[9] RFR(S): 8072398: assert fails in L1RGenerator::increment_event_counter_impl

[Vladimir Kozlov]

Good.

Thanks,
Vladimir

On 2/6/15 2:56 AM, Zoltán Majó wrote:
> Hi,
>
>
> the previous webrev excluded a one-line change that is needed to make this work.
>
> Here is the new webrev: http://cr.openjdk.java.net/~zmajo/8072398/webrev.01/
>
> The new change is in arguments.cpp at
>
> *** 3919,3929 ****
> --- 3925,3936 ----
>
> This change is analogous to the change at
>
> *** 1202,1213 ****
> --- 1207,1219 ----
>
> in the previous webrev.
>
> Sorry for the extra effort.
>
> Thank you and best regards,
>
>
> Zoltan
>
>
>
> On 02/06/2015 10:57 AM, Zoltán Majó wrote:
>> Thank you, Vladimir and Igor, for the review! I'll push this change today.
>>
>> Best regards,
>>
>>
>> Zoltan
>>
>> On 02/05/2015 11:52 PM, Zoltán Majó wrote:
>>> Hi Igor,
>>>
>>>
>>> thank you for the feedback.
>>>
>>>
>>> On 02/05/2015 09:59 PM, Igor Veresov wrote:
>>>> The code looks good. But I think the comment about CompileThresholdScaling == 0 being equivalent to -Xint is bit
>>>> misleading. The VM does not become interpreter-only, does it?
>>>
>>> It does.
>>>
>>> The reason is that CompileThreshold==0 has been historically equivalent to -Xint (see lines 3488-3493 in
>>> arguments.cpp). CompileThresholdScaling==0.0 scales down CompileThreshold to 0, so we decided to keep functionality
>>> consistent and make CompileThresholdScaling equivalent to -Xint as well when we introduced the
>>> CompileThresholdScaling flag in 8059604.
>>>
>>> I hope this makes sense.
>>>
>>> Thank you!
>>>
>>> Best regards,
>>>
>>>
>>> Zoltan
>>>
>>>
>>>
>>>>
>>>> igor
>>>>
>>>>> On Feb 5, 2015, at 8:18 AM, Zoltán Majó <zoltan.majo at oracle.com> wrote:
>>>>>
>>>>> Hi,
>>>>>
>>>>>
>>>>> please review the following small patch.
>>>>>
>>>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8072398
>>>>>
>>>>> Problem: The reported assert is triggered because the method LIRGenerator::increment_event_counter_impl is called
>>>>> with parameter 'frequency' set to -1.
>>>>>
>>>>> The value originates from Arguments::scaled_freq_log(): If both variables 'freq_log' and 'scale' are small,
>>>>> 'scaled_freq' can become 0 and the function log2_intptr returns -1 for the value 0. A similar problem appears if
>>>>> freq_log == 0.
>>>>>
>>>>>
>>>>> Solution:
>>>>>
>>>>> To treat the above mentioned cases graciously, this patch:
>>>>>
>>>>> - Changes Arguments::scaled_freq_log() and Arguments::scaled_compile_threshold() to allow scaling the
>>>>> frequency/threshold by 0.0 and return 0.0 in these cases without calling log2_intptr().
>>>>>
>>>>> - Adds a new condition to Arguments::set_tiered_flags(): Threshold values are *not scaled* if
>>>>> 'CompileThresholdScaling' is 0.0 (behavior equivalent to -Xint). Previously, this functionality was encoded in
>>>>> scaled_freq_log() and scaled_compile_threshold(). Handling this case in set_tiered_flags() results in clearer code.
>>>>>
>>>>> - Improves comments.
>>>>>
>>>>>
>>>>> Webrev: http://cr.openjdk.java.net/~zmajo/8072398/webrev.00/
>>>>>
>>>>> Testing:
>>>>> - manual testing with failing test cases
>>>>> - compiler/arguments/CheckCompileThresholdScaling.java
>>>>> - JPRT
>>>>>
>>>>> Thank you and best regards,
>>>>>
>>>>>
>>>>> Zoltan
>>>>>
>>>
>>
>