[Python-Dev] Optimization of the Year

["Martin v. Löwis"]

Guido van Rossum wrote:
> I'm all for this.  I imagine there are many subtleties though -- I
> know I consciously chose 'int' for this purpose in Python 0, and the
> assumption that it fits in an int is everywhere.  I would trust a
> decent compiler to find most cases, but not all.

The MS Win64 compiler (Itanium) is pretty good at finding size_t/int
conversions, as size_t is wider even than long on the system. I have
patches sitting on my disk to change all function signatures, but
not the structure layouts. They are (unfortunately) still incomplete,
and it is doubtful whether I can complete them before PyCon.

> Also, occasionally (size_t)(-1) may be used to indicate an error
> (e.g. PySequence_Size()).

This is the tricky part that I want to defer in the first round
of patches: I believe ob_size should be a signed type, so it should
be ssize_t where available. The tricky part is finding an appropriate
type on systems where ssize_t is not available.

(size_t)(-1) is not a problem per se; this is well-defined on all
ISO C compilers. However, some code may expect that the error marker
is smaller than any other value, which would not be the case for
(size_t)(-1).

Regards,
Martin