Re: Dependency cycle issues when using a Gexp-based snippet

[Ludovic Courtès]

Hi Maxim,

Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:

>>> Attempting a suggested fix by Ludovic in that same conversation [0],
>>> namely, making the snippet field of the <origin> record a thunked one:
>>>
>>> modified   guix/packages.scm
>>> @@ -250,7 +250,8 @@ as base32.  Otherwise, it must be a bytevector."
>>>    (patches   origin-patches                       ; list of file names
>>>               (default '()) (delayed))
>>>
>>> -  (snippet   origin-snippet (default #f))         ; sexp or #f
>>> +  (snippet   origin-snippet
>>> +             (default #f) (thunked))              ; sexp or #f
>>>    (patch-flags  origin-patch-flags                ; list of strings
>>>                  (default '("-p1")))
>>
>> We should check what this change costs in CPU and memory, but it’s
>> probably worth it.  As Marius noted before, the snippets for
>> ungoogled-chromium and linux-libre are contrived because of this
>> limitation.  (Perhaps we can use ‘delayed’ instead of ‘thunked’.)
>
> What is the difference between delayed and thunked? Would a thunked
> capture the closure of its environment while delayed not?  Is the
> closure useful to access record-bound values such as the version field
> of a package?

‘Thunk’ uses an actual thunk (zero-argument procedure) that’s called
each time the field is accessed; ‘delayed’ uses a promise, which is
similar except that the result is memoized (info "(guile) Delayed
Evaluation").

> I checked the usage at compilation and run time, using the 'time'
> command (aliased to time+ on my system), and didn't find any meaningful
> difference whether the snippet is made a thunked or delayed field, or
> none (current situation):
>
> On current master:
>
> time+ make -j8
> 2436.29user 56.47system 14:29.36elapsed 286%CPU (0avgtext+0avgdata 
> 870828maxresident)k
> 5480inputs+405952outputs (71major+320522minor)pagefaults 0swaps
>
> time+ ./pre-inst-env guix package -A | wc -l
> 9.87user 0.24system 0:06.51elapsed 155%CPU (0avgtext+0avgdata 
> 281564maxresident)k
> 0inputs+0outputs (0major+25636minor)pagefaults 0swaps
> 14702

What would be interesting is a comparison of the performance of
‘package-derivation’, which can be done with something like:

  time guix build -d --no-grafts libreoffice pandoc

For memory consumption, try:

  GUIX_PROFILING=gc guix build -d --no-grafts libreoffice pandoc

>> +              (snippet
>> +               (with-imported-modules '((guix build utils))
>> +                 #~(begin
>> +                     (use-modules (guix build utils))
>> +                     ;; corelib uses bundled harfbuzz, md4, md5, sha3
>> +                     (with-directory-excursion "src/3rdparty"
>> +                       (for-each delete-file-recursively
>> +                                 (list "double-conversion" "freetype" 
>> "harfbuzz-ng"
>> +                                       "libpng" "libjpeg" "pcre2" "sqlite" 
>> "xcb"
>> +                                       "zlib")))
>> +
>> +                     (let ((coreutils #+(canonical-package coreutils)))
>> +                       (substitute* "configure"
>> +                         (("/bin/pwd")
>> +                          (string-append coreutils "/bin/pwd")))
>> +                       (substitute* "src/corelib/global/global.pri"
>> +                         (("/bin/ls")
>> +                          (string-append coreutils "/bin/ls"))))
>> +                     #t)))))
>>
>> Such substitutions are system-dependent; thus, they should be made in a
>> phase, not in a snippet.  Perhaps we’ll sidestep the issue altogether?
>> :-)
>
> Indeed.  I didn't consider this aspect well.  Apart from being
> inefficient (the sources of a package would be different for each
> system) it would still technically work, no?

It would work, but it’s “not the right place” for that, aesthetically.

(Note that when there’s a snippet, we get different derivations for each
system anyway.)

Thanks,
Ludo’.