[PEP] auto keyword for automatic objects support in Python

[Manlio Perillo]

Hi.
This post follows "does python have useless destructors".
I'm not an expert, so I hope what I will write is meaningfull and
clear.


Actually in Python there is no possibility to write code that follows
C++ RAII pattern.
Of course Python objects are not statics like in C++, but in C++ the
auto_ptr class is used for enforcing this pattern for dynamical
allocated objects, by using a 'destructive' pointer semantic.


Now, here is a simple example of Python code:

.... afile.write('hello')



Internally (at least in CPython) when the function is called, a frame
object is created, with a dictionary of all local variables.
When the function terminates the frame is 'deleted' (where 'deleted'
means: 'its reference count is decremented').
The frame destructor then calls local variables destructors (again,
local variables references counter are decremented).

For our example this means that when use_file function exits, this
equivalent Python code is execuded:


Of course this not deletes afile, since there exists another reference
to it.


What I'm proposing is to add a new keyword 'auto'. Here its use:
.... afile.write('hello')



Simply, auto objects behaves like C++ auto_ptr but with important
differences and with the need of some support by the objects.

This equivalent code is now executed when the function exits:




With the use of auto there is the need(?) to think at object
destruction as a two fase operation (like object creation with __new__
and __init__).

In fact __del__ method is called when the object is being
deallocated/garbage collected (actually this is not 'deterministic').
The RAII pattern requires simply that as an object goes 'out of scope'
it should releases all the 'external resources' it has acquired.
So there is the need of a second special method (call it __deinit__)
that should be called for auto objects when they go 'out of scope'
(in a 'deterministic' way).


As an example, for a file object __deinit__ code is, of course, the
same as in the __del__ code.


More in detail:
as __init__ creates the class invariant, __deinit__ should be
'destroy' it; that is, it should be put the instance in a 'neutral'
state: all external resource released.

The difference between __deinit__ and __del__ is that after __del__ is
called the object is supposed to be deallocated/garbage collected;
after __deinit__ the object is still 'live' in a 'neutral' state.
So I think it is better to have two distinct methods.



Issues:
Q) Should be Frame objects be auto?
R) They should be auto if there are auto local variables in it.

Q) What about compound objects, ad example:
.... def __init__(self, afile): self.file = afile

R) self.file should be auto if an hold_file instance is auto.




Regards Manlio Perillo

 

[Christos TZOTZIOY Georgiou]

[snip]

With the use of auto there is the need(?) to think at object
destruction as a two fase operation (like object creation with __new__
and __init__).

In fact __del__ method is called when the object is being
deallocated/garbage collected (actually this is not 'deterministic').
The RAII pattern requires simply that as an object goes 'out of scope'
it should releases all the 'external resources' it has acquired.
So there is the need of a second special method (call it __deinit__)
that should be called for auto objects when they go 'out of scope'
(in a 'deterministic' way).

[snip]

I'm afraid your PEP's strongest adversary will be the "Explicit is
better than implicit". You suggest complications to the language
implementation that can be avoided just by user code.

For example, you could have a class Deallocator (untested improvised
code):

class Deallocator:
def __init__(self, *args):
self.args = args
def deallocate(self):
for obj in self.args:
obj.__deinit__()

then in your function start:
auto = Deallocator(obj1, obj2 ...)

and in your function end:
auto.deallocate()

If your function has multiple exit points, wrap its code in a try ...
finally sequence.


These are some of the obvious counter-arguments for your PEP, and
without looking I assume there have been already similar discussions in
the past. Good luck

 

[Manlio Perillo]

[snip]


[snip]

I'm afraid your PEP's strongest adversary will be the "Explicit is
better than implicit". You suggest complications to the language
implementation that can be avoided just by user code.

For example, you could have a class Deallocator (untested improvised
code):

class Deallocator:
def __init__(self, *args):
self.args = args
def deallocate(self):
for obj in self.args:
obj.__deinit__()

then in your function start:
auto = Deallocator(obj1, obj2 ...)

and in your function end:
auto.deallocate()

If your function has multiple exit points, wrap its code in a try ...
finally sequence.

The problem is that one have to use finally consistently.
The C++ RAII pattern is more simple.


Thanks and regards Manlio Perillo