This is my patch:

[ 1181301 ] make float packing copy bytes when they can

which hasn't been reviewed, despite numerous threats to check it in
anyway if noone reviews it.  Please read the diff on the checkin list,
at least!

The basic idea is to examine the bytes of some 'probe values' to see if
the current platform is a IEEE 754-ish platform, and if so
_PyFloat_{Pack,Unpack}{4,8} just copy bytes around.

The rest is hair for testing, and tests.

Include/floatobject.h

@@ -55,13 +55,18 @@ PyAPI_FUNC(void) PyFloat_AsString(char*, PyFloatObject *v);
  * routines produce a C double from such a string.  The suffix (4 or 8)
  * specifies the number of bytes in the string.
  *
- * Excepting NaNs and infinities (which aren't handled correctly), the 4-
- * byte format is identical to the IEEE-754 single precision format, and
- * the 8-byte format to the IEEE-754 double precision format.  On non-
- * IEEE platforms with more precision, or larger dynamic range, than
- * 754 supports, not all values can be packed; on non-IEEE platforms with
- * less precision, or smaller dynamic range, not all values can be
- * unpacked.  What happens in such cases is partly accidental (alas).
+ * On platforms that appear to use (see _PyFloat_Init()) IEEE-754 formats
+ * these functions work by copying bits.  On other platforms, the formats the
+ * 4- byte format is identical to the IEEE-754 single precision format, and
+ * the 8-byte format to the IEEE-754 double precision format, although the
+ * packing of INFs and NaNs (if such things exist on the platform) isn't
+ * handled correctly, and attempting to unpack a string containing an IEEE
+ * INF or NaN will raise an exception.
+ *
+ * On non-IEEE platforms with more precision, or larger dynamic range, than
+ * 754 supports, not all values can be packed; on non-IEEE platforms with less
+ * precision, or smaller dynamic range, not all values can be unpacked.  What
+ * happens in such cases is partly accidental (alas).
  */
 
 /* The pack routines write 4 or 8 bytes, starting at p.  le is a bool
@@ -70,8 +75,9 @@ PyAPI_FUNC(void) PyFloat_AsString(char*, PyFloatObject *v);
  * first, at p).
  * Return value:  0 if all is OK, -1 if error (and an exception is
  * set, most likely OverflowError).
- * Bug:  What this does is undefined if x is a NaN or infinity.
- * Bug:  -0.0 and +0.0 produce the same string.
+ * There are two problems on non-IEEE platforms:
+ * 1):  What this does is undefined if x is a NaN or infinity.
+ * 2):  -0.0 and +0.0 produce the same string.
  */
 PyAPI_FUNC(int) _PyFloat_Pack4(double x, unsigned char *p, int le);
 PyAPI_FUNC(int) _PyFloat_Pack8(double x, unsigned char *p, int le);
@@ -81,9 +87,8 @@ PyAPI_FUNC(int) _PyFloat_Pack8(double x, unsigned char *p, int le);
  * last, at p+3 or p+7), false if big-endian (exponent first, at p).
  * Return value:  The unpacked double.  On error, this is -1.0 and
  * PyErr_Occurred() is true (and an exception is set, most likely
- * OverflowError).
- * Bug:  What this does is undefined if the string represents a NaN or
- * infinity.
+ * OverflowError).  Note that on a non-IEEE platform this will refuse
+ * to unpack a string that represents a NaN or infinity.
  */
 PyAPI_FUNC(double) _PyFloat_Unpack4(const unsigned char *p, int le);
 PyAPI_FUNC(double) _PyFloat_Unpack8(const unsigned char *p, int le);

Include/pythonrun.h

@@ -105,6 +105,7 @@ PyAPI_FUNC(void) _PyExc_Init(void);
 PyAPI_FUNC(void) _PyImportHooks_Init(void);
 PyAPI_FUNC(int) _PyFrame_Init(void);
 PyAPI_FUNC(int) _PyInt_Init(void);
+PyAPI_FUNC(void) _PyFloat_Init(void);
 
 /* Various internal finalizers */
 PyAPI_FUNC(void) _PyExc_Fini(void);

Lib/test/test_float.py

+
+import unittest, struct
+from test import test_support
+
+class FormatFunctionsTestCase(unittest.TestCase):
+
+    def setUp(self):
+        self.save_formats = {'double':float.__getformat__('double'),
+                             'float':float.__getformat__('float')}
+
+    def tearDown(self):
+        float.__setformat__('double', self.save_formats['double'])
+        float.__setformat__('float', self.save_formats['float'])
+
+    def test_getformat(self):
+        self.assert_(float.__getformat__('double') in
+                     ['unknown', 'IEEE, big-endian', 'IEEE, little-endian'])
+        self.assert_(float.__getformat__('float') in
+                     ['unknown', 'IEEE, big-endian', 'IEEE, little-endian'])
+        self.assertRaises(ValueError, float.__getformat__, 'chicken')
+        self.assertRaises(TypeError, float.__getformat__, 1)
+
+    def test_setformat(self):
+        for t in 'double', 'float':
+            float.__setformat__(t, 'unknown')
+            if self.save_formats[t] == 'IEEE, big-endian':
+                self.assertRaises(ValueError, float.__setformat__,
+                                  t, 'IEEE, little-endian')
+            elif self.save_formats[t] == 'IEEE, little-endian':
+                self.assertRaises(ValueError, float.__setformat__,
+                                  t, 'IEEE, big-endian')
+            else:
+                self.assertRaises(ValueError, float.__setformat__,
+                                  t, 'IEEE, big-endian')
+                self.assertRaises(ValueError, float.__setformat__,
+                                  t, 'IEEE, little-endian')
+            self.assertRaises(ValueError, float.__setformat__,
+                              t, 'chicken')
+        self.assertRaises(ValueError, float.__setformat__,
+                          'chicken', 'unknown')
+
+BE_DOUBLE_INF = '\x7f\xf0\x00\x00\x00\x00\x00\x00'
+LE_DOUBLE_INF = ''.join(reversed(BE_DOUBLE_INF))
+BE_DOUBLE_NAN = '\x7f\xf8\x00\x00\x00\x00\x00\x00'
+LE_DOUBLE_NAN = ''.join(reversed(BE_DOUBLE_NAN))
+
+BE_FLOAT_INF = '\x7f\x80\x00\x00'
+LE_FLOAT_INF = ''.join(reversed(BE_FLOAT_INF))
+BE_FLOAT_NAN = '\x7f\xc0\x00\x00'
+LE_FLOAT_NAN = ''.join(reversed(BE_FLOAT_NAN))
+
+# on non-IEEE platforms, attempting to unpack a bit pattern
+# representing an infinity or a NaN should raise an exception.
+
+class UnknownFormatTestCase(unittest.TestCase):
+    def setUp(self):
+        self.save_formats = {'double':float.__getformat__('double'),
+                             'float':float.__getformat__('float')}
+        float.__setformat__('double', 'unknown')
+        float.__setformat__('float', 'unknown')
+        
+    def tearDown(self):
+        float.__setformat__('double', self.save_formats['double'])
+        float.__setformat__('float', self.save_formats['float'])
+
+    def test_double_specials_dont_unpack(self):
+        for fmt, data in [('>d', BE_DOUBLE_INF),
+                          ('>d', BE_DOUBLE_NAN),
+                          ('<d', LE_DOUBLE_INF),
+                          ('<d', LE_DOUBLE_NAN)]:
+            self.assertRaises(ValueError, struct.unpack, fmt, data)
+
+    def test_float_specials_dont_unpack(self):
+        for fmt, data in [('>f', BE_FLOAT_INF),
+                          ('>f', BE_FLOAT_NAN),
+                          ('<f', LE_FLOAT_INF),
+                          ('<f', LE_FLOAT_NAN)]:
+            self.assertRaises(ValueError, struct.unpack, fmt, data)
+
+
+# on an IEEE platform, all we guarantee is that bit patterns
+# representing infinities or NaNs do not raise an exception; all else
+# is accident (today).
+
+class IEEEFormatTestCase(unittest.TestCase):
+    if float.__getformat__("double").startswith("IEEE"):
+        def test_double_specials_do_unpack(self):
+            for fmt, data in [('>d', BE_DOUBLE_INF),
+                              ('>d', BE_DOUBLE_NAN),
+                              ('<d', LE_DOUBLE_INF),
+                              ('<d', LE_DOUBLE_NAN)]:
+                struct.unpack(fmt, data)
+
+    if float.__getformat__("float").startswith("IEEE"):
+        def test_float_specials_do_unpack(self):
+            for fmt, data in [('>f', BE_FLOAT_INF),
+                              ('>f', BE_FLOAT_NAN),
+                              ('<f', LE_FLOAT_INF),
+                              ('<f', LE_FLOAT_NAN)]:
+                struct.unpack(fmt, data)
+
+
+def test_main():
+    test_support.run_unittest(
+        FormatFunctionsTestCase,
+        UnknownFormatTestCase,
+        IEEEFormatTestCase)
+
+if __name__ == '__main__':
+    test_main()

Objects/floatobject.c

@@ -983,8 +983,139 @@ float_getnewargs(PyFloatObject *v)
 	return Py_BuildValue("(d)", v->ob_fval);
 }
 
+/* this is for the benefit of the pack/unpack routines below */
+
+typedef enum {
+	unknown_format, ieee_big_endian_format, ieee_little_endian_format
+} float_format_type;
+
+static float_format_type double_format, float_format;
+static float_format_type detected_double_format, detected_float_format;
+
+static PyObject *
+float_getformat(PyTypeObject *v, PyObject* arg)
+{
+	char* s;
+	float_format_type r;
+
+	if (!PyString_Check(arg)) {
+		PyErr_Format(PyExc_TypeError,
+	     "__getformat__() argument must be string, not %.500s",
+			     arg->ob_type->tp_name);
+		return NULL;
+	}
+	s = PyString_AS_STRING(arg);
+	if (strcmp(s, "double") == 0) {
+		r = double_format;
+	}
+	else if (strcmp(s, "float") == 0) {
+		r = float_format;
+	}
+	else {
+		PyErr_SetString(PyExc_ValueError,
+				"__getformat__() argument 1 must be "
+				"'double' or 'float'");
+		return NULL;
+	}
+	
+	switch (r) {
+	case unknown_format:
+		return PyString_FromString("unknown");
+	case ieee_little_endian_format:
+		return PyString_FromString("IEEE, little-endian");
+	case ieee_big_endian_format:
+		return PyString_FromString("IEEE, big-endian");
+	default:
+		Py_FatalError("insane float_format or double_format");
+		return NULL;
+	}
+}
+
+PyDoc_STRVAR(float_getformat_doc,
+"float.__getformat__(typestr) -> string\n"
+"\n"
+"You probably don't want to use this function.  It exists mainly to be\n"
+"used in Python's test suite.\n"
+"\n"
+"typestr must be 'double' or 'float'.  This function returns whichever of\n"
+"'unknown', 'IEEE, big-endian' or 'IEEE, little-endian' best describes the\n"
+"format of floating point numbers used by the C type named by typestr.");
+
+static PyObject *
+float_setformat(PyTypeObject *v, PyObject* args)
+{
+	char* typestr;
+	char* format;
+	float_format_type f;
+	float_format_type detected;
+	float_format_type *p;
+
+	if (!PyArg_ParseTuple(args, "ss:__setformat__", &typestr, &format))
+		return NULL;
+
+	if (strcmp(typestr, "double") == 0) {
+		p = &double_format;
+		detected = detected_double_format;
+	}
+	else if (strcmp(typestr, "float") == 0) {
+		p = &float_format;
+		detected = detected_float_format;
+	}
+	else {
+		PyErr_SetString(PyExc_ValueError,
+				"__setformat__() argument 1 must "
+				"be 'double' or 'float'");
+		return NULL;
+	}
+	
+	if (strcmp(format, "unknown") == 0) {
+		f = unknown_format;
+	}
+	else if (strcmp(format, "IEEE, little-endian") == 0) {
+		f = ieee_little_endian_format;
+	}
+	else if (strcmp(format, "IEEE, big-endian") == 0) {
+		f = ieee_big_endian_format;
+	}
+	else {
+		PyErr_SetString(PyExc_ValueError,
+				"__setformat__() argument 2 must be "
+				"'unknown', 'IEEE, little-endian' or "
+				"'IEEE, big-endian'");
+		return NULL;
+
+	}
+
+	if (f != unknown_format && f != detected) {
+		PyErr_Format(PyExc_ValueError,
+			     "can only set %s format to 'unknown' or the "
+			     "detected platform value", typestr);
+		return NULL;
+	}
+
+	*p = f;
+	Py_RETURN_NONE;
+}
+
+PyDoc_STRVAR(float_setformat_doc,
+"float.__setformat__(typestr, fmt) -> None\n"
+"\n"
+"You probably don't want to use this function.  It exists mainly to be\n"
+"used in Python's test suite.\n"
+"\n"
+"typestr must be 'double' or 'float'.  fmt must be one of 'unknown',\n"
+"'IEEE, big-endian' or 'IEEE, little-endian', and in addition can only be\n"
+"one of the latter two if it appears to match the underlying C reality.\n"
+"\n"
+"Overrides the automatic determination of C-level floating point type.\n"
+"This affects how floats are converted to and from binary strings.");
+
 static PyMethodDef float_methods[] = {
 	{"__getnewargs__",	(PyCFunction)float_getnewargs,	METH_NOARGS},
+	{"__getformat__",	(PyCFunction)float_getformat,	
+	 METH_O|METH_CLASS,		float_getformat_doc},
+	{"__setformat__",	(PyCFunction)float_setformat,	
+	 METH_VARARGS|METH_CLASS,	float_setformat_doc},
 	{NULL,		NULL}		/* sentinel */
 };
 
@@ -1078,6 +1209,56 @@ PyTypeObject PyFloat_Type = {
 	float_new,				/* tp_new */
 };
 
+void
+_PyFloat_Init(void)
+{
+	/* We attempt to determine if this machine is using IEEE
+	   floating point formats by peering at the bits of some
+	   carefully chosen values.  If it looks like we are on an
+	   IEEE platform, the float packing/unpacking routines can
+	   just copy bits, if not they resort to arithmetic & shifts
+	   and masks.  The shifts & masks approach works on all finite
+	   values, but what happens to infinities, NaNs and signed
+	   zeroes on packing is an accident, and attempting to unpack
+	   a NaN or an infinity will raise an exception.
+
+	   Note that if we're on some whacked-out platform which uses
+	   IEEE formats but isn't strictly little-endian or big-
+	   endian, we will fall back to the portable shifts & masks
+	   method. */
+
+#if SIZEOF_DOUBLE == 8
+	{
+		double x = 9006104071832581.0;
+		if (memcmp(&x, "\x43\x3f\xff\x01\x02\x03\x04\x05", 8) == 0)
+			detected_double_format = ieee_big_endian_format;
+		else if (memcmp(&x, "\x05\x04\x03\x02\x01\xff\x3f\x43", 8) == 0)
+			detected_double_format = ieee_little_endian_format;
+		else 
+			detected_double_format = unknown_format;
+	}
+#else
+	detected_double_format = unknown_format;
+#endif
+
+#if SIZEOF_FLOAT == 4
+	{
+		float y = 16711938.0;
+		if (memcmp(&y, "\x4b\x7f\x01\x02", 4) == 0)
+			detected_float_format = ieee_big_endian_format;
+		else if (memcmp(&y, "\x02\x01\x7f\x4b", 4) == 0)
+			detected_float_format = ieee_little_endian_format;
+		else 
+			detected_float_format = unknown_format;
+	}
+#else
+	detected_float_format = unknown_format;
+#endif
+
+	double_format = detected_double_format;
+	float_format = detected_float_format;
+}
+
 void
 PyFloat_Fini(void)
 {
@@ -1165,6 +1346,7 @@ PyFloat_Fini(void)
 int
 _PyFloat_Pack4(double x, unsigned char *p, int le)
 {
+	if (float_format == unknown_format) {
 		unsigned char sign;
 		int e;
 		double f;
@@ -1243,11 +1425,30 @@ _PyFloat_Pack4(double x, unsigned char *p, int le)
 		PyErr_SetString(PyExc_OverflowError,
 				"float too large to pack with f format");
 		return -1;
+	}
+	else {
+		float y = x;
+		const char *s = (char*)&y;
+		int i, incr = 1;
+
+		if ((float_format == ieee_little_endian_format && !le)
+		    || (float_format == ieee_big_endian_format && le)) {
+			p += 3;
+			incr = -1;
+		}
+		
+		for (i = 0; i < 4; i++) {
+			*p = *s++;
+			p += incr;
+		}
+		return 0;
+	}
 }
 
 int
 _PyFloat_Pack8(double x, unsigned char *p, int le)
 {
+	if (double_format == unknown_format) {
 		unsigned char sign;
 		int e;
 		double f;
@@ -1354,11 +1555,29 @@ _PyFloat_Pack8(double x, unsigned char *p, int le)
 		PyErr_SetString(PyExc_OverflowError,
 				"float too large to pack with d format");
 		return -1;
+	}
+	else {
+		const char *s = (char*)&x;
+		int i, incr = 1;
+
+		if ((double_format == ieee_little_endian_format && !le)
+		    || (double_format == ieee_big_endian_format && le)) {
+			p += 7;
+			incr = -1;
+		}
+		
+		for (i = 0; i < 8; i++) {
+			*p = *s++;
+			p += incr;
+		}
+		return 0;
+	}
 }
 
 double
 _PyFloat_Unpack4(const unsigned char *p, int le)
 {
+	if (float_format == unknown_format) {
 		unsigned char sign;
 		int e;
 		unsigned int f;
@@ -1380,6 +1599,14 @@ _PyFloat_Unpack4(const unsigned char *p, int le)
 		f = (*p & 0x7F) << 16;
 		p += incr;
 
+		if (e == 255) {
+			PyErr_SetString(
+				PyExc_ValueError,
+				"can't unpack IEEE 754 special value "
+				"on non-IEEE platform");
+			return -1;
+		}
+
 		/* Third byte */
 		f |= *p << 8;
 		p += incr;
@@ -1402,11 +1629,29 @@ _PyFloat_Unpack4(const unsigned char *p, int le)
 			x = -x;
 
 		return x;
+	}
+	else {
+		if ((float_format == ieee_little_endian_format && !le)
+		    || (float_format == ieee_big_endian_format && le)) {
+			char buf[8];
+			char *d = &buf[3];
+			int i;
+
+			for (i = 0; i < 4; i++) {
+				*d-- = *p++;
+			}
+			return *(float*)&buf[0];
+		}
+		else {
+			return *(float*)p;
+		}
+	}		
 }
 
 double
 _PyFloat_Unpack8(const unsigned char *p, int le)
 {
+	if (double_format == unknown_format) {
 		unsigned char sign;
 		int e;
 		unsigned int fhi, flo;
@@ -1421,6 +1666,7 @@ _PyFloat_Unpack8(const unsigned char *p, int le)
 		/* First byte */
 		sign = (*p >> 7) & 1;
 		e = (*p & 0x7F) << 4;
+		
 		p += incr;
 
 		/* Second byte */
@@ -1428,6 +1674,14 @@ _PyFloat_Unpack8(const unsigned char *p, int le)
 		fhi = (*p & 0xF) << 24;
 		p += incr;
 
+		if (e == 2047) {
+			PyErr_SetString(
+				PyExc_ValueError,
+				"can't unpack IEEE 754 special value "
+				"on non-IEEE platform");
+			return -1.0;
+		}
+
 		/* Third byte */
 		fhi |= *p << 16;
 		p += incr;
@@ -1454,7 +1708,6 @@ _PyFloat_Unpack8(const unsigned char *p, int le)
 		x = (double)fhi + (double)flo / 16777216.0; /* 2**24 */
 		x /= 268435456.0; /* 2**28 */
 
-	/* XXX This sadly ignores Inf/NaN */
 		if (e == 0)
 			e = -1022;
 		else {
@@ -1467,4 +1720,21 @@ _PyFloat_Unpack8(const unsigned char *p, int le)
 			x = -x;
 
 		return x;
+	}
+	else {
+		if ((double_format == ieee_little_endian_format && !le)
+		    || (double_format == ieee_big_endian_format && le)) {
+			char buf[8];
+			char *d = &buf[7];
+			int i;
+			
+			for (i = 0; i < 8; i++) {
+				*d-- = *p++;
+			}
+			return *(double*)&buf[0];
+		}
+		else {
+			return *(double*)p;
+		}
+	}
 }

Python/pythonrun.c

@@ -172,6 +172,8 @@ Py_InitializeEx(int install_sigs)
 	if (!_PyInt_Init())
 		Py_FatalError("Py_Initialize: can't init ints");
 
+	_PyFloat_Init();
+
 	interp->modules = PyDict_New();
 	if (interp->modules == NULL)
 		Py_FatalError("Py_Initialize: can't make modules dictionary");