Weird erratic error and illogical error message, could someone explain this?

On 09/03/2010 04:42 AM, Duncan Murdoch wrote:

Philippe Grosjean wrote:

Hello,

It's several days I try to track this bug, and even cannot cook a
reproducible example. Yet, it occurs consistently in a long-running
task after a variable period of time. Here is an example:

I would look closely at the other software that is running in your
long example.  Does it include C (or other external) code?  Look
closely at that, it might be writing outside it's own allocated
memory.  Also check for correct protection of intermediate results, if
you're producing SEXPs in the external code.  (Running under gctorture
might flush out the bug more quickly if the latter is the problem.)

If you're only running R code, then this looks like a bug in R, but it
might still be worth trying gctorture to make it reproducible.  We
won't be able to fix it if we can't reproduce it.

Duncan Murdoch

... my long-running code [as I said, cannot give something simple
that produces this bug in a reproducible manner]

Error in match(x, table, nomatch = 0L) :
      formal argument "nomatch" matched by multiple actual arguments

  >  traceback()

6: match(x, table, nomatch = 0L)
5: "factor" %in% attrib[["class", exact = TRUE]]
4: structure(.Internal(Sys.time()), class = c("POSIXt", "POSIXct"))
3: Sys.time()
2: chemTrigger() at chemostat_1.0-1.R#1132
1: chemRun()

I think this is a bug in R that has been fixed in the subversion commit
below, and so should be fixed in R-2.11.1.
What is your sessionInfo(), and does your error occur in the devel
version of R?

Martin

r51232 | falcon | 2010-03-09 13:59:48 -0800 (Tue, 09 Mar 2010) | 14 lines
Changed paths:
    M /trunk/src/main/match.c

Fix bug in matchArgs triggered by gc and finalizers

matchArgs was modifying the general purpose bits of SEXPs making up the
'formals' argument to keep track of ARGUSED.  When gc is triggered
inside matchArgs, finalizer code can end up calling matchArgs on the
same function (hence same formals) resulting in corruption of gp bits.

This patch uses an int array allocated on the stack to keep track of
ARGUSED and avoids modifying the SEXPs in formals.  In place modification
of SEXPs in supplied still occurs via ARGUSED/SET_ARGUSED, but should be
safe as long as no new allocating function calls are added to matchArgs.

A reproducible report of this bug was reported here:
https://stat.ethz.ch/pipermail/bioc-sig-sequencing/2010-March/000997.html

So, the culprid is a call inside `%in%` (from within structure() in
Sys.time()). But I can run millions times `%in%`, or structure(), or
Sys.time() on my machine without producing this bug. Arguments at 5:
are simple character strings. They don't hurt!

Also, I am lost because the message is totally illogical in the
context where it appears: I can understand this message here:

  >  match(1, 2, nomatch = 0L, nomatch = NA)

Error in match(1, 2, nomatch = 0L, nomatch = NA) :
    formal argument "nomatch" matched by multiple actual arguments

or here:

  >  test<- function (...) match(1, ..., nomatch = 0L)
  >  test(2, nomatch = NA)

Error in match(1, ..., nomatch = 0L) :
    formal argument "nomatch" matched by multiple actual arguments

but in the call "match(x, table, nomatch = 0L)" where x is the
character string "factor" and table is another character string
("numeric") extracted from a list, I don't understand why it produces
this error message. '.Internal(Sys.time())' uses do_systime c code
that returns a one-element double... not something that can hurt here?!

Can someone explain me, or give me an example where an argument is
NOT duplicated in the call (well, as I understand it here) and where
one gets such an error message? And why?

Many thanks, I am desperate :-(

I got this error on R 2.11.1 on Mac OS X 10.6.4, and on R 2.10.1 on
Windows XP SP3 (but it does not matter, since I cannot cook a
reproducible example).

Philippe

P.S.: seems related to this:
http://finzi.psych.upenn.edu/Rhelp10/2008-June/165101.html