[Rcpp-devel] coercion NULL to vector

[Serguei Sokol]

Le 18/04/2018 à 12:26, Georgi Boshnakov a écrit :
> I may be missing something here but
> it doesn't seem right to introduce silent coercion of NULL to vectors, etc.,
> especially if it would become imposed on everybody using Rcpp/Armadillo.
Nothing is imposed. Everyone will be free to use Nullable<T> mechanism as before.
Actually, when NULL is provided to where a vector or a matrix is expected
just an error is thrown. So we can hardly imagine that someone would have
his program broken or will be misled if such coercion is introduced.

> Even the "convenience" of this  is questionable.
Right. It's a matter of taste. But with a quick glance on SO
I have seen a discussion here
https://stackoverflow.com/questions/34718570/rcpp-pass-vector-of-length-0-null-to-cppfunction
and on this list here (and follow-ups)
http://lists.r-forge.r-project.org/pipermail/rcpp-devel/2014-February/007215.html
where some people are surprised not having such coercion and even
expressed a wish like "I would be super happy if in the future the special value
R_NilValue could be converted to an empty vector, it would make things
very easy and coherent with R's behaviour."
It's also my thought but I am ready to hear the arguments of those who
think otherwise.

Serguei.

> 
> Georgi Boshnakov
> 
> 
> -----Original Message-----
> From: rcpp-devel-bounces at lists.r-forge.r-project.org [mailto:rcpp-devel-bounces at lists.r-forge.r-project.org] On Behalf Of Serguei Sokol
> Sent: 18 April 2018 10:57
> To: Dirk Eddelbuettel
> Cc: Rcpp-devel
> Subject: Re: [Rcpp-devel] coercion NULL to vector
> 
> Le 17/04/2018 à 17:53, Dirk Eddelbuettel a écrit :
>>
>> On 17 April 2018 at 10:31, Dirk Eddelbuettel wrote:
>> |
>> | On 17 April 2018 at 15:09, Serguei Sokol wrote:
> ...
>> | | It will automatically coerce NULL to a declared vector type.
>> | | It's a more primitive solution than Nullable<T> but in many cases
>> | | it can be largely sufficient and make code look shorter and clearer.
>> | |
>> | | Will it break something in established usage habits? Is it compatible with your plans for Rcpp?
>> | | I can prepare a PR if you decide to include it.
>> |
>> | We could test that.  It may not do harm -- so I just turned on a
>> | rev.dep check for it.
>> |
>> | Can you open an issue for it on GH though?  Better visibility and
>> | easier 'per concrete topic' discussion.
> I was just about to do so when I saw your post here. So let decide if the semantic is useful or not and if yes, we'll switch to GH. OK?
> 
>>
>> I am not convinced that the semantics are useful.  Your example works
>> on the arma type (where we our very old design issue of always returning a matrix):
> What is returned: a matrix or a vector is not important here. By the way, it is already resolved by an optional RCPP_ARMADILLO_RETURN_COLVEC_AS_VECTOR macro, remember?
> The point here is how the /input/ NULL is interpreted. Is it coerced to smth or just an error is thrown?
> 
>>
>>> sourceCpp("/tmp/serguei.cpp")
>>> f(NULL)
>>        [,1]
>>> f(integer(0))
>>        [,1]
>>> f()
>> Error in f() : argument "x" is missing, with no default
> Normal. If you want to make the parameter optional, you have to declare vec f(Rcpp::NumericVector x=R_NilValue) {return x+1;}  > f()
>        [,1]
> It does not work yet with 'vec' type but it could with an appropriate patch.
> We are just discussing the usefulness of semantics not yet the patch.
> 
>>> class(f(NULL))
>> [1] "matrix"
> (not a big deal what is returned as discussed above).
> 
>>>
>>
>> But if I do the same with Rcpp types, say a matrix via
>>
>> // [[Rcpp::export]]
>> Rcpp::NumericMatrix g(Rcpp::NumericMatrix x) { return x+1; }
>>
>> then I get more restrictive behaviour (as NumericMatrix tests for
>> matrix)
> Right. I think that the next logical step would be to allow an automatic coercion of atomic vectors to matrices too. As in R  > as.matrix(integer(0))
>        [,1]
> we got a matrix of dims (0, 1) we could make the same available in Rcpp.
> More generally, a vector x could be coerced into a matrix of dims (length(x), 1) (once again just as in R as.matrix())
> 
>>
>>> sourceCpp("/tmp/serguei.cpp")
>>> g(NULL)
>> Error in g(NULL) : Not a matrix.
>>> g(matrix())
>>        [,1]
>> [1,]   NA
>>> g(vector())
>> Error in g(vector()) : Not a matrix.
>>> g(integer())
>> Error in g(integer()) : Not a matrix.
>>>    
>>
>>
>> Is this really useful, and can you not use Nullable<> instead?
> Sure I can, but as I already said, my thought is to make the code looking simpler and clearer in this kind of situations where automatic coercion does sufficient job. Nullable<T> is much more flexible than that but more expensive in coding on user's side.
> 
> Serguei.
> _______________________________________________
> Rcpp-devel mailing list
> Rcpp-devel at lists.r-forge.r-project.org
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
>