Why is a list comprehension so much faster than appending to a list?

Question

I was wondering why list comprehension is so much faster than appending to a list. I thought the difference is just expressive, but it's not.

>>> import timeit 
>>> timeit.timeit(stmt='''\
t = []
for i in range(10000):
    t.append(i)''', number=10000)
9.467898777974142

>>> timeit.timeit(stmt='t= [i for i in range(10000)]', number=10000)
4.1138417314859

The list comprehension is 50% faster. Why?

Answer 1

List comprehension is basically just a "syntactic sugar" for the regular for loop. In this case the reason that it performs better is because it doesn't need to load the append attribute of the list and call it as a function at each iteration. In other words and in general , list comprehensions perform faster because suspending and resuming a function's frame, or multiple functions in other cases, is slower than creating a list on demand.

Consider the following examples :

In [1]: def f1(): 
   ...:         l = [] 
   ...:         for i in range(5): 
   ...:             l.append(i) 
   ...:     
   ...:  
   ...: def f2(): 
   ...:     [i for i in range(5)] 
   ...:                                                                                                                                                                                                     

In [3]: import dis                                                                                                                                                                                          

In [4]: dis.dis(f1)                                                                                                                                                                                         
  2           0 BUILD_LIST               0
              2 STORE_FAST               0 (l)

  3           4 LOAD_GLOBAL              0 (range)
              6 LOAD_CONST               1 (5)
              8 CALL_FUNCTION            1
             10 GET_ITER
        >>   12 FOR_ITER                14 (to 28)
             14 STORE_FAST               1 (i)

  4          16 LOAD_FAST                0 (l)
             18 LOAD_METHOD              1 (append)
             20 LOAD_FAST                1 (i)
             22 CALL_METHOD              1
             24 POP_TOP
             26 JUMP_ABSOLUTE           12
        >>   28 LOAD_CONST               0 (None)
             30 RETURN_VALUE

In [5]:                                                                                                                                                                                                     

In [5]: dis.dis(f2)                                                                                                                                                                                         
  8           0 LOAD_CONST               1 (<code object <listcomp> at 0x7f397abc0d40, file "<ipython-input-1-45c11e415ee9>", line 8>)
              2 LOAD_CONST               2 ('f2.<locals>.<listcomp>')
              4 MAKE_FUNCTION            0
              6 LOAD_GLOBAL              0 (range)
              8 LOAD_CONST               3 (5)
             10 CALL_FUNCTION            1
             12 GET_ITER
             14 CALL_FUNCTION            1
             16 POP_TOP
             18 LOAD_CONST               0 (None)
             20 RETURN_VALUE

Disassembly of <code object <listcomp> at 0x7f397abc0d40, file "<ipython-input-1-45c11e415ee9>", line 8>:
  8           0 BUILD_LIST               0
              2 LOAD_FAST                0 (.0)
        >>    4 FOR_ITER                 8 (to 14)
              6 STORE_FAST               1 (i)
              8 LOAD_FAST                1 (i)
             10 LIST_APPEND              2
             12 JUMP_ABSOLUTE            4
        >>   14 RETURN_VALUE

In [6]:   

You can see that on offset 18 in the first function we have an append attribute while there's no such thing in second function using list comprehension. All those extra bytecodes will make the appending approach slower and since in this case you'll have loading of the append attribute in each iteration , in the end it will make the code to take approximately twice as slower as the second function using only list comprehension.

Answer 2

Even factoring out the time it takes to lookup and load the append function, the list comprehension is still faster because the list is created in C, rather than built up one item at a time in Python.

# Slow
timeit.timeit(stmt='''
    for i in range(10000):
        t.append(i)''', setup='t=[]', number=10000)

# Faster
timeit.timeit(stmt='''
    for i in range(10000):
        l(i)''', setup='t=[]; l=t.append', number=10000)

# Faster still
timeit.timeit(stmt='t = [i for i in range(10000)]', number=10000)

Answer 3

引用这篇文章，这是因为list的append属性不是作为函数查找、加载和调用的，这需要时间并且会在迭代中累加。