%ld instead of %d for LogCompilation counts?

[Vladimir Kozlov]

I filed RFE and we will fix it in jdk9:

https://bugs.openjdk.java.net/browse/JDK-8067374

Thanks,
Vladimir

On 12/12/14 12:31 AM, Chris Newland wrote:
> Hi Vladimir,
>
> I've tested changing %g to %f in those 2 places on OpenJDK9 and it looks
> good:
>
> <parse method='824' uses='194423.000000' stamp='0.698'>
>
> <branch target_bci='63' taken='1' not_taken='20000' cnt='20001.000000'
> prob='4.99975e-05'/>
>
> No accuracy is lost and I can parse using Double.valueOf().longValue() to
> get a cardinal number for the count.
>
> Kind regards,
>
> Chris
>
>
> On Mon, December 8, 2014 18:53, Vladimir Kozlov wrote:
>> Thank you, Chris
>>
>>
>> What do you think about using %f as John suggested? Will it solve your
>> problem?
>>
>> Both cases may have a fractional part and with 'long' it will be lost.
>>
>>
>> Thanks,
>> Vladimir
>>
>>
>> On 12/8/14 3:56 AM, Chris Newland wrote:
>>
>>> Hi Vladimir,
>>>
>>>
>>> The believe the only other LogCompilation output where %g is used for a
>>>   cardinal number is this one?
>>>
>>> share/vm/opto/parse1.cpp:443:
>>>
>>>
>>> log->begin_head("parse method='%d' uses='%g'",
>>> log->identify(parse_method), expected_uses);
>>>
>>>
>>> Kind regards,
>>>
>>>
>>> Chris
>>>
>>>
>>> On Fri, December 5, 2014 23:34, Vladimir Kozlov wrote:
>>>
>>>> Hi,
>>>>
>>>>
>>>>
>>>> The output was produced by following lines:
>>>>
>>>>
>>>>
>>>> C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'
>>>> cnt='%g' prob='%s'", iter().get_dest(), taken, not_taken, cnt,
>>>> prob_str);
>>>>
>>>> 'cnt' is declared as 'float' and defined as:
>>>>
>>>>
>>>>
>>>> cnt = (taken + not_taken) / FreqCountInvocations;
>>>>
>>>> I think it is artifact from long time ago when FreqCountInvocations
>>>> was
>>>>
>>>>> 1. And to avoid integer overflow, so we want to keep variable as
>>>>> float.
>>>>>
>>>>
>>>> But I am fine to change log output to long by converting it to
>>>> (jlong)cnt and using JLONG_FORMAT format.
>>>>
>>>>
>>>>
>>>> Do you know other places where we use %g for counters?
>>>>
>>>>
>>>>
>>>> Thanks,
>>>> Vladimir
>>>>
>>>>
>>>>
>>>> On 12/5/14 1:09 PM, Chris Newland wrote:
>>>>
>>>>
>>>>> Hi,
>>>>>
>>>>>
>>>>>
>>>>> I've encountered some LogCompilation counters in 8u25 that are
>>>>> getting printed in exponential form in hotspot.log :
>>>>>
>>>>> <branch target_bci='20' taken='23867' not_taken='2386581'
>>>>> cnt='2.38658e+06' prob='0.00990148'/>
>>>>>
>>>>> This is harder to parse and loses accuracy because %d is used as a
>>>>> formatter in hotspot/src/share/vm/opto/parse2.cpp:
>>>>>
>>>>> if (C->log() != NULL) { C->log()->elem("branch target_bci='%d'
>>>>> taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); }
>>>>>
>>>>>
>>>>>
>>>>> Would it be possible to modify some or all of the LogCompilation
>>>>> format strings to use %ld so that large integers can be printed
>>>>> without exponents?
>>>>>
>>>>> I don't mind going through the OpenJDK motions to submit this as a
>>>>> patch if it's a sensible idea.
>>>>>
>>>>> Thanks,
>>>>>
>>>>>
>>>>>
>>>>> Chris
>>>>>
>>>>>
>>>>>
>>>>
>>>
>>>
>>
>
>