[Python-3000] longobject.c and Windows x64

[Mark Dickinson]

On Mon, Apr 14, 2008 at 8:01 PM, Trent Nelson <tnelson at onresolve.com> wrote:

>
> Yeah, that's the 'wrong' part I was referring to.  I guess I wanted to
> bring that issue up as well as question the actual implementation.  For
> example, if we fixed the if statement, we'd having something looking like:
>
> PyObject *
> PyLong_FromSsize_t(Py_ssize_t ival)
> {
>       Py_ssize_t bytes = ival;
>       int one = 1;
>        if (ival < PyLong_BASE && ival > -PyLong_BASE)
>                return PyLong_FromLong(ival);
>       return _PyLong_FromByteArray(
>                       (unsigned char *)&bytes,
>                       SIZEOF_SIZE_T, IS_LITTLE_ENDIAN, 1);
> }
>
> I don't understand why we'd be interested in testing for the
> PyLong_FromLong() shortcut, then reverting to _PyLong_FromByteArray(), when
> we know we're always going to be dealing with a Py_ssize_t.  Speed?  Safety?
>  Cut and paste that went awry?  Why not just call the correct
> PyLong_FromLong(Long)() depending on sizeof(size_t) and be done with it?
>

The extra tests aren't in the trunk version of longobject.c;
It looks to me as though they're the result of merging
the 2.x longobject.c and intobject.c to produce the 3.0
longobject.c.

I also notice that  _PyLong_FromByteArray
doesn't do a CHECK_SMALL_INT, while PyLong_FromLong
does.  Perhaps this is the reason for the extra test?

I agree it would be simpler to just use PyLong_FromLong
or PyLong_FromLongLong.


>
> Surely, if we're guarding properly with #error in the case where
> sizeof(size_t) not in (sizeof(long), sizeof(Py_LONG_LONG)), reusing existing
> methods that do exactly what we want to do would be better than mimicking
> them?
>

Fair enough.  My twisted mind was trying to find ways that size_t
might be something other than long or long long, but that
seems unlikely...


>
> Ah, interesting.  Stepped through PyLong_FromLong via the following:
>
> Python 3.0a4+ (py3k, Apr 14 2008, 18:44:17) [MSC v.1500 64 bit (AMD64)] on
> win32
> Type "help", "copyright", "credits" or "license" for more information.
> >>> import _testcapi as t
> >>> t.getargs_l(t.LONG_MIN)
> -2147483648
> >>> t.getargs_l(t.LONG_MIN+1)
> -2147483647
>
> When ival == LONG_MIN, the 'ival = -ival' statement doesn't have any
> effect on the value of ival, it stays as LONG_MIN.  (With LONG_MIN+1 though,
> ival does correctly get cast back into the positive realm...)  This isn't
> causing a problem (at least not on Windows) as ival's cast to an unsigned
> long later in the method -- I wonder if all platforms/compilers silently
> ignore 'ival = -ival' when at LONG_MIN though...
>

Right:  I think it's technically undefined behaviour (a signed arithmetic
overflow)
that nevertheless ends up doing the right thing on most (all?) compilers.  I
think
it should be fixed.  Something like (untested)

if (ival < 0) {
    t = (unsigned long)(-1-ival) + 1;
}
else {
    t = (unsigned long)ival;
}

should be safe from overflow (including on machines with a ones' complement
or sign-magnitude representation of negative integers---do any such machines
exist any more?).

Shall I check in a fix?

Mark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.python.org/pipermail/python-3000/attachments/20080414/0114d8e8/attachment.htm