[bitbake-devel] Bitbake / oe-core anomaly with multilibs

[Richard Purdie]

On Tue, 2013-03-19 at 07:44 -0500, Mark Hatle wrote:
> On 3/19/13 6:35 AM, Richard Purdie wrote:
> > On Mon, 2013-03-18 at 14:12 -0500, Mark Hatle wrote:
> >> I found a strange situation today and I'm looking for an explanation to see if
> >> it's a bug or not.
> >>
> >> Between a .bb and .bbappend file, I end up with a situation where
> >> "reproducer.bb" and "reproducer.bbappend" get the following:
> >>
> >> RDEPENDS_${PN} += "hello1"
> >> RDEPENDS_reproducer += "hello2"
> >>
> >> When evaluatated (bitbake -e reproducer) I get:
> >>
> >> #
> >> # $RDEPENDS_reproducer [2 operations]
> >> #   append
> >> /home/mhatle/git/oss/oe-core/local/recipes-sample/hello/reproducer_1.0.bb:16
> >> #     "hello2"
> >> #   rename from RDEPENDS_${PN} data.py:161 [expandKeys]
> >> #     " hello1"
> >> # computed:
> >> #   " hello1"
> >> RDEPENDS_reproducer="hello1"
> >>
> >> So that tells me that it was initially set to "hello2", and then the ${PN}
> >> expansion occurred, causing it to be set to "hello1".
> >>
> >> So my first question is, should this be "hello1", "hello2", "hello1 hello2" or
> >> "hello2 hello1"?
> >
> > expandKeys tramples existing values so that result doesn't surprise me
> > even if its not what you expect.
> >
> >> The second issue is a bit stranger.. if I change the build from "reproducer" to
> >> "lib32-reproducer", I get a different result:
> >>
> >> # $RDEPENDS_lib32-reproducer [3 operations]
> >> #   rename from RDEPENDS_${PN} data.py:161 [expandKeys]
> >> #     " hello1"
> >> #   rename from RDEPENDS_reproducer classextend.py:95 [rename_package_variables]
> >> #     " hello2"
> >> #   set classextend.py:71 [map_depends_variable]
> >> #     "lib32-hello2"
> >> # computed:
> >> #   "lib32-hello2"
> >> RDEPENDS_lib32-reproducer="lib32-hello2"
> >>
> >> This time the system pulled in the ${PN} version first (obviously expanded it),
> >> and then turned out and found the RDEPENDS_reproducer, and remapped it to
> >> RDEPENDS_lib32-reproducer replacing the origin ${PN} version.
> >>
> >> So I have two concerns, the first is the value is 'different' from the
> >> non-multilib version, and second, what should the expected output be for this item?
> >
> > Again, I'm not surprised. The multilib remapping code gets executed at
> > the end, after expandKeys and in this case last wins. So I can see why
> > the system is doing it, even if you don't like the end result.
> >
> > Setting multiple conflicting keys like this is dangerous :(
> >
> > We've tried to fix this before and ended up just digging deeper holes
> > with more problems.
> 
> Is there any way we can issue a warning from the parser or something when it 
> finds conflicting keys?  My concern isn't that conflicting keys are causing a 
> problem -- but more that someone may not realize this is a problem, and they are 
> getting incorrect results.
> 
> I know since posting this query, I've had a few people comment to me that they 
> saw the same problem and it took them a very long time to diagnose it and figure 
> out it "just didn't work".

We could probably have data.expandKeys() check for existing values and
then error if its overwriting something else. I have no idea how often
the metadata does that for "correct" reasons...

Cheers,

Richard