Quirk difference between classes and functions

[Chris Angelico]

On Wed, Feb 27, 2019 at 2:21 PM <jfong at ms4.hinet.net> wrote:
>
> Chris Angelico於 2019年2月27日星期三 UTC+8上午9時25分11秒寫道：
> > On Wed, Feb 27, 2019 at 12:21 PM <jfong at ms4.hinet.net> wrote:
> > >
> > > Gregory Ewing at 2019/2/27 AM 5:26:49 wrote:
> > > > Thomas Jollans wrote:
> > > > > I imagine there's a justification for the difference in behaviour to do
> > > > > with the fact that the body of a class is only ever executed once, while
> > > > > the body of a function is executed multiple times.
> > > >
> > > > I suspect there isn't any deep reason for it, rather it's just
> > > > something that fell out of the implementation, in particular
> > > > the decision to optimise local variable access in functions
> > > > but not other scopes.
> > > >
> > > > When compiling a function, the compiler needs to know which
> > > > variables are local so that it can allocate slots for them in
> > > > the stack frame. But when executing a class body, the locals
> > > > are kept in a dict and are looked up dynamically.
> > >
> > > If the compiler can do any decision on the variable's name, when it goes to line of print, how it handle it?
> > >
> > > x = 0
> > > def f():
> > >     print(x)
> > >
> > > def g():
> > >     print(x)
> > >     x = 1
> > >     print(y)
> > >
> >
> > Not sure what you mean by "decision", but the definition is that since
> > x is assigned to in g(), it is local to g(). It's not local to f(), so
> > the global is visible.
>
> So, may I say that the Python compiler is a multi-pass one?
>

At this point, you're veering away from "Python-the-language" and
towards "CPython-the-implementation" (or whichever other
implementation you want to explore). But I would say that, yes, most
or all Python implementations will use multi-pass compilation. It's
the easiest way to guarantee correct behaviour. For instance, when you
reach a "try" statement, you don't know whether there'll be an
"except" block after it (there might just be a "finally"); the
bytecode that CPython produces is different for setting up a
finally-only block than for setting up a try/except. Easiest to find
out that sort of thing by running multiple passes.

Plus, depending on how you define "pass", there are additional steps
such as constant folding and peephole optimization that can be done
after everything else. So, yeah, it's highly unlikely that there are
any performant implementations of Python that run a single compilation
pass.

ChrisA