[Rd] Suggestion: default print method for S3 generics could offer some insights on '...' among registered methods

[Mikael Jagan]

On 2025-06-10 12:15 pm, Michael Chirico wrote:
> Thanks for the thoughtful reply Mikael.
> 
>> Any function F with '...' as a formal argument can pass '...' to another function G.
> 
> Yes, that's true. The difference is that in print(F) we can _usually_
> pick out at a glance how '...' is being used -- we can see which 'G'
> is getting '...'.
> 
> For S3 generics, we quickly reach the dead end of 'UseMethod' -- F
> being S3 generic is in fact _highly_ relevant.
> 

I don't really think of calls to UseMethod as dead ends.  I immediately do
something like

     for (nm in paste0(generic, ".", c(.class2(object), "default"))) {
         print(help(nm)) # or print(argsAnywhere(nm)) or ...
         if (satisfied) break
     }

In other words, I seek info about only those methods that might actually be
dispatched.  Details about other methods are really an unwanted distraction.

But, yes, the fact that '...' does not appear in calls to UseMethod is a
legitimate distinction, so thanks for clarifying.

As I see it, R provides many tools enabling users to efficiently and
programmatically interrogate the dispatch mechanism, available methods, their
formal arguments, etc.  These seem well documented but not enough advertised.
Hence I would sooner work to improve and promote the tools than work to change
print.default, especially if those changes risk overwhelming very new users of
R who may know nothing about OOP and for whom print.default is a primary mode
of interrogation.

Mikael

> Yes, the practical issues you raise are interesting & knotty (I
> especially have in mind [1] and [2]), but ultimately I think we could
> come up with something useful. Whether that becomes a default can
> depend on how useful it winds up being, and the empirical risk of
> back-incompatibility (which I suspect is low).
> 
> Mike C
> 
> [1] utils::isS3stdGeneric
> https://stat.ethz.ch/R-manual/R-devel/library/utils/html/isS3stdGen.html,
> which has a large # of false negatives
> [2] utils::nonS3methods
> https://stat.ethz.ch/R-manual/R-devel/library/tools/html/QC.html,
> which maintains an onerous list of S3 method lookalikes
> 
> On Mon, Jun 9, 2025 at 8:44 PM Mikael Jagan <jaganmn2 using gmail.com> wrote:
>>
>> I don't really understand the premise.  Any function F with '...' as a formal
>> argument can pass '...' to another function G.  The actual arguments matching
>> '...' in the call to F will be matched to the formal arguments of G.  So the
>> the maintainer of F may want to alert the user of F to the existence of G and
>> the user of F may want to consult the documentation of G.
>>
>> Whether F is S3 generic and G is registered as a method for F seems irrelevant.
>>
>> That is a conceptual issue.  There are practical issues, too:
>>
>>       * print.default is used "everywhere".  Backwards incompatible changes to
>>         default behaviour have the potential to break a lot of code out there.
>>
>>       * Testing that a function F is S3 generic seems nontrivial.  You have to
>>         deal with internally generic functions and for closures recurse through
>>         body(F) looking for a call to UseMethod.
>>
>>       * I would not want the output of print(F) to depend on details external to
>>         F or the method call, such as the state of the table of registered S3
>>         methods which changes as packages are loaded.  AFAIK, it is intended that
>>         options() is the only exception to the rule.
>>
>>       * More harmonious would be to implement the feature ("give me more
>>         information about S3 methods") as an option (disabled by default) of
>>         utils::.S3methods if not as a new function altogether.
>>
>> Mikael
>>
>>> Date: Fri, 6 Jun 2025 11:59:08 -0700
>>> From: Michael Chirico<michaelchirico4 using gmail.com>
>>>
>>> There is a big difference in how to think of '...' for non-generic
>>> functions like data.frame() vs. S3 generics.
>>>
>>> In the former, it means "any number of inputs" [e.g. columns]; in the
>>> latter, it means "any number of inputs [think c()], as well as any
>>> arguments that might be interpreted by class implementations".
>>>
>>> Understanding the difference for a given generic can require carefully
>>> reading lots of documentation. print(<generic>), which is useful for
>>> so many other contexts, can be a dead end.
>>>
>>> One idea is to extend the print() method to suggest to the reader
>>> which other arguments are available (among registered generics). Often
>>> ?<generic> will include the most common implementation, but not always
>>> so.
>>>
>>> For rbind (in a --vanilla session), we currently have one method,
>>> rbind.data.frame, that offers three arguments not present in the
>>> generic: make.row.names, stringsAsFactors, and factor.exclude. The
>>> proposal would be to mention this in the print(rbind) output somehow,
>>> e.g.
>>>
>>>> print(rbind)
>>> function (..., deparse.level = 1)
>>> .Internal(rbind(deparse.level, ...))
>>> <bytecode: 0x73d4fd824e20>
>>> <environment: namespace:base>
>>>
>>> +Other arguments implemented by methods
>>> +  factor.exclude: rbind.data.frame
>>> +  make.row.names: rbind.data.frame
>>> +  stringsAsFactors: rbind.data.frame
>>>
>>> I suggest grouping by argument, not generic, although something like
>>> this could be OK too:
>>>
>>> +Signatures of other methods
>>> +  rbind.data.frame(..., deparse.level = 1, make.row.names = TRUE,
>>> stringsAsFactors = FALSE,
>>> +      factor.exclude = TRUE)
>>>
>>> Where it gets more interesting is when there are many methods, e.g.
>>> for as.data.frame (again, in a --vanilla session):
>>>
>>>> print(as.data.frame)
>>> function (x, row.names = NULL, optional = FALSE, ...)
>>> {
>>>       if (is.null(x))
>>>           return(as.data.frame(list()))
>>>       UseMethod("as.data.frame")
>>> }
>>> <bytecode: 0x73d4fc1e70d0>
>>> <environment: namespace:base>
>>>
>>> +Other arguments implemented by methods
>>> +  base: as.data.frame.table
>>> +  check.names: as.data.frame.list
>>> +  col.names: as.data.frame.list
>>> +  cut.names: as.data.frame.list
>>> +  fix.empty.names: as.data.frame.list
>>> +  make.names: as.data.frame.matrix, as.data.frame.model.matrix
>>> +  new.names: as.data.frame.list
>>> +  nm: as.data.frame.bibentry, as.data.frame.complex, as.data.frame.Date,
>>> +    as.data.frame.difftime, as.data.frame.factor, as.data.frame.integer,
>>> +    as.data.frame.logical, as.data.frame.noquote, as.data.frame.numeric,
>>> +    as.data.frame.numeric_version, as.data.frame.ordered,
>>> +    as.data.frame.person, as.data.frame.POSIXct, as.data.frame.raw
>>> +  responseName: as.data.frame.table
>>> +  sep: as.data.frame.table
>>> +  stringsAsFactors: as.data.frame.character, as.data.frame.list,
>>> +    as.data.frame.matrix, as.data.frame.table
>>>
>>> Or
>>>
>>> +Signatures of other methods
>>> +  as.data.frame.aovproj(x, ...)
>>> +  as.data.frame.array(x, row.names = NULL, optional = FALSE, ...)
>>> +  as.data.frame.AsIs(x, row.names = NULL, optional = FALSE, ...)
>>> +  as.data.frame.bibentry(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.character(x, ..., stringsAsFactors = FALSE)
>>> +  as.data.frame.citation(x, row.names = NULL, optional = FALSE, ...)
>>> +  as.data.frame.complex(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.data.frame(x, row.names = NULL, ...)
>>> +  as.data.frame.Date(x, row.names = NULL, optional = FALSE, ..., nm =
>>> deparse1(substitute(x)))
>>> +  as.data.frame.default(x, ...)
>>> +  as.data.frame.difftime(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.factor(x, row.names = NULL, optional = FALSE, ..., nm
>>> = deparse1(substitute(x)))
>>> +  as.data.frame.ftable(x, row.names = NULL, optional = FALSE, ...)
>>> +  as.data.frame.integer(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.list(x, row.names = NULL, optional = FALSE, ...,
>>> cut.names = FALSE,
>>> +      col.names = names(x), fix.empty.names = TRUE, new.names =
>>> !missing(col.names),
>>> +      check.names = !optional, stringsAsFactors = FALSE)
>>> +  as.data.frame.logical(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.logLik(x, ...)
>>> +  as.data.frame.matrix(x, row.names = NULL, optional = FALSE,
>>> make.names = TRUE,
>>> +      ..., stringsAsFactors = FALSE)
>>> +  as.data.frame.model.matrix(x, row.names = NULL, optional = FALSE,
>>> make.names = TRUE,
>>> +      ...)
>>> +  as.data.frame.noquote(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.numeric(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.numeric_version(x, row.names = NULL, optional =
>>> FALSE, ..., nm = deparse1(substitute(x)))
>>> +  as.data.frame.ordered(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.person(x, row.names = NULL, optional = FALSE, ..., nm
>>> = deparse1(substitute(x)))
>>> +  as.data.frame.POSIXct(x, row.names = NULL, optional = FALSE, ...,
>>> nm = deparse1(substitute(x)))
>>> +  as.data.frame.POSIXlt(x, row.names = NULL, optional = FALSE, ...)
>>> +  as.data.frame.raw(x, row.names = NULL, optional = FALSE, ..., nm =
>>> deparse1(substitute(x)))
>>> +  as.data.frame.table(x, row.names = NULL, ..., responseName =
>>> "Freq", stringsAsFactors = TRUE,
>>> +      sep = "", base = list(LETTERS))
>>> +  as.data.frame.ts(x, ...)
>>>
>>> Obviously that's a bit more cluttered, but as.data.frame() should be a
>>> pretty unusual case. It also highlights better the differences in the
>>> two approaches: the former economizes on space and focuses on what
>>> sorts of arguments are available; the latter shows the defaults, does
>>> not hide the arguments shared with the generic, and will always
>>> produce as many lines as there are methods.
>>>
>>> There are other edge cases to think through (multiple registrations,
>>> interactions with S4, primitives, ...), but I want to first check with
>>> the list if this looks workable & valuable enough to pursue.
>>>
>>> Mike C
>>>
>>> ----
>>>
>>> Code that helped with the above:
>>>
>>> f = as.data.frame
>>> # NB: methods() and getAnywhere() require {utils}
>>> m = methods(f)
>>> generic_args = names(formals(f))
>>> f_methods = lapply(m, \(fn) getAnywhere(fn)$objs[[1L]])
>>> names(f_methods) = m
>>> new_args = sapply(f_methods, \(g) setdiff(names(formals(g)), generic_args))
>>> with( # group by argument name
>>>     data.frame(method = rep(names(new_args), lengths(new_args)), arg =
>>> unlist(new_args), row.names=NULL),
>>>     {tbl = tapply(method, arg, toString); writeLines(paste0(names(tbl),
>>> ": ", tbl))}
>>> )
>>> signatures=sapply(f_methods, \(g) paste(head(format(args(g)), -1),
>>> collapse="\n"))
>>> writeLines(paste0(names(signatures), gsub("^\\s*function\\s*", "", signatures)))
>>