Skip to content

Should as() find an ANY method even for an 'unregistered' S3 class?

3 messages · Martin Maechler, Michael Chirico

Original
Michael Chirico
# 
We have

setAs("ANY", "integer64", function(from) as.integer64(from))

I expected that would work to default as(x, "integer64") to use the S3
coercion, but not so for a different S3 class:

x = 1e10
class(x) = c("foo", "integer64")
as(x, "integer64")
# Error in as(x, "integer64") :
#   internal problem in as(): ?foo? is(object, "integer64") is TRUE, but
the metadata asserts that the 'is' relation is FALSE

This can be solved by registering the S3 inheritance with S4:

setOldClass(c("foo", "integer64"))
as(x, "integer64") # no error

But I still would have expected the "ANY" registration to have worked here.

Further context:
 - In {bit64} we refactored our c() method to use as(x, to) for generic
coercion of inputs
 - A downstream package extending the S3 class started erroring
 - So as long as the ANY coercion is not found we have two options (this
part bleeds into an r-package-devel topic)
    1. special-case the coercion to 'integer64', which is back-compatible
and solves the narrow issue for this particular coercion but may not be
global
    2. request the downstream (and future downstreams) to do the S4
registration

I also found that we can change our code:

as(x, "integer64", ext=TRUE)

But that directly goes against the documentation for ext=:
I don't see anything else in ?as or ?setAs covering this case.

Mike C
Martin Maechler
# 
> We have 
    >      setAs("ANY", "integer64", function(from) as.integer64(from))

Sorry, but  who is "we"? 
R does not know about "integer64" nor an as.integer64() 

I'd expect knowledgable posters to R-devel  to use reproducible
examples, please !?

Martin

    > I expected that would work to default as(x, "integer64")
    > to use the S3 coercion, but not so for a different S3
    > class:

    > x = 1e10 class(x) = c("foo", "integer64") as(x,
    > "integer64") # Error in as(x, "integer64") : # internal
    > problem in as(): ?foo? is(object, "integer64") is TRUE,
    > but the metadata asserts that the 'is' relation is FALSE

....
Michael Chirico
# 
OK, here is a self-contained version

setClass("integer64")
setAs("ANY", "integer64", function(from) {
  structure(from, class = "integer64")
})
x <- 1e10
class(x) <- c("foo", "integer64")
as(x, "integer64")
# Error in as(x, "integer64") :
#   internal problem in as(): ?foo? is(object, "integer64") is TRUE, but
the metadata asserts that the 'is' relation is FALSE

Mike C

On Thu, Mar 19, 2026 at 2:11?AM Martin Maechler <maechler at stat.math.ethz.ch>
wrote: