Re: Optimize Arm64 crc32c implementation in Postgresql

On 04/04/18 14:13, Thomas Munro wrote:
> On Wed, Apr 4, 2018 at 9:23 PM, Heikki Linnakangas <hlinnaka(at)iki(dot)fi> wrote:
>> Pushed, thanks everyone!
>
> On eelpout two test_decoding tests failed:
>
> test ddl ... FAILED
> test rewrite ... FAILED
>
> The output has several cases where pg_logical_slot_get_changes() is
> invoked and then fails like this:
>
> SELECT data FROM pg_logical_slot_get_changes('regression_slot',
> NULL, NULL, 'include-xids', '0', 'skip-empty-xacts', '1');
> ! ERROR: incorrect resource manager data checksum in record at 0/1BD6600
>
> Not sure why it would break when called by pg_logical_slot_get_changes()...
>
> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=eelpout&dt=2018-04-04%2009%3A58%3A56

Yep. I got the code before the main loop, to handle the first 1-7
unaligned bytes, wrong. Apparently those are the only tests that call
the CRC function with very short and unaligned input.

I've got a fix ready, and I'm running "make check-world" on my ARM box
now, to make sure there aren't any more failures. I'll push the fix once
that's finished. Should've run the full test suite before pushing in the
first place..

- Heikki