What speedups are you seeing?

Also,line 47 and 48 probably want the same change forcomplex.

For this version (38d024f):

sum_is_na<-Rcpp::cppFunction("int sum_is_na(NumericVector x) {  int n = 0;  for (R_xlen_t i = 0; i < x.size(); ++i) {    double xi = x[i];    if (Rcpp::traits::is_na<REALSXP>(xi)) ++n;  }  return n;}")N<-5e6x<- runif(N)x[x<0.5]<-NA_real_microbenchmark::microbenchmark(  sum(is.na(x)),  sum_is_na(x))#> Unit: milliseconds#>           expr      min       lq     mean   median       uq      max neval#>  sum(is.na(x)) 10.24388 10.62800 12.26077 11.25501 13.76406 18.73605   100#>   sum_is_na(x) 20.23770 20.65917 21.54281 21.25047 22.05458 25.35780   100#>  cld#>   a#>    b

Created on 2018-01-11 by thereprex package (v0.1.1.9000).

For a version usingstd::isnan() (e473015, currently in my local repo):

#> Unit: milliseconds#>           expr      min       lq     mean   median       uq      max neval#>  sum(is.na(x)) 10.32959 10.55353 12.03510 11.08237 13.15562 16.21278   100#>   sum_is_na(x) 13.08563 13.35475 13.72524 13.57882 14.02745 16.02345   100#>  cld#>   a#>    b

For master (f1ec6cf):

#> Unit: milliseconds#>           expr      min       lq     mean   median       uq      max neval#>  sum(is.na(x)) 10.31513 10.62298 12.18608 11.00471 13.85499 22.86626   100#>   sum_is_na(x) 49.11284 49.96843 51.05028 50.77832 52.05040 55.72882   100#>  cld#>   a#>    b

Bottom line: Faster by factor ~2.5, if we believe thatstd::isnan() will be the same asR_isnancpp() forever, we can make it faster by factor ~4. R is still faster even with the additional penalty of allocating and filling a logical vector (why???)

Will update complex implementation once we reach consensus.

2x speedup is good, even on a minuscule task. Just want to makereally sure the results are the same but I suppose our unit tests are comprehensive.

Can you please extend this to the complex case onlines 47 and 48 ?

Withstd::isnan() (x4) or withR_isnancpp() (x2.5)?

Sorry, you misunderstood what I meant. There was a commit missing. See#800 which I'll open in just a second as an extension of this. I made the change a few hours ago locally, and ran a rev.dep.

Do you have a Windows machine handy? I have been bisecting the last few hours trying to find what broke Windows. I get tests timeout on R Hub. Ithink it is#789 by@lionel but I am not quite sure.

Sure, I can run a few tests. Do you want me to run R CMD check with pre- and post-#789? On R-devel?

Yes, I just sent you a link to your "Kirill and R" handle at the mail box ... We can continue there.

I have some note, but I am mostly dizzy and need a break. A second set of eyes for checking would be good.

Also, look over at#789 -- I'll continue there.

I'll close this PR. Your commit is part of#800 though.