running with -client is still a workaround, but kind of insane one,
since it can *severely* degrade performance.
Words of wisdom: I'm using this command-line to print out GVN tracing:
$ ~/Install/jdk7u4/fastdebug/bin/java -XX:+PrintCompilation
-XX:+PrintDominators -XX:+PrintCompilation -XX:+PrintValueNumbering
-Xbatch -XX:CompileOnly=Test.$1,Test -client Test 2>&1 | tee asm.log
Can anyone more proficient in C1 code confirm this?
-Aleksey.
On Thu, May 17, 2012 at 12:41 AM, Aleksey Shipilev
<aleksey.shipilev at gmail.com> wrote:
> In my case, there are always two compiled versions for Test$1.run, one
> with cached $b, second one is with correct read for $b. I'd guess
> pastebin version had the second one.
>
> -Aleksey.
>
> On Thu, May 17, 2012 at 12:34 AM, Vitaly Davidovich <vitalyd at gmail.com> wrote:
>> I looked at the assembly on SO again (the pastebin link) and it seems to be
>> correct actually: after 'a' is cmp'ed against zero, 'b' is read from
>> memory. But now someone is saying there that it sometimes generates the
>> correct assembly and other times not - very strange.
>> 0x025bd2b9: cmp $0x0,%edx
>>
>> 30. 0x025bd2bc: je 0x025bd2a8 ;
>>
>> 32. 0x025bd2be: mov $0x147062e8,%edx ; {oop('test/TestVolatile')}
>>
>> 33. 0x025bd2c3: mov 0x1c4(%edx),%edx ;*getstatic b
>>
>> 34. ; - test.TestVolatile::run at 10 (line 17)
>>
>> 35. 0x025bd2c9: cmp $0x0,%edx
>>
>> Sent from my phone
>>
>> On May 16, 2012 3:55 PM, "Aleksey Shipilev" <aleksey.shipilev at gmail.com>
>> wrote:
>>>
>>> Update. GVN is clearly under suspicion -XX:-UseGlobalValueNumbering
>>> mitigates the bug in my setup. Digging through C1 codebase to see
>>> rules for volatiles.
>>>
>>> -Aleksey.
>>>
>>> On Wed, May 16, 2012 at 11:23 PM, Aleksey Shipilev
>>> <aleksey.shipilev at gmail.com> wrote:
>>> > All right, here's what is on the table.
>>> >
>>> > This bug is reproduced for me on Linux i686 with:
>>> > 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)
>>> >
>>> > It reproduces immediately only with -client.
>>> > Both -server and -Xint do NOT reproduce the bug.
>>> > The code is there in original SO post
>>> >
>>> > http://stackoverflow.com/questions/10620680/why-volatile-in-java-5-doesnt-synchronize-cached-copies-of-variables-with-main
>>> >
>>> > C1 seems to miscompile run(), and indeed does CSE for local:
>>> >
>>> > # {method} 'run' '()V' in 'Test$1'
>>> > [Verified Entry Point]
>>> > 0xb4a91e80: mov %eax,-0x4000(%esp)
>>> > 0xb4a91e87: push %ebp
>>> > 0xb4a91e88: sub $0x18,%esp ;*invokestatic access$000
>>> > ; - Test$1::run at 0 (line 11)
>>> > 0xb4a91e8b: mov $0xa09c4270,%edx ; {oop(a 'java/lang/Class' =
>>> > 'Test')}
>>> >>>>>>> 0xb4a91e90: mov 0x74(%edx),%edx ;*getstatic b <<<<<----
>>> >>>>>>> loads $b to %edx
>>> > ; - Test::access$000 at 0 (line 1)
>>> > ; - Test$1::run at 0 (line 11)
>>> > 0xb4a91e93: jmp 0xb4a91e9e ; OopMap{off=40}
>>> > ;*goto
>>> > ; - Test$1::run at 10 (line 13)
>>> > 0xb4a91e98: test %eax,0xb77a9100 ;*goto
>>> > ; - Test$1::run at 10 (line 13)
>>> > ; {poll}
>>> > 0xb4a91e9e: mov $0xa09c4270,%ecx ; {oop(a 'java/lang/Class' =
>>> > 'Test')}
>>> >>>>>> 0xb4a91ea3: mov 0x70(%ecx),%ecx ;*getstatic a <<<<<
>>> >>>>>> volatile read for $a
>>> > ; - Test::access$100 at 0 (line 1)
>>> > ; - Test$1::run at 4 (line 13)
>>> > 0xb4a91ea6: cmp $0x0,%ecx // <---- $a is at %ecx
>>> > 0xb4a91ea9: je 0xb4a91e98 ;*ifne
>>> > ; - Test$1::run at 7 (line 13)
>>> > >>>> 0xb4a91eab: cmp $0x0,%edx // <<<<<<---- $b is cached in
>>> > %edx here
>>> > 0xb4a91eae: jne 0xb4a91ed8 ;*ifne
>>> > ; - Test$1::run at 16 (line 17)
>>> > 0xb4a91eb4: nopl 0x0(%eax)
>>> > 0xb4a91eb8: jmp 0xb4a91f0e ; {no_reloc}
>>> > 0xb4a91ebd: xchg %ax,%ax
>>> > 0xb4a91ec0: jmp 0xb4a91f28 ; implicit exception:
>>> > dispatches to 0xb4a91f18
>>> > 0xb4a91ec5: nop ;*getstatic out
>>> > ; - Test$1::run at 19 (line 18)
>>> > 0xb4a91ec6: cmp (%ecx),%eax ; implicit exception:
>>> > dispatches to 0xb4a91f32
>>> > 0xb4a91ec8: mov $0xa09c6488,%edx ;*invokevirtual println
>>> > ; - Test$1::run at 24 (line 18)
>>> > ; {oop("error")}
>>> >
>>> >
>>> > Thanks,
>>> > Aleksey.
>>> >
>>> > On Wed, May 16, 2012 at 11:12 PM, Vitaly Davidovich <vitalyd at gmail.com>
>>> > wrote:
>>> >> It can be a compiler (mis)optimization that causes this, and not x86
>>> >> memory
>>> >> ordering.
>>> >>
>>> >> Someone posted the assembly output in the comments on SO and it does
>>> >> seem
>>> >> like there's a place that loads 'b' from the stack rather than memory.
>>> >> Hans' theory of CSE sounds plausible - can someone repro this without
>>> >> that
>>> >> "int tt = b;" line?
>>> >>
>>> >> Adding hotspot compiler guys in case they want to chime in.
>>> >>
>>> >> Sent from my phone
>>> >>
>>> >> On May 16, 2012 3:07 PM, "Aleksey Shipilev"
>>> >> <aleksey.shipilev at gmail.com>
>>> >> wrote:
>>> >>>
>>> >>> On Wed, May 16, 2012 at 10:40 PM, Boehm, Hans <hans.boehm at hp.com>
>>> >>> wrote:
>>> >>> > A JDK bug AND a serious test suite omission?
>>> >>>
>>> >>> Stress tests would probably JIT-compile the code in question. See
>>> >>> below.
>>> >>>
>>> >>> > But is the problem real? Can it be reproduced on a mainstream JVM?
>>> >>>
>>> >>> Same question.
>>> >>>
>>> >>> > Note that the example in the original posting also read b before the
>>> >>> > loop,
>>> >>> > so naïve common subexpression elimination would cause the bug.
>>> >>> > Hopefully
>>> >>> > nobody does CSE in cases like this.
>>> >>>
>>> >>> FWIW, the test case in SO would probably not hit any compilation
>>> >>> threshold in HotSpot, so it could be executed in interpreter. Then,
>>> >>> assuming the interpreter does not reorder Java code, and assuming
>>> >>> original SO poster runs Windows, and hence x86, and hence has TSO,
>>> >>> this bug seems very unlikely. I would be surprised if it actually
>>> >>> *can* be reproduced. That makes the whole story rather interesting.
>>> >>>
>>> >>> -Aleksey.
>>> >>>
>>> >>> _______________________________________________
>>> >>> Concurrency-interest mailing list
>>> >>> Concurrency-interest at cs.oswego.edu
>>> >>> http://cs.oswego.edu/mailman/listinfo/concurrency-interest
More information about the hotspot-compiler-dev
mailing list