[Rd] class(<matrix>) |--> c("matrix", "arrary") -- and S3 dispatch

Martin Maechler m@ech|er @end|ng |rom @t@t@m@th@ethz@ch
Thu Nov 21 17:57:51 CET 2019

TLDR: This is quite technical, still somewhat important: 
     1)  R 4.0.0 will become a bit more coherent: a matrix is an array
     2)  Your package (or one you use) may be affected.

>>>>> Martin Maechler 
>>>>>     on Fri, 15 Nov 2019 17:31:15 +0100 writes:

>>>>> Pages, Herve 
>>>>>     on Thu, 14 Nov 2019 19:13:47 +0000 writes:

    >> On 11/14/19 05:47, Hadley Wickham wrote:
    >>> On Sun, Nov 10, 2019 at 2:37 AM Martin Maechler ... wrote:

    >>>>> Note again that both "matrix" and "array" are special [see ?class] as
    >>>>> being of  __implicit class__  and I am considering that this
    >>>>> implicit class behavior for these two should be slightly
    >>>>> changed ....
    >>>>> And indeed I think you are right on spot and this would mean
    >>>>> that indeed the implicit class
    >>>>> "matrix" should rather become c("matrix", "array").
    >>>> I've made up my mind (and not been contradicted by my fellow R
    >>>> corers) to try go there for  R 4.0.0   next April.

    >>> I can't seem to find the previous thread, so would you mind being a
    >>> bit more explicit here? Do you mean adding "array" to the implicit
    >>> class?

    >> It's late in Europe ;-)

    >> That's my understanding. I think the plan is to have class(matrix()) 
    >> return c("matrix", "array"). No class attributes added to matrix or 
    >> array objects.

    >> It's all what is needed to have inherits(matrix(), "array") return TRUE 
    >> (instead of FALSE at the moment) and S3 dispatch pick up the foo.array 
    >> method when foo(matrix()) is called and there is no foo.matrix method.

    > Thank you, Hervé!  That's exactly the plan.

BUT it's wrong what I (and Peter and Hervé and ....) had assumed: 

If I just change the class
     (as I already did a few days ago, but you must activate the change 
      via environment variable, see below),

S3 dispatch does *NOT* at all pick it up:
"matrix" (and "array") are even more special here (see below),
and from Hadley's questions, in hindsight I now see that he's been aware
of that and I hereby apologize to Hadley for not having thought
and looked more, when he asked ..

Half an hour ago, I've done another source code commit (svn r77446),
to "R-devel" only, of course, and the R-devel NEWS now starts as




    •  .... intention that the next non-patch release should be 4.0.0.

    • R now builds by default against a PCRE2 library ........

    • For now only active when environment variable
      _R_CLASS_MATRIX_ARRAY_ is set to non-empty, but planned to be the
      new unconditional behavior when R 4.0.0 is released:

      Newly, matrix objects also inherit from class "array", namely,
      e.g., class(diag(1)) is c("matrix", "array") which invalidates
      code (wrongly) assuming that length(class(obj)) == 1, a wrong
      assumption that is less frequently fulfilled now.  (Currently
      only after setting _R_CLASS_MATRIX_ARRAY_ to non-empty.)

      S3 methods for "array", i.e., <someFun>.array(), are now also
      dispatched for matrix objects.

(where only the very last 1.5 lines paragraph is new.)

Note the following
(if you use a version of R-devel, with svn rev >= 77446; which
 you may get as a binary for Windows in about one day; everyone
 else needs to compile for the sources .. or wait a bit, maybe
 also not much longer than one day, for a docker image) :

> Sys.unsetenv("_R_CLASS_MATRIX_ARRAY_") # ==> current R behavior
> class(m <- diag(1))
[1] "matrix"
> Sys.setenv("_R_CLASS_MATRIX_ARRAY_" = "BOOH !") # ==> future R behavior
> class(m)
[1] "matrix" "array" 
> foo <- function(x) UseMethod("foo")
> foo.array <- function(x) "made in foo.array()"
> foo(m)
[1] "made in foo.array()"
> Sys.unsetenv("_R_CLASS_MATRIX_ARRAY_")# ==> current R behavior
> foo(m)
Error in UseMethod("foo") : 
  no applicable method for 'foo' applied to an object of class "c('matrix', 'double', 'numeric')"

> Sys.setenv("_R_CLASS_MATRIX_ARRAY_" = TRUE) # ==> future R behavior
> foo(m)
[1] "made in foo.array()"
> foo.A <- foo.array ; rm(foo.array)
> foo(m)
Error in UseMethod("foo") : 
  no applicable method for 'foo' applied to an object of class "c('matrix', 'array', 'double', 'numeric')"

So, with my commit 77446, the  _R_CLASS_MATRIX_ARRAY_
environment variable also changes the 

   "S3 dispatch determining class"

mentioned as 'class' in the error message (of the two cases, old
and new) above,  which in R <= 3.6.x for a numeric matrix is

    c('matrix', 'double', 'numeric')

and from R 4.0.0 on  will be

    c('matrix', 'array', 'double', 'numeric')

Note that this is *not* (in R <= 3.6.x, nor very probably in R 4.0.0)
the same as  R's  class().
Hadley calls this long class vector the  'implicit class' -- which
is a good term but somewhat conflicting with R's (i.e. R-core's)
"definition" used in the  ?class  help page (for ca. 11 years).

R's internal C code has a nice function class R_data_class2()
which computes this 'S3-dispatch-class' character (vector) for
any R object, and R_data_class2() is indeed called from (the
underlying C function of)  R's UseMethod().

Using the above fact of an error message,
I wrote a nice (quite well tested) function  my.class2()  which
returns this S3_dispatch_class() also in current versions of R:

my.class2 <- function(x) { # use a fn name not used by any sane ..
    foo.7.3.343 <- function(x) UseMethod("foo.7.3.343")
    msg <- tryCatch(foo.7.3.343(x), error=function(e) e$message)
    clm <- sub('"$', '', sub(".* of class \"", '', msg))
    if(is.language(x) || is.function(x))
    else {
        cl <- str2lang(clm)
        if(is.symbol(cl)) as.character(cl) else eval(cl)

## str2lang() needs R >= 3.6.0:
if(getRversion() < "3.6.0") ## substitute for str2lang(), good enough here:
    str2lang <- function(s) parse(text = s, keep.source=FALSE)[[1]]

Now you can look at such things yourself:

## --------------------- the "interesting" cases : ---
## integer and double
my.class2( pi) 	# == c("double",  "numeric")
my.class2(1:2) 	# == c("integer", "numeric")
## matrix and array [also combined with int / double ] :
my.class2(matrix(1L, 2,3))   	# == c(matrixCL, "integer", "numeric")  <<<
my.class2(matrix(pi, 2,3))   	# == c(matrixCL,  "double", "numeric")  <<<
my.class2(array("A", 2:3))   	# == c(matrixCL,  "character")          <<<
my.class2(array(1:24, 2:4))  	# == c("array",  "integer", "numeric")
my.class2(array( pi , 2:4))  	# == c("array",   "double", "numeric")
my.class2(array(TRUE, 2:4))  	# == c("array", "logical")
my.class2(array(letters, 2:4))	# == c("array", "character")
my.class2(array(1:24 + 1i, 2))	# == c("array", "complex")

## other cases
my.class2(NA) 	# == class(NA) : "logical"
my.class2("A") 	# == class("B"): "character"
my.class2(as.raw(0:2)) 	# == "raw"
my.class2(1 + 2i) 	# == "complex"
my.class2(USJudgeRatings)#== "data.frame"
my.class2(class) 	# == "function" # also for a primitive
my.class2(globalenv()) 	# == "environment"
my.class2(quote(sin(x)))# == "call"
my.class2(quote(sin) )  # == "name"
my.class2(quote({}))	# == class(*) == "{"
my.class2(quote((.)))	# == class(*) == "("


note that of course, the lines marked "<<<" above, contain
'matrixCL'  which is "matrix" in "old" (i.e. current) R,
	    and is c("matrix", "array") in "new" (i.e. future) R.

Last but not least: It's quite trivial (only few words need to
be added to the sources; more to the documentation)  to add an R
function to base R which provides the same as my.class2() above,
(but much more efficiently, not via catching error messages !!),
and my current proposal for that function's name is  .class2()
{it should start with a dot ("."), as it's not for the simple
 minded average useR ... and you know how I'm happy with
 function names that do not need one single [Shift] key ...}

The current plan contains

1)  Notify CRAN package maintainers (ca 140) whose packages no
    longer pass R CMD check  when the feature is turned on
    (via setting the environment variable) in R-devel.

2a) (Some) CRAN team members set _R_CLASS_MATRIX_ARRAY_ (to non-empty),
    as part of the incoming checks, at least for all new CRAN submissions

2b) set the  _R_CLASS_MATRIX_ARRAY_ (to non-empty), as part of
    ' R CMD check --as-cran <pkg>'

3)  Before the end of 2019, change the R sources (for R-devel)
    such that it behaves as it behaves currently when the environment
    variable is set *AND* abolish this environment variable from
    the sources.  {read on to learn *why*}

Consequently (to 3), R 4.0.0 will behave as indicated, unconditionally.

Note that (as I've shown above in the first example set) this is
set up in such a manner that you can change the environment
variable during a *running* R session, and observe the effect immediately.
This however lead to some slow down of quite a bit of the R
code, because actually the environment variable has to be
checked quite often (easily dozens of times for simple R calls).

For that reason, we want to do "3)" as quickly as possible.

Please do not hesitate to ask or comment
-- here, not on Twitter, please --  noting that I'll be
basically offline for an extended weekend within 24h, now.

I hope this will eventually to lead to clean up and clarity in
R, and hence should be worth the pain of broken
back-compatibility and having to adapt your (almost always only
sub-optimally written ;-)) R code, 
see also my Blog   http://bit.ly/R_blog_class_think_2x

Martin Maechler
ETH Zurich and R Core team

More information about the R-devel mailing list