User-defined augmented assignment

[Pierre Barbier de Reuille]

Hello,

a discussion began on python-dev about this. It began by a bug report,
but is shifted and it now belongs to this discussion group.

The problem I find with augmented assignment is it's too complex, it's
badly explained, it's error-prone. And most of all, I don't see any
use-case for it !

The most common error is to consider that :

a += b <==> a.__iadd__(b)

when the truth is :

a += b <==> a = a.__iadd__(b)

which can be very confusing, as the two "a" are not necessarily the
same. It then leads to subtle errors like:
2

Also, the following behavior is pretty confusing :

a = [1]
b = [a]
c = (a,)
b[0] += [2] # Ok, no pb
print a [1,2]
c[0] += [3]

Traceback (most recent call last):
[1,2,3]

Then, in the standard library, there is no use-case of user-defined
augmented assignment I could find. Of course, I find the augmented
assignement itself very useful ! I use it a lot with immutable objects
(strings, numbers, tuples, ...) but I tend to avoid it with mutables,
and so it seems in the standard library that uses extensively the
"extend" method of lists and very seldom the "+=" operator with lists.
And even where the "a+=b" is used, it could be replaced with either
"a.extend(b)" or "a = a+b" without bugs.

So, what I would suggest is to drop the user-defined augmented
assignment and to ensure this equivalence :

a X= b <=> a = a X b

with 'X' begin one of the operators.

Pierre

 

[Reinhold Birkenfeld]

So, what I would suggest is to drop the user-defined augmented
assignment and to ensure this equivalence :

a X= b <=> a = a X b

with 'X' begin one of the operators.

It can be done, but it's unnecessary for mutable objects like
sets or lists. A new object must be created in these cases where
one would suffice.

Reinhold

 

[Pierre Barbier de Reuille]

Reinhold Birkenfeld a écrit :

Pierre Barbier de Reuille wrote:




It can be done, but it's unnecessary for mutable objects like
sets or lists. A new object must be created in these cases where
one would suffice.

Well, my point is: the benefit is too small compared to the
disadvantage. If you really have a mutable (let say a list with +=) then
you do:

and there is no interpretation error possible. BTW, that's what's done
in the standard library ...

Pierre

 

[Paddy]

I thought along these lines:
It is an augmented ASSIGNMENT. (It even has an equals sign in it).
tuples are immutable so you should not be able to assign to one of
its elements.

- So there is no problem for me - I shouldn't be messing with an
element of an
immutable type!

- Cheers, Paddy.

 

[Tom Anderson]

a discussion began on python-dev about this. It began by a bug report,
but is shifted and it now belongs to this discussion group.

The problem I find with augmented assignment is it's too complex, it's
badly explained, it's error-prone. And most of all, I don't see any
use-case for it !

The most common error is to consider that :

a += b <==> a.__iadd__(b)

when the truth is :

a += b <==> a = a.__iadd__(b)

which can be very confusing, as the two "a" are not necessarily the
same.

Indeed. I certainly didn't realise that was how it worked.

So, what I would suggest is to drop the user-defined augmented
assignment and to ensure this equivalence :

a X= b <=> a = a X b

with 'X' begin one of the operators.

That seems quite an odd move. Your proposal would lead to even more
surprising behaviour; consider this:

a = [1, 2, 3]
b = a
a += [4, 5, 6]
print b

At present, this prints [1, 2, 3, 4, 5, 6]; if we were to follow your
suggestion, it would be [1, 2, 3].

So, -1, i'm afraid.

I think the right solution here is staring us in the face: if everyone
seems to think that:

a += b <==> a.__iadd__(b)

Then why not make it so that:

a += b <==> a.__iadd__(b)

Principle of Least Surprise and all that.

Since not everything that should support += is mutable (integers, for
example), how about saying that if the recipient of a += doesn't have an
__iadd__ method, execution falls back to:

a = a + b

I say 'a + b', because that means we invoke __add__ and __radd__
appropriately; indeed, the __add__ vs __radd__ thing is a precedent for
this sort of fallback.

Doesn't that leave everyone happy?

tom