[R] question about dll crashing R

[Prof Brian Ripley]

On Thu, 3 Aug 2006, Thomas Lumley wrote:

> On Thu, 3 Aug 2006, Benn Fine wrote:
> 
> > I have ported some R code to C to make it faster.
> >
> > I can perform .Call("foobar",....) once and it
> > works fine. Absolutely correct answer.
> >
> > If I put a loop inside foobar and run the  main code
> > routine more than 100 times, it crashes R.
> >
> > Or if I call .Call("foobar"....) seperately more than
> > two tims it crashes R.
> >
> <snip>
> > SEXP mvntest (SEXP mean, SEXP cov, SEXP temp)
> >
> >
> > { int nrx , ncx , nry , ncy ,info,mode;
> > SEXP xdims , ydims , ans;
> >
> > int i,j, one=1;
> > info = 1;
> >
> > xdims = getAttrib (mean , R_DimSymbol ) ;
> > ydims = getAttrib (cov , R_DimSymbol ) ;
> > mode = REALSXP;
> > nrx = INTEGER( xdims ) [ 0 ] ;
> > ncx = INTEGER( xdims ) [ 1 ] ;
> > nry = INTEGER( ydims ) [ 0 ] ;
> > ncy = INTEGER( ydims ) [ 1 ] ;
> >
> >
> >
> > /* create the upper trianglular matrix A */
> >
> > /* such that t(A) %*% A = Sigma */
> >
> > GetRNGstate();
> >
> > F77_CALL(dpofa) ( REAL( cov ), &nry , &ncy , &info);
> > Rprintf("Info = %d\n",info);
> >
> >
> > for(i=0;i<nry;i++)
> >  for(j=0;j<i;j++)
> >    REAL(cov)[i+j*ncy] = 0.0;
> >
> >
> > PROTECT( ans = allocMatrix (mode, nrx , one ) ) ;
> > for(i=0;i<nry;i++)
> >  REAL(temp)[i] = rnorm(0,1);
> > ans = tmatrixprod(cov,temp);
> 
> ^^^^^^^^^
> Here you are returning a new SEXP from tmatrixprod but not protecting it. 
> Remember, PROTECT() operates on the *value* of its argument, so it 
> protects the thing that ans points to. When ans points to a new thing, it 
> is still the old thing that is protected.

and the old value of 'ans' is never used.

> > for(i=0;i<nry;i++)
> >  REAL(ans)[i] = REAL(ans)[i]+REAL(mean)[i];
> > UNPROTECT( 1 ) ;
> > PutRNGstate();
> > return( ans ) ;

but the next two lines do not do any allocations.  However, PutRNGstate 
does, so you need to UNPROTECT *after* that call.

While we are at it, REAL(ans) is a function call, and you will do better 
to assign the value once outside the loop.

> > I have a feeling I am messing up memory usage
> > somewhere but haven't a clue. Do I need to do garbage
> > collecting inside the C program ?The fact that the
> > code
> > runs a few times before R crashes is driving me nuts.
> > I send most of what I need into the C routine from R,
> > so I am not creating that many SEXP objects within the
> > program.
> >
> > Any hints or ideas ?

See the debugging sections in `Writing R Extensions'.  Using 
gctorture(TRUE) and valgrind (if possible) will find these things faster.

-- 
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