How come unpacking is faster than accessing by index?

Question

I’m referring to this question, and especially the comments to the first answer from @David Robinson and @mgilson:
Sum the second value of each tuple in a list

The original question was to sum the second value of each tuble:

structure = [('a', 1), ('b', 3), ('c', 2)]

First Answer:

sum(n for _, n in structure)

Second Answer:

sum(x[1] for x in structure)

According to discussion, the first answer is 50% faster.

Once I figured out what the first answer does (coming from Perl, I Googled for the special _ variable means in python), I got wondering how come what appears as a pure subset task (getting only the second element of each tuple vs. getting and binding into variables both elements) is actually slower? Is it a missing opportunity to optimize index access in Python? Am I missing something the second answer does which takes time?

Asked By: Uri

||

Source

Answer 1

Indexing goes through the __getitem__ special method, which thus has to do function lookup and execution for each item. That means that for a list of n items, you wind up doing n lookups/calls.

Unpacking doesn’t have to deal with that when working with native lists/tuples; it just goes through __iter__ which is a single call, and then unpacks the resulting sequence in C.

Answered By: Amber

Answer 2

If you take a look at the python bytecode, it becomes quite obvious very quickly why unpacking is faster:

>>> import dis
>>> def unpack_or_index(t=(0, 1)):
...     _, x = t
...     x = t[1]
... 
>>> dis.dis(unpack_or_index)
  2           0 LOAD_FAST                0 (t)
              3 UNPACK_SEQUENCE          2
              6 STORE_FAST               1 (_)
              9 STORE_FAST               2 (x)

  3          12 LOAD_FAST                0 (t)
             15 LOAD_CONST               1 (1)
             18 BINARY_SUBSCR       
             19 STORE_FAST               2 (x)
             22 LOAD_CONST               0 (None)
             25 RETURN_VALUE

The tuple unpacking operation is a simple bytecode (UNPACK_SEQUENCE), while the indexing operation has to call a method on the tuple (BINARY_SUBSCR). The unpack operation can take place, inline, in the python evaluation loop, while the subscription call requires looking up of the function on the tuple object to retrieve the value, using PyObject_GetItem.

The UNPACK_SEQUENCE opcode source code special-cases a python tuple or list unpack where the the sequence length matches the argument length exactly:

        if (PyTuple_CheckExact(v) &&
            PyTuple_GET_SIZE(v) == oparg) {
            PyObject **items = 
                ((PyTupleObject *)v)->ob_item;
            while (oparg--) {
                w = items[oparg];
                Py_INCREF(w);
                PUSH(w);
            }
            Py_DECREF(v);
            continue;
        } // followed by an "else if" statement for a list with similar code

The above code reaches into the native structure of the tuple and retrieves the values directly; no need to use heavy calls such as PyObject_GetItem which have to take into account that the object could be a custom python class.

The BINARY_SUBSCR opcode is only optimized for python lists; anything that isn’t a native python list requires a PyObject_GetItem call.

Answered By: Martijn Pieters

How come unpacking is faster than accessing by index?

Question:

Answers: