I'm running into an issue where code that ran previously becomes so slow it doesn't finish. I've traced the slowness back to the function .[integer64 in this package.

Or more concretely to this line:

na_idx <- is.na(rep(TRUE, length(x))[i])

In my case I have a duckdb database I'm using DBI to fetch results of a query in batches. In my case i is 100k integers and x is almost 500 million integer64's. Materializing a vector of size x seems to be the issue. Using tag 4.0.5 does not have this issue while 4.5.2 and later have it. From the changelog it there was a bugfix I think is responsible: ""[.integer64"(x,i) can now cope with i longer than x"

The following code should demonstrate the difference when running with tags 4.0.5 and 4.6.0-1

# remotes::install_github("r-lib/[email protected]") or remotes::install_github("r-lib/[email protected]")
library(bit64)
x <- as.integer64(rep(1, 1e8))
i <- sample(c(NA, 1:1e4, 1e7 + 1), size = 1e5, replace = TRUE)
microbenchmark(
  bit64 = x[i],
  times = 10L
)

4.0.5:

Unit: microseconds
  expr     min      lq     mean  median      uq      max neval
 bit64 672.959 894.574 916.0243 945.992 961.139 1027.361    10

4.6.0-1:

Unit: milliseconds
  expr      min       lq     mean   median       uq      max neval
 bit64 194.0301 200.1372 213.4576 204.3696 233.6707 235.2293    10

current main:

Unit: milliseconds
  expr      min       lq     mean  median       uq      max neval
 bit64 193.7276 200.4751 214.2612 205.575 232.2011 238.9536    10

Roughly a difference of ~333x. I thought about replacing the offending line with

 na_idx <- is.na(i) | (i > length(x))

But unfortunately that doesn't work when i is boolean and length(i) > length(x) (untested case btw).