RFR(xs): 8153876: Replace 4K stack allocations with Resource allocations

[David Holmes]

On 4/06/2016 9:10 AM, Calvin Cheung wrote:
> Hi Harold, David,
>
> I've updated the webrev with ResourceMark and reverted the changes in
> metaspaceShared.cpp.
> http://cr.openjdk.java.net/~ccheung/8153876/webrev.01/

Looks good.

Thanks,
David

>
> On 6/3/16, 12:06 PM, harold seigel wrote:
>> Hi Calvin,
>>
>> Methods that allocate memory using NEW_RESOURCE_ARRAY and do not have
>> ResourceMarks probably want to return the allocated memory to their
>> caller.  I think that most of the places where you added
>> NEW_RESOURCE_ARRAY allocations also should have ResourceMarks.
> Except for ClassPathDirEntry::open_stream(). The buffer will be added to
> the ClassFileStream but the path is a local variable. So I've added
> FREE_RESOURCE_ARRAY for path in several places.
>>
>> ResourceMarks should be added to classLoader.cpp at lines 346, 511,
>> and 753.  If THREAD is available then please add the ResourceMarks as
>> ResourceMark rm(THREAD);.
> Yes, they've been added in the updated webrev.
>
> I'll rerun some tests.
>
> thanks,
> Calvin
>
>>
>> Thanks, Harold
>>
>>
>> On 6/3/2016 2:06 PM, Calvin Cheung wrote:
>>> Hi David,
>>>
>>> Thanks for your review.
>>>
>>> On 6/2/16, 6:05 PM, David Holmes wrote:
>>>> Hi Calvin,
>>>>
>>>> On 3/06/2016 9:13 AM, Calvin Cheung wrote:
>>>>>
>>>>> Please review this small fix for replacing some of the array
>>>>> declarations with the resource allocation macros.
>>>>
>>>> If you add a resource allocation don't you also need to ensure there
>>>> is a ResourceMark to clean it up again?
>>> I looked at some of the existing usage and none of them has a
>>> ResourceMark.
>>> One example is in Arguments::build_resource_string().
>>>
>>>>
>>>> And what is the overhead of using resource allocation versus stack
>>>> allocation? A lot of this seems to be in startup code.
>>> Since my fix touches the ctw_visitor() in classLoader.cpp, I tried to
>>> use the -XX:+CompileTheWorld to see if there's any difference in perf
>>> before and after the fix.
>>>
>>> I modified the function so that it only ran a limited number of
>>> iterations. For a full run, it took more than 7 hrs.:
>>> CompileTheWorld : Done (41453 classes, 518964 methods, 25723180 ms)
>>>
>>> 100 iterations
>>> ------------
>>> no fix
>>> ======
>>> CompileTheWorld : Done (103 classes, 2096 methods, 66605 ms)
>>> CompileTheWorld : Done (103 classes, 2096 methods, 67171 ms)
>>> CompileTheWorld : Done (103 classes, 2096 methods, 66376 ms)
>>>
>>> with fix
>>> ========
>>> CompileTheWorld : Done (103 classes, 2096 methods, 66694 ms)
>>> CompileTheWorld : Done (103 classes, 2096 methods, 66358 ms)
>>> CompileTheWorld : Done (103 classes, 2096 methods, 66622 ms)
>>>
>>>
>>>
>>> 300 iterations
>>> ------------
>>> no fix
>>> ------
>>> CompileTheWorld : Done (321 classes, 4484 methods, 145322 ms)
>>> CompileTheWorld : Done (321 classes, 4484 methods, 144472 ms)
>>> CompileTheWorld : Done (321 classes, 4484 methods, 144229 ms)
>>>
>>> with fix
>>> --------
>>> CompileTheWorld : Done (321 classes, 4484 methods, 144748 ms)
>>> CompileTheWorld : Done (321 classes, 4484 methods, 145070 ms)
>>> CompileTheWorld : Done (321 classes, 4484 methods, 144773 ms)
>>>
>>> So the times are pretty much the same with and without the fix.
>>>
>>> Below is the disassembly of the NEW_RESOURCE_ARRAY:
>>>  511                  char* path = NEW_RESOURCE_ARRAY(char,
>>> JIMAGE_MAX_PATH);
>>> 00007ffff6736f7d:   mov $0x1000,%eax
>>> 00007ffff6736f82:   mov $0x0,%esi
>>> 00007ffff6736f87:   mov %rax,%rdi
>>> 00007ffff6736f8a:   callq 0x7ffff7079ca0
>>> <_Z23resource_allocate_bytesmN17AllocFailStrategy13AllocFailEnumE>
>>> 00007ffff6736f8f:   mov %rax,-0x10(%rbp)
>>>>
>>>>> bug: https://bugs.openjdk.java.net/browse/JDK-8153876
>>>>> webrev: http://cr.openjdk.java.net/~ccheung/8153876/webrev.00/
>>>>
>>>> src/share/vm/memory/metaspaceShared.cpp
>>>>
>>>> This change seems unnecessary - we don't have stackoverflow concerns
>>>> when doing a dump.
>>> I was thinking the same. I will revert the change.
>>>
>>> thanks,
>>> Calvin
>>>>
>>>> Thanks,
>>>> David
>>>>
>>>>> Testing:
>>>>>     JPRT
>>>>>     non-colo "quick" tests
>>>>>
>>>>> thanks,
>>>>> Calvin
>>