In more sane languages, the expression [] means give me an empty list value, nothing more and nothing less. So then whatever the syntax for [[]]*4 is in such sane languages, it gives you four empty list values. If you want something more imperative, you can get it, but it most cases, you don't want or even need such things.
Python makes no distinction, at a level able to be visualized by the user, between immutable and mutable data. There are a lot of bizarre situations in Python that will bite you like this. And in most cases, it's strange that such behaviors were chosen as the deault.
I don't buy this. Either the syntax to give you four empty list values is some kind of comprehension, which Python has and does exactly what you like, or it must have lazy evaluation somehow. Otherwise, how can the language know you don't want to copy the existing reference but create a new empty list? If [[]]*4 were to create new list on the fly, then `[print('hello')]*4` would have to print hello 4 times, i.e., the print statement is lazily executed, which does not fit into Python at all.
Overall it's a very useful feature and I don't think it's that surprising. It's pretty obvious [x]4 will reuse the same reference. Just like manually adding "x" four times will.
I don't think that makes the complexity tradeoff worth it. This is the kind of tradeoff that makes the language implementation, reference and semantics more and more complex. Unfortunately Python has a lot of such complexities already. We don't need more of it. Such complexity hurts the creation of completing implementations.
IMHO programming language specifications and implementations should be simple and consistent so that competing implementations can be developed easily.
> people often using it subtly wrong?
I think that... you know... before writing serious software, these people need to roll up their sleeves and just learn the damn language. Just learn what references are and how they work in Python. It's not that hard. It is most definitely easier than adding a sphagetti monster to the compiler make it bend to the will of programmers who can't be bothered to read the tutorial. I mean this gotcha about references is taught in the 4th chapter of the tutorial. Can't developers even bother to RTFM these days before developing serious software with a programming language?
You really have no choice but to do that. But the critique here is that some languages make this hard. And some languages, like Python, appear deceptively simple and consistent, when they are anything but. And as I pointed out, these decisions were not really required or designed to solve certain problems. They just kind of came about in Python's development, one historically and intentionally ignorant of pre-existing languages and their good ideas.
When everybody just says "learn the language" or "there's list comprehensions" or "there's for loops", then why does the [...]*n syntax exist? What problems is it solving that require the confusion that it generates?
I think I answered that already. It keeps the language spec consistent and simpler.
Imagine the complexity you have to add to the language spec to say that when we write [] we deal with the reference to this list except in the multiplication syntax a = [[]] * 5 where the inner [] is not a reference to the list but the list value! Such special case will make the language both inconsistent and harder to understand for experienced programmers.
I prefer simplicity and consistency in a programming language grammar, syntax and semantics as I advocated above. So yes I'd be happy to lose that multiplication syntax. It is not worth it.
Research Rank Polymorphism and Array Languages:
I certainly don't want to tell you how to design your language.
I'm just trying to say, don't rule the pattern out because it's perceived as a type error; you may have other reasons that are good, but my citations provide a basis of how to make the typing work (if you were inclined).
If you have other reasons, even if it's just aesthetic, that's fine.
I think the motivation was being able to do scalar multiplication with a vector, but it gets ugly when operating on anything but numbers imo
> In more sane languages, the expression [] means give me an empty list value, nothing more and nothing less
What is a “list value”? It’s either a pointer to some other space where the lists values are stored, or it’s the values stored themselves. If you say it’s the second (I can’t think of any language that does), then `a = […]; b = a` is an expensive copy by value assignment. Is that worth that trade off to get less surprising results in the articles case? I’m not saying the answer is “obviously no”, but it definitely isn’t “obviously yes”.
The interpreter is slightly more clever and instead implements this as a copy-on-writr mechanism, but that's just an implementation detail .
In Swift collections are value types, so b = a conceptually is a copy. It’s not expensive, though, because Swift does copy-on-write.
You pay for that by a combination of having more expensive collection updates (they, at least conceptually, have to check whether to make that copy) and a more complex compiler (it tries to avoid having to actually do those checks at runtime)
var a: [Int]=[]
var b = [a,a,a,a]
b[0].append(1)
print(b)
[[1],[],[],[]]Actually, if you really want to have re-evaluting (sub)expressions, then Python does give you an option: comprehensions.
List.init 4 (fun _ -> [])
[for _ in 1..4 -> []]
> Actually, if you really want to have re-evaluting (sub)expressions, then Python does give you an option: comprehensions.
That's the complaint, that [[]]*4 is not shorthand for list comprehensions and that it's something entirely different. This is common in Python where two things you'd expect to be the same are indeed different in subtle ways.
List comprehensions in Python are closer to sane, but they still have unexpected behavior because of the way scoping works in Python.
There are lots of reasons you would want items in a list repeated rather than having, essentially, the result of separate executions of the same constructor -- [] is just shorthand for list() -- so its good that [x()]4 is not shorthand for [x() for _ in range(4)]. Especially since that would mean it would have to be special-case sytax for list construction, and not a list operator that you could use on existing lists. The mildly annyoing edge case you need to learn about isn’t worth wiping out the whole broadly-useful language feature, or creating a great special-case inconsistency which would be more cognitive load to deal with, to eliminate.
[list() for _ in range(4)]I'm not. The fun _ -> [] is needed because the argument is the element's index. That's just what List.init happens to expect, a function that takes an index and initializes the element, but the index is not needed here.
Ah I see, you’re saying that the construction of a list in F# mandates a lambda, thereby reducing one possible point of confusion. I could agree with that. But then it comes at a cost of comfort: [0]*4 has less boilerplate than the F# version. So it comes down to trade-offs again.
Well, of course. It's a shorthand for duplicating the list elements four times. In this case, the list contains a single reference to an empty list, do the result contains four references to an empty list.
It would be bad for this case, though. As confusing as the particular case of:
[[]] * 4
Besides, that still doesn't help with the case of
x = function_that_returns_a_list()
x = x * 4
I think the small example is misguiding the intuition here.
If I write `my_list = f(expensive operation); long_list = my_list * 4`, I don't expect python to recreate the whole thing four times. I justly assume `f` will be evaluated once, and then copied.
The first code is equivalent to:
list_multiply([x], 4)
(the function is actually called list.__mul__, but that's an unimportant detail). The latter is a syntax sugar for:
out = [] for _ in range(4): out.append(x)
(actually a bit more complex, but close enough).
When you understand it, the behavior is obvious in both cases.
There’s a separate type called tuple specifically designed to represent lists-as-values.
If you do [()]*4 then you get exactly what you want.
tuples are also references, but to an immutable object.
> If you do [()]*4 then you get exactly what you want.
I think what is actually wanted here is four different mutable lists rather than the same mutable list four times, something like what you would get from:
[[] for _ in range(4)]I can't think of anything that's a value? At least since Java, the idea of "everything is an object" seems pretty dominant in most languages.
There are a few immutable types, like ints, strings and tuples, but you still refer to them by reference.
It is always clear, and it is 100% always a reference, with a set of cases where that reference is to a singleton or immutable object, and a subset of those cases where, in particular Python implementations, that singleton may effectively be implemented as a special case value.
If you treat it as a reference, you will never, ever be wrong.
The critique here is that it's a badly designed language in that it's design decisions often get in the way of writing software that works as you expect. We're not discussing about whether you can learn such a system. The discussion is a lament on why it had to be this way.
Maybe. But that's largely irrelevant since we are discussing the impact of a syntax choice regarding an operator on an explicitly-mutable container type, and specifically the behavior of it used when the container also contains a member of the same type.
And the alleged surprising thing is not the mutability of anything, but thar after the operation, all the members of the other container are the same instance of the inner mutable container type.
Python being designed ground up with immutability by default and explicitly opt-in mutability would not directly have any bearing on this issue, as long as it still had the capacity to express immutable lists and the capacity to apply something like the list-reptition operation to them. Except it would probably make all of the involved examples more verbose.
> In more sane languages, the expression [] means give me an empty list value, nothing more and nothing less
What is an “empty list value”?
a = []
b = [a]*4
a.append(1)
print(b)
Lists pass by reference, tuples pass by value, if you want to understand Python you're just going to have to get used to that distinction. There are other options, but most are worse.
My argument is that that distinction isn't needed, by default, in a high-level language, and that plenty of people get it wrong, understandably. It is simply complexity that is not needed.
These simple examples can be toyed with, but sooner or later one is going to want to solve real problems, and unintuitive, default behavior doesn't help with that. Python's mode of operation tends to favor implicit over explicit. And the syntax [...]*n is implicit.
If you want mutable values in your language, you need to expect values to mutate. So what should list multiplication do?
- There is the current implementation, where a change to the first entry of [f(x)] * 4 will be reflected in all entries of the list (because they're all references to the same value). - It could only work on list literals. So [f(x)] * 4 would call f four times and collect the values in a list. Then, a = [f(x)]; a * 4 would not work, which would be surprising. - It could work differently on list literals and lists. So [f(x)] * 4 would call f four times collecting the results, but a = [f(x)]; a * 4 would just return a list of four references to the one result of f(x). - It could copy the value in the list. How would you do that if the list contained values like file handlers or network connections that can't be copied?
Within the context of Python (an imperative language with mutable values), the current implementation is the most sensible one. If you work in such a language, you must always exercise some care when working with mutable values. Learn your tool!
Python could have been designed with different decisions. But who knows if we would even be talking about it today in that case?
bitmc can correct me if I'm misinterpreting them, but I feel like what they're saying is that multiplying a list by a number doesn't make sense, so it should throw a type error. That feels reasonable to me (and it's the choice I'm making in the interpreter I'm writing).
Everything in Python is an object with an id. Assignment and parameter passing is by object id.
There are some mutable objects ("objects", lists, dicts) and some immutable objects (numbers, strings, tuples).
The []*n syntax is treated as assignment, hence the id is duplicated. This isn't an issue with immutable objects since there's no way to change the underlying object - you can only change the id of each object in the list. Obviously there are some gotchas like with mutable objects here where the succinct syntax can be confusing. But I think the simplicity of everything being an id has deep intuitive value and is one reason Python has been so successful, even if many users never really think much about ids.
Is that true? I haven't looked at the implementation so I can't say it's impossible, but it would be extremely surprising to me if tuples passed by value, because tuples longer than the size operated upon my individual processor instructions would become expensive to pass around.
From a correctness perspective the difference between pass-by-reference and pass-by-value (pass-by-copy) aren't usually important, but for performance it can be a very large difference.
If you compare their ids it does look like python makes multiple instances and passes them around by reference, but without a way to modify them in place it's hard to tell the difference, also Python tries to avoid making unnecessary copies of the same thing if possible.
You can try to test it a bit with something like the following:
t = (1,2)
def f(x):
return x is t
f(t), f((1,2)), (1,2) is (1,2) # -> True, False, True
Ahh, you were talking about the semantics, not the implementation. Understood.
Another way to think of this might be in terms of mutability, i.e. tuples are immutable. Pass-by-value versus pass-by-reference for immutable objects doesn't make much difference in the behavior of objects in a language like Python which doesn't have explicit pointers, but it's a big difference for performance.
> If you compare their ids it does look like python makes multiple instances and passes them around by reference
I didn't think of answering my own question this way. Nice!
> but the results might be implementation dependent.
It's possible since Python isn't standardized, but I strongly doubt you'll find a mature implementation that passes potentially-large, variable-sized objects like tuples around by value, for two reasons:
1. Once you get above the size operated on by processor instructions (64 bits on most modern desktop/laptop processors) pass by value requires a memcopy, which is orders of magnitude slower. Tuples can grow past the L1 or L2 cache size, which corresponds to another two orders of magnitude in performnance loss. Comparing this to passing around a 64-bit pointer which is about as fast as passing around any other value.
2. Virtual machines usually implement their own stack. This is easiest to do as an array of same-sized values. Implementing a performant stack with variable-sized elements is a nightmare I wouldn't wish on my worst enemy.
That's not what you're asking for, though. You're asking for one bucket, and then you're putting it in another bucket, and then you're duplicating the contents of the second bucket four times.
This isn't whitewashing, this is just what those expressions mean.
It's not like it's hard to get four buckets in Python, BTW:
[[] for _ in range(4)]I don't get how this is an example of python mixing up mutable/immutable data, and I'm unaware of a situation where python does. Conceptually I'm not even sure what you mean by a language 'mixing up mutable and immutable data' unless it's literally just a buggy/non hardened implementation.
A = (1, ); A[0] = 2 # this raises an exception
> a = [0]
and
> b = [0]
these two expressions produces different 'value' objects for `[0]`, and modifying `a.append(a[-1] + 1)` does not change b.
so it's unintuitive that using `[0]` as a default value in a function argument can cause that argument to retain the same value across multiple function calls, when programmers would assume the argument to a function is not the same across multiple invocations.
The issue really here is the usage of an array object, which is not a value object (immutable), but a reference object. It would not have been a problem if the default argument was a string, or a number.
This is a reason not to use arrays in default arguments.
On the other hand, if we abstract from the technical minutiae, yes, it would have been entirely possible for Python to treat default arguments in the function definitions as expressions, by using machinery similar to what gets used for comprehensions: after all, as you point out, function's arguments (and local variables) do get re-assigned on every function call.
No, I didn't know that, and I'd guess that the vast majority of Python developers don't, because people don't generally learn Python by reading the docs front to back and finding out the official names of everything.
> It's a value that's reused for every function call, not an expression that re-evaluated at every function call.
No one is confused about that. I'm aware that's how it is, but is that how it should be? I.e. if we could choose for this to be reused for every function call or re-evaluated every function call, which results in more intuitive behavior?
They are part of function definition, they aren't in the function body. You should expect them to be "executed" when the function is defined.
Sure, there are a lot of subjective aesthetics that go into the spec, but in this case, there are objective reasons for not liking this. It's a well-known footgun that causes bugs. And it's almost never what you want, so you end up doing something like this:
def f(xs = None):
# Are these two lines actually faster than the
# interpreter creating defaults at call time?
if xs is None:
xs = []
...
# Are these two lines actually faster than the
# interpreter creating defaults at call time?
No, that's not what I'm proposing. Why would it check anything? Just evaluate the given default expression at call time. If you don't want the overhead of an expression, don't put a default.
You can also do defaults like:
LIST_OF_X = []
def foo(xs = LIST_OF_X):
...
The code I posted is showing how you can explicitly get the mutable behavior if you want it when the default expressions are evaluated at call time.
That is to say, if you evaluate default expressions at call time, you can get either behavior by being explicit with minimal loss in performance:
def foo(xs = []):
xs.append(1)
print(xs) # always prints [1]
DEFAULT = []
def bar(xs = DEFAULT):
xs.append(1)
print(xs) # prints [1], [1,1], [1,1,1], etc.
You're clearly not as knowledgeable as you think you are on this and you're just cherry picking things you don't understand to feel smart.
Alternatively, the issue is that Python does not distinguish between value and reference types.
In this case, the default arguments get compiled as part of the function code and are part of the function's environment. As a result every time that function is called it is always having the same environment, i.e. the default argument values are not reinterpreted on every invocation. So the mutations to the environment are visible across function calls.
Yeah, I understand that; I'm saying that was a choice they made, and I think having the defaults be evaluated at each call time would have been a better choice.
The more unintuitive example of this to me is in function defaults, because those feel more like part of the function signature--i.e. I'd expect them to be assigned at function call time. But apparently, function defaults get called once when the function is defined.