bug#50349: [PATCH] packages: Add 'define-package' syntax.

[Sarah Morgensen <iskarian@mgsn.dev>]

Hi Taylan,

Taylan Kammer <taylan.kammer@gmail.com> writes:

> On 04.09.2021 16:44, Tobias Geerinckx-Rice wrote:
>> Taylan Kammer 写道：
>>> To me the most obvious thing to do seems
>>>
>>>   (define-package foo ...)  ;no explicit name needed
>>>
>>> to bind the variable 'foo' and use symbol->string for the name of the
>>> package, with the possibility to override the name like
>>>
>>>   (define-package foo (name "foobar") ...)
>>>
>>> which would bind the variable 'foo' to a package named "foobar".
>> 
>> Right, that's what I meant, and it's how I read bug #15284, and it looks remarkably like the form I use in my personal channels (and I'm sure I'm not the first! :-).
>> 
>> You're much better at the language/implementation side of things than I am, Taylan.  Would this negatively affect performance (including ‘guix pull’ compilation)?
>> 
>> Kind regards,
>> 
>> T G-R

I agree; if we added that magic, that's definitely how it should be.

>
> I'm flattered, but don't really know the answer, so I decided to attempt
> some sort of benchmark. :-P
>
> test1.scm uses the syntax-case macro, test2.scm just define-public.
>
> I don't actually import the Guix modules, so the (package ...) form isn't
> macro-expanded, regardless of whether it's used directly or results from
> expanding define-package.
>
> This way, the impact of define-package should dominate the time difference.
>
> The results are... interesting.  I started out with 256 definitions in the
> files, and the define-package one would take about 4.2s to compile while
> the regular one about 3.9s.  There was some jitter in the results though
> after running it several times so I thought, let's increase the number of
> packages to reduce noise.
>
> With 512 packages, the one using regular define-public takes a whole
> minute to compile, whereas the define-package one just ~14 seconds!
>
> So no idea what's going on there, and how the use of this macro in the
> actual (gnu packages ...) modules would affect performance. :-)

Thanks for running some benchmarks. Were both those latter runs with a
warm cache?

If so, is it possible that due to a compilation optimization, many of
the global symbol lookups only happen once with the define-package
macro?

If you were really interested, I suppose you could test with compilation
optimization off, but I'd be more interested in the performance impact
with (guix packages) imported.

--
Sarah