Re: set TESTDIR from perl rather than Makefile

On Mon, Feb 21, 2022 at 07:00:54AM -0500, Andrew Dunstan wrote:
> On 2/19/22 18:53, Justin Pryzby wrote:
> > On Sat, Feb 19, 2022 at 05:41:49PM -0600, Justin Pryzby wrote:
> >> I rebased and fixed the check-guc script to work, made it work with vpath
> >> builds, and cleaned it up some.
> > I also meant to also attach it.
> 
> This is going to break a bunch of stuff as written.
> 
> First, it's not doing the same thing. The current system sets TESTDIR to
> be the parent of the directory that holds the test. e.g. for
> src/bin/pg_ctl/t/001_start_stop.pl it's src/bin/pg_ctl in the build
> tree, not the 't' subdirectory. This patch apparently sets it to the 't'
> subdirectory. That will break anything that goes looking for log files
> in the current location, like the buildfarm client, and possibly some CI
> setups as well.

Yes, thanks.

I changed the patch to use ENV{CURDIR} || dirname(dirname($0)).  If I'm not
wrong, that seems to be doing the right thing.

> Also, unless I'm mistaken it appears to to the wrong thing for vpath
> builds:
> 
>     my $test_dir = File::Spec->rel2abs(dirname($0));
> 
> is completely wrong for vpaths, since that will place it in the source
> tree, not the build tree.
> 
> Last, and for no explained reason that I can see, the patch undoes
> commit f4ce6c4d3a, but only for msvc builds. Even if that's justified it
> appears to have no relevance to this patch.

Andres' idea is that perl should set TESTDIR and PATH.  Here I commented out
PATH, and had the improbable issue that nothing seemed to be breaking,
including the pipeline test under msvc.  It'd be helpful to know what
configuration that breaks so I can test that it's broken and then test that
it's fixed when set from within perl...

I got busy here, and may not be able to progress this for awhile.