gh-136006: fix Py_NAN expansion on Solaris systems (#136575)

On Solaris, `Py_NAN` may expand as a function address instead of a floating-point number.
This amends commit 7a3b03509e.
This commit is contained in:
Bénédikt Tran 2025-09-10 11:11:04 +02:00 committed by GitHub
parent b9c50b4988
commit d54b1091d4
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 18 additions and 1 deletions

View file

@ -57,9 +57,24 @@
/* Py_NAN: Value that evaluates to a quiet Not-a-Number (NaN). The sign is
* undefined and normally not relevant, but e.g. fixed for float("nan").
*
* Note: On Solaris, NAN is a function address, hence arithmetic is impossible.
* For that reason, we instead use the built-in call if available or fallback
* to a generic NaN computed from strtod() as a last resort.
*
* See https://github.com/python/cpython/issues/136006 for details.
*/
#if !defined(Py_NAN)
# define Py_NAN ((double)NAN)
# if defined(__sun)
# if _Py__has_builtin(__builtin_nanf)
# define Py_NAN ((double)__builtin_nanf(""))
# else
# include <stdlib.h>
# define Py_NAN (strtod("NAN", NULL))
# endif
# else
# define Py_NAN ((double)NAN)
# endif
#endif
#endif /* Py_PYMATH_H */

View file

@ -0,0 +1,2 @@
On Solaris, the :c:macro:`!Py_NAN` macro now expands to a :c:type:`!double`
instead of a function address. Patch by Bénédikt Tran.