Why is an inline if/else faster than a .get() in Python?

Question

I was reviewing some code earlier and the developer wrote an inline if/else rather than a get() to retrieve an element from a list if it exists (otherwise give it a default value). I decided to spring some timeit code on repl and was pretty confused by the result. The if/else takes 1/3 the time of the get().

Here is the repl code, and below is the code in the repl as well as the result for posterity:

import timeit

D = {"a": 1, "b": 2, "c": 3}

def ef(): return D['a'] if 'a' in D else 1

def gt(): return D.get('a', 1)

print "gt1", timeit.timeit(gt, number=10000)
print "ef1", timeit.timeit(ef, number=10000)
print "ef2", timeit.timeit(ef, number=10000)
print "gt2", timeit.timeit(gt, number=10000)

and the results:

gt1 0.0659999847412
ef1 0.0239999294281
ef2 0.0249998569489
gt2 0.0539999008179

and a visual of 10 iterations of the above timeit calls, where the result has been multiplied by 10000 for representation purposes

Answer 1

The D.get() path includes an attribute lookup, and a method call:

>>> import dis
>>> D = {"a": 1, "b": 2, "c": 3}
>>> def gt(): return D.get('a', 1)
... 
>>> dis.dis(gt)
  1           0 LOAD_GLOBAL              0 (D)
              3 LOAD_ATTR                1 (get)
              6 LOAD_CONST               1 ('a')
              9 LOAD_CONST               2 (1)
             12 CALL_FUNCTION            2
             15 RETURN_VALUE        

The attribute lookup (LOAD_ATTR) especially slows things down.

If you remove the attribute lookup (and give the in test a local to work with), the field is evened out:

>>> def gt_fast(D_get=D.get): return D_get('a', 1)
... 
>>> def ef_fast(D=D): return D['a'] if 'a' in D else 1
... 
>>> timeit.timeit(gt_fast)
0.2174091339111328
>>> timeit.timeit(ef_fast)
0.2139298915863037