RFC: Slots on operations, and bundle filenames

[Faré]

On Wed, Jul 27, 2016 at 6:28 PM, Robert Goldman <rpgoldman at sift.net> wrote:
> Once upon a time in old ASDF, there was an idea that one might put slots
> on operations.  IIRC, this worked oddly, at best, because it wasn't
> clear how to propagate these slot values around in the plans that ASDF
> built.
>
> OTOH, we could really use the ability to add information to some of the
> ASDF operations.  For example, now that we have an IMAGE-OP, we really
> don't want the images hidden away in the .cache directory where they are
> hard to find.
>
> Justification: it's fine to have fasl files hidden away there, because
> the fasls are loaded by ASDF:LOAD-OP based on computed fasl file
> locations.  But if you want to load an image, you need to find that
> image *before* you have ASDF -- even before you have lisp -- so an
> invisible file, whose location is computed, is really inappropriate.
>
> Note also that by Faré's design principle -- the one who knows should
> specify -- this information cannot be somehow programmed into the system
> -- it must be supplied by the user.
>
> I'm inclined to think we could allow the user to set an output-file for
> some of the bundle operations with a keyword argument to OPERATE, and
> then jam the value of the keyword argument into the bundle-op for it to
> be written.
>
> This works, I think, because it's a trivial case.  If we try to put some
> properties on an operation and need them to be propagated around the
> plan, I think we just lose.
>
> Comments?
>
I generally agree with your concerns and solutions. A few thoughts:

* ASDF 3 kept from ASDF 1, for backward compatibility, the idea that
the arguments to OPERATE are passed as initarg to OPERATION. This is
completely backward. ASDF 1 was trying to conflate into the OPERATION
object many concerns, most of which ASDF 3 has moved to the new PLAN
object. The ASDF action traversal model actually depends on any slots
of OPERATION being *not meaningful*, only cosmetic at best, or global
constants that will never change, etc. That was true implicitly in
ASDF 1 and ASDF 3 made it somewhat more explicit. Because they don't
really work, about no one relies on operation slots anyway, except
that SLIME uses operation-forced for some obscure hack, despite my
repeated attempts to get them to stop (I never had a good rapport with
SLIME maintainers, except for João Távora who eventually left to fork
SLY). The operation slots were also preserved to maintain backward
compatibility with the old ECL bundle operations -- but a different
interface was provided (moving the slots to components instead, where
they belong), so hopefully anyone who still uses the old interface can
be told to just use the new one.

* If we altogether remove this non-working operation initargs thing,
then we free the keyword args of OPERATE (and its front-ends
load-system, etc.) for more useful things. Among them, a :output-file
and/or :output-files keyword could populate an asdf-cache entry for
the output-files of the requested action (pair of operation and
component) and select the output location.

* Since ASDF3, bundle operations currently magically use
:build-pathname if the requested operation is the :build-operation
(though even that is broken in some versions of ASDF3). If you provide
a better mechanism, this one should probably go away, since it fails
proper division of labor, indeed, and I don't think anyone relies on
it but a few of my own systems, that I will happily update.

So, once again, we are being held up by some unfortunate choices of
ASDF 1, and there's a transition ahead. My recommendation would be to
get SLIME to stop using operation-forced, make sure the ECL folks are
not using the old interface (unlikely since we renamed the bundle
operations recently; at least if they could update their code for
that, they presumably can update their code to put linker information
in components, not operations). Then, include these
backward-incompatible changes in a 3.2 branch.

—♯ƒ • François-René ÐVB Rideau •Reflection&Cybernethics• http://fare.tunes.org
If there's a book you really want to read but it hasn't been written yet,
then you must write it. — Toni Morrison [same with software, —♯f]