Consider masking definition of Rf_error()?

Start with the Rf_error() definition used in C++ contexts and the Rcpp::stop() alternative described in the issue. Review the 35-comment thread to determine whether masking is technically viable and what behavior should replace direct Rf_error() use; done means reaching a decided approach that addresses destructor safety.

Somewhat related to a recent post on R-devel, some users will still try to use Rf_error() in C++ code using Rcpp. However, (as is documented) C++ destructors won't run when a longjmp occurs in cases like this. For example:

#include <Rcpp.h>
using namespace Rcpp;

struct A {
    A() { Rprintf("A()\n"); }
    ~A() { Rprintf("~A()\n"); }
};

// [[Rcpp::export]]
SEXP uhoh() {
    A a;
    Rf_error("Oops!");
}

/*** R
uhoh()
*/

In this example, ~A() is never printed.

We could consider masking the definition of Rf_error() in these contexts. For example, something like:

#define Rf_error(...) \
    static_assert(false, "Use of Rf_error() in C++ contexts is unwise: consider using Rcpp::stop() instead.");

There might also be a way to provide our own definition of Rf_error() that "masks" the version provided by R, but I wasn't able to find something immediately obvious.