specifying constants for a function (WAS: generator expressions:performance anomaly?)

[Steven Bethard]

I wrote:

If you really want locals that don't contribute to arguments, I'd be
much happier with something like a decorator, e.g.[1]:

@with_consts(i=1, deftime=time.ctime())
def foo(x, y=123, *args, **kw):
return x*y, kw.get('which_time')=='now' and time.ctime() or deftime

Then you don't have to mix parameter declarations with locals
definitions.

Steve

[1] I have no idea how implementable such a decorator would be. I'd
just like to see function constants declared separate from arguments
since they mean such different things.

I played around with this, and I think it's basically implementable:

py> import new
py> class with_consts(object):
.... def __init__(self, **consts):
.... self.consts = consts
.... def __call__(self, func):
.... return new.function(func.func_code,
.... dict(globals(), **self.consts))
....
py> @with_consts(y=123)
.... def f(x):
.... return x*y, str
....
py> f(2)
(246, <type 'str'>)

I just update the function globals with the keywords passed to the
decorator. The only problem is that updates to globals() aren't
reflected in the function's globals:

py> str = 5
py> f(2)
(246, <type 'str'>)

Steve

[Nick Coghlan]

Steven Bethard wrote:

I wrote:

> If you really want locals that don't contribute to arguments, I'd be
> much happier with something like a decorator, e.g.[1]:
>
> @with_consts(i=1, deftime=time.ctime())
> def foo(x, y=123, *args, **kw):
> return x*y, kw.get('which_time')=='now' and time.ctime() or deftime
>
> Then you don't have to mix parameter declarations with locals
> definitions.
>
> Steve
>
> [1] I have no idea how implementable such a decorator would be. I'd
> just like to see function constants declared separate from arguments
> since they mean such different things.

I played around with this, and I think it's basically implementable:

Raymond's constant binding decorator is probably a good model for how to do it:
http://aspn.activestate.com/ASPN/Coo.../Recipe/277940

Cheers,
Nick.

--
Nick Coghlan | nc******@email.com | Brisbane, Australia
---------------------------------------------------------------
http://boredomandlaziness.skystorm.net

[Bengt Richter]

On Sun, 23 Jan 2005 13:14:10 -0700, Steven Bethard <st************@gmail.com> wrote:

Bengt Richter wrote:

On Sat, 22 Jan 2005 16:22:33 +1000, Nick Coghlan <nc******@iinet.net.au> wrote:

[Steven Bethard]

> If you really want locals that don't contribute to arguments, I'd be
> much happier with something like a decorator, e.g.[1]:
>
> @with_consts(i=1, deftime=time.ctime())
> def foo(x, y=123, *args, **kw):
> return x*y, kw.get('which_time')=='now' and time.ctime() or deftime
>
> Then you don't have to mix parameter declarations with locals
> definitions. [...]

[Nick Coghlan]Raymond's constant binding decorator is probably a good model for how to do it:
http://aspn.activestate.com/ASPN/Coo.../Recipe/277940

[Bengt Richter]

I thought so too. I modified it to accept **presets as a keyword argument
and generate constants and assignments from its values matching assignment names
when the rhs was __frompresets__, e.g.,

>>> from makeconstpre import make_constants as pre
>>> import time
>>> @pre(verbose=True, deftime=time.ctime(), a=1, b=2, c=3, pi=__import__('math').pi)

... def foo():
... deftime = __frompresets__
... b, a = __frompresets__
... c = __frompresets__
... pi = __frompresets__
... return locals()
...
__frompresets__ : deftime --> Sat Jan 22 20:18:09 2005
__frompresets__ : ('b', 'a') --> (2, 1)
__frompresets__ : c --> 3
__frompresets__ : pi --> 3.14159265359
locals --> <built-in function locals>

[Steven Bethard]Hmm... Having to state
deftime = __frompresets__
when you've already stated
deftime=time.ctime()
seems a little redundant to me...

Yeah, but as a quick hack, it made the byte code twiddling all changing byte codes in-place.
I didn't want to verify that byte code wouldn't need relocation fixups if I inserted something,
(unless of course within a suite with branches or loops) but hopefully it's easily so at the
beginning.

BTW, unlike default arguments, varname = __frompresets__ will work anywhere in the code,
(for the current decorator version) and the generated byte code will have a load of
an appropriate constant (the same one if the same name is being assigned). This would
be tricky to get right if done in a hidden just-in-time manner only in branches that
used the values, but hidden initialization of the whole presets set at the beginning
should be relatively easy if there are no relocation problems.
So, e.g., for

presets = dict(a=1, b=2, deftime=__import__('time').ctime())
in the decorator args, the next version will act as if the decorated
function had the source code
print '%s = __frompresets__' % ', '.join(sorted(presets))

a, b, deftime = __frompresets__

for the current version, except that it will be hidden.

Regards,
Bengt Richter

[Steven Bethard]

Bengt Richter wrote:

So, e.g., for

>>> presets = dict(a=1, b=2, deftime=__import__('time').ctime())
in the decorator args, the next version will act as if the decorated
function had the source code
>>> print '%s = __frompresets__' % ', '.join(sorted(presets))

a, b, deftime = __frompresets__

for the current version, except that it will be hidden.

Cool! Keep us posted. I assume you'll put this into the Cookbook?

Steve

[Bengt Richter]

On Mon, 24 Jan 2005 00:31:17 -0700, Steven Bethard <st************@gmail.com> wrote:

Bengt Richter wrote:

So, e.g., for

>>> presets = dict(a=1, b=2, deftime=__import__('time').ctime())

in the decorator args, the next version will act as if the decorated
function had the source code

>>> print '%s = __frompresets__' % ', '.join(sorted(presets))

a, b, deftime = __frompresets__

for the current version, except that it will be hidden.

Cool! Keep us posted. I assume you'll put this into the Cookbook?

Not ready for prime time, I guess, but I made a module that does presets
and also adjusts parameters and count to do currying without nested function
calling in the curried function.

from presets import presets, curry
@presets(verbose=True, a=1, b=2, deftime=__import__('time').ctime()) ... def foo():
... print (a, b)
... print deftime
...

presets: -- "name(?)" means name may be unused
a = 1
b = 2
deftime = 'Mon Jan 24 12:16:07 2005'
foo() (1, 2)
Mon Jan 24 12:16:07 2005 @curry(verbose=True, y=444) ... def bar(x=3, y=4):
... return x*y
...

presets: -- "name(?)" means name may be unused
y = 444
bar(2) 888 bar(2, 3) Traceback (most recent call last):
File "<stdin>", line 1, in ?
TypeError: bar() takes at most 1 argument (2 given) bar() 1332 import dis
dis.dis(bar) 1 0 LOAD_CONST 1 (444)
3 STORE_FAST 1 (y)

3 6 LOAD_FAST 0 (x)
9 LOAD_FAST 1 (y)
12 BINARY_MULTIPLY
13 RETURN_VALUE dis.dis(foo) 1 0 LOAD_CONST 1 ((1, 2, 'Mon Jan 24 12:16:07 2005'))
3 UNPACK_SEQUENCE 3
6 STORE_FAST 0 (a)
9 STORE_FAST 1 (b)
12 STORE_FAST 2 (deftime)

3 15 LOAD_FAST 0 (a)
18 LOAD_FAST 1 (b)
21 BUILD_TUPLE 2
24 PRINT_ITEM
25 PRINT_NEWLINE

4 26 LOAD_FAST 2 (deftime)
29 PRINT_ITEM
30 PRINT_NEWLINE
31 LOAD_CONST 0 (None)
34 RETURN_VALUE

Now I have to see whether I can optimize the result with Raymond's make_constants
as two decorators in a row... (pre without keyword args is an alias for Raymond's make_constants,
just handy for the moment)
from pre import pre
@pre(verbose=True) ... @presets(verbose=True, a=1, b=2, deftime=__import__('time').ctime())
... def foo():
... print (a, b)
... print deftime
... print __name__
...

presets: -- "name(?)" means name may be unused
a = 1
b = 2
deftime = 'Mon Jan 24 12:29:21 2005'

__name__ --> __main__ foo() (1, 2)
Mon Jan 24 12:29:21 2005
__main__ dis.dis(foo)

1 0 LOAD_CONST 1 ((1, 2, 'Mon Jan 24 12:29:21 2005'))
3 UNPACK_SEQUENCE 3
6 STORE_FAST 0 (a)
9 STORE_FAST 1 (b)
12 STORE_FAST 2 (deftime)

3 15 LOAD_FAST 0 (a)
18 LOAD_FAST 1 (b)
21 BUILD_TUPLE 2
24 PRINT_ITEM
25 PRINT_NEWLINE

4 26 LOAD_FAST 2 (deftime)
29 PRINT_ITEM
30 PRINT_NEWLINE

5 31 LOAD_CONST 2 ('__main__')
34 PRINT_ITEM
35 PRINT_NEWLINE
36 LOAD_CONST 0 (None)
39 RETURN_VALUE

Regards,
Bengt Richter

[Bengt Richter]

On Mon, 24 Jan 2005 20:35:09 GMT, bo**@oz.net (Bengt Richter) wrote:

On Mon, 24 Jan 2005 00:31:17 -0700, Steven Bethard <st************@gmail.com> wrote:

Bengt Richter wrote:

So, e.g., for

>>> presets = dict(a=1, b=2, deftime=__import__('time').ctime())

in the decorator args, the next version will act as if the decorated
function had the source code

>>> print '%s = __frompresets__' % ', '.join(sorted(presets))
a, b, deftime = __frompresets__

for the current version, except that it will be hidden.

Cool! Keep us posted. I assume you'll put this into the Cookbook?

Not ready for prime time, I guess, but I made a module that does presets
and also adjusts parameters and count to do currying without nested function
calling in the curried function.

from presets import presets, curry
@presets(verbose=True, a=1, b=2, deftime=__import__('time').ctime()) ... def foo():
... print (a, b)
... print deftime
...

presets: -- "name(?)" means name may be unused
a = 1
b = 2
deftime = 'Mon Jan 24 12:16:07 2005'
foo() (1, 2)
Mon Jan 24 12:16:07 2005 @curry(verbose=True, y=444) ... def bar(x=3, y=4):
... return x*y
...

presets: -- "name(?)" means name may be unused
y = 444
bar(2) 888 bar(2, 3) Traceback (most recent call last):
File "<stdin>", line 1, in ?
TypeError: bar() takes at most 1 argument (2 given) bar()

1332

Not much reaction, so I guess I won't see it if any, since I am going off line for a while.
BTW, this exercise makes me realize that you could do some other interesting things too,
if you put your mind to it. E.g., it might be interesting to inject properties into
the local namespace of a function, so that e.g., x=1 would would trigger an effect
like someobj.x = 1, where type(someobj).x was a property. It wouldn't be
too hard to make

@addprop(verbose=True, x=property(getx, setx, delx))
def foo():
x = 123
return x

etc. Maybe could be used to monitor access to specific locals for debug,
or other uses. Haven't even begun to think about that.

Another thing is to turn selected local defs into constants, since executing
the aggregation-of-parts code that accomplished the def may well be unchanging.
This would reduce the execution cost of nested functions. I guess that is
a specific case of a more general sticky-evaluate-rhs-once kind of directive
for locals that are only assigned in one place. special sentinels could
be used in the decorator keyword values to indicate this treatment of the
associated name, e.g.
from presets import presets
@presets(x=presets.STICKY, y=123)
def foo(y):
x = initx() # done once on the first call, making x references constant after that
return x*y

Another thing that could be done is to inject pre and post things like atexit stuff.
Also maybe some adaptation things that is just a buzzword so far that I haven't explored.

BTW, is there another op that needs relocation besides JUMP_ABSOLUTE?

I was wondering about doing code mods by messing with an ast, but a decorator would
have to chase down the source and reparse I guess, in order to do that. It would be
interesting to be able to hook into ast stuff with some kind of metacompile hook (hand waving ;-)

Bye for a while.

Regards,
Bengt Richter

[Steven Bethard]

Bengt Richter wrote:

On Mon, 24 Jan 2005 20:35:09 GMT, bo**@oz.net (Bengt Richter) wrote:

>from presets import presets, curry
>@presets(verbose=True, a=1, b=2, deftime=__import__('time').ctime())

... def foo():
... print (a, b)
... print deftime
...

presets: -- "name(?)" means name may be unused
a = 1
b = 2
deftime = 'Mon Jan 24 12:16:07 2005'

>foo()

(1, 2)
Mon Jan 24 12:16:07 2005

Not much reaction

Sorry, it's definitely cool -- I was just waiting for you to post the
Cookbook link so I could see how you did it. ;)

Steve