[Rd] Custom C finalizers for .Call
murdoch.duncan at gmail.com
Wed Nov 25 00:58:45 CET 2015
On 24/11/2015 6:20 PM, Andrew Piskorski wrote:
> On Tue, Nov 24, 2015 at 12:10:12AM +0100, Jeroen Ooms wrote:
>> Currently it is all to easy for package authors to introduce a memory
>> leak or stack imbalance by calling Rf_error() or
>> R_CheckUserInterrupt() in a way that skips over the usual cleanup
> I have a more modest request: Please improve the documentation of
> exactly what Rf_error() does and how it should be used! E.g., does
> calling it terminate execution of the C function from which it is
> called, or not? The docs below don't actually say:
> The docs imply that these are "C-level equivalents" to the R stop()
> function, which of course DOES terminate execution. I believe
> Rf_error() does also, but I really wish it was explicitly stated one
> way or the other.
I'd imagine the author of that section thought it *was* explicit. The
documentation can't answer every question, or clarify every
misunderstanding. That's why R-devel exists. fortune("source")
is also relevant. But now that you know the truth, you could suggest
rewording, as an svn patch against the trunk version of R-exts.texi.
> Grepping the R source, I never did find where Rf_error() is actually
> implemented. What am I missing?
You probably forgot to grep the .h files:
include\R_ext\Error.h(48): #define error Rf_error
The implementation of the function calls it error(), and uses the define
to remap that to Rf_error(). You can find the implementation in (where
> In src/include/Rinternals.h, the implementation of error_return()
> first calls Rf_error(msg) and then does a separate "return R_NilValue"
> in the next statement. So you might think that Rf_error() must NOT
> terminate execution, otherwise why the extra return statement? But it
> also has a comment:
> /* return(.) NOT reached : for -Wall */
> Meaning that that return statement is just to silence compiler
> warnings, and is not supposed to be reached because Rf_error()
> terminates execution. But this is a ridiculously roundabout way to
> infer what the behavior of Rf_error() is supposed to be...
I won't disagree that you took a roundabout route to the right conclusion.
More information about the R-devel