Re: abi-compliance-checker

On 30.05.23 06:32, Peter Eisentraut wrote:
> I think the way to use this would be to compute the ABI for the .0 
> release (or rc1 or something like that) and commit it into the tree. And 
> then compute the current ABI and compare that against the recorded base 
> ABI.
> 
> Here is the workflow:
> 
> # build REL_11_0
> abidw src/backend/postgres > src/tools/postgres-abi-REL_11_0.xml
> # build REL_11_20
> abidw src/backend/postgres > src/tools/postgres-abi.xml
> abidiff --no-added-syms src/tools/postgres-abi-REL_11_0.xml 
> src/tools/postgres-abi.xml

Here is a demo patch that implements this.

Right now, I have only added support for libpq and postgres.  For 
completeness, the ecpg libraries should be covered as well.

(Unlike the above example, I did not use src/tools/ but each component's 
own subdirectory.)

The patch as currently written will actually fail the tests because it 
contains only one base ABI file to compare against, but the build_32 
task on cirrus will of course produce a different ABI.  But I left it 
for now to able to see the results.  Eventually, the base ABI file names 
should include something from host_system.cpu_family().

Also, I commented out the abidiff test for postgres, because the base 
file is 8 MB and I don't want to send that around.


Various findings while playing with these tools:

* Different Linux distributions produce slightly different ABI reports. 
In some cases, symbols like 'optarg@GLIBC_2.17' leak out.

* PostgreSQL compilation options affect the exposed ABI.  This is 
perhaps expected to some degree, but there are some curious details.

* For example, --enable-cassert exposes additional symbols, and it's 
maybe not impossible for those to leak into an extension.

* Also, --with-openssl actually *removes* symbols from the ABI (such as 
pg_md5_init).

So it's probably not sensible to try to get some universal ABI 
definition that works everywhere.  Instead, I think it would be better 
to get one specific case working, which would be the one tested on the 
cirrus linux tasks and/or some equivalent buildfarm machine.