GH-48912: [R] Configure C++20 in conda R on continuous benchmarking

[thisisnic]

Rationale for this change

Benchmark failing since C++20 upgrade due to lack of C++20 configuration

What changes are included in this PR?

Changes entirely from 🤖 (Claude) with discussion from me regarding optimal approach.

Description as follows:

conda-forge's R package doesn't have CXX20 configured in Makeconf, even though the compiler (gcc 14.3.0) supports C++20. This causes Arrow R package installation to fail with "a C++20 compiler is required" because R CMD config CXX20 returns empty.

This PR adds CXX20 configuration to R's Makeconf before building the Arrow R package in the benchmark hooks, if not already present.

Are these changes tested?

I got 🤖 to try it locally in a container but I'm not convinced we'll know for sure til we try it out properly.

Tested in Docker container with Amazon Linux 2023 + conda-forge R - confirmed R CMD config CXX20 returns empty before patch and g++ after patch.

The only thing we didn't test end-to-end was actually building Arrow R, but that would have taken much longer and the configure check (R CMD config CXX20 returning non-empty) is exactly what Arrow's configure script tests before proceeding.

Are there any user-facing changes?

Nope

• GitHub Issue: [R] Compilation errors on continuous benchmarking #48912

[github-actions]

⚠️ GitHub issue #48912 has been automatically assigned in GitHub to PR creator.

[thisisnic]

I did some more digging with 🤖 and ended up here @jonkeane - I spent some time asking questions and getting verification from code, so it's a bit over my head, so let me know if this sounds feasible to you!

Problem: conda-forge builds R with gcc 10, which has incomplete C++20 support. R's configure tests for C++20, fails, and leaves CXX20 empty in Makeconf. Even though benchmark machines install gcc 14.3.0 (which does have C++20 support) via the compilers package, R doesn't know to use it for C++20.

Our fix: Check if R CMD config CXX20 returns empty, and if so, append CXX20 settings to Makeconf pointing to g++ (which resolves to gcc 14.3.0 in the conda environment).

Why it works: The compiler is there and capable - R's Makeconf just wasn't configured to use it. Our patch fills in that missing config.

[jonkeane]

That sounds ok to me if it works!

[pitrou]

@ursabot please benchmark

[rok]

Benchmark runs are scheduled for commit f164e4d. Watch https://buildkite.com/apache-arrow and https://conbench.arrow-dev.org for updates. A comment will be posted here when the runs are complete.

[conbench-apache-arrow]

Thanks for your patience. Conbench analyzed the 3 benchmarking runs that have been run so far on PR commit f164e4d.

There were 2 benchmark results indicating a performance regression:

• Pull Request Run on amd64-c6a-4xlarge-linux at 2026-01-26 15:35:45Z
  • UniqueInt64 (C++) with params=0, source=cpp-micro, suite=arrow-compute-vector-hash-benchmark
  • HashIntegers (C++) with source=cpp-micro, suite=arrow-hashing-benchmark

The full Conbench report has more details.

[thisisnic]

Looks like the benchmarks are building - I think the one which failed is a timeout though unsure why.

[pitrou]

Thanks a lot @thisisnic . This looks good to me! Just one question.

[pitrou]

Did we report an issue to the corresponding conda-forge feedstock, so that we can get rid of this workaround later?

[thisisnic]

Wasn't sure it was a bug but reported it at conda-forge/r-base-feedstock#401

[pitrou]

Looks like the benchmarks are building - I think the one which failed is a timeout though unsure why.

We'll see if that reproduces in later benchmark runs, but this PR can certainly be merged anyway.