[
https://issues.apache.org/jira/browse/ARROW-14804?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17448452#comment-17448452
]
Antoine Pitrou commented on ARROW-14804:
----------------------------------------
Not only dedicated pointer types may be "more natural", but {{double}} is
actually _insufficient_. A double precision floating-point number has 53 bits
of mantissa, which is not enough to represent every possible 64-bit pointer.
I suppose pointer values are often clustered around 0, which makes them exactly
representable as a {{double}}, but that's just luck and there's no guarantee
that pointers will always be like that.
> [R] import_from_c() / export_to_c() methods should accept external pointers
> ---------------------------------------------------------------------------
>
> Key: ARROW-14804
> URL: https://issues.apache.org/jira/browse/ARROW-14804
> Project: Apache Arrow
> Issue Type: Improvement
> Components: R
> Reporter: Dewey Dunnington
> Priority: Minor
>
> The import and export to C methods for Field, Shema, DataType, Array,
> RecordBatch, and RecordBatchReader all require the pointer to be passed as a
> {{double}} (casted from {{uintptr_t}} casted from {{*}}. I imagine there's
> good reason for this path to exist; however, the R externalptr type and
> 64-bit integer are more natural types and I think it would not be hard to
> allow pointers to be passed in this way. A [short
> discussion|https://github.com/apache/arrow/pull/11679#discussion_r755178800]
> of this occurred when updating the Python docs on how to pass Array pointers
> to R.
> The implementation lives in
> [arrow_cpp11.h|https://github.com/apache/arrow/blob/eb2fa2df6dedf2743047df9dd19d7ed4f4482c22/r/src/arrow_cpp11.h#L49-L54]
> and it seems like it would be easy to allow a user to pass something more
> R-like while still keeping the double route open for the packages that need
> it.
> I see that this is getting used for pointers that R6 objects are storing for
> the underlying C++ objects...I imagine there's a good reason for this but I
> wonder if there's a way to make the places where this interacts with other
> packages more intuitive.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
