You are correct, the patch should use fabs
I would go with a standard, cross-platform definition of Py_NAN so that pickled objects could be opened by other platforms. Would this patch be better? It's more complicated as I needed to cast the repr of Py_NAN to a unsigned char[]. It passes the tests in test.test_math and handles the copysign in a more intuitive way

>>> math.copysign(1., float('nan')) => 1. on win32, microsoft compiler

diff -r efeca6ff2751 Include/pymath.h
--- a/Include/pymath.h  Thu Apr 05 22:51:00 2012 +0200
+++ b/Include/pymath.h  Sun Apr 08 22:20:16 2012 +0300
@@ -152,8 +152,13 @@
  * doesn't support NaNs.
  */
 #if !defined(Py_NAN) && !defined(Py_NO_NAN)
+#if DBL_MANT_DIG == 53 /* ieee 754 doubles */
+extern double * _Py_NAN;
+#define Py_NAN (*_Py_NAN)
+#else
 #define Py_NAN (Py_HUGE_VAL * 0.)
 #endif
+#endif

 /* Py_OVERFLOWED(X)
  * Return 1 iff a libm function overflowed.  Set errno to 0 before calling
diff -r efeca6ff2751 Modules/_math.c
--- a/Modules/_math.c   Thu Apr 05 22:51:00 2012 +0200
+++ b/Modules/_math.c   Sun Apr 08 22:20:16 2012 +0300
@@ -24,6 +24,10 @@
 static const double two_pow_p28 = 268435456.0; /* 2**28 */
 static const double zero = 0.0;

+#if DBL_MANT_DIG == 53 /* ieee 754 doubles */
+static const unsigned char _Py_NAN_as_char[8] = {0, 0, 0, 0, 0, 0, 0xf8, 0x7f};
+extern double * _Py_NAN = (double *)(_Py_NAN_as_char);
+#endif
 /* acosh(x)
  * Method :
  *      Based on