JIT breaks PostGIS

From: Darafei "Komяpa" Praliaskouski <me(at)komzpa(dot)net>
To: PostgreSQL Developers <pgsql-hackers(at)lists(dot)postgresql(dot)org>
Subject: JIT breaks PostGIS
Date: 2018-07-21 20:14:47
Message-ID: CAC8Q8tLLP-_jSS6k8hz4jd2CsRWtRZ72sWk0zJJZ5Qv517AiJw@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

Hi,

Today I spent some time closing PostGIS tickets in preparation to Monday's
release.

One of the blockers, https://trac.osgeo.org/postgis/ticket/4125, was filed
by Postgres APT repository maintainer Christoph Berg who noticed a test
suite failure on Debian Stretch with Postgres 11.

Upon investigation we found:
- A build for Ubuntu Bionic installed on Debian Stretch passes the suite,
requiring llvm6;
- A build for Debian Stretch fails the suite on a call to external library
GEOS, showing no traces of JIT in the stacktrace;
- Setting jit=off lets the suite pass;
- The query called in clean session by itself does not crash Postgres.
Queries above it are required to reproduce the crash;
- The crash affects not only Stretch, but customly collected Postgres 11 /
clang 3.9 on Travis CI running Ubuntu Trusty:
https://github.com/postgis/postgis/pull/262.

I suspect that a fix would require to bisect llvm/clang version which stops
showing this behavior and making it a new minimum for JIT, if this is not a
symptom of bigger (memory management?) problem.

Thank you!
--
Darafei Praliaskouski
Support me: http://patreon.com/komzpa

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Tom Lane 2018-07-21 20:19:51 Re: Possible performance regression in version 10.1 with pgbench read-write tests.
Previous Message Charles Cui 2018-07-21 20:08:10 Re: project updates