RFR: 8234735: InstanceKlass:find_method_index regression after JDK-8231610
Claes Redestad
claes.redestad at oracle.com
Mon Dec 9 15:06:22 UTC 2019
Coleen, Ioi,
thank you for reviewing!
/Claes
On 2019-12-09 15:37, coleen.phillimore at oracle.com wrote:
> +1
> Coleen
>
> On 11/25/19 1:50 PM, Ioi Lam wrote:
>> Hi Claes,
>>
>> This change looks good to me.
>>
>> Thanks!
>> - Ioi
>>
>> On 11/25/19 5:51 AM, Claes Redestad wrote:
>>> Hi,
>>>
>>> JDK-8231610 refactored InstanceKlass::binary_search to ::quick_search,
>>> which now has a linear slow-path version in case the VM is currently
>>> dumping the dynamic CDS archive. This cause the methods to grow enough
>>> that ::quick_search is not inlined into find_method_index, which cause a
>>> local regression.
>>>
>>> Since the linear search is not as performance sensitive as the default
>>> binary search, then outlining the slow-path to a different method makes
>>> sense. Tests show that this allows the performance critical path to be
>>> completely inlined again, recuperating about half of the performance
>>> regression (still adds a branch to check for whether we should do a
>>> linear search to the fast path, compared to pre-8231610)
>>>
>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8234735
>>> Webrev: http://cr.openjdk.java.net/~redestad/8234735/open.00/
>>>
>>> Testing: tier1-3
>>>
>>> Thanks!
>>>
>>> /Claes
>>
>
More information about the hotspot-runtime-dev
mailing list