R devel - Jan 2023 - Object are not destroy while using error (Rf_error)

Dear all,

I try to understand why on my computer I do not clear all data with this
code:


#include <R.h>
static int count = 0;

class A {
public:
  A(){  printf("c %d\n", count);
    count++;  }

  ~A(){count--;
    printf("d %d\n", count);  }
};

extern "C" {
  void testAL(){
    A a;
    {
      A b;
    }
    error("does not write [d 0]");
  }
}

To run with R: I build  gcc -shared -I/opt/R-202301/lib/R/include/
myError.cpp -o myError.so

then in R: dyn.load("myError.so")
.C("testAL")

This writes c0, c1, d1 but not d0.
If I comment line "error", I does write latest d0.

How could I get all my objects destroy while sending en error message to R ?

Regards,

Antoine Lucas.

	[[alternative HTML version deleted]]

Antoine,

I think there are few things going on here.  One is that actual _R_ objects
may only get destroyed when gc() gets called.  Which we applications writer
do not control.  Another thing that may have an effect is the use of .C()
which we all more or less moved away from.  Anyway, if we make your example a
proper C++ one with proper C++ scope (by trivial Rcpp wrapping for ease, but
that should not be a requirement) we see the four message you desire:


#### Code

#include <Rcpp/Lightest>

static int count = 0;

class A {
public:
  A(){  printf("c %d\n", count);
    count++;  }

  ~A(){count--;
    printf("d %d\n", count);  }
};

// [[Rcpp::export]]
void testAL(){
  A a;
  {
    A b;
  }
  Rcpp::stop("does not write [d 0]");
}

/*** R
testAL()
*/


#### Output
> Rcpp::sourceCpp("/tmp/ctorDemo.cpp")
> testAL()
c 0
c 1
d 1
d 0
Error in eval(ei, envir) : does not write [d 0]
> 

Hope this helps,  Dirk

-- 
dirk.eddelbuettel.com | @eddelbuettel | edd at debian.org

On 1/21/23 16:55, Antoine Lucas wrote:
> Dear all,
>
> I try to understand why on my computer I do not clear all data with this
> code:
>
>
> #include <R.h>
> static int count = 0;
>
> class A {
> public:
>    A(){  printf("c %d\n", count);
>      count++;  }
>
>    ~A(){count--;
>      printf("d %d\n", count);  }
> };
>
> extern "C" {
>    void testAL(){
>      A a;
>      {
>        A b;
>      }
>      error("does not write [d 0]");
>    }
> }
>
> To run with R: I build  gcc -shared -I/opt/R-202301/lib/R/include/
> myError.cpp -o myError.so
>
> then in R: dyn.load("myError.so")
> .C("testAL")
>
> This writes c0, c1, d1 but not d0.
> If I comment line "error", I does write latest d0.
>
> How could I get all my objects destroy while sending en error message to R
?
The problem is that Rf_error() uses a long jump, which unwinds the C++ 
frames without calling destructors. Using the C API of R from C++ code 
is tricky because of this incompatibility of C long jumps with C++ 
destructors.

Technically, you need to protect all calls from C++ to R against long 
jumps. There is an API for that, see R_UnwindProtect in 6.12 of "Writing 
R Extensions". So, your wrapper will get called in case of a long jump 
and will handle it in a way compatible with your C++ frames (possibly by 
throwing a C++ exception).

Similarly, you should not throw C++ exceptions over R frames on the 
stack, because R's meta-data about long jump targets on its stacks would 
go out of sync, causing a crash later. You again would have to convert 
C++ exceptions to long jumps.

So, in your example where you use the R API to only trigger an R error, 
you could instead throw a C++ exception, and catch it later, below all 
your C++ code but before unwinding could potentially reach R frames. And 
there you could call Rf_error().

Best
Tomas

(this is related: 
https://blog.r-project.org/2019/03/28/use-of-c-in-packages )
>
> Regards,
>
> Antoine Lucas.
>
> 	[[alternative HTML version deleted]]
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel