On Fri, Mar 19, 2010 at 10:01:05AM +0000, Ciaran McCreesh wrote:
> On Fri, 19 Mar 2010 02:56:08 -0700
> Brian Harring <ferringb@...> wrote:
> > > We are waiting on ABI dependencies (and extended support for
> > > multiple ABIs in package manager), which will provide some needed
> > > functionality.
> > 
> > You can do it now w/out waiting on ABI dependencies- I'm not saying 
> > the dependencies would be pretty, but it's doable to get abi level 
> > depdencies per slotting via expanding out the use combinations.
> > 
> > Note that's a step beyond what's in place now- converting over to the 
> > ruby abuse of USE_EXPAND hands over better control to users now w/
> > the same dep gurantees.
> > 
> > So... yeah, it's not reliant on EAPI.  An EAPI extension *would* make 
> > it easier, but it's not required to do it (especially since the deps 
> > are already autogenerated to a decent degree).
> 
> How would do it and deal with existing packages not having the required
> things in IUSE without (+)/(-) use deps? I don't see a way of doing it
> legally without those.

Roughly,

PYTHON_DEPS="pkg1 pkg2 pkg3"
SUPPORTED_PYTHONS="2.6 2.7 3.1"
inherit insanely-unfriendly-trickery

w/in said eclass, it does a few things

1) IUSE addition of the USE_EXPAND targets for the supported abis
2) take the enabled USE_EXPAND'd flags intersected against 
SUPPORTED_PYTHON, then set deps/rdeps to 
"""
python_2.6? ( pkg1[python_2.6] pkg2[python_2.6] pkg3[python2.6] )
python_2.7? ( pkg1[python_2.7] pkg2[python_2.7] pkg3[python2.7] )
python_3.1? ( pkg1[python_3.1] pkg2[python_3.1] pkg3[python3.1] )
"""

Yes, that is horrible (ciaran you knew it was going to be).  Few flaws 
with it also-

1) edge case when the user turns off all use flags needs addressing- 
worst case, python eclass forces whatever we consider the 'default' 
(aka 2.6)
2) python_2.6 isn't actually a valid use flag- it would have to be 
python_2-6 or python_26 since periods aren't allowed (arfie pointed 
this out)
3) this can be ugly for users if the PM doesn't treat use flags as 
tristate- specifically 'explicitly-set', 'explicitly-unset', 
'indeterminant'.  If the ocnfiguration forces an explicit and it 
conflicts w/ the use dep, ok, configuration needs to be changed.  If 
the use flag is indeterminant, then the PM needs to flip the flag on 
it's own in that case- I know pkgcore should do this, I don't think 
portage/paludis do (please correct me if wrong).


Thing to note, the deps *would* be accurate- further at the vdb level 
they'd actually be usable.  A dev-lang/python in the vdb is basically 
unusable since implicitly the pkg that has the dep is built against 
whatever slottings of python were available at the time- so if you 
take a pkg from now, a year down the line when py3.2 is stabled as far 
as the PM can tell that pkg *still* would work if <=py3.1 were removed 
(this obviously is not true).

Note also that what I laid out above is as far as I know, going a 
couple of steps beyond what the ruby eclass does (same for what the 
python eclass does).  I'm not necessarily advocating the approach 
above, but for the raw dev-lang/python dependency we *really* should 
be using use_expand there- it'll hand folk a fair amount of control as 
to what abi's get installed into.

~harring