[Rd] help with rchk warnings on Rf_eval(Rf_lang2(...))

[Tomas Kalibera]

On 3/23/20 8:39 PM, Ben Bolker wrote:
> Dear r-devel folks,
>
>    [if this is more appropriate for r-pkg-devel please let me know and
> I'll repost it over there ...]
>
> I'm writing to ask for help with some R/C++ integration idioms that are
> used in a package I'm maintaining, that are unfamilar to me, and that
> are now being flagged as problematic by Tomas Kalibera's 'rchk'
> machinery (https://github.com/kalibera/rchk); results are here
> https://raw.githubusercontent.com/kalibera/cran-checks/master/rchk/results/lme4.out
>
> The problem is with constructions like
>
> ::Rf_eval(::Rf_lang2(fun, arg), d_rho)
>
> I *think* this means "construct a two-element pairlist from fun and arg,
> then evaluate it within expression d_rho"
>
> This leads to warnings like
>
> "calling allocating function Rf_eval with argument allocated using Rf_lang2"
>
> Is this a false positive or ... ? Can anyone help interpret this?
This is a true error. You need to protect the argument of eval() before 
calling eval, otherwise eval() could destroy it before using it. This is 
a common rule: whenever passing an argument to a function, that argument 
must be protected (directly or indirectly). Rchk tries to be smart and 
doesn't report a warning when it can be sure that in that particular 
case, for that particular function, it is safe. This is easy to fix, 
just protect the result of lang2() before the call and unprotect (some 
time) after.
> Not sure why this idiom was used in the first place: speed? (e.g., see
> https://stat.ethz.ch/pipermail/r-devel/2019-June/078020.html ) Should I
> be rewriting to avoid Rf_eval entirely in favor of using a Function?
> (i.e., as commented in
> https://stackoverflow.com/questions/37845012/rcpp-function-slower-than-rf-eval
> : "Also, calling Rf_eval() directly from a C++ context is dangerous as R
> errors (ie, C longjmps) will bypass the destructors of C++ objects and
> leak memory / cause undefined behavior in general. Rcpp::Function tries
> to make sure that doesn't happen.")

Yes, eval (as well as lang2) can throw an error, this error has to be 
caught via R API and handled (e.g. by throwing as exception or something 
else, indeed that exception then needs to be caught and possibly 
converted back when leaving again to C stack frames). An R/C API you can 
use here is R_UnwindProtect. This is of course a bit of a pain, and one 
does not have to worry when programming in plain C.

I suppose Rcpp provides some wrapper around R_UnwindProtect, that would 
be a question for Rcpp experts/maintainers.

Best
Tomas

>
>   Any tips, corrections, pointers to further documentation, etc. would be
> most welcome ... Web searching for this stuff hasn't gotten me very far,
> and it seems to be deeper than most of the introductory material I can
> find (including the Rcpp vignettes) ...
>
>    cheers
>     Ben Bolker
>
> ______________________________________________
> R-devel using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel