Merge pull request #1004 from esc/release0.41

Cherry-picks and release notes for 0.41.1

.readthedocs.yaml

@@ -15,5 +15,6 @@ sphinx:
 
 python:
   install:
+    - requirements: docs/rtd-requirements.txt
     - method: pip
       path: .

CHANGE_LOG

@@ -1,3 +1,23 @@
+v0.41.1 (Oct 17, 2023)
+----------------------
+
+This is a maintenance release that includes a workaround in the test suite for
+ORCJit issues on the ``aarch64`` platform. Also, this is the last release to
+support the Windows 32-bit platform (``win32``).
+
+Pull-Requests:
+
+* PR `#996 <https://github.com/numba/llvmlite/pull/996>`_: fix typos found by codespell (`esc <https://github.com/esc>`_)
+* PR `#997 <https://github.com/numba/llvmlite/pull/997>`_: Fix issue #880 by ensuring all sources are compiled under FreeBSD. (`ke6jjj <https://github.com/ke6jjj>`_)
+* PR `#998 <https://github.com/numba/llvmlite/pull/998>`_: adding sphinx_rtd_theme to RTD build to fix build (`esc <https://github.com/esc>`_)
+* PR `#1001 <https://github.com/numba/llvmlite/pull/1001>`_: Fix / workaround for OrcJIT blocking issues (`gmarkall <https://github.com/gmarkall>`_)
+
+Authors:
+
+* `esc <https://github.com/esc>`_
+* `ke6jjj <https://github.com/ke6jjj>`_
+* `gmarkall <https://github.com/gmarkall>`_
+
 v0.41.0 (Sept 20, 2023)
 -----------------------
 

README.rst

@@ -51,7 +51,7 @@ Key Benefits
   ctypes (no need to wrestle with Python's compiler requirements and C++ 11
   compatibility).
 * The Python binding layer has sane memory management.
-* llvmlite is quite faster than llvmpy's thanks to a much simpler architeture
+* llvmlite is faster than llvmpy thanks to a much simpler architecture
   (the Numba_ test suite is twice faster than it was).
 
 Compatibility

docs/rtd-requirements.txt

@@ -0,0 +1 @@
+sphinx_rtd_theme

docs/source/user-guide/ir/ir-builder.rst

@@ -530,7 +530,7 @@ Function call
        optimized, but may be ignored.
      * ``"musttail"``: indicates that the call must be tail-call optimized for
        program correctness.
-     * ``"notail"``: indicate thats the call should never be tail-call
+     * ``"notail"``: indicates that the call should never be tail-call
        optimized.
 
      For backwards compatibility with previous versions, the following values

ffi/Makefile.freebsd

@@ -11,7 +11,7 @@ LIBS = $(LLVM_LIBS)
 INCLUDE = core.h
 SRC = assembly.cpp bitcode.cpp core.cpp initfini.cpp module.cpp value.cpp \
 	executionengine.cpp transforms.cpp passmanagers.cpp targets.cpp dylib.cpp \
-	linker.cpp object_file.cpp orcjit.cpp
+	linker.cpp object_file.cpp orcjit.cpp custom_passes.cpp
 OUTPUT = libllvmlite.so
 
 all: $(OUTPUT)

llvmlite/tests/test_binding.py

@@ -112,6 +112,9 @@ def no_de_locale():
     }}
     """
 
+if platform.python_implementation() == 'PyPy':
+    asm_getversion = asm_getversion.replace('Py_GetVersion', 'PyPy_GetVersion')
+
 # `fadd` used on integer inputs
 asm_parse_error = r"""
     ; ModuleID = '<string>'
@@ -1202,6 +1205,11 @@ def jit(self, mod, target_machine=None):
         return llvm.create_mcjit_compiler(mod, target_machine)
 
 
+# There are some memory corruption issues with OrcJIT on AArch64 - see Issue
+# #1000. Since OrcJIT is experimental, and we don't test regularly during
+# llvmlite development on non-x86 platforms, it seems safest to skip these
+# tests on non-x86 platforms.
+@unittest.skipUnless(platform.machine().startswith("x86"), "x86 only")
 class TestOrcLLJIT(BaseTest):
 
     def jit(self, asm=asm_sum, func_name="sum", target_machine=None,