Henrik Bengtsson
on Wed, 19 Mar 2025 10:58:46 -0700 writes:
> Hello.
> I just (re-)discovered that structure(sum, init = 100) is "sticky",
> i.e. it stays with base::sum(). Here's an minimal example:
>> void <- structure(sum, some_attr = TRUE)
>> str(sum)
> function (..., na.rm = FALSE)
> - attr(*, "some_attr")= logi TRUE
>> From my very basic troubleshooting, it looks like this is happening
> for primitive functions. I think I understand that this comes down to
> primitive functions cannot be copied and baseenv() being special, i.e.
> in structure() there will be no copy made of the primitive function,
> and then attributes()<- ends up modifying the original primitive
> function. Even if this is a documented feature, I believe, it is a
> silent feature with risky side effects. We might already have code out
> there that silently produces incorrect results, because of this
> behavior. For example, I was about to a custom reduce() function where
> I control the initial value via an "init" attribute of the reducer
> function, e.g.
> x <- 1:10
> sum1 <- reduce(x, `+`)
> sum2 <- reduce(x, structure(`+`, init = 100)) # == 100 + sum1
> the 'init' attribute that was set in the sum2 statement will affect
> sum3 such that sum3 == sum2, not sum3 == sum1 as one would expect.
> If this is a bug, then I think it needs to be fixed. If it cannot be
> fixed, maybe this could be protected against, e.g.
>> void <- structure(sum, some_attr = TRUE)
> Error: You must not set attributes on a primitive function: sum
> Maybe it's sufficient to implement a protection against this in
> attr()<-, attributes()<-, and class()<-.
Yes, this is a bug -- a version of your code above:
void <- structure(sum, foo = TRUE)
identical(void, sum)
function (..., na.rm = FALSE) .Primitive("sum")
attr(,"foo")
[1] TRUE
Above, you are already looking at ways to deal with it; and that
*is* more delicate, indeed:
One thing we (R core) found previously (a couple of years ago)
was that the `structure(..)` function was
already too "slow" for it to be used in some base R functions,
and adding another if(..) to it will not help making it faster ...
OTOH, structure() being a pure R function (no direct .Internal(), .Call() ..)
is also important I think, as it keeps its code nicely self documenting.
I'm pretty convinced we should fix it by checking for primitive
functions inside the C code of `attributes<-` :
arguably the bug is really there, rather than in structure().
Patches are welcome (via R's Bugzilla or just here).
Martin