Move to 4.0.3 breaks ABI compatability ?

[amckinstry]

This is an issue i'm seeing on Debian unstable, which has recently moved to 4.0.3

Our regression tests show a failure with mpigrafic:

mpgrafic: symbol lookup error: mpgrafic: undefined symbol: ompi_f08_mpi_double_precision

This is with a move from 4.0.2 -> 4.0.3:
https://ci.debian.net/data/autopkgtest/testing/amd64/m/mpgrafic/4897659/log.gz

This appears to be because symbol(s) were removed:

51e23f8#diff-8c7fd97572800eac0b55c84f29ee9451

I think the unused symbols were not as unused as you expected.

[gpaulsen]

Added blocker label to block a v4.0.x release until we understand this, and hopefully fix this.

[gpaulsen]

@jsquyres said he'd take a look at this. We probably just want a solution on the v4.0.x branch, and leave master as-is for upcoming v5.0.

FYI: v4.0.x PR looks like this came into v4.0.x on #7116

[ggouaillardet]

@amckinstry I am now double checking in my environment.

FWIW, the commit you are pointing is only for the master branch.
In order to workaround a flang bug/feature, some constants were revamped (e.g. moved from a BIND(C, ... symbol to a parameter), so that clearly broke the ABI and we are aware of it.
That being said, extra care was taken when back porting this to the v4.0.x branch.
Newly compiled programs will use the parameter, but the previous symbols were kept not to break the ABI.

What if you

nm libmpi_usempif08.so | grep double_precision

could the symbol be there but not visible?
(Debian's linker is way stricter than the one in RedHat I mainly use)

[ggouaillardet]

the root cause is ompi/mpi/fortran/use-mpi-f08/base/constants.c has been removed from Makefile.am, so the symbols are no more slurped into the libmpi_usempif08.so library.

It seems that was caused by subsequent backport(s), I am now working on a fix.

[ggouaillardet]

@amckinstry you can manually download and apply the patch at https://github.com/open-mpi/ompi/pull/7622.patch.
If you apply it on the v4.0.3 tarball, you will also need to run ./autogen.pl --force

[jsquyres]

Thanks @ggouaillardet

[jsquyres]

@amckinstry The fix for this was merged in #7622. Can you try the latest nightly v4.0.x snapshot to make sure we got it right?

https://www.open-mpi.org/nightly/v4.0.x/

[gpaulsen]

@amckinstry, Thanks. Based on the sounds of it, this should possibly drive an earlier v4.0.4 release, would you agree?

[amckinstry]

The patch appears to be missing constants.c (not present in 4.0.3) ? Adding constants.[ch] makes the patch work.
Meanwhile my test case has disappeared, as mpgrafic in its latest version (0.3.18) had changed to avoid using the symbols. No urgency on at 4.0.4 ?