On Wednesday, 5 December 2012 22:10:51 UTC+5:30, Bruno Dupuis wrote:
> On Wed, Dec 05, 2012 at 04:15:59PM +0000, Neil Cerutti wrote:
>
> > On 2012-12-05, Bruno Dupuis <python.ml.bruno.dup...@lisael.org> wrote:
>
> > > Hi,
>
> > >
>
> > > I'm interested in compilers optimizations, so I study python
>
> > > compilation process
>
> > >
>
> > > I ran that script:
>
> > >
>
> > > import timeit
>
> > >
>
> > > def f(x):
>
> > > return None
>
> > >
>
> > > def g(x):
>
> > > return None
>
> > > print(x)
>
> > >
>
> > > number = 10000
>
> > >
>
> > > print(timeit.timeit('f(1)',setup="from __main__ import f",
> > > number=number))
>
> > > print(timeit.timeit('g(1)',setup="from __main__ import g",
> > > number=number))
>
> > >
>
> > > print(dis.dis(f))
>
> > > print(dis.dis(g))
>
> > >
>
> > > It gives this output:
>
> > >
>
> > > 0.003460251959040761
>
> > > 0.004164454061537981
>
> > > 17 0 LOAD_CONST 0 (None)
>
> > > 3 RETURN_VALUE
>
> > > None
>
> > > 20 0 LOAD_GLOBAL 0 (None)
>
> > > 3 RETURN_VALUE
>
> > >
>
> > > 21 4 LOAD_GLOBAL 1 (print)
>
> > > 7 LOAD_FAST 0 (x)
>
> > > 10 CALL_FUNCTION 1 (1 positional, 0 keyword
> > > pair)
>
> > > 13 POP_TOP
>
> > > None
>
> > >
>
> > > I do not understand why the dead code `print(x)` takes time (~20% in
>
> > > that case). As we see in the opcode, a call to g(1) returns immediately,
> > > so
>
> > > there should be no delay at all. Where am i wrong?
>
> > >
>
> > > mmhh... it comes to me now that the gap must be in function loading
> > > time...
>
> > > I'll check ceval.c
>
> > >
>
> > > However, isn't there a room for a slight optim here? (in this case, the
>
> > > dead code is obvious, but it may be hidden by complex loops and
>
> > > conditions)
>
> >
>
> > Maybe it's the difference between LOAD_CONST and LOAD_GLOBAL. We
>
> > can wonder why g uses the latter.
>
>
>
> Good point! I didn't even noticed that. It's weird... Maybe the
>
> difference comes from a peehole optim on f which is not possible on g as
>
> g is to complex.
>
>
>
> >
>
> > --
>
> > Neil Cerutti
>
> > --
>
> > http://mail.python.org/mailman/listinfo/python-list
>
>
>
> --
>
> Bruno Dupuis
peehole haha
--
http://mail.python.org/mailman/listinfo/python-list
