[llvm-testresults] Grawp nightly tester results

[Evan Cheng]

On Sep 5, 2009, at 7:30 PM, Daniel Dunbar wrote:

> On Sat, Sep 5, 2009 at 5:37 PM, Evan Cheng<evan.cheng at apple.com>  
> wrote:
>> On Sep 5, 2009, at 4:50 PM, Evan Cheng wrote:
>>> On Sep 5, 2009, at 4:45 PM, Daniel Dunbar wrote:
>>>> On Sat, Sep 5, 2009 at 4:11 PM, Evan Cheng<evan.cheng at apple.com>
>>>> wrote:
>>>>>
>>>>> On Sep 5, 2009, at 2:58 PM, Daniel Dunbar wrote:
>>>>>
>>>>>> On Fri, Sep 4, 2009 at 6:28 PM, Evan Cheng<evan.cheng at apple.com>
>>>>>> wrote:
>>>>>>>
>>>>>>> What JIT regressions? All of the following failures are crashing
>>>>>>> in the
>>>>>>> loop
>>>>>>> unswitching pass. Devang is looking into it.
>>>>>>
>>>>>> The performance regressions, a large number of JIT tests seem to
>>>>>> have
>>>>>> dropped by maybe an average of 5-10%.
>>>>>
>>>>> You're right. It looks real too. But none of the patches look
>>>>> suspicious.
>>>>
>>>> That's why I asked if you can remember if the JIT callback patch  
>>>> was
>>>> in this round or the previous one. Not sure if just changing the
>>>> compilation callback could account for these swings, but I don't
>>>> remember any other obvious JIT changes.
>>>
>>> No. It's not due to that change. I just tried r80904 on 253.perlbmk.
>>> It took 11.271s vs. TOT 18.325s.
>>
>> Ok, this doesn't really make sense to me, but numbers don't lie:
>>
>> 80923: 12.982s
>> 80926: 19.403s
>>
>> Dan, please take a look.
>
> Thanks for tracking it down. Just based on the commit comment it
> sounds like perhaps we were previously optimizing based on invalid
> loop information? I can see how that could make things go faster. :)

I dunno. If that's the case LLC performance would have been affected  
as well.

Evan

>
> - Daniel