On Windows, exceptions thrown from RCPP modules bypass try(..., silent = TRUE)

[venelin]

Hello, I am trying to find an answer to the following question:

What is the recommended way of handling exceptions thrown from C++ methods exported using RCPP_MODULEs?

I notice that these exceptions are handled differently on Windows vs Linux/Mac OSX systems. On Linux Mac OSX, it is possible to wrap the call to the C++ method in a try(..., silent = TRUE) and to examine whether the class of the result is a 'try-error'. On Windows, this mechanism seems to fail and the R-process crashes. The package PCMBaseCpp uses Rcpp modules and is compiled using a C++11 compiler with compiler options:

PKG_CXXFLAGS=$(SHLIB_OPENMP_CXXFLAGS) -DARMA_DONT_USE_OPENMP
PKG_LIBS=$(SHLIB_OPENMP_CXXFLAGS) $(LAPACK_LIBS) $(BLAS_LIBS) $(FLIBS)

CXX_STD = CXX11

Below is an example code reproducing the problem (copied from this issue).

install.packages(c("TreeSim", "PCMBase", "PCMBaseCpp"))

library(TreeSim)
library(PCMBase)
library(PCMBaseCpp)

modelBM <- PCMBase::PCM(model = "BM", k = 1)
phyltree<-TreeSim::sim.bd.taxa(n=5,numbsim=1,lambda=1,mu=0)[[1]]
phyltree$root.edge<-NULL
mTraits<-matrix(0,ncol=1,nrow=5)
loglik_PCMBase<-PCMBase::PCMLik(t(mTraits), phyltree, modelBM)
loglik_PCMBase

# This line of code returns an object of class 'try-error' on
# Linux and Mac OSX, but crashes the R process on Windows:
loglik_PCMBaseCpp_PCMLik<-PCMBase::PCMLik(t(mTraits), phyltree, modelBM, metaI = PCMBaseCpp::PCMInfoCpp)

Here's an excerpt of the stack-trace (links pointing to lines of code on github).
try() -->
...
PCMBaseCpp::PCMLmr.PCMInfoCpp -->
...
SPLITT::TraversalSpecification::TraverseTree() -->
...
std::rethrow_exception().

The C++ method TraverseTree is exported via an RCPP_MODULE defined here.

The try(..., silent=TRUE) statement is in the R-package PCMBase here.

[eddelbuettel]

Didn't read all of this but I think the gist of it is that exception handling across shared libraries is difficult to impossible, as are long jumps, across OS. So what you observed is ... all you can get.

(Also: nobody claimed you could catch something in R that was thrown in C++. You can catch [some/most/it's still tricky/...] C++ exception via Rcpp and return to R as an error. That is even as automated as it gets if you use Rcpp Attributes.)

[venelin]

Thanks.

[jjallaire]

It is indeed a known limitation that Rcpp module methods cannot stop() on Windows (as the crash you observed occurs). This is ultimately because of the fact that R links to libgcc statically (see https://stackoverflow.com/questions/2424836/exceptions-are-not-caught-in-gcc-program).

@eddelbuettel You may recall that we had to re-write Rcpp attributes entirely at the beginning. That was because of this issue (see https://github.com/RcppCore/Rcpp/blob/master/inst/NEWS.Rd#L1544-L1547).

One possible piece of good news is that we did do the gymnastics in Rcpp attributes to work around this, so you could either: (1) Expose your underlying C++ code as a set of top-level R functions using attributes; and/or (2) Wrap those functions in an S4 ref class or R6 ref class to provide the user with object/reference semantics.

[venelin]

Thanks @jjallaire and @eddelbuettel for the quick replies!

(1) Expose your underlying C++ code as a set of top-level R functions using attributes; and/or
(2) Wrap those functions in an S4 ref class or R6 ref class to provide the user with object/reference semantics.

I am not familiar with the above techniques, yet, I am not convinced that any of these would be easy to implement, because the package relies heavily on Rcpp modules, see e.g. this code. Therefore, I think the only feasible approach would be to handle the exceptions on the C++ level and return a valid object to R each time.

[eddelbuettel]

@jjallaire Much appreciate the follow-up and link. Let's earmark this for a Rcpp FAQ addition as I have to add an entry for the upcoming 'no modules' and 'no rtti' options. I'll draft a paragraph and have you check. Deal?

[eddelbuettel]

I think the only feasible approach would be to handle the exceptions on the C++ level and return a valid object to R each time.

Correct. "Been there. done that". And mayne returning a NULL in case of error / exception. Rcpp makes it easy for you turn a C++ exception into an R error.

[jjallaire]

Sounds good!

…

On Fri, Oct 18, 2019 at 9:58 AM Dirk Eddelbuettel ***@***.***> wrote: @jjallaire <https://github.com/jjallaire> Much appreciate the follow-up and link. Let's earmark this for a Rcpp FAQ addition as I have to add an entry for the upcoming 'no modules' and 'no rtti' options. I'll draft a paragraph and have you check. Deal? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1001?email_source=notifications&email_token=AAAZPR5L4P3YME6AGEXNQWLQPG6HRA5CNFSM4JCGWGTKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBUR3SQ#issuecomment-543759818>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAZPRYCTQZ44QZU645BDZ3QPG6HRANCNFSM4JCGWGTA> .

[eddelbuettel]

See #1009 for a start.