Python's += Is Weird

Python

Here's a Python gotcha I've hit often enough to merit a blog post: x += 1 is weird in Python. It's compiled roughly like x = x + 1, with two surprising consequences. One is this familiar pitfall:

>>> x = 0
>>> def f():
...     x += 1
... 
>>> f()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 2, in f
UnboundLocalError: local variable 'x' referenced before assignment

The compiler thinks of x += 1 similarly to x = x + 1, so it considers x to be a local variable bound in the scope of f. But x is referenced before it's assigned to. Let's look at the bytecode:

>>> dis.dis(f)
  2           0 LOAD_FAST                0 (x)
              3 LOAD_CONST               1 (1)
              6 INPLACE_ADD         
              7 STORE_FAST               0 (x)
             10 LOAD_CONST               0 (None)
             13 RETURN_VALUE

The first opcode, LOAD_FAST, fails to load x because it's not in scope. Obviously, we need to declare global x:

>>> def f():
...     global x
...     x += 1
... 
>>> dis.dis(f)
  3           0 LOAD_GLOBAL              0 (x)
              3 LOAD_CONST               1 (1)
              6 INPLACE_ADD         
              7 STORE_GLOBAL             0 (x)
             10 LOAD_CONST               0 (None)
             13 RETURN_VALUE

Now LOAD_FAST is replaced with LOAD_GLOBAL, which correctly locates x.

The second pitfall of += is lost updates. If we run f ten thousand times in parallel, sometimes x is incremented less than ten thousand times:

>>> def go():
...     global x
...     x = 0
...
...     def f():
...         global x
...         x += 1
...
...     ts = [threading.Thread(target=f)
...         for _ in range(10000)]
...
...     for t in ts:
...         t.start()
...
...     for t in ts:
...         t.join()
...
...     print x
... 
>>> go()
10000
>>> go()
10000
>>> go()
9998

Again, the problem is clear if we look at the bytecode. f is compiled as a series of opcodes that load the global integer referenced by x onto the stack, add 1 to it, and store the new integer back into x:

>>> dis.dis(f)
  3           0 LOAD_GLOBAL              0 (x)
              3 LOAD_CONST               1 (1)
              6 INPLACE_ADD         
              7 STORE_GLOBAL             0 (x)
             10 LOAD_CONST               0 (None)
             13 RETURN_VALUE

The interpreter can switch threads anywhere between LOAD_FAST, which loads the global value of x onto this thread's stack frame, and STORE_FAST, which saves it back to the global x.

Say x is 17 and two threads execute f. Thread A loads the integer 17 onto its stack, adds one to it, and gets interrupted. Now Thread B also loads 17 onto its stack and adds one. No matter the order the threads now complete, the final value of x will be 18, although we expect 19.

The solution is to protect += statements with a Lock.