& performance difference (lazy evaluation) from R 4.2.0 to R 4.3.1

Question

This code below benchmarks the & operation given opposite conditions that benefit or not from lazy evaluation conditions in the vectors.

set.seed(1)

N <- 1e6
V <- runif(N)
v1 <- V > 0.1
v2 <- V > 0.4
v3 <- V > 0.9

mb_res_le <- microbenchmark::microbenchmark(
  times = 100, unit = "ms",
  v1 & v2 & v3, v3 & v2 & v1
)

ggplot2::autoplot(mb_res_le)

I understand the result from R 4.2.0. i.e. Having more FALSEes to the LHS of & performs quicker because of lazy evaluation. But I don't get how it is the other way aound for R 4.3.1 with the same comparison.

There are a few versions between 4.2.0 and 4.3.1, and this could be introduced somewhere in between, but I can't find anything in the release notes that could explain this.

I was surprised by the result from R 4.3.1. Which is why I tried earlier versions of R for which I remember being able to interpret the performance differences.

[updated plots after adding set.seed(1)]

Answer 1

Docker allows us to compare easily. I happen to have R 4.3.1 (current) and R 4.2.3 (final of the previous release cycle) here. I simply added install.packages("microbenchmark") and printed the result summary.

Then for R 4.3.1:

Unit: milliseconds
         expr       min        lq     mean    median       uq      max neval
 v1 & v2 & v3  8.815923  9.328674 10.49479  9.990183 11.14268 16.22374   100
 v3 & v2 & v1 10.906409 11.338012 12.42658 12.030553 12.90834 18.53381   100

and for R 4.2.3

Unit: milliseconds
         expr       min        lq     mean   median       uq      max neval
 v1 & v2 & v3  8.869552  9.362924 10.41865  9.83979 11.12159 14.79225   100
 v3 & v2 & v1 10.700630 11.142198 12.26546 11.61229 13.11973 17.68791   100

and modulo normal variation during benchmarking seem identical.

Modified code

set.seed(1)

N <- 1e6
V <- runif(N)
v1 <- V > 0.1
v2 <- V > 0.4
v3 <- V > 0.9

install.packages("microbenchmark")

mb_res_le <- microbenchmark::microbenchmark(times = 100, unit = "ms", 
                                            v1 & v2 & v3, v3 & v2 & v1)
print(mb_res_le)

#ggplot2::autoplot(mb_res_le)

Call

This is the one for 4.2.3, and 4.3.1 is obviously similar:

docker run --rm -ti -v $PWD:/mnt -w /mnt rocker/r-base:4.2.3 Rscript question.R