R devel - Dec 2005 - .Call not counting parameters consistently (PR#8450)

The R_registerRoutines C function allows the number of parameters to a
.Call function to be registered.  For example, the tools package
function md5sum() calls "Rmd5", which has been registered to require
just one parameter.

But if it is called with the wrong number of parameters, only the first
error gets caught:

  > library(tools)
  > .Call("Rmd5",1,2,PACKAGE="tools")
Error: Incorrect number of arguments (2), expecting 1 for Rmd5
  > .Call("Rmd5",1,2,PACKAGE="tools")
Error: argument 'files' must be character

This happens in Windows versions of R-patched and R-devel.  Charlie
Geyer was bitten by this in his fuzzyRankTests package (version 0.3),
which fails check because of it on Windows, but not on Unix.  I haven't
done any testing on Unix.

I've traced into the do_dotcall function in src/main/dotcode.c, and I
see that on the second call the "symbol.symbol.call" member is NULL,
so
no test is done, but I don't see why this is happening.  The
resolveNativeRoutine function does some strange stuff.

Duncan Murdoch

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Coincidentally, I am in the process of working on a related aspect of
symbol resolution.

The issue is likely to be the caching of native symbols
that we do. We do not cache the registration information,
just the address of the routine.  And so the test is not
repeated.  We can fix this, but it will make things more
complex. This is a good example of where using the new mechanism
within namespaces to resolve the symbols when the package
is loaded and making them into regular R objects that
are passed to .Call/.C/... makes things more rational.
In that case, the caching is done in R and type information
is more directly available.


Off hand, I don't see why this would behave differently
on Windows and Unix, so I'd appreciate if you could let
us know the specific call that raises the error on Windows
and not on Unix.


D.


murdoch at stats.uwo.ca wrote:
> The R_registerRoutines C function allows the number of parameters to a
> .Call function to be registered.  For example, the tools package
> function md5sum() calls "Rmd5", which has been registered to
require
> just one parameter.
> 
> But if it is called with the wrong number of parameters, only the first
> error gets caught:
> 
>   > library(tools)
>   > .Call("Rmd5",1,2,PACKAGE="tools")
> Error: Incorrect number of arguments (2), expecting 1 for Rmd5
>   > .Call("Rmd5",1,2,PACKAGE="tools")
> Error: argument 'files' must be character
> 
> This happens in Windows versions of R-patched and R-devel.  Charlie
> Geyer was bitten by this in his fuzzyRankTests package (version 0.3),
> which fails check because of it on Windows, but not on Unix.  I haven't
> done any testing on Unix.
> 
> I've traced into the do_dotcall function in src/main/dotcode.c, and I
> see that on the second call the "symbol.symbol.call" member is
NULL, so
> no test is done, but I don't see why this is happening.  The
> resolveNativeRoutine function does some strange stuff.
> 
> Duncan Murdoch
> 
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
- --
Duncan Temple Lang                    duncan at wald.ucdavis.edu
Department of Statistics              work:  (530) 752-4782
4210 Mathematical Sciences Building   fax:   (530) 752-7099
One Shields Ave.
University of California at Davis
Davis,
CA 95616,
USA
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2 (Darwin)

iD8DBQFDsu8Q9p/Jzwa2QP4RAuqGAJ9oHP7lhmb7IN9c/3WoRL5qimCTSACfaNhr
WFY6NLnh2dfDTOGcaStbgNo=rCZg
-----END PGP SIGNATURE-----

Duncan Temple Lang wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Coincidentally, I am in the process of working on a related aspect of
> symbol resolution.
> 
> The issue is likely to be the caching of native symbols
> that we do. We do not cache the registration information,
> just the address of the routine.  And so the test is not
> repeated.  We can fix this, but it will make things more
> complex. This is a good example of where using the new mechanism
> within namespaces to resolve the symbols when the package
> is loaded and making them into regular R objects that
> are passed to .Call/.C/... makes things more rational.
> In that case, the caching is done in R and type information
> is more directly available.
> 
> 
> Off hand, I don't see why this would behave differently
> on Windows and Unix, so I'd appreciate if you could let
> us know the specific call that raises the error on Windows
> and not on Unix.

In principle, the code where this came up in the tests for the new CRAN 
package fuzzyRankTests is:

   library(fuzzyRankTests)
   set.seed(123)
   mu <- 0
   x <- rnorm(10)
   y <- rnorm(10) + 1.5
   try(.Call("fpvranksum", y, x, mu, "great",
      PACKAGE = "fuzzyRankTests"))

Uwe


> 
> D.
> 
> 
> murdoch at stats.uwo.ca wrote:
> 
>>The R_registerRoutines C function allows the number of parameters to a
>>.Call function to be registered.  For example, the tools package
>>function md5sum() calls "Rmd5", which has been registered to
require
>>just one parameter.
>>
>>But if it is called with the wrong number of parameters, only the first
>>error gets caught:
>>
>>  > library(tools)
>>  > .Call("Rmd5",1,2,PACKAGE="tools")
>>Error: Incorrect number of arguments (2), expecting 1 for Rmd5
>>  > .Call("Rmd5",1,2,PACKAGE="tools")
>>Error: argument 'files' must be character
>>
>>This happens in Windows versions of R-patched and R-devel.  Charlie
>>Geyer was bitten by this in his fuzzyRankTests package (version 0.3),
>>which fails check because of it on Windows, but not on Unix.  I
haven't
>>done any testing on Unix.
>>
>>I've traced into the do_dotcall function in src/main/dotcode.c, and
I
>>see that on the second call the "symbol.symbol.call" member is
NULL, so
>>no test is done, but I don't see why this is happening.  The
>>resolveNativeRoutine function does some strange stuff.
>>
>>Duncan Murdoch
>>
>>______________________________________________
>>R-devel at r-project.org mailing list
>>https://stat.ethz.ch/mailman/listinfo/r-devel
> 
> 
> - --
> Duncan Temple Lang                    duncan at wald.ucdavis.edu
> Department of Statistics              work:  (530) 752-4782
> 4210 Mathematical Sciences Building   fax:   (530) 752-7099
> One Shields Ave.
> University of California at Davis
> Davis,
> CA 95616,
> USA
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.2 (Darwin)
> 
> iD8DBQFDsu8Q9p/Jzwa2QP4RAuqGAJ9oHP7lhmb7IN9c/3WoRL5qimCTSACfaNhr
> WFY6NLnh2dfDTOGcaStbgNo> =rCZg
> -----END PGP SIGNATURE-----
> 
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel