3.5: Include/pyatomic.h is incompatible with OpenMP (compilation of the third-party module fails on Python 3.5)

[axh]

BPO	25150
Nosy	@arekm, @pitrou, @scoder, @vstinner, @benjaminp
Files	error.txt: error message pyatomic.patch pyatomic-2.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2020-06-05.20:42:45.704>
created_at = <Date 2015-09-17.12:58:59.262>
labels = ['extension-modules', '3.8', 'build', '3.7']
title = '3.5: Include/pyatomic.h is incompatible with OpenMP (compilation of the third-party module fails on Python 3.5)'
updated_at = <Date 2020-06-05.20:42:45.701>
user = 'https://bugs.python.org/axh'

bugs.python.org fields:

activity = <Date 2020-06-05.20:42:45.701>
actor = 'vstinner'
assignee = 'none'
closed = True
closed_date = <Date 2020-06-05.20:42:45.704>
closer = 'vstinner'
components = ['Extension Modules']
creation = <Date 2015-09-17.12:58:59.262>
creator = 'axh'
dependencies = []
files = ['40491', '40493', '40497']
hgrepos = ['318']
issue_num = 25150
keywords = ['patch', '3.5regression']
message_count = 30.0
messages = ['250884', '250886', '250888', '250889', '250901', '250919', '250940', '250944', '250945', '250961', '250964', '250972', '250976', '250977', '251157', '251160', '253754', '253755', '254103', '254104', '255229', '255231', '321824', '321825', '321828', '321834', '321860', '321865', '321866', '370783']
nosy_count = 8.0
nosy_names = ['arekm', 'pitrou', 'scoder', 'vstinner', 'benjamin.peterson', 'python-dev', 'axh', 'gul916']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'compile error'
url = 'https://bugs.python.org/issue25150'
versions = ['Python 3.6', 'Python 3.7', 'Python 3.8']

[axh]

trying to install the yt package, most recent version 3.2.1 on python 3.5.0 using pip

pip3 install -U yt

error output attached. The same package installs fine with python 3.4.3, same compiler and also build from scratch, so the suspicion is that it is related to the new 3.5.0 version. The system installed is the current Fedora 22, gcc 5.1.1-4.

[vstinner]

It looks like the problem is that yt uses OpenMP whereas OpenMP doesn't support atomic operations:

In file included from /home/alex/Python/include/python3.5m/pyatomic.h:12:0,
                 from /home/alex/Python/include/python3.5m/Python.h:53,
                 from build/src.linux-x86_64-3.5/yt/utilities/lib/geometry_utils.c:4:
/usr/lib/gcc/x86_64-redhat-linux/5.1.1/include/stdatomic.h:40:1: sorry, unimplemented: ‘_Atomic’ with OpenMP
 typedef _Atomic _Bool atomic_bool;
 ^

I'm not sure that Include/pyatomic.h should be included by the main Include/Python.h header. We already skipped Include/pyatomic.h on C++ because this header is incompatible with C++.

[axh]

When I just comment out the

#include "pyatomic.h" 

line, python 3.5.0 will no longer compile

gcc -pthread -c -Wno-unused-result -Wsign-compare -Wunreachable-code -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -Werror=declaration-after-statement -I. -IInclude -I./Include -DPy_BUILD_CORE -o Programs/python.o ./Programs/python.c
In file included from Include/traceback.h:8:0,
from Include/Python.h:97,
from ./Programs/python.c:3:
Include/pystate.h:186:8: error: unknown type name ‘_Py_atomic_address’
PyAPI_DATA(_Py_atomic_address) _PyThreadState_Current;
^
Makefile:747: recipe for target 'Programs/python.o' failed
make: *** [Programs/python.o] Error 1

pystate.h:

/* Assuming the current thread holds the GIL, this is the
PyThreadState for the current thread.

   Issue python/cpython#67832: pyatomic.h is incompatible with C++ (yet). Disable
   PyThreadState_GET() optimization: declare it as an alias to
   PyThreadState_Get(), as done for limited API. */
#if !defined(Py_LIMITED_API) && !defined(__cplusplus)
PyAPI_DATA(_Py_atomic_address) _PyThreadState_Current;
#endif

[vstinner]

"When I just comment out the << #include "pyatomic.h" >> line, python 3.5.0 will no longer compile"

Sure. My idea is to "disable" the header with we are not building Python itself. There is a nice define for that: Py_BUILD_CORE. See attached patch.

Since all symbols in pyatomic.h are prefixed by _Py, this header is fully part of the Python private API and so it's fine to modify it in a bugfix release (3.5.0 => 3.5.1).

[brettcannon]

If everything in the header is a _Py definition then I agree we should hide it in Python.h from the public.

[axh]

if I just include this patch, some modules don't build:

(...)

gcc -pthread -fPIC -Wno-unused-result -Wsign-compare -Wunreachable-code -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -Werror=declaration-after-statement -Ibuild/temp.linux-x86_64-3.5/libffi/include -Ibuild/temp.linux-x86_64-3.5/libffi -I/home/alex/Python-3.5.0/Modules/_ctypes/libffi/src -I./Include -I/home/alex/Python/include -I. -IInclude -I/usr/local/include -I/home/alex/Python-3.5.0/Include -I/home/alex/Python-3.5.0 -c /home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c -o build/temp.linux-x86_64-3.5/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.o -Wall -fexceptions
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c: In function ‘PyCSimpleType_from_param’:
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2062:15: error: ‘_PyThreadState_Current’ undeclared (first use in this function)
         if (Py_EnterRecursiveCall("while processing _as_parameter_")) {
               ^
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2062:15: note: each undeclared identifier is reported only once for each function it appears in
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2062:28: error: ‘__atomic_load_ptr’ undeclared (first use in this function)
         if (Py_EnterRecursiveCall("while processing _as_parameter_")) {
                            ^
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2062:66: error: argument 1 of ‘__atomic_load’ must be a non-void pointer type
         if (Py_EnterRecursiveCall("while processing _as_parameter_")) {
                                                                  ^
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2062:19: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast]
         if (Py_EnterRecursiveCall("while processing _as_parameter_")) {
                   ^
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2067:62: error: argument 1 of ‘__atomic_load’ must be a non-void pointer type
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2067:21: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast]
         Py_LeaveRecursiveCall();
                     ^
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2067:62: error: argument 1 of ‘__atomic_load’ must be a non-void pointer type
/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2067:139: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast]

Failed to build these modules:
_ctypes _decimal _json
_pickle

[axh]

So, apparently, more than just one spot needs to be fixed. I also tried just modifying the Python.h after install, but that does not do the trick either.

[vstinner]

/home/alex/Python-3.5.0/Modules/_ctypes/_ctypes.c:2062:15: error: ‘_PyThreadState_Current’ undeclared (first use in this function)
         if (Py_EnterRecursiveCall("while processing _as_parameter_")) {
               ^

Ah yes, if you check the fix for C++ (changeset cb05b6d7aacd), I also had to modify pystate.h.

Please try pyatomic-2.patch which hides more CPython internals in public headers.

[vstinner]

I created a venv with a Python patched with pyatomic-2.patch: I successfully installed yt. yt depends on numpy & Cython: good news, numpy & Cython were compiled correctly. These two libraries are well known users of the Python C API. It's not enough to check if this change breaks modules on PyPI, but it's still a good news :-)

[scoder]

Would there be a way to expose these internals rather than hiding them?

[vstinner]

Would there be a way to expose these internals rather than hiding them?

Here is the issue is that pyatomic.h cannot be compiled on OpenMP. We had the same issue with C++. In fact, it doesn't make sense to compile pyatomic.h differently to access an atomic variable from an extension module. We must always use exactly the same implementation, otherwise bad things will happen.

A solution for that is to hide the implementation details and only expose high level APIs.

For example, pyatomic.h must be completly hidden.

A consequence is that the _PyThreadState_Current variable must be hidden to. _PyThreadState_Current is an implementation detail, you must not access it directly.

The PyThreadState_GET() macro uses directly the _PyThreadState_Current variable. So the solution to expose the "PyThreadState_GET" symbol (not necessary as a macro) is to define it as an alias to the PyThreadState_Get() *function*.

The advantage of using a function is that we don't expose implementation details to third-party extensions, it avoids the risk of ABI incompatibilies.

[scoder]

Understood and agreed. Second patch looks good to me.

Cython calls PyThreadState_GET() in pretty much every helper function that deals with exceptions, but I doubt that the potential speed difference is going to be relevant in the real world. And we target CPython's API level anyway, not the ABI, so the C code will just adapt at compile time.