Why is if True slower than if 1?

Question

Why is if True slower than if 1 in Python? Shouldn't if True be faster than if 1?

I was trying to learn the timeit module. Starting with the basics, I tried these:

>>> def test1():
...     if True:
...         return 1
...     else:
...         return 0

>>> print timeit("test1()", setup = "from __main__ import test1")
0.193144083023


>>> def test2():
...     if 1:
...         return 1
...     else:
...         return 0

>>> print timeit("test2()", setup = "from __main__ import test2")
0.162086009979


>>> def test3():
...     if True:
...             return True
...     else:
...             return False

>>> print timeit("test3()", setup = "from __main__ import test3")
0.214574098587

>>> def test4():
...     if 1:
...             return True
...     else:
...             return False

>>> print timeit("test4()", setup = "from __main__ import test4")
0.160849094391

I am confused by these things:

According to the response from Mr. Sylvain Defresne in this question, everything is implicitly converted to a bool first and then checked. So why is if True slower than if 1?
Why is test3 slower than test1 even though only the return values are different?
Like Question 2, but why is test4 a little faster than test2?

NOTE: I ran timeit three times and took the average of the results, then posted the times here along with the code.

This question does not relate to how to do micro benchmarking(which I did in this example but I also understand that it is too basic) but why checking a 'True' variable is slower than a constant.

I think your tests are too small. And the average of three runs isn't enough :p — keyser, Aug 08 '13 at 10:48
I understand :) even the test case is too basic to think. However, we need to start somewhere right :) — thiruvenkadam, Aug 08 '13 at 10:50
possible duplicate of [How long does a microbenchmark need to run?](http://stackoverflow.com/questions/2857470/how-long-does-a-microbenchmark-need-to-run) — , May 20 '15 at 17:28
@JarrodRoberson Thanks for the link :-) However, my question is not related to generic micro bench marks. It is about why does using a keyword to check for boolean is slower than check for a constant. — thiruvenkadam, May 21 '15 at 05:14

Kabie · Accepted Answer · 2013-08-08T10:59:55.070

43

True and False are not keywords in Python 2.

They must resolve at runtime. This has been changed in Python 3

Same test on Python 3:

>>> timeit.timeit('test1()',setup="from __main__ import test1", number=10000000)
2.806439919999889
>>> timeit.timeit('test2()',setup="from __main__ import test2", number=10000000)
2.801301520000038
>>> timeit.timeit('test3()',setup="from __main__ import test3", number=10000000)
2.7952816800000164
>>> timeit.timeit('test4()',setup="from __main__ import test4", number=10000000)
2.7862537199999906

Time error is in 1%, which is acceptable.

edited Aug 08 '13 at 10:59

answered Aug 08 '13 at 10:51

Kabie

10,489
1
38
45

7

try `True,False=False,True` – Kabie Aug 08 '13 at 10:53

score 21 · Answer 2 · answered Aug 08 '13 at 11:01

Bytecode disassembly makes difference obvious.

>>> dis.dis(test1)
  2           0 LOAD_GLOBAL              0 (True)
              3 JUMP_IF_FALSE            5 (to 11)
              6 POP_TOP             

  3           7 LOAD_CONST               1 (1)
             10 RETURN_VALUE        
        >>   11 POP_TOP             

  5          12 LOAD_CONST               2 (0)
             15 RETURN_VALUE        
             16 LOAD_CONST               0 (None)
             19 RETURN_VALUE

As Kabie mentioned, True and False are globals in Python 2. Lots of stuff is going on to access them.

>>> dis.dis(test2)
  3           0 LOAD_CONST               1 (1)
              3 RETURN_VALUE

Python compiler was able to recognize 1 as a constantly "truthy" expression and optimize redundant condition away!

>>> dis.dis(test3)
  2           0 LOAD_GLOBAL              0 (True)
              3 JUMP_IF_FALSE            5 (to 11)
              6 POP_TOP             

  3           7 LOAD_GLOBAL              0 (True)
             10 RETURN_VALUE        
        >>   11 POP_TOP             

  5          12 LOAD_GLOBAL              1 (False)
             15 RETURN_VALUE        
             16 LOAD_CONST               0 (None)
             19 RETURN_VALUE

Pretty much the same as test1, with one more LOAD_GLOBAL.

>>> dis.dis(test4)
  3           0 LOAD_GLOBAL              0 (True)
              3 RETURN_VALUE

See test2. But LOAD_GLOBAL is a bit more costly than LOAD_CONST.

Why is if True slower than if 1?

2 Answers2

Linked