FIX: building on Windows-ARM64 (#391)

* Narrow scopes where 3.12 hacks apply

* Update CI to target and test ARM64 Windows

* Fixed expressions and unset ${VSCMD_ARG_TGT_ARCH}

* pre-include `pycore_atomic.h` to circumvent bad funcdefs

* temporarily truncate job matrix to focus on Windows-ARM64

* sidestep inclusion of `pycore_atomic.h`, just mock what we need

* temporarily stop building Windows-AMD64 and ABI3 wheels

* Added missing #define

* Try moving Python internal includes into the C file

* simplify macros

* roll back #392

* vendor in `_Py_atomic_int` from `pycore_atomic.h`

* Skip unavailable platforms in CI

* Allow Cython tests to XFAIL on Windows-ARM64

* Fixed attempted import on the failing path

* Undo eda8d0b and f407d00, restoring the rest of the pipeline

* Temp: truncate pipeline to debug test on Win-ARM64

This reverts commit 3d47c90.

* Set MSVC arch for ARM runners

* Revert soft-XFAIL of Cython test

This partially reverts commit 0a16f5c
but keeps the typing fixes introduced therein.

* Un-truncate pipeline

This reverts commit 072106d.

* Streamline pipeline, added comments

* changelog

* Make suggested changes

Co-authored-by: Jon Crall <erotemic@gmail.com>

---------

Co-authored-by: da-woods <dw-git@d-woods.co.uk>
Co-authored-by: Jon Crall <erotemic@gmail.com>

Author: 3 people

.github/workflows/tests.yml

@@ -128,16 +128,20 @@ jobs:
         - ubuntu-latest
         - macOS-latest
         - windows-latest
+        # Since the oldest Python available on windows-11-arm on GitHub
+        # is 3.11.0 (see https://raw.githubusercontent.com/actions/\
+        # python-versions/main/versions-manifest.json), older wheels
+        # cannot be tested... so just skip pre-building them
         cibw_skip:
-        - '*-win32 *-win32 cp313-musllinux_i686'
+        - '*-win32 cp3{9,10}-win_arm64 cp313-musllinux_i686'
         arch:
         - auto
     steps:
     - name: Checkout source
       uses: actions/checkout@v4.2.2
     - name: Enable MSVC 64bit
       uses: ilammy/msvc-dev-cmd@v1
-      if: matrix.os == 'windows-latest' && ${{ contains(matrix.cibw_skip, '*-win32') }}
+      if: ${{ startsWith(matrix.os, 'windows-') }} && ${{ contains(matrix.cibw_skip, '*-win32') }}
     - name: Set up QEMU
       uses: docker/setup-qemu-action@v3.0.0
       if: runner.os == 'Linux' && matrix.arch != 'auto'
@@ -150,9 +154,17 @@ jobs:
         config-file: pyproject.toml
       env:
         CIBW_SKIP: ${{ matrix.cibw_skip }}
+        # We're building on Windows-x64, so ARM64 wheels can't be tested
+        # locally by `cibuildwheel` (don't worry, we're testing them
+        # later though in `test_binpy_wheels`)
+        CIBW_TEST_SKIP: '*-win_arm64'
         CIBW_ARCHS_LINUX: ${{ matrix.arch }}
         CIBW_ENVIRONMENT: PYTHONUTF8=1
         PYTHONUTF8: '1'
+        # `msvc-dev-cmd` sets this envvar, which interferes with
+        # cross-architecture building...
+        # just let `cibuildwheel` handle that
+        VSCMD_ARG_TGT_ARCH: ''
     - name: Show built files
       shell: bash
       run: ls -la wheelhouse
@@ -224,6 +236,7 @@ jobs:
           install-extras: tests-strict,runtime-strict
           os: windows-latest
           arch: auto
+        # Note: cibuildwheel can't target 3.8 on Window ARM64
         - python-version: '3.13'
           install-extras: tests-strict,runtime-strict,optional-strict
           os: ubuntu-latest
@@ -236,6 +249,10 @@ jobs:
           install-extras: tests-strict,runtime-strict,optional-strict
           os: windows-latest
           arch: auto
+        - python-version: '3.13'
+          install-extras: tests-strict,runtime-strict,optional-strict
+          os: windows-11-arm
+          arch: auto
         - python-version: '3.13'
           install-extras: tests
           os: macOS-latest
@@ -244,6 +261,10 @@ jobs:
           install-extras: tests
           os: windows-latest
           arch: auto
+        - python-version: '3.13'
+          install-extras: tests
+          os: windows-11-arm
+          arch: auto
         - python-version: '3.8'
           install-extras: tests,optional
           os: ubuntu-latest
@@ -328,12 +349,35 @@ jobs:
           install-extras: tests,optional
           os: windows-latest
           arch: auto
+        # Again, cibuildwheel can't target 3.8 on Window ARM64, and
+        # GitHub doesn't have anything below Python 3.11 on their ARM64
+        # machines, so just test the built wheels from 3.11+
+        - python-version: '3.11'
+          install-extras: tests,optional
+          os: windows-11-arm
+          arch: auto
+        - python-version: '3.12'
+          install-extras: tests,optional
+          os: windows-11-arm
+          arch: auto
+        - python-version: '3.13'
+          install-extras: tests,optional
+          os: windows-11-arm
+          arch: auto
+        - python-version: 3.14.0-rc.1
+          install-extras: tests,optional
+          os: windows-11-arm
+          arch: auto
     steps:
     - name: Checkout source
       uses: actions/checkout@v4.2.2
     - name: Enable MSVC 64bit
       uses: ilammy/msvc-dev-cmd@v1
-      if: matrix.os == 'windows-latest'
+      if: ${{ startsWith(matrix.os, 'windows-') }}
+      # As noted in msvc-dev-cmd #90 (and the Action docs), it currently
+      # # assumes `arch=x64`, so we have to manually set it here...
+      with:
+        arch: ${{ contains(matrix.os, 'arm') && 'arm64' || 'x64' }}
     - name: Set up QEMU
       uses: docker/setup-qemu-action@v3.0.0
       if: runner.os == 'Linux' && matrix.arch != 'auto'

CHANGELOG.rst

@@ -10,6 +10,7 @@ Changes
 * FIX: ref-count leaks #372
 * FIX: mitigate speed regressions introduced in 5.0.0 #376
 * FIX: Use import system to locate module file run by ``kernprof -m`` #389
+* FIX: Fixed build on Windows-ARM64 and now building wheels therefor in CI #391
 
 5.0.0
 ~~~~~

line_profiler/c_trace_callbacks.h

@@ -14,6 +14,12 @@
  *   and causes problems in 3.12 (see CPython #105268, #105350, #107348)
  * - Undefine the `HAVE_STD_ATOMIC` macro, which causes problems on
  *   Linux in 3.12 (see CPython #108216)
+ * - Set `Py_ATOMIC_H` to true to circumvent the #include of
+ *   `include/pycore_atomic.h` (in `include/pycore_interp.h`, so that
+ *   problematic function definitions therein are replaced with dummy
+ *   ones (see #390); note that we still need to vendor in parts
+ *   therefrom which are used by `pycore_interp.h`, and its dependencies
+ *   `pycore_ceval_state.h` and `pycore_gil.h` (or at least mock them)
  * Note in any case that we don't actually use `PyInterpreterState`
  * directly -- we just need its memory layout so that we can refer to
  * its `.last_restart_version` member
@@ -24,17 +30,45 @@
 #   ifndef Py_BUILD_CORE
 #       define Py_BUILD_CORE 1
 #   endif
-#   ifdef _PyGC_FINALIZED
+#   if PY_VERSION_HEX < 0x030d0000  // 3.13
 #       undef _PyGC_FINALIZED
+#       ifdef __linux__
+#           undef HAVE_STD_ATOMIC
+#       endif
+#       if (defined(_M_ARM) || defined(_M_ARM64)) && (! defined(Py_ATOMIC_H))
+#           define Py_ATOMIC_H
+            // Used in `pycore_interp.h`
+            typedef struct _Py_atomic_address {
+                volatile uintptr_t _value;
+            } _Py_atomic_address;
+            // Used in `pycore_gil.h` and `pycore_ceval_state.h`
+            typedef struct _Py_atomic_int {
+                volatile int _value;
+            } _Py_atomic_int;
+            /* Stub out macros in `pycore_atomic.h` used in macros in
+             * `pycore_interp.h` (which aren't related to the
+             * `struct _is` we need).
+             * If any stub is referenced, fail the build with an
+             * unresolved external.
+             * This ensures we never ship wheels that "use" these
+             * placeholders. */
+#           ifdef _MSC_VER
+                __declspec(dllimport) void lp_link_error__stubbed_cpython_atomic_LOAD_relaxed_was_used_this_is_a_bug(void);
+                __declspec(dllimport) void lp_link_error__stubbed_cpython_atomic_STORE_relaxed_was_used_this_is_a_bug(void);
+#           else
+                extern void lp_link_error__stubbed_cpython_atomic_LOAD_relaxed_was_used_this_is_a_bug(void);
+                extern void lp_link_error__stubbed_cpython_atomic_STORE_relaxed_was_used_this_is_a_bug(void);
+#           endif
+#           define _LP_ATOMIC_PANIC_LOAD_EXPR()  (lp_link_error__stubbed_cpython_atomic_LOAD_relaxed_was_used_this_is_a_bug(), 0)
+#           define _LP_ATOMIC_PANIC_STORE_STMT() do { lp_link_error__stubbed_cpython_atomic_STORE_relaxed_was_used_this_is_a_bug(); } while (0)
+            // Panic-on-use shims (expression/statement forms)
+#           undef  _Py_atomic_load_relaxed
+#           undef  _Py_atomic_store_relaxed
+#           define _Py_atomic_load_relaxed(obj)       ((void)(obj), _LP_ATOMIC_PANIC_LOAD_EXPR())
+#           define _Py_atomic_store_relaxed(obj, val)  do { (void)(obj); (void)(val); _LP_ATOMIC_PANIC_STORE_STMT(); } while (0)
+#       endif
 #   endif
-#   ifdef HAVE_STD_ATOMIC
-#       undef HAVE_STD_ATOMIC
-#   endif
-#   if PY_VERSION_HEX >= 0x030900a6  // 3.9.0a6
-#      include "internal/pycore_interp.h"
-#   else
-#      include "internal/pycore_pystate.h"
-#   endif
+#   include "internal/pycore_interp.h"
 #endif
 
 typedef struct TraceCallback

pyproject.toml

@@ -40,6 +40,8 @@ test-extras = ["tests-strict", "runtime-strict"]
 # https://cibuildwheel.readthedocs.io/en/stable/options/#archs
 [tool.cibuildwheel.macos]
 archs = ["x86_64", "universal2", "arm64"]
+[tool.cibuildwheel.windows]
+archs = ['AMD64', 'ARM64']
 
 
 [tool.mypy]

tests/test_cython.py

@@ -18,7 +18,8 @@
 
 from line_profiler._line_profiler import (
     CANNOT_LINE_TRACE_CYTHON, find_cython_source_file)
-from line_profiler.line_profiler import get_code_block, LineProfiler
+from line_profiler.line_profiler import (  # type:ignore[attr-defined]
+    get_code_block, LineProfiler)
 
 
 def propose_name(prefix: str) -> Generator[str, None, None]:
@@ -71,7 +72,8 @@ def _install_cython_example(
 
 @pytest.fixture(scope='module')
 def cython_example(
-        tmp_path_factory: pytest.TempPathFactory) -> Tuple[Path, ModuleType]:
+    tmp_path_factory: pytest.TempPathFactory,
+) -> Generator[Tuple[Path, ModuleType], None, None]:
     """
     Install the example Cython module, yield the path to the Cython
     source file and the corresponding module, uninstall it at teardown.
@@ -83,7 +85,7 @@ def cython_example(
         yield (path, import_module(mod_name))
 
 
-def test_recover_cython_source(cython_example: Tuple[Path, str]) -> None:
+def test_recover_cython_source(cython_example: Tuple[Path, ModuleType]) -> None:
     """
     Check that Cython sources are correctly located by
     `line_profiler._line_profiler.find_cython_source_file()` and
@@ -104,16 +106,17 @@ def test_recover_cython_source(cython_example: Tuple[Path, str]) -> None:
     CANNOT_LINE_TRACE_CYTHON,
     reason='Cannot line-trace Cython code in version '
     + '.'.join(str(v) for v in sys.version_info[:3]))
-def test_profile_cython_source(cython_example: Tuple[Path, str]) -> None:
+def test_profile_cython_source(cython_example: Tuple[Path, ModuleType]) -> None:
     """
     Check that calls to Cython functions (built with the appropriate
     compile-time options) can be profiled.
     """
     prof_cos = LineProfiler()
     prof_sin = LineProfiler()
 
-    cos = prof_cos(cython_example[1].cos)
-    sin = prof_sin(cython_example[1].sin)
+    _, module = cython_example
+    cos = prof_cos(module.cos)
+    sin = prof_sin(module.sin)
     assert pytest.approx(cos(.125, 10)) == math.cos(.125)
     assert pytest.approx(sin(2.5, 3)) == 2.5 - 2.5 ** 3 / 6 + 2.5 ** 5 / 120