[Rd] RFC: Declaring "foo.bar" as nonS3method() ?!
Henrik Bengtsson
henrik.bengtsson at ucsf.edu
Fri Jun 12 21:18:59 CEST 2015
Analogously to how S4 methods are declared in the code, cf.
methods::setMethod(), I'd find it more natural to also declare S3
methods in the code and note in the NAMESPACE. For example:
# S3 method summary() for class 'aov':
summary.aov <- function(x, ...) {
# something
}
S3class(summary.aov) <- "aov"
with
`S3class<-` <- function(x, value) {
attr(x, "S3class") <- value
x
}
For backward compatibility, if 'S3class' is not set, the default
could/should be to infer it using the current strategy, i.e. the part
after the last period/dot plus other bells'n'whistles discussed. If
all S3 methods had attribute 'S3class' set, there would be no need to
declare the non-S3 case.
Finally, to explicitly declare a function _not_ to be a S3 method, one
could allow for
S3class(all.effects) <- FALSE
At this point, I need to bring up the wish of have a core R function,
again cf. setMethod(), doing the above for us, e.g.
setMethodS3("summary", "aov", function(x, ...) {
# something
})
It can be extremely light weight and would resemble what setMethod()
does for S4 methods.
Also, with the 'S3class' attribute set, one could imagine not having
to declare them as S3method("summary", "aov") in the NAMESPACE. This
could be fully automatic (and backward compatible for migration).
Absolutely not a rant, but from a developers point of view I always
found it a bit ad hoc to have to declare S3 methods in the NAMESPACE
rather than in the code. We're not doing it for S4 methods, so why
for S3 ones? (BTW, I think I understand why). For the same reason,
I'd would think adding NAMESPACE declaration nonS3method() would just
add another workaround.
The above would be backward compatible, allow for a long-term
migration, while allowing folks to use periods/dots however they wish.
It would also allow code inspections such as 'R CMD check --as-cran'
to avoid false positives.
/Henrik
On Fri, Jun 12, 2015 at 8:30 AM, <luke-tierney at uiowa.edu> wrote:
> The notes available off the devloper page
> https://developer.r-project.org/ describe some of the rationale for
> the S3 method search design. One thing that has changed since then is
> that all packages now have name spaces. We could change the search
> algorithm to skip attached package exports (and package imports and
> base), which would require methods defined in packages that are to be
> accessible outside the package to be declared. Methods defined inside
> a package for internal use or methods defined in scripts not in
> packages would still be found. Packages not currently registering
> their methods would have to do so -- not sure how many that would
> affect. Testing on CRAN/Bioc should show how much of an effect this
> would have and whether there are any other issues.
>
> Best,
>
> luke
>
>
> On Fri, 12 Jun 2015, Duncan Murdoch wrote:
>
>> On 12/06/2015 10:53 AM, Hadley Wickham wrote:
>>>
>>> To me, it seems like there's actually two problems here:
>>>
>>> 1) Preventing all() from dispatching to all.effects() for objects of
>>> class effects
>>> 2) Eliminating the NOTE in R CMD check
>>>
>>> My impression is that 1) actually causes few problems, particularly
>>> since people are mostly now aware of the problem and avoid using `.`
>>> in function names unless they're S3 methods. Fixing this issue seems
>>> like it would be a lot of work for relatively little gain.
>>>
>>> However, I think we want to prevent people from writing new functions
>>> with this confusing naming scheme, but equally we want to grandfather
>>> in existing functions, because renaming them all would be a lot of
>>> work (I'm looking at you t.test()!).
>>>
>>> Could we have a system similar to globalVariables() where you could
>>> flag a function as definitely not being an S3 method? I'm not sure
>>> what R CMD check should do - ideally you wouldn't be allow to use
>>> method.class for new functions, but still be able suppress the note
>>> for old functions that can't easily be changed.
>>
>>
>> We have a mechanism for suppressing the warning for existing functions,
>> it's just not available to users to modify. So it would be possible to
>> add effects::all.effects to the stop list, and this might be the easiest
>> action here.
>>
>> This isn't perfect because all.effects() would still act as a method.
>> However, it does give the deprecated message if you ever call it, so
>> nobody would do this unknowingly. The only real risk is that if anyone
>> ever wrote an all.effects function that *was* supposed to be an S3
>> method, it might be masked by the one in effects.
>>
>> Duncan Murdoch
>>
>>>
>>> Hadley
>>>
>>> On Fri, Jun 12, 2015 at 6:52 AM, Kurt Hornik <Kurt.Hornik at wu.ac.at>
>>> wrote:
>>>>>>>>>
>>>>>>>>> Duncan Murdoch writes:
>>>>
>>>>
>>>>> On 12/06/2015 7:16 AM, Kurt Hornik wrote:
>>>>>>>>>>>
>>>>>>>>>>> Duncan Murdoch writes:
>>>>>>
>>>>>>
>>>>>>> On 12/06/2015 4:12 AM, Martin Maechler wrote:
>>>>>>>>
>>>>>>>> This is a topic ' "apparent S3 methods" note in R CMD check '
>>>>>>>> from R-package-devel
>>>>>>>> https://stat.ethz.ch/pipermail/r-package-devel/2015q2/000126.html
>>>>>>>>
>>>>>>>> which is relevant to here because some of us have been thinking
>>>>>>>> about extending R because of the issue.
>>>>>>>>
>>>>>>>> John Fox, maintainer of the 'effects' package has enquired about
>>>>>>>> the following output from 'R CMD check effects'
>>>>>>>>
>>>>>>>>> * checking S3 generic/method consistency ... NOTE
>>>>>>>>> Found the following apparent S3 methods exported but not
>>>>>>>>> registered:
>>>>>>>>> all.effects
>>>>>>>>
>>>>>>>>
>>>>>>>> and added
>>>>>>>>
>>>>>>>>> The offending function, all.effects(), is deprecated in favour of
>>>>>>>>> allEffects(), but I'd rather not get rid of it for backwards
>>>>>>>>> compatibility.
>>>>>>>>> Is there any way to suppress the note without removing
>>>>>>>>> all.effects()?
>>>>>>>>
>>>>>>>>
>>>>>>>> and I had agreed that this was a "False Positive" in this case.
>>>>>>>>
>>>>>>>> [.......]
>>>>>>>>
>>>>>>>> and then
>>>>>>>>
>>>>>>>>> Now I agree .. and have e-talked about this with another R core
>>>>>>>>> member .. that it would be desirable for the package author to
>>>>>>>>> effectively declare the fact that such a function is not an S3
>>>>>>>>> method even though it "looks like it" at least if looked from far.
>>>>>>>>
>>>>>>>>
>>>>>>>>> So, ideally, you could have something like
>>>>>>>>
>>>>>>>>
>>>>>>>>> nonS3method("all.effects")
>>>>>>>>
>>>>>>>>
>>>>>>>>> somewhere in your package source ( in NAMESPACE or R/*.R )
>>>>>>>>> which would tell the package-checking code -- but *ALSO* all the
>>>>>>>>> other S3
>>>>>>>>> method code that all.effects should be treated as a regular R
>>>>>>>>> function.
>>>>>>>>
>>>>>>>>
>>>>>>>>> I would very much like such a feature in R, and for that reason,
>>>>>>>>> I'm cross posting this (as one of the famous exceptions that
>>>>>>>>> accompany real-life rules!!) to R-devel.
>>>>>>>>
>>>>>>>>
>>>>>>>> and actually I did *not* cross post, but have now moved the
>>>>>>>> relevant part of the thread to R-devel.
>>>>>>>>
>>>>>>
>>>>>>> It sounds like a good idea. It's a nontrivial amount of work,
>>>>>>> because
>>>>>>> of the "all the other S3 method code" part. There's the question of
>>>>>>> functions defined outside of packages: presumably they are still S3
>>>>>>> methods, with no way to suppress that.
>>>>>>
>>>>>>
>>>>>> I am not sure this is the right solution: S3 dispatch will still occur
>>>>>> because we first look at foo.bar exports and then in the S3 registry,
>>>>>> afaicr (the "all the other S3 method code" part).
>>>>>>
>>>>>> If we could move to only looking at the registry for dispatch, there
>>>>>> would be no need to declare situations where we should not dispatch on
>>>>>> foo.bar exports.
>>>>>>
>>>>
>>>>> I think that would break a lot of existing scripts. I think the logic
>>>>> should be something like this.
>>>>
>>>>
>>>>> For each class in the class list:
>>>>
>>>>
>>>>> Search backwards through the environment chain. If the current
>>>>> location
>>>>> is a package environment or namespace, look only in the registry. If
>>>>> not, look at all functions.
>>>>
>>>>
>>>>> If that search failed, try the next class.
>>>>
>>>>
>>>> Yep---that's what I meant. I forgot to write the "if namespace
>>>> semantics applies" part :-)
>>>>
>>>> Best
>>>> -k
>>>>
>>>>> A variation on the test is: If there's a registry in the current
>>>>> location, look there. But I think the registry is not on the search
>>>>> list, so I'm not sure that would work.
>>>>
>>>>
>>>>> This assumes that we keep separate registries in each package; if we
>>>>> merge them into one big registry, it gets harder. I'm not familiar
>>>>> enough with the code to know which way we do it.
>>>>
>>>>
>>>>> Duncan Murdoch
>>>>
>>>>
>>>> ______________________________________________
>>>> R-devel at r-project.org mailing list
>>>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>>
>>>
>>>
>>>
>>
>> ______________________________________________
>> R-devel at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>
>
> --
> Luke Tierney
> Ralph E. Wareham Professor of Mathematical Sciences
> University of Iowa Phone: 319-335-3386
> Department of Statistics and Fax: 319-335-3017
> Actuarial Science
> 241 Schaeffer Hall email: luke-tierney at uiowa.edu
> Iowa City, IA 52242 WWW: http://www.stat.uiowa.edu
>
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
More information about the R-devel
mailing list