RFR: 8176593: Throwable::getStackTrace performance regression

[Claes Redestad]

Hi,

On 2017-03-14 19:20, Jiangli Zhou wrote:
> Hi Claes,
>
> Looks good! Thank you for making the changes.

Thanks!

>
> I just noticed you used 4-space for indentation at line 212 and 231. Hotspot coding style is 2-space for indentation.

I guess I'll never get used to that. Fixed. :-)

/Claes

>
> Thanks!
> Jiangli
>
>> On Mar 14, 2017, at 11:09 AM, Claes Redestad <claes.redestad at oracle.com> wrote:
>>
>> Hi Jiangli,
>>
>> On 2017-03-14 18:22, Jiangli Zhou wrote:
>>> Hi Claes,
>>>
>>> Thank you for looking into this! The changes look good to me. I have
>>> some small suggestions below.
>>>
>>> I’d suggest to add a comment explaining the
>>> java_lang_String::hash_code() call before lookup_shared(). For example,
>>> “shared table requires java_lang_String::hash_code()”.
>>
>> Sure!
>>
>>>
>>> How about changing the following hash_string() calls to
>>> AltHashing::murmur3_32() directly? That would avoid the second
>>> use_alternate_hashcode() check in hash_string().
>>>
>>> 206 if (use_alternate_hashcode()) {
>>> 207 hash = hash_string(name, len);
>>> 208 }
>>>
>>> 224 if (use_alternate_hashcode()) {
>>> 225 hashValue = hash_string(name, len);
>>> 226 }
>>
>> I'm not so sure manually inlining matters here, but I haven't inspected
>> the disassembly, but with a little refactoring and it arguably makes the
>> intent clearer:
>>
>> http://cr.openjdk.java.net/~redestad/8176593/hotspot.02/
>> http://cr.openjdk.java.net/~redestad/8176593/hotspot.01.02.diff/
>>
>> Thanks!
>>
>> /Claes
>>
>>>
>>> Thanks,
>>> Jiangli
>>>
>>>> On Mar 14, 2017, at 9:57 AM, Claes Redestad <claes.redestad at oracle.com
>>>> <mailto:claes.redestad at oracle.com>> wrote:
>>>>
>>>> Hi,
>>>>
>>>> current implementation of StringTable::lookup_shared computes hash
>>>> values twice even when the same hash algorithm is used, which stand out
>>>> in profiles as the primary cause for a performance regression on things
>>>> like Throwable::getStackTrace.
>>>>
>>>> By refactoring the logic to not compute hashes using a specific
>>>> algorithm more than once we recuperate the regression and actually beat
>>>> JDK 8 on reported microbenchmarks.
>>>>
>>>> Webrev: http://cr.openjdk.java.net/~redestad/8176593/hotspot.01/
>>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8176593
>>>>
>>>> Testing: RBT jdk-tier1,hs-runtime-pre-checkin,hs-tier2
>>>>
>>>> Thanks!
>>>>
>>>> /Claes
>>>>
>>>
>