asSEXP may need 'PreserveObject' about rcall.jl

[dmbates]

You're right. I'll make the corresponding changes.

[dmbates]

This is getting more complicated than I first imagined. Right now I define the SEXP type in Julia as immutable and consisting only of a Ptr{Void}. But you can't put a finalizer on such an object, because it corresponds to a bitstype. One way out is to declare SEXP as a type, not an immutable type. There will be some cost in performance but probably not a measurable difference.

[randy3k]

I am thinking there should be two versions of asSEXP. One does the preserving and releasing whereas the other variant does not. The reason I suggest this is because releasing an R object could take a lot of time when the number of preserved objects is large. (it is based on some pairlist in R, and checking is done recursively).

For example, if one runs x = RCall.asSEXP([1:10]) 10 thousands times. If the preserving procedure is included, than there will be ten thousands of which x in the memory. Of course, 9999 of them will be collected when GC, but you can imagine that this step could be time consuming.

I don't really know the best way to resolve this issue. Perhaps we need more experts to give advices.

[dmbates]

Perhaps create an abstract type AbstractSEXP with a concrete immutable type SEXP and a concrete type protectedSEXP?

For the time being I suggest keeping the current scheme (everything is the mutable type SEXP and is protected) until we see if there are performance problems. I doubt very much that R_ReleaseObject is going to be the bottleneck but we can check. I looked and R_ReleaseObject does do a linear search on a cons list.

[randy3k]

Actually, the bottleneck can be easily seen by

julia> @time for i in 1:100000 RCall.asSEXP([1:10]) end
elapsed time: 0.44483907 seconds (26656472 bytes allocated, 68.26% gc time)

julia> @time for i in 1:100000 RCall.asSEXP([1:10]) end
elapsed time: 144.245652407 seconds (28945552 bytes allocated, 99.90% gc time)

The second run is extremely slow since GC is freeing the memory allocated by the first run.

[dmbates]

Right now I am working on documentation. If you want to create a PR, please do.

[randy3k]

Perhaps we can add an additional argument own in asSEXP.

When own is true, the R object is persevered, otherwise, it is unprotected.

How does it sound? I could do this easy PR if you agree.

[randy3k]

I think it could be closed as the PR #59 is coming.