[R][CI] rtools35 job failing on 32-bit build tests

[asfimport]

See https://github.com/apache/arrow/actions/workflows/r.yml?query=branch%3Amaster, this started when ARROW-9697 (CountRows for Scanner) merged. It's only failing on rtools35 (aka gcc 4.9), and only on the 32-bit build (i386). Since there's no output about what failed, it's probably a segfault. The easiest way to get more information is to flip this if: false to true and let it print detailed output about where it was when it died https://github.com/apache/arrow/blob/master/.github/workflows/r.yml#L186

Reporter: Neal Richardson / @nealrichardson
Assignee: David Li / @lidavidm

PRs and other links:

• GitHub Pull Request #10282

_{Note: This issue was originally created as ARROW-12694. Please see the migration documentation for further details.}

[asfimport]

Antoine Pitrou / @pitrou:
Isn't it soon dead?

[asfimport]

David Li / @lidavidm:
It cuts off after Start test: dim() correctly determine numbers of rows and columns on arrow_dplyr_query object (logs) which is just calling CountRows on a FileSystemDataset. I've been trying to replicate the build locally, so far without much success.

[asfimport]

Neal Richardson / @nealrichardson:
@pitrou it is true, to my knowledge we won't have to support (for CRAN) the rtools35 toolchain much longer, I believe the next R release (which makes R 3.6 no longer be "oldrel" and thus not required to pass on CRAN) is coming in May. That said, I know David had found another problem with CountRows after merging, and just because we won't have to support this exact compiler version much longer doesn't mean there isn't an issue there.

It does probably make sense to timebox the work looking into this if there's not an obvious issue. Past experience setting up gdb to debug further on this toolchain has been fruitless.

[asfimport]

David Li / @lidavidm:
I think it's a miscompilation or something: the test passes if I decompose this ARROW_ASSIGN_OR_RAISE into simpler but otherwise equivalent code. I can submit a PR to do something like that (maybe with a #ifdef) to get things passing again.

[asfimport]

Neal Richardson / @nealrichardson:
Issue resolved by pull request 10282
#10282