Repository navigation
Conversation
|
|
|
@github-actions crossbow submit r-binary-packages |
|
Revision: 2834b54 Submitted crossbow builds: ursacomputing/crossbow @ actions-e5eea6767f
|
a9ce60c to
75cc517
Compare
|
Test failures not related to this PR. Will re-run CI to check no subsequent changes have broken anything. I've iterated on this with AI, but I don't quite trust my understanding enough to stand by every change here myself. |
|
@github-actions crossbow submit r-binary-packages |
|
Revision: 75cc517 Submitted crossbow builds: ursacomputing/crossbow @ actions-611a18301d
|
|
This PR only changes what per-PR CI builds and tests. The Windows libarrow binary that users download is still built with Rtools40 by the crossbow |
There was a problem hiding this comment.
🔵 Needs a closer look
The Rtools45 migration changes the Windows ABI and static dependencies, requiring human review of CI linking/loading and timezone/cloud-storage behavior because the actual toolchain artifacts cannot be verified here.
No open issues 🎉
What changed in this PR
Migrates the R package’s Windows libarrow build to Rtools45, addressing the archived Rtools40 toolchain while retaining the legacy build path.
Changes:
- Adds a direct CMake build and packaging path for Rtools42+, with curl dependency flags included in the archive.
- Updates R’s linker configuration to support both legacy and new archives.
- Switches Windows CI to Rtools45 and adds compatibility handling for MinGW headers and R-provided timezone data.
| File | Description |
|---|---|
r/configure.win |
Detects new archives and adjusts dependency link flags. |
cpp/src/arrow/util/chrono_internal.h |
Allows overriding standard chrono usage for R’s timezone support. |
cpp/cmake_modules/ThirdpartyToolchain.cmake |
Detects missing BCrypt macros before applying the MinGW workaround. |
ci/scripts/r_windows_build.sh |
Defaults to Rtools45 with direct CMake builds, retaining the Rtools40 path. |
.github/workflows/r.yml |
Uses Rtools45 build artifacts and explicitly installs ccache. |
Review effort: Balanced
| # Build the static libarrow that r/configure.win downloads on Windows. | ||
| # | ||
| # * Rtools40 (MSYS2): makepkg-mingw with ci/scripts/PKGBUILD. | ||
| # * Rtools42+ (MXE, no pacman/makepkg): call CMake directly against the |
There was a problem hiding this comment.
| # * Rtools42+ (MXE, no pacman/makepkg): call CMake directly against the | |
| # * Rtools42+ (libraries are a fixed set built with MXE): call CMake directly against the |
I'll apply this change before merging, not worth ending up rerunning CI right this sec for a comment
| # specific language governing permissions and limitations | ||
| # under the License. | ||
|
|
||
| # Build the static libarrow that r/configure.win downloads on Windows. |
There was a problem hiding this comment.
AFAICT reading the diff, there is basically nothing shared between the old version and the new one, so maybe it would be cleaner and more readable to just have two different scripts.
|
|
||
| # cURL's link dependencies vary by Rtools version, so read them from | ||
| # libcurl.pc, ship the toolchain ones, and record the flags in libcurl.libs | ||
| # for r/configure.win. Static link order matters, so: |
There was a problem hiding this comment.
IDK the answer to this but I wonder whether this logic belongs better in configure.win
| # Workaround missing BCRYPT_RSA_ALG_HANDLE macro in older MinGW-w64 headers. | ||
| # google-cloud-cpp v3+ uses it without guards in sign_using_sha256.cc. | ||
| # Workaround missing BCRYPT_RSA_ALG_HANDLE and BCRYPT_MD5_ALG_HANDLE macros | ||
| # in older MinGW-w64 headers. google-cloud-cpp v3+ uses them without guards. |
There was a problem hiding this comment.
Would turning off google-cloud-cpp in the older Windows builds simplify this?
| matrix: | ||
| config: | ||
| - { rtools: 40, arch: 'ucrt64' } | ||
| - { rtools: 45, arch: 'ucrt64' } |
There was a problem hiding this comment.
Forgive my ignorance, but if we only build rtools45 here, where does the rtools40 get built?
There was a problem hiding this comment.
The plan is to build RTools45 on every PR and do both RTools40 and RTools45 for binaries to distribute (in a follow-up PR). So at the moment, after this PR, we'd build with RTools40 in the r-binary-packages job like it is at the moment (which is why that file doesn't show up in this PR).
Rationale for this change
Upgrading from old toolchain for R Windows builds
What changes are included in this PR?
Bump default RTools version to 4.5 but keep 4.0 on in CI so we can support past few R versions still.
I'm filing a follow-up issue for actually distributing the binaries, to keep this PR smaller and more focused.
Are these changes tested?
On CI, aye
Are there any user-facing changes?
Nope
Was AI used for this PR?
Yeah, for most of it, but I prodded it and got it to reconsider a few decisions. That said, it'll need a deeper look to make sure we're not doing anything daft here.
PR code and description written by:
Reviewed before submission by: