C code validation

Dear R-devel,

I'm trying to validate the results from a C function, against a (trial
and tested) older R function. For reasons unknown to me, the C
function seems to give different result sometimes at each trial, even
with the very same data.

These are the relevant outputs from R:

library(QCA)

Loading required package: lpSolve

benchmark <- function(x, y) {

+     index <- 0
+     while((index <- index + 1) < length(x)) {
+         x <- setdiff(x, findSubsets(y + 1, x[index], max(x)))
+     }
+     return(x)
+ }

set.seed(12345)
exprnec <- sort(sample(13:19683, 500))
noflevels <- rep(2, 9)
mbase <- rev(c(1, cumprod(rev(noflevels + 1))))[-1]
foo1 <- benchmark(exprnec, noflevels)
foo2 <- .Call("removeRedundants", exprnec, noflevels, mbase)

all.equal(foo1, foo2)

[1] "Numeric: lengths (184, 181) differ"

set.seed(12345)
exprnec <- sort(sample(13:19683, 500))
foo2 <- .Call("removeRedundants", exprnec, noflevels, mbase)
all.equal(foo1, foo2)

[1] TRUE

set.seed(12345)
exprnec <- sort(sample(13:19683, 500))
foo2 <- .Call("removeRedundants", exprnec, noflevels, mbase)
all.equal(foo1, foo2)

[1] TRUE

set.seed(12345)
exprnec <- sort(sample(13:19683, 500))
foo2 <- .Call("removeRedundants", exprnec, noflevels, mbase)
all.equal(foo1, foo2)

[1] "Numeric: lengths (184, 173) differ"


The relevant C code can be found here:
http://codeviewer.org/view/code:287e

As I am too baffled by these (to me strange) results, any hint is much
appreciated.

Thanks very much in advance,
Adrian

[...]
See the advice in 'Writing R Extensions'.  In particular, the most common
cause is the use of initialized values, so run under valgrind.

Also check that your C code does not change any of its arguments: if it does
you may end up with self-modifying R code.  (That's probably rare, but I
spent a couple of hours on such an example last night.)