Re: [PATCH] Remove symlinks created in argv0-symlink.exp and general cleanup

[Pedro Alves <palves at redhat dot com>]

On 08/11/2015 06:35 PM, Doug Evans wrote:
> On Tue, Aug 4, 2015 at 10:21 AM, Joel Brobecker <brobecker@adacore.com> wrote:
>>> Personally, I'd vote for making in-tree builds, or at least testing,
>>> officially unsupported.  I think gcc already does so...
>>
>> We should seriously discuss this, one day, as we keep having to
>> spend time fixing issues specific to that mode.
> 
> I don't have a strong opinion on making in-tree testing unsupported,
> but I do have a strong opinion on another thing that can make the
> in-tree testing issue moot.
> 
> Let's remove all of testsuite/*/Makefile.in, they're always
> out of date anyway (i.e., no one updates gdb.foo/Makefile.in to
> remove binaries added by new tests),
> and except for "make clean" aren't really used for anything.
> The "make clean" in testsuite/Makefile.in can just "rm -rf foo bar ..."
> [where "foo bar ..." are *not* the gdb.* testsuite dirs, but rather
> the outputs,etc. directories of check-parallel]
> IOW, testing would create the needed directories on the fly,
> even in serial mode,
> and to simplify "make clean" they'd always be put in a fixed
> subdir of testsuite (just like check-parallel does now).

*nod*

I believe Yao had patches a while ago that went in this direction.
Don't know what happened to them.  [+Yao].

> 
> As for whether to always have one directory per test
> (in serial and parallel modes), that *could* be treated as a separate
> issue, but if it reduces complexity by doing the same thing
> for serial and parallel then great.

I believe this would be worthwhile.

Taking this a step further, if we always ran in parallel mode (the
equivalent of make check -j1 FORCE_PARALLEL=1), then variable bleeding
between tests would no longer be an issue, as then we'd invoke a
dejagnu/runtest per test.  That would mean making GNU Make a
requirement for testing (which IMO, should be OK).

> [And while in-tree testing could still be unsupported,
> I think(!) the current issue with it would be fixed.]

Thanks,
Pedro Alves