ARROW-3604: [R] Support to collect int64s as R integers

[javierluraschi]

Fix for, https://issues.apache.org/jira/browse/ARROW-3604

Enabled collecting int64s as R integers by replacing overflows and underflows with NAs and triggering a waring.

In sparklyr, this enables:

> sdf_len(sc, 10) %>% dplyr::transmute(new = cast("123456789123451234" %as% BIGINT))
# Source: spark<?> [?? x 1]
     new
 * <int>
 1    NA
 2    NA
 3    NA
 4    NA
 5    NA
 6    NA
 7    NA
 8    NA
 9    NA
10    NA
# ... with more rows
Warning message:
In RecordBatch__to_dataframe(x) :
  Integer overflow, 10 values replaced with NAs. Consider using 'options(arrow.int64 = "bit64")'.

CC: @romainfrancois

[wesm]

This one probably needs some tests

[wesm]

change these to constexpr?

[romainfrancois]

Sure. Does that mean they have to be functions ?

Otherwise, since ::minuend ::maxare themselvesconstexpr` perhaps we don't even need the constants here.

Also, not sure what MIN_INT32 is used for, but we need to be careful because it's also the NA sentinel for int in R.

[wesm]

yes they are constexpr, e.g.

arrow/cpp/src/arrow/util/hash.h

Line 32 in b165c86

static constexpr hash_slot_t kHashSlotEmpty = std::numeric_limits<int32_t>::max();

[romainfrancois]

This, and #2819 should be discussed more broadly. I'm not a fan of relying on a global options here.

I'd rather have those lossy conversions to be explicit.

[wesm]

Yeah I would suggest passing an options list to the conversion entry point

[javierluraschi]

From the decimals PR, looks like we would preffer users to explicitly cast, as opposed to configuring with options the desired behavior. I'll close this one then.