Re: Windows regress fails (latest HEAD)

On 11/10/20 6:16 PM, Russell Foster wrote:
> On Tue, Nov 10, 2020 at 12:03 PM Ranier Vilela <ranier(dot)vf(at)gmail(dot)com> wrote:
>>
>> Em sex., 26 de jun. de 2020 às 08:21, Ranier Vilela <ranier(dot)vf(at)gmail(dot)com> escreveu:
>>>
>>> Em qui., 11 de jun. de 2020 às 10:28, Ranier Vilela <ranier(dot)vf(at)gmail(dot)com> escreveu:
>>>>
>>>> Em qui., 11 de jun. de 2020 às 10:01, Andrew Dunstan <andrew(dot)dunstan(at)2ndquadrant(dot)com> escreveu:
>>>>>
>>>>>
>>>>> On 6/11/20 8:52 AM, Ranier Vilela wrote:
>>>>>> Hi,
>>>>>> Latest HEAD, fails with windows regress tests.
>>>>>>
>>>>>> float8 ... FAILED 517 ms
>>>>>> partition_prune ... FAILED 3085 ms
>>>>>>
>>>>>>
>>>>>
>>>>> The first thing you should do when you find this is to see if there is a
>>>>> buildfarm report of the failure. If there isn't then try to work out why.
>>>>
>>>> Sorry, I will have to research the buildfarm, I have no reference to it.
>>>>
>>>>>
>>>>>
>>>>> Also, when making a report like this, it is essential to let us know
>>>>> things like:
>>>>>
>>>>> * which commit is causing the failure (git bisect is good for finding
>>>>> this)
>>>>
>>>> Thanks for hit (git bisect).
>>>>
>>>>>
>>>>> * what Windows version you're testing on
>>>>
>>>> Windows 10 (2004)
>>>>
>>>>> * which compiler you're using
>>>>
>>>> msvc 2019 (64 bits)
>>>
>>>
>>> Only for registry, if anyone else is using msvc 2019.
>>> I'm using latest msvc 2019 64 bits (16.6.0)
>>> Problably this is a compiler optimization bug.
>>> vcregress check with build DEBUG, pass all 200 tests.
>>
>> With the current HEAD, the regression float8 in release mode (msvc 2019 64 bits) is gone.
>> Maybe it's this commit:
>> https://github.com/postgres/postgres/commit/0aa8f764088ea0f36620ae2955fa6c54ec736c46
>>
>> But (partition_prune) persists.
>> partition_prune ... FAILED
>>
>> regards,
>> Ranier Vilela
>
> I am also experiencing this issue on one of my Windows machines (x64)
> using 12.4. I believe this is new, possibly since 12.2. It doesn't
> occur on another machine though, which is strange. It appears to be
> the same diff output. Is it possible that the given result is also
> valid for this test?
>

That's unlikely, I think. The regression tests are constructed so that
the estimates are stable. It's more likely this is some difference in
rounding behavior, for example. I wonder which msvc builds are used on
the machines that fail/pass the tests, and if the compiler flags are the
same.

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company