Re: [PATCH RFC 0/4] Getting rid of input labels?

[Ludovic Courtès]

Hello,

Nicolas Goaziou <mail@nicolasgoaziou.fr> skribis:

> Ludovic Courtès <ludo@gnu.org> writes:

[...]

>>   • Packages such as ‘tzdata’ use labels to refer to non-package
>>     inputs.  These cannot be converted to the automatic labeling
>>     style, or not without extra changes.
>
> Would it be possible to write something like
>
>   (inputs (let ((tzcode (origin ...)))
>             (list ... tzcode ...)))
>
> ?

Yes, but the problem is that the automatically-assigned label for
<origin> records is “_” (a placeholder), because origins have no name,
unlike packages.

Thus, this phase:

    (replace 'unpack
       (lambda* (#:key source inputs #:allow-other-keys)
         (invoke "tar" "xvf" source)
         (invoke "tar" "xvf" (assoc-ref inputs "tzcode"))))

… needs to be written differently, because there’s no “tzcode” label.

One option on ‘core-updates’ is to use gexps:

    #~(modify-phases %standard-phases
        ;; …
        (replace 'unpack
           (lambda* (#:key source inputs #:allow-other-keys)
             (invoke "tar" "xvf" source)
             (invoke "tar" "xvf" #$tzcode))))

However, this style breaks common uses of ‘inherit’ and uses of the
inputs field: ‘tzcode’ here need not even be listed in ‘inputs’, and
consequently, you cannot easily inherit from ‘tzdata’ and give it a
different ‘tzcode’.

We need to find and encourage appropriate idioms for corner cases like
this.  One option is the status quo: keep using labels in those rare
cases.

A crazier option would be to interpret input lists, when possible, as
both input lists and formal parameter lists for ‘arguments’.  Assume a
package with:

  (inputs (list foo bar baz))

The ‘arguments’ field, which is currently a thunk, would magically be
turned into:

  (lambda (foo bar baz)
    …)

That way, arguments could refer to #$foo etc., and that’d just do the
right thing when inputs are overridden.

This would be reaching a level of magic that may not be wise, and it may
be hard to implement.

> Could the new syntax accept both variables and specifications, e.g.,
>
>    (list "glib:bin" foo "bar@2.3")
>
> ?

No!  I mean, yes it could, but no, I don’t think that’s a good idea.
:-)

In terms of API, I prefer clarity; in this case, I think inputs should
be a list of packages or other “lowerable” objects, rather than a list
of “anything” that could be magically interpreted at run time.

More importantly, I think package specs are a UI feature, and that
packages should be separate from UI concerns.  Specs are implemented in
(gnu packages), not in (guix …) for a reason: it’s a feature that makes
sense in the distro, not in the core Guix machinery where there’s no
“package set” to traverse.

I hope that makes sense!

Ludo’.