RFR (S) 8216136: Don't take Compile_lock for SystemDictionary::_modification_counter

[coleen.phillimore at oracle.com]

After some internal discussion, Dean convinced me that removing the 
Compile_lock here might be too dangerous.   So for these asserts and the 
error condition, the compiler thread goes to VM from native to check the 
SystemDictionary::modification_counter under the Compile_lock, with 
safepoint checking always.

Tested with tier1,2,6 and 8.

open webrev at http://cr.openjdk.java.net/~coleenp/8216136.02/webrev
bug link https://bugs.openjdk.java.net/browse/JDK-8216136

Thanks,
Coleen

On 1/17/19 7:15 AM, coleen.phillimore at oracle.com wrote:
>
>
> On 1/16/19 10:53 PM, dean.long at oracle.com wrote:
>> Hi Coleen.  You still can't safely call notice_modification() outside 
>> of Compile_lock, (at least not without other changes), so this:
>>
>> - static inline void notice_modification() { 
>> assert_locked_or_safepoint(Compile_lock); ++_number_of_modifications; }
>> + static inline void notice_modification() { 
>> Atomic::inc(&_number_of_modifications); }
>>
>> should be:
>>
>> static inline void notice_modification() { 
>> assert_locked_or_safepoint(Compile_lock); 
>> Atomic::inc(&_number_of_modifications); }
>>
>>
>> Are you trying to eventually remove Compile_lock completely?  If so, 
>> then notice_modification() would have to be called *before* the
>> class hierarchy is changed, not after, and probably other changes 
>> would be needed as well.
> Dean, Thank you for looking at this and your comments.
>
> No, I'm not trying to remove Compile_lock entirely and I can assert 
> that notice_modification has the Compile_lock as above. The class 
> hierarchy code has been changed to be lock free rather than requiring 
> the Compile_lock, although I think the Compile_lock still protects 
> some of this code.
>
> There are also some Compile_lock free ways of getting to dependencies, 
> because putting notice_modification after flush_dependencies caused 
> bugs that I'll ask to you offline about.
>
> Thanks for your help.  I was just trying to peel off one place where 
> Compile_lock seemed wrong.
>
> Thanks,
> Coleen
>>
>> dl
>>
>>
>> On 1/16/19 8:43 AM, coleen.phillimore at oracle.com wrote:
>>> Summary: make SystemDictionary::modification_counter atomic so not 
>>> to require Compile_lock.
>>>
>>> I moved updating the modification counter when the class is defined 
>>> and added to the hierarchy.  I didn't remove the Compile_lock 
>>> completely because there may be other code currently under the lock 
>>> that needs it (flush_dependencies). Can someone from the compiler 
>>> area also review this?
>>>
>>> Made Compile_lock an always safepointing lock.
>>>
>>> Tested with mach5 tier1-6.
>>>
>>> open webrev at http://cr.openjdk.java.net/~coleenp/8216136.01/webrev
>>> bug link https://bugs.openjdk.java.net/browse/JDK-8216136
>>>
>>> Thanks,
>>> Coleen
>>
>