Join with NA can alter the type of the column

[tdeenes]

Consider the following scenario:

• create a data.table, having an integer and a logical column with NAs ('x' and 'y', respectively)
• case 1: join on list(x = NA) (note: originally x is integer, but i.x is now logical)
• case 2: join on list(y = NA_integer_) (note: originally y is logical, but i.y is now integer)

d <- data.table(x = c(NA_integer_, 2L), y = c(TRUE, NA))
str(x)
# Classes ‘data.table’ and 'data.frame':	2 obs. of  2 variables:
# $ x: int  NA 2
# $ y: logi  TRUE NA

dx <- d[list(x = NA), on = "x"]
dy <- d[list(y = NA_integer_), on = "y"]

One expects that the type of the columns in 'dx' and 'dy' will be the same the respective columns in the original data.table. Indeed, this is what we get until data.table_1.12.2:

typeof(dx$x)
# [1] "integer"
typeof(dy$y)
# [1] "logical"

However, from 1.12.4 on (tested also with the most recent devel version), this is what we get:

typeof(dx$x)
# [1] "logical"
typeof(dy$y)
# [1] "integer"

[jangorecki]

During join, coercion is properly applied, which can be seen with verbose=TRUE, but then when constructing answer at

data.table/R/data.table.R

Line 1237 in 12586af

ans[[target]] = .Call(CsubsetVector,i[[source]],ii) # i.e. i[[source]][ii]

the whole column from i is copied as is, with no coercion.
Not sure how to address that in a robust way.
We could use memrecycle which do coercion nicely, but then we need to pre-allocate and copy into it. Or coerceAs from #4491 which is wrapper around memrecycle. Any other ideas @mattdowle ?