R devel - Jun 2012 - package slot of generic "[" and missing env@target?

Hi,

Since upgraded to R 2.15, I have a problem with duplicate S4 class name no
longer works (the reason for having duplicate S4 class names is just software
forks - they are largely identical but don't have an inheritence
relationship, and will never have such).

This is happening with "library(mypackage2)" :

Error: package slot missing from signature for generic ?[?
and classes myclass, ANY, ANY, ANY
cannot use with duplicate class names (the package may need to be re-installed)

I have done "debug(methods:::.pkgMethodLabel)" and found that
curiously, it is invoked for the two copies of the "[" generic with
signature "classes myclass, ANY, ANY, ANY", ones with
packageSlot(method at target) sets to c("mypackage2",
"methods", "methods", "methods"), the other just
c("", "", "",""), instead of
c("mypackage1", "methods", "methods",
"methods").

So in the R debugger, I modified the packageSlot() directly, and get
library(mypacakge2) to finish loading, a few times (once in a while R craps out
with some error message about LANGSXP - I know it is probably a bad idea to
modify data structures within the debugger, so that's probably expected).

So my next step was simply to patch R/library/methods to downgrade the stop() to
a warning(), as well as inserting the missing packageSlot values there and then.

This got both packages loaded. But then, some actual use of "[", it
gives this error instead:

Error in .methodPackageSlots(def, label, table, dupl, isCurrent) : 
  trying to get slot "target" from an object of a basic class
("environment") with no slots

and when I downgrade this error to a warning again, modifying
(src/main/attrib.c) I got "[" dispatched to the base class (as well as
stack imbalance - the latter is expected and "asking for it"...).

Anyway, I am thinking:

- doing "debug(methods:::.pkgMethodLabel)", I see two "[" of
signature "classes myclass, ANY, ANY, ANY", one comes with packageSlot
with the correct stuff and one with just 4 "", that seems to be a bug?

- from the warning at the beginning: 
"A specification for class ?myclass? in package ?mypackage1? seems
equivalent to one from package ?mypackage2? and is not turning on duplicate
class definitions for this class"

It would appear that there is a "correct" way of turning on duplicate
class definition, but I have not found any documentation to that effect.
I see some discussion about class unions, and inheritances, but in this case,
neither are appropriate - the two packages are just forks and have no union nor
inheritance relationship; but it is desirable to load both, to check for
differences and bugs.

- I am also curious why only "[" triggers the problem, but not
"[<-" or another of the other duplicated methods. Obviously
subsetting is being treated specially, as is mentioned a few times in R
internals; but I was expecting subassignment to be affected also, and that does
not seems to be the case.


Anyhow, it is an extremely useful feature to be able to load a past version of a
package simultaneously just by renaming the package name and namespace, to check
for regression and change-in-behavior, i.e. newly added bugs; so I am hoping to
get this working again.

TIA.

--- On Thu, 21/6/12, Hin-Tak Leung <htl10 at users.sourceforge.net> wrote:
> Hi,
> 
> Since upgraded to R 2.15, I have a problem with duplicate S4
> class name no longer works (the reason for having duplicate
> S4 class names is just software forks - they are largely
> identical but don't have an inheritence relationship, and
> will never have such). 
> 
> This is happening with "library(mypackage2)" :
> 
> Error: package slot missing from signature for generic
> ?[?
> and classes myclass, ANY, ANY, ANY
> cannot use with duplicate class names (the package may need
> to be re-installed)
> 
> I have done "debug(methods:::.pkgMethodLabel)" and found
> that curiously, it is invoked for the two copies of the "["
> generic with signature "classes myclass, ANY, ANY, ANY",
> ones with 
> packageSlot(method at target) sets to c("mypackage2",
> "methods", "methods", "methods"), the other
just c("", "",
> "",""), instead of? 
> c("mypackage1", "methods", "methods",
"methods").
> 
> So in the R debugger, I modified the packageSlot() directly,
> and get library(mypacakge2) to finish loading, a few times
> (once in a while R craps out with some error message about
> LANGSXP - I know it is probably a bad idea to modify data
> structures within the debugger, so that's probably
> expected).
> 
> So my next step was simply to patch R/library/methods to
> downgrade the stop() to a warning(), as well as inserting
> the missing packageSlot values there and then.
> 
> This got both packages loaded. But then, some actual use of
> "[", it gives this error instead:
> 
> Error in .methodPackageSlots(def, label, table, dupl,
> isCurrent) : 
> ? trying to get slot "target" from an object of a basic
> class ("environment") with no slots
> 
> and when I downgrade this error to a warning again,
> modifying (src/main/attrib.c) I got "[" dispatched to the
> base class (as well as stack imbalance - the latter is
> expected and "asking for it"...).
> 
> Anyway, I am thinking:
> 
> - doing "debug(methods:::.pkgMethodLabel)", I see two
"[" of
> signature "classes myclass, ANY, ANY, ANY", one comes with
> packageSlot with the correct stuff and one with just 4 "",
> that seems to be a bug?
> 
> - from the warning at the beginning: 
> "A specification for class ?myclass? in package
> ?mypackage1? seems equivalent to one from package
> ?mypackage2? and is not turning on duplicate class
> definitions for this class"
> 
> It would appear that there is a "correct" way of turning on
> duplicate class definition, but I have not found any
> documentation to that effect.
> I see some discussion about class unions, and inheritances,
> but in this case, neither are appropriate - the two packages
> are just forks and have no union nor inheritance
> relationship; but it is desirable to load both, to check for
> differences and bugs.
> 
> - I am also curious why only "[" triggers the problem, but
> not "[<-" or another of the other duplicated methods.
> Obviously subsetting is being treated specially, as is
> mentioned a few times in R internals; but I was expecting
> subassignment to be affected also, and that does not seems
> to be the case.
> 
> 
> Anyhow, it is an extremely useful feature to be able to load
> a past version of a package simultaneously just by renaming
> the package name and namespace, to check for regression and
> change-in-behavior, i.e. newly added bugs; so I am hoping to
> get this working again.
> 
> TIA.
> 
Found a small work-around - whenever I want to switch, just do

detach("package:package1", unload=TRUE)
library(package2, quietly=TRUE)

or vice versa. A bit tedious, also documentation says detach() doesn't
always work as planned.  (with 'echo=FALSE' in Sweave, the resulting tex
doesn't even change at all). There are also a dozen of:

Warning in FUN(X[[1L]], ...) :
  Created a package name, ?2012-06-22 10:58:02?, when none found

whenever detach() + library() is done. Still  would appreciate clarification of
the missing package slot in '[' generic (and whether it is a bug) and
other matters.