Sporadic failure to build IO in the macOS CI job

[wilfwilson]

I first mentioned this #4372 (comment) and #4373 (comment).

Starting recently (I first noticed it today, I've seen it about four times so far) the Gather coverage data step sometimes fails in the macOS job (docomp testinstall - BOOTSTRAP_MINIMAL=yes - macos-latest), because it cannot load the profiling package. @ChrisJefferson has looked a little further and noticed that this is because IO fails to build. From #4373 (comment):

Looking at the full log, it's io which is failing to build. I'm not 100% sure what the issue is here:

==== Checking io
2021-04-08T11:49:56.3701800Z + mkdir -p gen
2021-04-08T11:49:56.3722930Z + autoconf -Wall -f
2021-04-08T11:49:56.6746330Z configure.ac:4: warning: The macro `AC_CONFIG_HEADER' is obsolete.
2021-04-08T11:49:56.6752650Z configure.ac:4: You should run autoupdate.
2021-04-08T11:49:56.6753690Z ./lib/autoconf/status.m4:719: AC_CONFIG_HEADER is expanded from...
2021-04-08T11:49:56.6754580Z configure.ac:4: the top level
2021-04-08T11:49:56.6756010Z configure.ac:14: warning: The macro `AC_HEADER_STDC' is obsolete.
2021-04-08T11:49:56.6757030Z configure.ac:14: You should run autoupdate.
2021-04-08T11:49:56.6757980Z ./lib/autoconf/headers.m4:704: AC_HEADER_STDC is expanded from...
2021-04-08T11:49:56.6758860Z configure.ac:14: the top level
2021-04-08T11:49:56.6760160Z configure.ac:24: warning: The macro `AC_HEADER_TIME' is obsolete.
2021-04-08T11:49:56.6761140Z configure.ac:24: You should run autoupdate.
2021-04-08T11:49:56.6762080Z ./lib/autoconf/headers.m4:743: AC_HEADER_TIME is expanded from...
2021-04-08T11:49:56.6763050Z configure.ac:24: the top level
2021-04-08T11:49:56.7688140Z + autoheader -Wall -f
2021-04-08T11:50:01.4933090Z Running './configure' '--with-gaproot=/Users/runner/work/gap/gap' 
2021-04-08T11:50:01.9169440Z configure: error: cannot find required auxiliary files: config.guess config.sub
2021-04-08T11:50:01.9748500Z Running 'make' 'clean' 
2021-04-08T11:50:01.9846550Z make: *** No rule to make target `clean'.  Stop.
2021-04-08T11:50:01.9948710Z Running 'make' 
2021-04-08T11:50:02.0050840Z make: *** No targets specified and no makefile found.  Stop.
2021-04-08T11:50:02.0152060Z 
2021-04-08T11:50:02.0253320Z WARNING: Failed to build io

Failure instances:

• https://github.com/gap-system/gap/runs/2295959507
• https://github.com/wilfwilson/gap/runs/2295971288
• https://github.com/wilfwilson/gap/actions/runs/729428204
• https://github.com/wilfwilson/gap/actions/runs/729054339
• https://github.com/wilfwilson/gap/actions/runs/728945155

[wilfwilson]

It actually seems to be very frequent, indeed I haven't managed to get this job to pass on #4372 for quite a while now. I'll also note that GAP is compiled without readline in the macOS job:

 ┌───────┐   GAP 4.12dev built on 2021-04-08 12:01:19+0000
 │  GAP  │   https://www.gap-system.org
 └───────┘   Architecture: x86_64-apple-darwin19.6.0-default64-kv8
 Configuration:  gmp 6.2.1, GASMAN, KernelDebug
 Loading the library and packages ...
 Packages:   GAPDoc 1.6.4, IO 4.7.0dev, PrimGrp 3.4.1, SmallGrp 1.4.2, 
             TransGrp 3.0
 Try '??help' for help. See also '?copyright', '?cite' and '?authors'
Architecture: x86_64-apple-darwin19.6.0-default64-kv8

(that's of course an example of a build where IO did actually compile).

[fingolfin]

This is caused by the fact that autoconf recently was update to 2.71 (after years of no activity at all), and apparently they decided it's a good idea to suddenly deprecate a ton of stuff. And homebrew now ships it. The fact that it only hits us sometimes likely is due to difference in the various macOS VMs used by GitHub; I assume the issue will happen more and more often...

I can work on a fix for io, but it is likely that similar issues will affect more packages...

[wilfwilson]

Ah, thank you for diagnosing this @fingolfin.

Temporarily, we could remove the code coverage from the Mac build - we would lose coverage data for Mac-specific code, but (if such a job doesn’t already exist) we could replicate the rest of the coverage in another Ubuntu job. That would stop IO from needing to be built. What do you think?

[ChrisJefferson]

I tried temporarily fixing this based on https://bugs.ruby-lang.org/issues/17723 ( which said run autoreconf -f -i). That actually seemed to fix it on my machine -- still lots of warnings, but at least it built. In https://github.com/ChrisJefferson/gap/tree/fix-autoconf271 (only change is in BuildPackages.sh, where I add:

    # (Hopefully) temporary fix for Autoconf 2.71
    if [[ $(which autoreconf) && ! -x config.sub ]]
    then
      autoreconf -f -i
    fi

[fingolfin]

I believe this issue has been resolved by my changes to io. Can it be closed?