[Rd] unsplit() mangles attributes

[Steve Martin]

Hello,

Unsplitting a named vector that's been split sets all the names as missing.

x <- 1:12
names(x) <- letters[x]
f <- gl(2, 6)

unsplit(split(x, f), f)
<NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA> <NA>
   1    2    3    4    5    6    7    8    9   10   11   12

The unsplit() function correctly deals with row names when unsplitting
a split data frame, and the same approach preserves regular names as
well. Here's a stripped-down version of unsplit() that keeps names:

unsplit_with_names <- function(value, f) {
  len <- length(f)
  x <- value[[1L]][rep(NA_integer_, len)] # names get lost here...
  split(x, f) <- value
  has_names <- !is.null(names(value[[1L]]))
  if (has_names) {
    split(names(x), f) <- lapply(value, names) # so add them back here
  }
  x
}

unsplit_with_names(split(x, f), f)
 a  b  c  d  e  f  g  h  i  j  k  l
 1  2  3  4  5  6  7  8  9 10 11 12

I plan on reporting this on bugzilla, with a more general fix, but
would first like to see if I'm missing anything, and check that my
reasoning is clear.

It seems that names are the only attribute for unclassed vectors that
survive the default method of split(), and so I think the above
version of unsplit() replaces all the attributes it can for unclassed
vectors.

I'm less confident about classed vectors, as unsplit() isn't generic
and potentially needs to deal with objects. Dates and factors work
fine, as it seems they can only lose names; this is addressed with the
above version of unsplit(). But are there other attributes for classed
objects that may get lost with unsplit? Can my fix above cause
problems for certain classes? (Note that I didn't use the recursion
that unsplit() uses for data frames, as that relies on names not
themselves having names.)

The real challenge is that unsplit need not have all the information
about the original object it's trying to put back together. Take the
case of a vector with a dim attribute.

y <- matrix(x, 3, 4, dimnames = list(letters[1:3], letters[1:4]))

unsplit(split(y, f), f)
[1]  1  2  3  4  5  6  7  8  9 10 11 12

A possible solution is for split() to record the attributes of its
argument for later use by unsplit(). Again, consider some
stripped-down alternatives:

split_with_attr <- function(x, f) {
  res <- split(x, f)
  structure(res, original.attr = attributes(x))
}

unsplit_with_attr <- function(value, f) {
  len <- length(f)
  x <- value[[1L]][rep(NA_integer_, len)]
  split(x, f) <- value
  attributes(x) <- attr(value, "original.attr")
  x
}

unsplit_with_attr(split_with_attr(y, f), f)
  a b c  d
a 1 4 7 10
b 2 5 8 11
c 3 6 9 12

But this seems complicated, and may muck up existing code. It would be
much easier if I can just restrict attention to restoring lost names
for unclassed vectors :)

Any thoughts are much appreciated.

Thanks,
Steve