On 2025-04-25 10:57, Ludovic Courtès wrote:
> Hello,
>
> Nicolas Graves <ngra...@ngraves.fr> writes:
>
>> Actually the thunk was not necessary because args were already passed to
>> the build-bag procedure, and modules and imported-modules were already
>> used in every bag-build procedures, except for trivial and raw
>> build-systems.
>>
>> Patch should look like the one attached, overall pretty simple.
I'll probably additionally drop defaults in bag-build procedures, to
make it clear that the defaults are coming from <build-system>-modules
rather than resulting from default function argument.
>> However it does break the build-system API so channels that define a
>> build-system will have to update too.
>
> Maybe we the field could default to the empty list?
Yes, I was unsure what to do there.
The issue is that for dependent channels, it's probably better to have a
clear failure (e.g. "missing required #:imported-modules") rather than
having '() silently passed in modules and imported-modules.
There are only raw, trivial and channel that use '(), so IMO not a clear
requirement. I guess it depends on which error is clearer for dependent
channels. IMHO, it might be better to keep no defaults, as to not
silently override current dependent channels modules/imported-modules,
and making them fail with a clear error message instead.
I could use #:allow-other-keys instead of adding unused
(imported-)modules keys to functions, if it makes it clearer.
WDYT ?
>
>> I'll investigate now if we can simply do away without imported-modules
>> at all.
>
> Yes, we should definitely do that.
Actually I think the 68315 would be a requirement to do that properly
(brainstorming here).
Let's say we want some arbitrary way to extend imported modules on the
build-system level (i.e. not the package level). We already explored
two solutions :
- lower procedure properties
- build-system field
If we remove build-system field, I don't know how I can do the generic
build-system transformation in my use-case.
But with 68315 the returned value from the bag-build is a monadic gexp (rather
than a monadic derivation), I probably can do something like:
(define (make-new-lower old-lower other-args)
(lambda* args
(let ((old-bag (apply old-lower args)))
(bag
(inherit old-bag)
(build
(lambda* (name inputs #:key (outputs '("out"))
#:allow-other-keys #:rest rest)
(mlet %store-monad
((builder (apply (bag-build old-bag)
name inputs #:outputs outputs rest)))
(return
(with-imported-modules additional-imported-modules
#~(begin
(use-modules ,@additional-modules)
(more-code)
(any-wrapper
#$builder))))))))))
which is my use-case (and the reason why I needed those in the first
place).
I don't know how to do that kind of thing if we only drop the
default-imported-modules field without 68315. Am I missing something
here?
(The reason I need that is that I want guix local instantiate (see next
answer) to support a --git option and not drop the .git directory, but
still apply guix package-source patches/snippet. IIUC, I some
additional code thus some additional imported-modules to support that.)
>> + (default-modules build-system-default-modules)
>
> … and probably change this one to just ‘modules’.
OK.
>> 1) some patches are improvements independent of wherever I try to do with
>> partial builds. They can already be reviewed now independently from the
>> rest.
>
> “Partial build”, interesting. :-)
I actually have something working pretty well for trivial packages
without patches nor snippets. Workflow is, in any arbitrary directory:
$ guix local instantiate hello
$ cd hello* (I have to figure the proper chdir after gnu:unpack, and do
that in the previous call guix local instantiate).
$ guix local build hello
That would build hello locally (in said dir, and install each output in
./$output dir, with all guix-rich info: arguments, store-injected
shebangs etc), and keep compilation caches. There an additional caveat:
I drop the 'install-license-files phase, because it is likely dependent
on having the output in the store for some reason.
>> 2) some patches are standardizing improvements for the names :
>> - %XXX-build-system-modules --> %default-XXX-imported-modules
>> - %XXX-modules or %default-modules --> %default-XXX-modules
>> - is that a fine standardizing change or should I go through a GCD?
>
> Well, there was already a name change in this area that left me
> unconvinced and that we’re still adjusting to, many months later. So my
> advice would be to think twice before renaming bindings with a lot of
> users. A GCD might sound overkill but OTOH it could help think through
> the implications.
I'm not really that much into it either, just thought some
standardization is welcome based on the current state of build-systems.
I'll see if it's easy to separate patches where I create the
%default-XXX-modules (rarely used but still cleaner in the file) to be
applicable before the %default-XXX-imported-modules ones, which would
probably require a GCD, and are not a requirement in my use-case.
>> 3) some patches are about the core of 68315 (allowing monadic bag-build).
Read "monadic gexp" here.
--
Best regards,
Nicolas Graves
