[Rcpp-devel] RCPP_USE_UNWIND_PROTECT by default

[Jeroen Ooms]

On Sun, Feb 6, 2022 at 5:56 PM Dirk Eddelbuettel <edd at debian.org> wrote:
>
>
> On 6 February 2022 at 17:40, Jeroen Ooms wrote:
> | We can try to take V8 out of the equation, and see what actually
> | causes the change. V8 uses (and tests!) the Rcpp feature to call an R
> | function from C++. This behaves quite differently when using
> | RCPP_UNWIND_PROTECT.
> |
> | Here is a dummy package to demonstrate this: https://github.com/jeroen/uptest
> |
> | The use case is simple: we want to call some R function from C++, and
> | if it fails, catch the error message and deal with it in C++. I
> | suspect there are more packages doing this, but perhaps they are not
> | testing this case.
> |
> | The example from uptest above show that when we compile with
> | RCPP_UNWIND_PROTECT, any R error is always printed directly to the
> | console. Even if we catch the Rcpp::LongjumpException in C++, the R
> | error still seems to bubble up (as Iñaki also noted). I think this is
> | at least surprising.
> |
> | Then I am not sure what we do in C++ now with this
> | Rcpp::LongjumpException if we catch it. We certainly don't want to
> | unwind all the way in the middle of running the javascript code. The
> | JavaScript code should be able to catch the R error, and continue
> | running the script.
> |
> | Anyway if you want to make RCPP_UNWIND_PROTECT the default, I can
> | obviously opt-out in V8, but as an Rcpp user I am not convinced this
> | is an improvement.
>
> Thanks for the feedback. I think there is an easy-to-understand, easy-to-make
> error in a sample package (but kudos for creating one!).  Because Rcpp will
> always wrap try/catch around the function you create with its mechanism, the
> layer inside is redundant.  So I suggest the C++ function file becomes this
> simpler / shorter variant:
>
>   #include <Rcpp.h>
>   using namespace Rcpp;
>
>   // [[Rcpp::export]]
>   Rcpp::NumericVector call_r_from_rcpp() {
>     Rcpp::Function callback = Rcpp::Environment::namespace_env("uptest")["callback"];
>     callback();
>     return Rcpp::NumericVector::create(42);
>   }
>
> which behaves fine for me in both cases.

Well not really, this kind of misses my point that it the
unwind-protect makes it impossible to meaningfully catch the R error
in C++, handle it, and continue running the C++ code, without aborting
the entire mission and throwing the user back in the console.