[R] R/C++ interfaces: crashes when using .c(), followed by correct results when R restarted

[Prof Brian Ripley]

First, the posting guide asks you to send questions about compiled code 
to the R-devel list.

These are all classic symptoms of the use of uninitialized variables or 
writing out of bounds, but of which you can catch by using valgrind: see 
'Writing R Extensions'.  They are much harder to find on Windows (which 
although you did not tell us your OS, is what this looks like).


On 20/09/2012 10:19, Franckx Laurent wrote:
> Dear all
>
>
> I have written a function in C++  , equil_distC, that I am calling from an R script.
>
> In the last few days, R has repeatedly crashed when calling this function, or delivered obviously wrong outputs. However, when I restarted R after the crash, the results turned out to be OK most of the time although I had not modified the C++ code.
>
> Although the code runs correctly now, I am not sure why it does work right now, and didn't just an hour ago.
>
> Let me summarize the main steps I undertook:
>
> * After having modified the script with the C++ code, I compiled it using R CMD SHLIB mylatestversion.cpp - no problems occurred then.
> * Next, I loaded the dll into R with
> dyn.load("C:\\Ccodefortransortmodel\\ mylatestversion.dll")
> * Next, I called equil_distC with
> testres2 <- .C("equil_distC", as.integer(rowmaxfiets),as.integer(colmaxfiets), as.double(nodes), as.integer(links_dist_calc), as.integer(cyclepathsfirstline), as.integer(numzones), as.integer(numnodes), as.integer(numlinks), result = double(6))
> * About half of the time, whenever I had changed the C++ code, the results were either nonsensical or R crashed.
> * When R did not crash, I unloaded the dll with:
> dyn.unload("C:\\Ccodefortransortmodel\\equildistCforodpair_2012_09_20v8.dll")
> * Next, I deleted the .dll and the .o of the most recent version, I modified the C++ code and I ran the above procedure again.
> * In most cases when R crashed, a new run of the procedure above resulted in correct results, even though the C++ code was not modified.
>
> I have also noticed that, in several cases, apparently irrelevant changes to the code would lead to a crash. This occurred for instance when I replaced %d with %f in Rprintf()) - my first conjecture was actually that this was the root of the problem, but the code runs correctly now with %f.
>
> My own conjecture is that, some way or another, old versions of equil_distC kept on circulating in memory.
>
> For instance, even after a dyn.unload of all previous versions of the dll, "is.loaded("equil_distC")" would still lead to a TRUE.
>
> I would like to understand the structural reason for these problems. Again, the problem is not the C++ code - it does run correctly now, although it had crashed in the previous run.
>
>
> Best regards
>
>
>
> Laurent Franckx, PhD
> VITO NV
> Boeretang 200, 2400 MOL, Belgium
> Tel. + 32 14 33 58 22
> Skype: laurent.franckx
> laurent.franckx at vito.be



-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595