Meson vs. Solaris
I wondered why Greg Burd's new buildfarm animal "icarus" hasn't
yet managed to complete a successful run. The symptoms suggest
that getopt() isn't doing what we want:
/scratch/build/HEAD/pgsql.build/tmp_install/scratch/build/HEAD/inst/bin/postgres: illegal option -- cluster-name=node1
Try "postgres --help" for more information.
But we have other BF animals running on OpenIndiana (Solaris)
without trouble.
On investigation, it seems that Greg is the first to try to use
the meson build system on Solaris, and it doesn't work. Our
Autoconf infrastructure knows that we should use our own getopt()
because Solaris's doesn't do what we want. meson.build did try
to copy that logic:
always_replace_getopt = host_system in ['windows', 'cygwin', 'openbsd', 'solaris']
However, as seen elsewhere in that file, 'solaris' is not the
value of host_system on this platform! It's 'sunos'. So we
are trying to use the system's getopt(), and then the failure is
completely expected.
I've verified on a local OpenIndiana VM that the attached patch
makes it work. Also, a quick grep doesn't find any other cases
of the same error in our meson.build files.
regards, tom lane
Attachments:
solaris.patchtext/x-diff; charset=us-ascii; name=solaris.patchDownload+1-1
On Sat, Feb 7, 2026, at 4:21 PM, Tom Lane wrote:
I wondered why Greg Burd's new buildfarm animal "icarus" hasn't
yet managed to complete a successful run.
Hi Tom!
Wow, thanks for taking a minute to dig into this. I have been putting off digging into my "animals" for too long, apologies for that. I had mentioned to Thomas that this was annoying and I needed to dig into it ASAP, but neither of us had great ideas off the top of our head so the fact that this was immediately recognizable to you is amazing.
The symptoms suggest that getopt() isn't doing what we want:
/scratch/build/HEAD/pgsql.build/tmp_install/scratch/build/HEAD/inst/bin/postgres:
illegal option -- cluster-name=node1
Try "postgres --help" for more information.But we have other BF animals running on OpenIndiana (Solaris)
without trouble.On investigation, it seems that Greg is the first to try to use
the meson build system on Solaris, and it doesn't work.
I tend to be a glutton for punishment of this variety...
Our
Autoconf infrastructure knows that we should use our own getopt()
because Solaris's doesn't do what we want. meson.build did try
to copy that logic:always_replace_getopt = host_system in ['windows', 'cygwin', 'openbsd',
'solaris']However, as seen elsewhere in that file, 'solaris' is not the
value of host_system on this platform! It's 'sunos'. So we
are trying to use the system's getopt(), and then the failure is
completely expected.I've verified on a local OpenIndiana VM that the attached patch
makes it work. Also, a quick grep doesn't find any other cases
of the same error in our meson.build files.regards, tom lane
I'll find a bit of time to validate this on icarus ASAP. Thank you for the patch. I really have a sentimental/irrational attachment to Sun/SPARC/Solaris... er Illumos/OpenIndiana/whatever (says the ex-Sun employee).
I'll have you know that I'm likely to fool around with swapping out GCC for Clang 22 once the build starts working, just because... glutton for punishment. :)
best.
-greg
PS: greenfly and unicorn are on this weeks list for me too.
Show quoted text
diff --git a/meson.build b/meson.build index df907b62da3..96b3869df86 100644 --- a/meson.build +++ b/meson.build @@ -2911,7 +2911,7 @@ gnugetopt_dep = cc.find_library('gnugetopt', required: false) # (i.e., allow '-' as a flag character), so use our version on those platforms # - We want to use system's getopt_long() only if the system provides struct # option -always_replace_getopt = host_system in ['windows', 'cygwin', 'openbsd', 'solaris'] +always_replace_getopt = host_system in ['windows', 'cygwin', 'openbsd', 'sunos'] always_replace_getopt_long = host_system in ['windows', 'cygwin'] or not cdata.has('HAVE_STRUCT_OPTION')# Required on BSDs
Attachments:
* solaris.patch
Hi,
On 2026-02-07 16:21:31 -0500, Tom Lane wrote:
I wondered why Greg Burd's new buildfarm animal "icarus" hasn't
yet managed to complete a successful run. The symptoms suggest
that getopt() isn't doing what we want:/scratch/build/HEAD/pgsql.build/tmp_install/scratch/build/HEAD/inst/bin/postgres: illegal option -- cluster-name=node1
Try "postgres --help" for more information.But we have other BF animals running on OpenIndiana (Solaris)
without trouble.On investigation, it seems that Greg is the first to try to use
the meson build system on Solaris, and it doesn't work. Our
Autoconf infrastructure knows that we should use our own getopt()
because Solaris's doesn't do what we want. meson.build did try
to copy that logic:always_replace_getopt = host_system in ['windows', 'cygwin', 'openbsd', 'solaris']
However, as seen elsewhere in that file, 'solaris' is not the
value of host_system on this platform! It's 'sunos'. So we
are trying to use the system's getopt(), and then the failure is
completely expected.I've verified on a local OpenIndiana VM that the attached patch
makes it work. Also, a quick grep doesn't find any other cases
of the same error in our meson.build files.
Ugh. At some point I had solaris working, but apparently I broke it again at
some point. Thanks for noticing & fixing.
Greetings,
Andres Freund