What causes [*a] to overallocate?

问题

Apparently list(a) doesn't overallocate, [x for x in a] overallocates at some points, and [*a] overallocates all the time?

Here are sizes n from 0 to 12 and the resulting sizes in bytes for the three methods:

0 56 56 56
1 64 88 88
2 72 88 96
3 80 88 104
4 88 88 112
5 96 120 120
6 104 120 128
7 112 120 136
8 120 120 152
9 128 184 184
10 136 184 192
11 144 184 200
12 152 184 208

Computed like this, reproducable at repl.it, using Python 3.8:

from sys import getsizeof

for n in range(13):
    a = [None] * n
    print(n, getsizeof(list(a)),
             getsizeof([x for x in a]),
             getsizeof([*a]))

So: How does this work? How does [*a] overallocate? Actually, what mechanism does it use to create the result list from the given input? Does it use an iterator over a and use something like list.append? Where is the source code?

(Colab with data and code that produced the images.)

Zooming in to smaller n:

Zooming out to larger n:

回答1:

[*a] is internally doing the C equivalent of:

1. Make a new, empty list
2. Call newlist.extend(a)
3. Returns list.

So if you expand your test to:

from sys import getsizeof

for n in range(13):
    a = [None] * n
    l = []
    l.extend(a)
    print(n, getsizeof(list(a)),
             getsizeof([x for x in a]),
             getsizeof([*a]),
             getsizeof(l))

Try it online!

you'll see the results for getsizeof([*a]) and l = []; l.extend(a); getsizeof(l) are the same.

This is usually the right thing to do; when extending you're usually expecting to add more later, and similarly for generalized unpacking, it's assumed that multiple things will be added one after the other. [*a] is not the normal case; Python assumes there are multiple items or iterables being added to the list ([*a, b, c, *d]), so overallocation saves work in the common case.

By contrast, a list constructed from a single, presized iterable (with list()) may not grow or shrink during use, and overallocating is premature until proven otherwise; Python recently fixed a bug that made the constructor overallocate even for inputs with known size.

As for list comprehensions, they're effectively equivalent to repeated appends, so you're seeing the final result of the normal overallocation growth pattern when adding an element at a time.

To be clear, none of this is a language guarantee. It's just how CPython implements it. The Python language spec is generally unconcerned with specific growth patterns in list (aside from guaranteeing amortized O(1) appends and pops from the end). As noted in the comments, the specific implementation changes again in 3.9; while it won't affect [*a], it could affect other cases where what used to be "build a temporary tuple of individual items and then extend with the tuple" now becomes multiple applications of LIST_APPEND, which can change when the overallocation occurs and what numbers go into the calculation.

回答2:

Full picture of what happens, building on the other answers and comments (especially ShadowRanger's answer, which also explains why it's done like that).

Disassembling shows that BUILD_LIST_UNPACK gets used:

>>> import dis
>>> dis.dis('[*a]')
  1           0 LOAD_NAME                0 (a)
              2 BUILD_LIST_UNPACK        1
              4 RETURN_VALUE

That's handled in ceval.c, which builds an empty list and extends it (with a):

        case TARGET(BUILD_LIST_UNPACK): {
            ...
            PyObject *sum = PyList_New(0);
              ...
                none_val = _PyList_Extend((PyListObject *)sum, PEEK(i));

_PyList_Extend uses list_extend:

_PyList_Extend(PyListObject *self, PyObject *iterable)
{
    return list_extend(self, iterable);
}

Which calls list_resize with the sum of the sizes:

list_extend(PyListObject *self, PyObject *iterable)
    ...
        n = PySequence_Fast_GET_SIZE(iterable);
        ...
        m = Py_SIZE(self);
        ...
        if (list_resize(self, m + n) < 0) {

And that overallocates as follows:

list_resize(PyListObject *self, Py_ssize_t newsize)
{
  ...
    new_allocated = (size_t)newsize + (newsize >> 3) + (newsize < 9 ? 3 : 6);

Let's check that. Compute the expected number of spots with the formula above, and compute the expected byte size by multiplying it with 8 (as I'm using 64-bit Python here) and adding an empty list's byte size (i.e., a list object's constant overhead):

from sys import getsizeof
for n in range(13):
    a = [None] * n
    expected_spots = n + (n >> 3) + (3 if n < 9 else 6)
    expected_bytesize = getsizeof([]) + expected_spots * 8
    real_bytesize = getsizeof([*a])
    print(n,
          expected_bytesize,
          real_bytesize,
          real_bytesize == expected_bytesize)

Output:

0 80 56 False
1 88 88 True
2 96 96 True
3 104 104 True
4 112 112 True
5 120 120 True
6 128 128 True
7 136 136 True
8 152 152 True
9 184 184 True
10 192 192 True
11 200 200 True
12 208 208 True

Matches except for n = 0, which list_extend actually shortcuts, so actually that matches, too:

        if (n == 0) {
            ...
            Py_RETURN_NONE;
        }
        ...
        if (list_resize(self, m + n) < 0) {

回答3:

These are going to be implementation details of the CPython interpreter, and so may not be consistent across other interpreters.

That said, you can see where the comprehension and list(a) behaviors come in here:

https://github.com/python/cpython/blob/master/Objects/listobject.c#L36

Specifically for the comprehension:

 * The growth pattern is:  0, 4, 8, 16, 25, 35, 46, 58, 72, 88, ...
...

new_allocated = (size_t)newsize + (newsize >> 3) + (newsize < 9 ? 3 : 6);

Just below those lines, there is list_preallocate_exact which is used when calling list(a).

来源：https://stackoverflow.com/questions/60549865/what-causes-a-to-overallocate