NOTE: stopifnot() comes with some overhead and is getting slower in R (>= 3.5.0)

[HenrikBengtsson]

Pasting in my old private notes to myself here:

Date: Sun, Sep 11, 2016, 22:16
Subject: R SPEEDUP: stopifnot()

stopifnot() is always calling match.call() even when there are no error.  It adds unnecessary overhead.

stopifnot <- function (...)
{
    n <- length(ll <- list(...))
    if (n == 0L)
        return(invisible())
    mc <- match.call()
    for (i in 1L:n) if (!(is.logical(r <- ll[[i]]) && !anyNA(r) && all(r))) {
        ch <- deparse(mc[[i + 1]], width.cutoff = 60L)
        if (length(ch) > 1L)
            ch <- paste(ch[1L], "....")
        stop(sprintf(ngettext(length(r), "%s is not TRUE", "%s are not all TRUE"),
            ch), call. = FALSE, domain = NA)
    }
    invisible()
}


stopifnot2 <- function (...)
{
    n <- length(ll <- list(...))
    if (n == 0L)
        return(invisible())
    for (i in 1L:n) if (!(is.logical(r <- ll[[i]]) && !anyNA(r) &&
        all(r))) {
        mc <- match.call()
        ch <- deparse(mc[[i + 1]], width.cutoff = 60L)
        if (length(ch) > 1L)
            ch <- paste(ch[1L], "....")
        stop(sprintf(ngettext(length(r), "%s is not TRUE", "%s are not all TRUE"),
            ch), call. = FALSE, domain = NA)
    }
    invisible()
}


stopifnot3 <- function (...)
{
    n <- length(ll <- list(...))
    if (n == 0L)
        return(invisible())
    for (i in 1L:n) if (!(is.logical(r <- .subset2(ll, i)) && !anyNA(r) && all(r))) {
        mc <- match.call()
        ch <- deparse(mc[[i + 1]], width.cutoff = 60L)
        if (length(ch) > 1L)
            ch <- paste(ch[1L], "....")
        stop(sprintf(ngettext(length(r), "%s is not TRUE", "%s are not all TRUE"),
            ch), call. = FALSE, domain = NA)
    }
    invisible()
}

library("microbenchmark")
stats <- microbenchmark(
  stopifnot(TRUE),
  stopifnot2(TRUE),
  stopifnot3(TRUE)
)
print(stats)
# Unit: microseconds
#              expr   min     lq      mean median    uq       max neval cld
#   stopifnot(TRUE) 4.150 4.3555  95.19823 4.4855 4.682  9040.550   100   a
#  stopifnot2(TRUE) 1.246 1.3605 115.90333 1.4555 1.549 11443.605   100   a
#  stopifnot3(TRUE) 1.248 1.3560  99.56975 1.4170 1.515  9811.166   100   a

library("microbenchmark")
x <- 1:1e6
stats <- microbenchmark(
  stopifnot(is.numeric(x), length(x) == 1e6),
  stopifnot2(is.numeric(x), length(x) == 1e6), ## clearly faster than stopifnot()
  stopifnot3(is.numeric(x), length(x) == 1e6) ## same speed as stopifnot2()
)
print(stats)
# Unit: microseconds
#                                           expr   min     lq    mean median     uq    max neval cld
#   stopifnot(is.numeric(x), length(x) == 1e+06) 4.613 4.8780 5.62728 5.0145 5.3120 43.986   100   b
#  stopifnot2(is.numeric(x), length(x) == 1e+06) 1.841 2.0095 2.18218 2.0720 2.2250  5.159   100  a 
#  stopifnot3(is.numeric(x), length(x) == 1e+06) 1.831 1.9770 2.09460 2.0455 2.1575  4.376   100  a 

[HenrikBengtsson]

... and follow up note from 2018-03-15: Ouch... in R-devel, stopifnot() has become yet 4-5 times slower;

> source("stopifnot.R")
Unit: microseconds
                                          expr    min      lq     mean  median      uq    max
  stopifnot(is.numeric(x), length(x) == 1e+06) 15.308 15.7010 17.50978 15.9335 16.3295 71.253
 stopifnot2(is.numeric(x), length(x) == 1e+06)  1.489  1.6425  1.90277  1.7815  1.9605  8.348
 neval
   100
   100

which is due to a complete rewrite using tryCatch() and withCallingHandlers().

[HenrikBengtsson]

I've moved to using a private, simpler barebone stop_if_not() implementation, e.g. https://github.com/HenrikBengtsson/listenv/blob/251138d4576cfd0a531570698c338778af5fc76b/R/utils.R#L47-L61

[mmaechler]

Well, that version uses list(...) as pre R-3.5.0 did too ... not delivering what its documentation promised:
Expressions are evaluated "one by one" and the evaluation stops with an error as soon as one of the results is not all true. Suharto Anggono's proposals on the R-devel list will lead to a fast and good stopifnot() from R-3.6.0 on .. i.e., in less than 2 months from now.