[Rd] Multiple Assignment built into the R Interpreter?

Pavel Krivitsky p@kr|v|t@ky @end|ng |rom un@w@edu@@u
Mon Mar 13 03:36:59 CET 2023


Dear All,

As a maintainer of large, complex packages, I can think of many places
in which deconstructing assignment would simplify the code, as well as
facilitate readability by breaking up larger functions into helpers, so
I would be very glad to see this incorporated somehow.

I think the crux of the matter is that while there is a number of ways
to implement deconstructing assignment within R, there is no mechanism
to tell R CMD check about it without also suppressing checks for every
other instance of that variable name. This is particularly problematic
because those variable names are likely to be used elsewhere in the
package.

Workarounds that have been suggested all defeat the conciseness and
clarity of the deconstructing assignment and introduce potential for
subtle bugs.

The check warnings are something that can only be addressed in
'codetools', with a finer API than what utils::globalVariables()
provides. Perhaps this would have a lower hurdle than modifying R
language itself?

From skimming through the relevant 'codetools' code, one idea for such
an API would be a function, along the lines of

utils::alternativeAssignment(op, assigned)

that sets up a callback assigned = function(op, e) that given the
operator (as string) and the expression it's embedded in, returns a
list of three elements:
 * a character vector containing a list of variables assigned to that
   might not otherwise be detected
 * a character vector containing a list of variables referenced that
   might not otherwise be detected
 * expression e with potentially "offending" elements removed, which
   will then be processed by the rest of the checking code

Then, say, 'zeallot' could implement zeallot::zeallot_assign_detect(),
and a package developer using it could put

utils::alternativeAssignment("%<-%", zeallot::zeallot_assign_detect)

in their .onLoad() function. Similarly, users of 'dotty' could set up
callbacks for all standard assignment operators to inform the code
about the nonstandard assignment.

Best Regards,Pavel

On Sun, 2023-03-12 at 14:05 +0200, Sebastian Martin Krantz wrote:
> Kevins package is very nice as a proof of concept, no doubt about
> that, but
> it is not at the level of performance or convenience that a native R
> implementation would offer. I would probably not use it to translate
> matlab
> routines into R packages placed on CRAN, because it’s an additional
> dependency, I have a performance burden in every iteration, and
> utils::globalVariables() is everything but elegant. From that
> perspective
> it would be more convenient for me right now to stick with
> collapse::%=%,
> which is already written in C, and also call
> utils::globalVariables().
> 
> But again my hope in starting this was that R Core might see that the
> addition of multiple assignment would be a significant enhancement to
> the
> language, of the same order as the base pipe |> in my opinion.
> 
> I think the discussion so far has at least brought forth a way to
> implement
> this in a way that does not violate fundamental principles of the
> language.
> Which could form a basis for thinking about an actual addition to the
> language.
> 
> Best regards,
> 
> Sebastian
> 
> 
> On Sun 12. Mar 2023 at 13:18, Duncan Murdoch
> <murdoch.duncan using gmail.com>
> wrote:
> 
> > On 12/03/2023 6:07 a.m., Sebastian Martin Krantz wrote:
> > > Thinking more about this, and seeing Kevins examples at
> > > https://github.com/kevinushey/dotty
> > > <https://github.com/kevinushey/dotty>, I think this is the most
> > > R-like
> > > way of doing it,
> > > with an additional benefit as it would allow to introduce the
> > > useful
> > > data.table semantics DT[, .(a = b, c, d)] to more general R. So I
> > > would
> > > propose to
> > > introduce a new primitive function . <- function(...)
> > > .Primitive(".") in
> > > R with an assignment method and the following features:
> > 
> > I think that proposal is very unlikely to be accepted.  If it was a
> > primitive function, it could only be maintained by R Core.  They
> > are
> > justifiably very reluctant to take on extra work for themselves.
> > 
> > Kevin's package demonstrates that this can be done entirely in a
> > contributed package, which means there's no need for R Core to be
> > involved.  I don't know if he has plans to turn his prototype into
> > a
> > CRAN package.  If he doesn't, then it will be up to some other
> > interested maintainer to step up and take on the task, or it will
> > just
> > fade away.
> > 
> > I haven't checked whether your proposals below represent changes
> > from
> > the current version of dotty, but if they do, the way to proceed is
> > to
> > fork that project, implement your changes, and offer to contribute
> > them
> > back to the main branch.
> > 
> > Duncan Murdoch
> > 
> > 
> > 
> > > 
> > >   * Positional assignment e.g. .[nr, nc] <- dim(x), and named
> > > assignment
> > >     e.g. .[new = carb] <- mtcars or .[new = log(carb)] <- mtcars.
> > > All
> > >     the functionality proposed by Kevin at
> > >     https://github.com/kevinushey/dotty
> > >     <https://github.com/kevinushey/dotty> is useful, unambiguous
> > > and
> > >     feasible.
> > >   * Silent dropping of RHS values e.g. .[mpg_new, cyl_new] <-
> > > mtcars.
> > >   * Mixing of positional and named assignment e.g .[mpg_new,
> > > carb_new =
> > >     carb, cyl_new] <- mtcars. The inputs not assigned by name are
> > > simply
> > >     the elements of RHS in the order they occur, regardless of
> > > whether
> > >     they have been used previously e.g. .[mpg_new, cyl_new = cyl,
> > >     log_cyl = log(cyl), cyl_new2] <- mtcars is feasible. RHS here
> > > could
> > >     be any named vector type.
> > >   * Conventional use of the function as lazy version of of
> > > list(), as in
> > >     data.table: .(A = B, C, D) is the same as list(A = B, C = C,
> > > D = D).
> > >     This would also be useful, allowing more parsimonious code,
> > > and
> > >     avoid the need to assign names to all return values in a
> > > function
> > >     return, e.g. if I already have matrices A, C, Q and R as
> > > internal
> > >     objects in my function, I can simply end by return(.(A, C, Q,
> > > R))
> > >     instead of return(list(A = A, C = C, Q = Q, R = R)) if I
> > > wanted the
> > >     list to be named with the object names.
> > > 
> > > The implementation of this in R and C should be pretty
> > > straightforward.
> > > It would just require a modification to R CMD Check to recognize
> > > .[<- as
> > > assignment.
> > > 
> > > Best regards,
> > > 
> > > Sebastian
> > > -
> > > 2.)
> > > 
> > > On Sun, 12 Mar 2023 at 09:42, Sebastian Martin Krantz
> > > <sebastian.krantz using graduateinstitute.ch
> > > <mailto:sebastian.krantz using graduateinstitute.ch>> wrote:
> > > 
> > >     Thanks Gabriel and Kevin for your inputs,
> > > 
> > >     regarding your points Gabriel, I think Python and Julia do
> > > allow
> > >     multiple sub-assignment, but in-line with my earlier
> > > suggestion in
> > >     response to Duncan to make multiple assignment an
> > > environment-level
> > >     operation (like collapse::%=% currently works),  this would
> > > not be
> > >     possible in R.
> > > 
> > >     Regarding the [a] <- coolest_function() syntax, yeah it would
> > > mean
> > >     do multiple assignment and set a equal to the first element
> > > dropping
> > >     all other elements. Multiple assignment should be positional
> > > loke in
> > >     other languages, enabling flexible renaming of objects on the
> > > fly.
> > >     So it should be irrelevant whether the function returns a
> > > named or
> > >     unnamed list or vector.
> > > 
> > >     Thanks also Kevin for this contribution. I think it’s a
> > > remarkable
> > >     effort, and I wouldn’t mind such semantics e.g. making it a
> > > function
> > >     call to ‘.[‘ or any other one-letter function, as long as
> > > it’s coded
> > >     in C and recognized by the interpreter as an assignment
> > > operation.
> > > 
> > >     Best regards,
> > > 
> > >     Sebastian
> > > 
> > > 
> > > 
> > > 
> > > 
> > >     On Sun 12. Mar 2023 at 01:00, Kevin Ushey
> > > <kevinushey using gmail.com
> > >     <mailto:kevinushey using gmail.com>> wrote:
> > > 
> > >         FWIW, it's possible to get fairly close to your proposed
> > semantics
> > >         using the existing metaprogramming facilities in R. I put
> > together a
> > >         prototype package here to demonstrate:
> > > 
> > >         https://github.com/kevinushey/dotty
> > >         <https://github.com/kevinushey/dotty>
> > > 
> > >         The package exports an object called `.`, with a special
> > >         `[<-.dot` S3
> > >         method which enables destructuring assignments. This
> > > means you
> > can
> > >         write code like:
> > > 
> > >              .[nr, nc] <- dim(mtcars)
> > > 
> > >         and that will define 'nr' and 'nc' as you expect.
> > > 
> > >         As for R CMD check warnings, you can suppress those
> > > through the
> > >         use of
> > >         globalVariables(), and that can also be automated within
> > > the
> > >         package.
> > >         The 'dotty' package includes a function 'dotify()' which
> > automates
> > >         looking for such usages in your package, and calling
> > >         globalVariables()
> > >         so that R CMD check doesn't warn. In theory, a similar
> > > technique
> > >         would
> > >         be applicable to other packages defining similar
> > > operators
> > (zeallot,
> > >         collapse).
> > > 
> > >         Obviously, globalVariables() is a very heavy hammer to
> > > swing for
> > >         this
> > >         issue, but you might consider the benefits worth the
> > > tradeoffs.
> > > 
> > >         Best,
> > >         Kevin
> > > 
> > >         On Sat, Mar 11, 2023 at 2:53 PM Duncan Murdoch
> > >        
> > > <murdoch.duncan using gmail.com <mailto:murdoch.duncan using gmail.com>>
> > wrote:
> > >          >
> > >          > On 11/03/2023 4:42 p.m., Sebastian Martin Krantz
> > > wrote:
> > >          > > Thanks Duncan and Ivan for the careful thoughts. I'm
> > > not
> > >         sure I can
> > >          > > follow all aspects you raised, but to give my
> > > limited take
> > >         on a few:
> > >          > >
> > >          > >> your proposal violates a very basic property of the
> > >         language, i.e. that all statements are expressions and
> > > have a
> > >         value.  > What's the value of 1 + (A, C =
> > > init_matrices()).
> > >          > >
> > >          > > I'm not sure I see the point here. I evaluated 1 +
> > > (d =
> > >         dim(mtcars); nr
> > >          > > = d[1]; nc = d[2]; rm(d)), which simply gives a
> > > syntax
> > error,
> > >          >
> > >          >
> > >          >    d = dim(mtcars); nr = d[1]; nc = d[2]; rm(d)
> > >          >
> > >          > is not a statement, it is a sequence of 4 statements.
> > >          >
> > >          > Duncan Murdoch
> > >          >
> > >          >   as the
> > >          > > above expression should. `%=%` assigns to
> > >          > > environments, so 1 + (c("A", "C") %=%
> > > init_matrices())
> > returns
> > >          > > numeric(0), with A and C having their values
> > > assigned.
> > >          > >
> > >          > >> suppose f() returns list(A = 1, B = 2) and I do  >
> > > B, A <-
> > >         f() > Should assignment be by position or by name?
> > >          > >
> > >          > > In other languages this is by position. The feature
> > > is not
> > >         meant to
> > >          > > replace list2env(), and being able to rename objects
> > > in the
> > >         assignment
> > >          > > is a vital feature of codes
> > >          > > using multi input and output functions e.g. in
> > > Matlab or
> > Julia.
> > >          > >
> > >          > >> Honestly, given that this is simply syntactic
> > > sugar, I
> > >         don't think I would support it.
> > >          > >
> > >          > > You can call it that, but it would be used by almost
> > > every
> > >         R user almost
> > >          > > every day. Simple things like nr, nc = dim(x);
> > > values,
> > >         vectors =
> > >          > > eigen(x) etc. where the creation of intermediate
> > > objects
> > >          > > is cumbersome and redundant.
> > >          > >
> > >          > >> I see you've already mentioned it ("JavaScript-
> > > like"). I
> > >         think it would  fulfil Sebastian's requirements too, as
> > > long as
> > >         it is considered "true assignment" by the rest of the
> > > language.
> > >          > >
> > >          > > I don't have strong opinions about how the issue is
> > > phrased
> > or
> > >          > > implemented. Something like [t, n] = dim(x) might
> > > even be
> > >         more clear.
> > >          > > It's important though that assignment remains by
> > > position,
> > >          > > so even if some output gets thrown away that should
> > > also be
> > >         positional.
> > >          > >
> > >          > >>  A <- 0  > [A, B = A + 10] <- list(1, A = 2)
> > >          > >
> > >          > > I also fail to see the use of allowing this.
> > > something like
> > >         this is an
> > >          > > error.
> > >          > >
> > >          > >> A = 2
> > >          > >> (B = A + 1) <- 1
> > >          > > Error in (B = A + 1) <- 1 : could not find function
> > > "(<-"
> > >          > >
> > >          > > Regarding the practical implementation, I think
> > >         `collapse::%=%` is a
> > >          > > good starting point. It could be introduced in R as
> > > a
> > >         separate function,
> > >          > > or `=` could be modified to accommodate its
> > > capability. It
> > >         should be
> > >          > > clear that
> > >          > > with more than one LHS variables the assignment is
> > > an
> > >         environment level
> > >          > > operation and the results can only be used in
> > > computations
> > >         once assigned
> > >          > > to the environment, e.g. as in 1 + (c("A", "C") %=%
> > >         init_matrices()),
> > >          > > A and C are not available for the addition in this
> > >         statement. The
> > >          > > interpretor then needs to be modified to read
> > > something
> > >         like nr, nc =
> > >          > > dim(x) or [nr, nc] = dim(x). as an environment-level
> > >         multiple assignment
> > >          > > operation with no
> > >          > > immediate value. Appears very feasible to my limited
> > >         understanding, but
> > >          > > I guess there are other things to consider still.
> > >         Definitely appreciate
> > >          > > the responses so far though.
> > >          > >
> > >          > > Best regards,
> > >          > >
> > >          > > Sebastian
> > >          > >
> > >          > >
> > >          > >
> > >          > >
> > >          > >
> > >          > > On Sat, 11 Mar 2023 at 20:38, Duncan Murdoch
> > >        
> > > <murdoch.duncan using gmail.com <mailto:murdoch.duncan using gmail.com>
> > >          > > <mailto:murdoch.duncan using gmail.com
> > >         <mailto:murdoch.duncan using gmail.com>>> wrote:
> > >          > >
> > >          > >     On 11/03/2023 11:57 a.m., Ivan Krylov wrote:
> > >          > >      > On Sat, 11 Mar 2023 11:11:06 -0500
> > >          > >      > Duncan Murdoch <murdoch.duncan using gmail.com
> > >         <mailto:murdoch.duncan using gmail.com>
> > >          > >     <mailto:murdoch.duncan using gmail.com
> > >         <mailto:murdoch.duncan using gmail.com>>> wrote:
> > >          > >      >
> > >          > >      >> That's clear, but your proposal violates a
> > > very
> > >         basic property
> > >          > >     of the
> > >          > >      >> language, i.e. that all statements are
> > > expressions
> > >         and have a value.
> > >          > >      >
> > >          > >      > How about reframing this feature request from
> > >         multiple assignment
> > >          > >      > (which does go contrary to "everything has
> > > only one
> > >         value, even
> > >          > >     if it's
> > >          > >      > sometimes invisible(NULL)") to "structured
> > > binding"
> > >         / "destructuring
> > >          > >      > assignment" [*], which takes this single
> > > single
> > >         value returned by the
> > >          > >      > expression and subsets it subject to certain
> > > rules?
> > >         It may be
> > >          > >     easier to
> > >          > >      > make a decision on the semantics for
> > > destructuring
> > >         assignment (e.g.
> > >          > >      > languages which have this feature typically
> > > allow
> > >         throwing unneeded
> > >          > >      > parts of the return value away), and it
> > > doesn't seem
> > >         to break as much
> > >          > >      > of the rest of the language if implemented.
> > >          > >      >
> > >          > >      > I see you've already mentioned it
> > >         ("JavaScript-like"). I think it
> > >          > >     would
> > >          > >      > fulfil Sebastian's requirements too, as long
> > > as it
> > >         is considered
> > >          > >     "true
> > >          > >      > assignment" by the rest of the language.
> > >          > >      >
> > >          > >      > The hard part is to propose the actual
> > > grammar of
> > >         the new feature (in
> > >          > >      > terms of src/main/gram.y, preferably without
> > introducing
> > >          > >     conflicts) and
> > >          > >      > its semantics (including the corner cases,
> > > some of
> > >         which you have
> > >          > >      > already mentioned). I'm not sure I'm up to
> > > the task.
> > >          > >      >
> > >          > >
> > >          > >     If I were doing it, here's what I'd propose:
> > >          > >
> > >          > >         '[' formlist ']' LEFT_ASSIGN expr
> > >          > >         '[' formlist ']' EQ_ASSIGN expr
> > >          > >         expr RIGHT_ASSIGN  '[' formlist ']'
> > >          > >
> > >          > >     where `formlist` has the syntax of the formals
> > > list for
> > >         a function
> > >          > >     definition.  This would have the following
> > > semantics:
> > >          > >
> > >          > >          {
> > >          > >            *tmp* <- expr
> > >          > >
> > >          > >            # For arguments with no "default"
> > > expression,
> > >          > >
> > >          > >            argname1 <- *tmp*[[1]]
> > >          > >            argname2 <- *tmp*[[2]]
> > >          > >            ...
> > >          > >
> > >          > >            # For arguments with a default listed
> > >          > >
> > >          > >            argname3 <- with(*tmp*, default3)
> > >          > >          }
> > >          > >
> > >          > >
> > >          > >     The value of the whole thing would therefore be
> > >         (invisibly) the
> > >          > >     value of
> > >          > >     the last item in the assignment.
> > >          > >
> > >          > >     Two examples:
> > >          > >
> > >          > >         [A, B, C] <- expr   # assign the first three
> > >         elements of expr to A,
> > >          > >     B, and C
> > >          > >
> > >          > >         [A, B, C = a + b] <- expr  # assign the
> > > first two
> > >         elements of expr
> > >          > >                                    # to A and B,
> > >          > >                                    # assign
> > > with(expr, a +
> > >         b) to C.
> > >          > >
> > >          > >     Unfortunately, I don't think this could be done
> > entirely by
> > >          > >     transforming
> > >          > >     the expression (which is the way |> was done),
> > > and that
> > >         makes it a lot
> > >          > >     harder to write and to reason about.  E.g. what
> > > does
> > >         this do?
> > >          > >
> > >          > >         A <- 0
> > >          > >         [A, B = A + 10] <- list(1, A = 2)
> > >          > >
> > >          > >     According to the recipe above, I think it sets A
> > > to 1
> > >         and B to 12, but
> > >          > >     maybe a user would expect B to be 10 or 11.  And
> > >         according to that
> > >          > >     recipe this is an error:
> > >          > >
> > >          > >         [A, B = A + 10] <- c(1, A = 2)
> > >          > >
> > >          > >     which probably isn't what a user would expect,
> > > given
> > >         that this is fine:
> > >          > >
> > >          > >         [A, B] <- c(1, 2)
> > >          > >
> > >          > >     Duncan Murdoch
> > >          > >
> > >          >
> > >          > ______________________________________________
> > >          > R-devel using r-project.org <mailto:R-devel using r-project.org>
> > > mailing
> > list
> > >          > https://stat.ethz.ch/mailman/listinfo/r-devel
> > >         <https://stat.ethz.ch/mailman/listinfo/r-devel>
> > > 
> > 
> > 
> 
>         [[alternative HTML version deleted]]
> 
> ______________________________________________
> R-devel using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel



More information about the R-devel mailing list