[gentoo-dev] Choosing GUI toolkits from multiple choices

[Mart Raudsepp]

Ühel kenal päeval, R, 03.06.2016 kell 22:40, kirjutas Daniel Campbell:
> You touched on the part that I'm most concerned about: choosing. If
> the
> 'GUI' USE_EXPAND gets in, do we maintainers check that variable and
> if
> there's no preference just build whatever? Will we not be expected to
> emit an ewarn or something similar to clarify *why* the package is
> being
> built a certain way? Granted, if a user has a problem and reports a
> bug,
> their make.conf's GUI variable should be present in emerge -v output,
> easily explaining the issue.

Said 'GUI' USE_EXPAND is outside the intended scope of the USE=gui
thread, but anticipating discussion will happen regardless now, I've
cut the thread and named it something else.
Many discussions have happened on IRC on this as well, which is where
Daniel got this from.


I think it's actually a rather corner-case to have an optional GUI, and
then that GUI being buildable against a selection of toolkits. Arguably
in some of those cases it might be more ideal to have the GUI parts in
a separate package too, but usually upstream sources aren't so
accommodating to our source-based case (while binary distributions
split it up into multiple binary packages, built from one source in one
go).
In most cases when there's a choice, a GUI imho isn't usually optional.
You choose e.g qt4 or qt5, or gtk3 or qt5 and having both shipped is
not so common from the same package. Transmission is an example where
it is, because they have a multiple frontend system (arguably it would
be neat to have these in separate transmission-qt, transmission-gtk and
so on packages), and one of these can be a web service UI instead of a
dedicated graphics toolkit.

USE=gui would replace a ton of USE=gtk's at least, where it's mostly
about simple extra GUI tools added with a gtk2 dep, but I've seen it in
other toolkit cases too, but I don't know how common it is there.
Ideally USE=gui can be agreed upon while ignoring the corner-case of
multiple choices; in some of these cases it might make sense to ignore
USE=gui at the start, until the multiple choice case gets some
resolution though, e.g emacs and perhaps transmission could just keep
the current way until we agree how to express multiple choice cases
universally.

I drafted that mentioned USE_EXPAND idea as a means to get some 'design
from the scratch' discussion going and flesh out this way of
potentially doing it (such a USE_EXPAND was idly mentioned at start by
others as something that was deemed too crazy, but I didn't find any
references). It is currently still as a draft over at
http://piratepad.net/iwvgjB1P5d -  but I didn't want the original
USE=gui thread to discuss this, as it's a separate and much more
complex matter really, and would work in tandem with USE=gui when
appropriate.

> It's the implementation that gets me here, not the idea. The idea
> could
> be neat and make Gentoo management easier at the expense of some
> (hopefully) minor ebuild bloat. Another issue that hasn't been
> covered
> well yet is how are we going to select DEPENDs? I was told DEPEND
> doesn't support exactly-one-of, and we don't want extraneous
> dependencies.

The part of it being not clear (due to the intentional lack of
REQUIRED_USE usage in that design) what is getting built is probably
the main drawback of that drafted idea, and some QA members tell me
this would be outright vetoed as a QA violation.

This uncertainty also echoes in the need for these complex DEPEND atoms
as well then, based on maintainer chosen preference, combined with user
set flags.

So it isn't ideal at all indeed, and we'd need to really do the
suggested EAPI/package manager improvements first, to express this
maintainer order of preference to the user and filter the flags somehow
to what will actually be chosen then before the DEPEND atoms get
processed, which would make the *DEPENDs feasible, as you could
simplify the conditionals, because the unused (but set by user) USE
flags are already filtered out then (or one added when user didn't have
any, and one would be chosen by default).

I ruled out REQUIRED_USE because I don't like it at all when it is used
together with common global USE flags, as opposed to just some local
flags. In my opinion it tends to results in users disabling or enabling
something globally, instead of locally for the package in question. And
with that having made the choice unknowingly for a ton of future
packages to be installed as well. pkg_pretend is a bit better, because
you can customize the error message to be readable, but it's still
something that takes away my choice of not having to care.. just give
me a GUI, preferably a gtk3 one, ok? thx.

> Transmission is a good example: supports gtk3, qt4, qt5. Let's say
> the
> maintainer prefers the qt5 version. Would we do this?:
> 
> DEPEND="gui_qt5? ( dev-qt/qtcore:5 ) gui_qt4? ( dev-qt/qtcore:4 )
> gui_gtk3? ( x11-libs/gtk+:3 )"
> 
> or this?:
> 
> DEPEND="gui_qt5? ( !gui_qt4? ( !gui_gtk3? ( dev-qt/qtcore:5 ) ) )"
> 
> (snipping because I don't feel like writing it all out)

Transmissions is a good example indeed, but not in the way you describe
here, because transmission can be built with gtk and qt in parallel. So
it's a good example of what a hypothetical EAPI-next markup for
describing this order of choices would have to support. qt4 or qt5 or
neither is one choice. gtk2 or gtk3 or neither another. And none of
them have to be there, so if USE=-gui, GUI="..." is ignored, but with
USE=gui and GUI="<empty>" a qt5 choice would be made.
I wonder if the same syntax as REQUIRED_USE can be used for expressing
this, sans order of priority (unless it can be implicit from the order
of things listed).

If we had some way to express this, similar to REQUIRED_USE, that part
could take care of simplifying the DEPEND to
"gui? (
 gui_qt5? ( dev-qt/qtgui:5 )
 gui_qt4? ( dev-qt/qtgui:4 )
 gui_gtk3? ( x11-libs/gtk+:3 )
 gui_gtk2? ( x11-libs/gtk+:2 ) )"
if it would override the flags to what will be chosen for us already at
this point via some markup like
GUI_PREFS="gui? ( || ( ?? ( gui_qt5 gui_qt4 )  ?? ( gui_gtk3 gui_gtk2 )
) )"
if we can default to the first of each branch of choice, or something
more complex to sequence these choices.

This could also probably be achieved one way or another via an eclass
(or maybe not - the rules of global variables like *DEPEND are rather
strict and I'm not sure you can influence them sufficiently via an
eclass), but that doesn't solve the issues of package manager not
really knowing what happened (apparently bad for binary packages), and
the user not knowing what's going on without reading the ebuild.

Brainstorming welcome on this, but a medium like IRC might be more
constructive at this point for that purpose.


Mart