python - Why does dict.get(key) run slower than dict[key]

Question

While running a numerical integrator, I noticed a noticeable difference in speed depending on how I extract the value of the field in a dictionary

import numpy as np

def bad_get(mydict):
    '''Extract the name field using get()'''
    output = mydict.get('name', None)
    return output

def good_get(mydict):
    '''Extract the name field using if-else'''
    if 'name' in mydict:
        output = mydict['name']
    else:
        output = None
    return output


name_dict = dict()
name_dict['name'] = np.zeros((5000,5000))

On my system, I notice the following difference (using iPython)

%%timeit
bad_get(name_dict) 

The slowest run took 7.75 times longer than the fastest. This could mean that an intermediate result is being cached.
1000000 loops, best of 3: 247 ns per loop

Compared to

%%timeit
good_get(name_dict)  

1000000 loops, best of 3: 188 ns per loop

This may seem like a small difference, but in for some arrays the difference appears to be even more dramatic. What causes this behavior, and is there some way I should alter my use of the get() function?

Answer 1

Python has to do more work for dict.get():

• get is an attribute, so Python has to look this up, and then bind the descriptor found to the dictionary instance.
• () is a call, so the current frame has to be pushed on the stack, a call has to be made, then the frame has to be popped again from the stack to continue.

The [...] notation, used with a dict, doesn't require a separate attribute step or frame push and pop.

You can see the difference when you use the Python bytecode disassembler dis:

>>> import dis
>>> dis.dis(compile('d[key]', '', 'eval'))
  1           0 LOAD_NAME                0 (d)
              3 LOAD_NAME                1 (key)
              6 BINARY_SUBSCR
              7 RETURN_VALUE
>>> dis.dis(compile('d.get(key)', '', 'eval'))
  1           0 LOAD_NAME                0 (d)
              3 LOAD_ATTR                1 (get)
              6 LOAD_NAME                2 (key)
              9 CALL_FUNCTION            1
             12 RETURN_VALUE

so the d[key] expression only has to execute a BINARY_SUBSCR opcode, while d.get(key) adds a LOAD_ATTR opcode. CALL_FUNCTION is a lot more expensive than BINARY_SUBSCR on a built-in type (custom types with __getitem__ methods still end up doing a function call).

If the majority of your keys exist in the dictionary, you could use try...except KeyError to handle missing keys:

try:
    return mydict['name']
except KeyError:
    return None

Exception handling is cheap if there are no exceptions.