PKG-17264 — Python 3.15.0rc1 (master-replay) by skupr-anaconda · Pull Request #249 · AnacondaRecipes/python-feedstock

skupr-anaconda · 2026-08-26T14:36:15Z

Destination channel: ad-testing/label/py315

Links

Jira
Upstream: https://github.com/python/cpython/tree/v3.15.0rc1

What changed

This PR moves python-feedstock from 3.14.6 (current master) to CPython 3.15.0rc1 in a single branch.

It does not include the conda-forge alignment from #233.

Version and patches

Recipe version, dev tag, and sha256 now point at Python-3.15.0rc1.
Our Anaconda patches were re-applied on that tarball (including the Windows xxlimited / stable-ABI patch).

Build settings (conda_build_config.yaml)

Python 3.15
Tk 9.0 from pkgs/main (Tk 9 ships libtcl9tk9.0, not libtk9.0)
macOS Clang/LLVM 21, so the JIT and LTO tools are the same major version (mixing 20 and 21 breaks the macOS build)

Build scripts

Unix: link Tcl/Tk 9 with the tcl9tk name; on macOS use llvm-ar-21.
Windows: pass the Tk version from the recipe into MSBuild; win-arm64 and free-threading use the arm64 / amd64t / arm64t output folders.
CMake test: allow a free-threaded interpreter (Python_FIND_ABI). Extra components are a string, not a CMake option() (that turned into OFF and broke find_package).

CI (abs.yaml)

Turn on ANACONDA_ROCKET_ENABLE_PY315.
Upload only to ad-testing/label/py315.

Out of scope

Enabling Visual Studio 2026
conda-forge alignment (PKG-15601 + PKG-2624 + PKG-17264 — CF-align freethreading/debug on 3.15.0rc1 (main) #233)

Please do not merge until we know where PBP puts artifacts on merge commits (we do not want this on pkgs/main yet).

Apply 0001–0025 against v3.15.0rc1 and add 0026 xxlimited stable-ABI link. Keep 0008 d1trimfile hotspot and AR-only 0023 branding, 0024 zlib-ng, 0026. PKG-17264

Override aggregate tk 8.6 and clang 20 so JIT/LTO match CPython 3.15. Keep gil_type; do not CF-align. PKG-17264

Use -ltcl${TCLTK_VER} -ltcl${TK_MAJOR_VER}tk${TCLTK_VER} and $HOST-llvm-ar -> llvm-ar-$LLVM_VER so osx LTO does not load LLVM 20. PKG-17264

Pass /p:TclVersion=%tk%, use CONDA_PYTHON_EXE, PLATFORM=ARM64, and amd64t/arm64t for freethreading. Leave cli-64.exe as-is. PKG-17264

Use STRING cache vars; option() made EXTRA_COMPONENTS boolean OFF and broke find_package. PKG-17264

Set version/dev/sha256, list the refreshed patches, gate 0026 to win+gil_type disabled, and depend on clang-21/llvm-tools-21. PKG-17264

Enable ANACONDA_ROCKET_ENABLE_PY315. PKG-17264

Windows prefix now uses lib/python and include/python. Cross-compile and shared sysconfig follow conda-forge. Keep gil_type; skip FT/debug Jinja. PKG-17264

Drop CF-only test ! -f libpython*.so/.dylib (we ship the shared lib in python, not a libpython output). Keep Lib/ on Windows — getpath landmarks Lib\os.py. SYSCONFIG from the static build so --embed links .a, not the dylib (osx libpython-static dyld abort). PKG-17264

Without -c, python.exe treats print('Hello') as a filename. findstr then matches Hello in the error path and the check falsely passes. PKG-17264

cbouss

It's worth comparing our current patches with conda-forge. Especially, we should take https://github.com/conda-forge/python-feedstock/blob/dev/recipe/patches/0027-More-unixy-layout-for-windows.patch

cbouss · 2026-08-28T19:57:12Z

@@ -47,7 +47,7 @@ index b89d7427e3f..21df0efac12 100644
         # {HKCU,HKLM}\Software\Python\PythonCore\{winver}\PythonPath
 diff --git a/PC/getpathp.c b/PC/getpathp.c


This is adding a whole file that doesn't look to be used.

It's already too complicated to align with the conda-forge recipe in one PR. conda-forge seems to adopt many layers we didn't have in our recipe. I've already tried to port some layers (the commit) after the live session with @ifitchet from 777b793, but we didn't notice that there are more layers like this patch and installation scripts.

I would do that full alignment in #233

This patch is a real change. Let's do it as part of rc2 though.

cbouss · 2026-08-28T19:58:32Z

On conda-forge, they have additional files being patched. This warrant a review.

Oh, as in the CF one affects 50 files (including Serhii's xxlimited files) and ours affects 41 files.

I don't think I've ever even understood what the d1trimfile patch was meant to be doing...

cbouss · 2026-08-28T19:59:39Z

Could rename the file to align with conda-forge. Easier comparison.

cbouss · 2026-08-28T20:03:46Z

-     <tclExternalTommath Condition="$(TclMajorVersion) == '9'">TCL_WITH_EXTERNAL_TOMMATH;</tclExternalTommath>
+     <tcltkSuffix Condition="'$(TclMajorVersion)' == '8'">t</tcltkSuffix>
+     <tkPrefix Condition="'$(TclMajorVersion)' == '9'">tcl9</tkPrefix>
+     <TclshNativeFlag Condition="$(Platform) != 'Win32'">TCLSH_NATIVE="$(tcltkDir)\..\win32\bin\tclsh$(TclMajorVersion)$(TclMinorVersion)$(tcltkSuffix).exe"</TclshNativeFlag>


To more keys in conda-forge: tclExternalTommath and tixDir
To check if it is something to add.

Tix (Tk Interface Extension) was last seen in Python in 3.12, https://github.com/python/cpython/blob/3.12/Doc/library/tkinter.tix.rst, after being deprecated in 3.6.

Of interest, the tixDir variables existed until python/cpython@a989b73

We don't have libtommath (CF do) so the tk build uses its internal tommath package.

We should maybe maintain that tcltk.props line, though, if only to ensure that we do NOT set TCL_WITH_EXTERNAL_TOMMATH

ifitchet

I want to believe that much of the issues relate to alignment with CF rather than faults per se with our recipe...

Stop adding lib/python/site-packages via conda-site.pth; it put a version-less dir on sys.path and broke numba. Rename the Darwin cross-compile patch to 0009-cross-compile-darwin.patch to match CF.

skupr-anaconda added 7 commits August 26, 2026 10:49

feat(patches): refresh AR series for CPython 3.15.0rc1

880df8c

Apply 0001–0025 against v3.15.0rc1 and add 0026 xxlimited stable-ABI link. Keep 0008 d1trimfile hotspot and AR-only 0023 branding, 0024 zlib-ng, 0026. PKG-17264

feat(cbc): enable Python 3.15, tk 9.0, and osx LLVM 21

70cfe2d

Override aggregate tk 8.6 and clang 20 so JIT/LTO match CPython 3.15. Keep gil_type; do not CF-align. PKG-17264

feat(build): link tk 9 as tcl9tk and wrap llvm-ar-21 on osx

d4578fc

Use -ltcl${TCLTK_VER} -ltcl${TK_MAJOR_VER}tk${TCLTK_VER} and $HOST-llvm-ar -> llvm-ar-$LLVM_VER so osx LTO does not load LLVM 20. PKG-17264

feat(build): drive Win Tcl/Tk from CBC and stage win-arm64

7abe818

Pass /p:TclVersion=%tk%, use CONDA_PYTHON_EXE, PLATFORM=ARM64, and amd64t/arm64t for freethreading. Leave cli-64.exe as-is. PKG-17264

fix(test): find freethreading Python via Python_FIND_ABI

06b0242

Use STRING cache vars; option() made EXTRA_COMPONENTS boolean OFF and broke find_package. PKG-17264

feat(recipe): bump CPython to 3.15.0rc1

531a2c6

Set version/dev/sha256, list the refreshed patches, gate 0026 to win+gil_type disabled, and depend on clang-21/llvm-tools-21. PKG-17264

ci(abs): upload 3.15.0rc1 to ad-testing/label/py315

06ef1e1

Enable ANACONDA_ROCKET_ENABLE_PY315. PKG-17264

skupr-anaconda self-assigned this Aug 26, 2026

skupr-anaconda added 3 commits August 27, 2026 15:03

feat(recipe): port CF layout and test fixes from 777b793

b853e78

Windows prefix now uses lib/python and include/python. Cross-compile and shared sysconfig follow conda-forge. Keep gil_type; skip FT/debug Jinja. PKG-17264

fix(build): pass -c so Win Hello smoke test runs code

5ea1b5b

Without -c, python.exe treats print('Hello') as a filename. findstr then matches Hello in the error path and the check falsely passes. PKG-17264

cbouss self-requested a review August 28, 2026 19:40

cbouss reviewed Aug 28, 2026

View reviewed changes

cbouss approved these changes Aug 31, 2026

View reviewed changes

skupr-anaconda requested a review from ifitchet September 1, 2026 08:46

ifitchet approved these changes Sep 1, 2026

View reviewed changes

fix(build): drop experimental noarch site-packages pth

89f8040

Stop adding lib/python/site-packages via conda-site.pth; it put a version-less dir on sys.path and broke numba. Rename the Darwin cross-compile patch to 0009-cross-compile-darwin.patch to match CF.

skupr-anaconda merged commit 73af5de into master Sep 1, 2026
38 of 39 checks passed

skupr-anaconda deleted the PKG-17264-py315-master branch September 1, 2026 13:33

skupr-anaconda mentioned this pull request Sep 4, 2026

PKG-17802 — Python 3.15.0rc2 #257

Open

Navigation Menu

Uh oh!

PKG-17264 — Python 3.15.0rc1 (master-replay) - #249

PKG-17264 — Python 3.15.0rc1 (master-replay)#249
skupr-anaconda merged 11 commits into
masterfrom
PKG-17264-py315-master

skupr-anaconda commented Aug 26, 2026 •

edited

Loading

Uh oh!

cbouss left a comment

Uh oh!

cbouss Aug 28, 2026

Uh oh!

skupr-anaconda Aug 31, 2026

Uh oh!

cbouss Aug 31, 2026

Uh oh!

cbouss Aug 28, 2026

Uh oh!

ifitchet Sep 1, 2026

Uh oh!

cbouss Aug 28, 2026

Uh oh!

cbouss Aug 28, 2026

Uh oh!

ifitchet Sep 1, 2026

Uh oh!

ifitchet Sep 1, 2026

Uh oh!

ifitchet left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		@@ -47,7 +47,7 @@ index b89d7427e3f..21df0efac12 100644
		# {HKCU,HKLM}\Software\Python\PythonCore\{winver}\PythonPath
		diff --git a/PC/getpathp.c b/PC/getpathp.c

Uh oh!

Conversation

skupr-anaconda commented Aug 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Links

What changed

Out of scope

Uh oh!

cbouss left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ifitchet left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

skupr-anaconda commented Aug 26, 2026 •

edited

Loading