[Python-Dev] Third milestone of FAT Python

[Victor Stinner]

Le mercredi 16 décembre 2015, Franklin? Lee <leewangzhong+python at gmail.com>
a écrit :
>
> I am confident that the time overhead and the savings will beat the
> versioning dict. The versioning dict method has to save a reference to
> the variable value and a reference to the name, and regularly test
> whether the dict has changed.


The performance of guards matters less than the performance of regular
usage of dict. If we have to make a choice, I prefer "slow" guard but no
impact on regular dict methods. It's very important that enabling FAT mode
doesn't kill performances. Remember that FAT python is a static optimizer
and so can only optimize some patterns, not all Python code.

In my current implementation, a lookup is only needed when aguard is
checked if the dict was modified. The dict version doesn't change if a
mutable object was modified in place for example. I didn't benchmark, but I
expect that the lookup is avoided in most cases. You should try FAT python
and implement statistics before going too far in your idea.


> I've read it again. By subclass, I mean that it implements the same
> interface. But at the C level, I want to have it be a fork(?) of the
> current dict implementation. As for `exec`, I think it might be okay
> for it to be slower at the early stages of this game.


Be careful, dict methods are hardcoded in the C code. If your type is not a
subtype, there is risk of crashes. I fixed issues in Python/ceval.c, but
it's not enough. You may also have to fix issues in third party C
extensions why only expect dict for namespaces.

Victor
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/python-dev/attachments/20151216/8bb08216/attachment.html>