Investigate switching from Rcpp to cpp11

[mcol]

Package Rcpp is very convenient but also quite bloated. Perhaps using cpp11 can provide the same convenience but with a smaller footprint. We should try to convert some of our functions and see if it's a viable approach.

[mcol]

I've converted create_UID.cpp and the result is this:

#include <R_ext/Random.h>
#include <cpp11.hpp>
#include <time.h>

// [[cpp11::register]]
cpp11::strings create_UID() {

  //define variables
  time_t rawtime;
  struct tm * timeinfo;
  char timestamp [80];

  //set date + timestamp (code snippet taken from C++ reference page)
  time (&rawtime);
  timeinfo = localtime (&rawtime);
  strftime (timestamp,80,"%Y-%m-%d-%I:%M.",timeinfo);

  // get a 4-digit random number
  //according to the CRAN policy the standard C-function, rand(), even sufficient here, is not allowed
  GetRNGstate();
  unsigned int random = (unsigned int) (unif_rand() * 10000);
  PutRNGstate();

  //combine and return results
  sprintf(timestamp, "%s%04u", timestamp, random);
  return cpp11::writable::strings( { timestamp } );
}

Here I shortened the random number to 4 digits (which may be sufficient for our purposes) as cpp11 doesn't provide an automatic way of converting a float to a string as was done previously.

Overall it was not too hard, but it took me a bit of back and forth to get the string operations to compile (the docs barely talk about strings). However I'm still getting a couple of C++ warnings for the string concatenation:

   create_UID.cpp: In function ‘cpp11::strings create_UID()’:
   create_UID.cpp:34:25: warning: ‘%04u’ directive writing between 4 and 10 bytes into a region of size between 1 and 80 [-Wformat-overflow=]
      34 |   sprintf(timestamp, "%s%04u", timestamp, random);
         |                         ^~~~
   create_UID.cpp:34:10: note: ‘sprintf’ output between 5 and 90 bytes into a destination of size 80
      34 |   sprintf(timestamp, "%s%04u", timestamp, random);
         |   ~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
   create_UID.cpp:34:10: warning: ‘sprintf’ argument 3 overlaps destination object ‘timestamp’ [-Wrestrict]
   create_UID.cpp:20:8: note: destination object referenced by ‘restrict’-qualified argument 1 was declared here
      20 |   char timestamp [80];
         |        ^~~~~~~~~

Major differences:

• We need to explicitly deal with the RNG state (trivial)
• String concatenation is not quite as straightforward (before it was just timestamp + Rcpp::as<std::string>(random))
• Registration of the cpp function is seamless as before, once it compiles it just works

So overall it's hard to say from this if it's worth continuing. I'll probably have to try a more involved function to see.

[RLumSK]

Sorry, I don't have good advice here, I would need to sit down with the code myself; currently I am little bit under time constraints.

[mcol]

No worries, I was just summarising my current thoughts. As I mentioned, I'll get back to it at some point to see if it's worth carrying on with this.

[mcol]

This may be blocked by the performance implication for analyse_IRSAR.RF(): the src_analyse_IRSARRF_SRS() C++ function uses RcppArmadillo both for its arma::vec container and for RcppArmadillo::sample(), and changing these may bring performance penalties.

I had tried to use NumericVector() instead of arma::vec, but already this seemingly innocuous change caused a little slow down; replacing RcppArmadillo::sample() with Rcpp::sample() seemed pretty neutral, but for the move to cpp11, we would ideally move the sampling to R, and that may also cause a slowdown.

[RLumSK]

OK, then we keep it as it is for now.