Remove the Py_NO_NAN macro: require NAN to build Python 3.11

[mdickinson]

BPO	46656
Nosy	@mdickinson, @vstinner
PRs	bpo-46656: Remove Py_NO_NAN macro #31160 bpo-46656: Building Python now requires a C11 compiler #31557

^{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 2022-02-25.00:45:52.191>
created_at = <Date 2022-02-06.09:54:34.993>
labels = ['interpreter-core', 'type-bug', 'build', '3.11']
title = 'Remove the Py_NO_NAN macro: require NAN to build Python 3.11'
updated_at = <Date 2022-03-10.16:52:54.326>
user = 'https://github.com/mdickinson'

bugs.python.org fields:

activity = <Date 2022-03-10.16:52:54.326>
actor = 'vstinner'
assignee = 'none'
closed = True
closed_date = <Date 2022-02-25.00:45:52.191>
closer = 'vstinner'
components = ['Build', 'Interpreter Core']
creation = <Date 2022-02-06.09:54:34.993>
creator = 'mark.dickinson'
dependencies = []
files = []
hgrepos = []
issue_num = 46656
keywords = ['patch']
message_count = 13.0
messages = ['412620', '412621', '412625', '412629', '412630', '412632', '412636', '412640', '412642', '412864', '413944', '413945', '414864']
nosy_count = 2.0
nosy_names = ['mark.dickinson', 'vstinner']
pr_nums = ['31160', '31557']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue46656'
versions = ['Python 3.11']

[mdickinson]

The macro Py_NAN may or may not be defined: in particular, a platform that doesn't have NaNs is supposed to be able to define Py_NO_NAN in pyport.h to indicate that.

But not all of our uses of Py_NAN are guarded by suitable #ifdef conditionals. As a result, compilation fails if Py_NAN is not defined.

[mdickinson]

Here's the first point of failure on my machine. Fixing this shows up more failures.

gcc -c -Wsign-compare -Wunreachable-code -DNDEBUG -g -fwrapv -O3 -Wall -std=c99 -Wextra -Wno-unused-parameter -Wno-missing-field-initializers -Wstrict-prototypes -Werror=implicit-function-declaration -fvisibility=hidden -I./Include/internal -I. -I./Include -DPy_BUILD_CORE -o Objects/complexobject.o Objects/complexobject.c
Objects/complexobject.c:120:27: error: use of undeclared identifier 'Py_NAN'
r.real = r.imag = Py_NAN;
^
Objects/complexobject.c:206:16: error: use of undeclared identifier 'Py_NAN'
return Py_NAN;
^
2 errors generated.
make: *** [Objects/complexobject.o] Error 1

[vstinner]

Python uses Py_NAN without "#ifdef Py_NAN" guard since 2008. Building Python with Py_NO_NAN never worked. Nobody reported such build failure in the last 14 years...

IMO it's time to stop supporting platforms without NaN support.

Objects/complexobject.c:120:27: error: use of undeclared identifier 'Py_NAN'

I only try a few Python versions: I reproduce this issue in Python 3.11, 3.5, 3.2 and... even Python 2.7!

The Py_NO_NAN macro was introduced in Python 2.6 (2007) by bpo-1635 "Float patch for inf and nan on Windows (and other platforms)":

commit 0a8143f
Author: Christian Heimes <christian@cheimes.de>
Date: Tue Dec 18 23:22:54 2007 +0000

Applied patch bpo-1635: Float patch for inf and nan on Windows (and other platforms).

The patch unifies float("inf") and repr(float("inf")) on all platforms.

(...)

+#if !defined(Py_NAN) && !defined(Py_NO_NAN)
+#define Py_NAN (Py_HUGE_VAL * 0.)
+#endif
---

The following change started to use Py_NAN in many C files:

• Modules/cmathmodule.c
• Objects/complexobject.c
• Objects/floatobject.c
• Python/pymath.c

---
commit 6f34109 (HEAD)
Author: Christian Heimes <christian@cheimes.de>
Date: Fri Apr 18 23:13:07 2008 +0000

I finally got the time to update and merge Mark's and my trunk-math branch. The patch is collaborated w

ork of Mark Dickinson and me. It was mostly done a few months ago. The patch fixes a lot of loose ends and
edge cases related to operations with NaN, INF, very small values and complex math.

The patch also adds acosh, asinh, atanh, log1p and copysign to all platforms. Finally it fixes differen

ces between platforms like different results or exceptions for edge cases. Have fun :)
---

At this commit, floatobject.c and pymath.c use "#ifdef Py_NAN" carefully, whereas cmathmodule.c and complexobject.c use Py_NAN with no "#ifdef Py_NAN" guard.

[vstinner]

If someone suddenly requires platforms without NaN support, they can maintain a (downstream) patch, or we can revert the change in Python.

[vstinner]

See also bpo-46640 "Python can now use the C99 NAN constant or __builtin_nan()".

[mdickinson]

Okay, the comments I made on bpo-46640 still apply (even though they didn't properly apply on that issue). I think this needs a python-dev discussion before it can be moved forward - requiring the existence of NaNs is very close to requiring IEEE 754 floating-point, and that's something we've been historically reluctant to do.

[vstinner]

Is the macro PY_NO_SHORT_FLOAT_REPR also related to platforms which don't support IEEE 754?

In 2022, which platforms don't support IEEE 754?