This is a patch to Tom Lee's AST optimization branch.

Python bytecode includes at least one operation which is not directly
accessible from Python code: LIST_APPEND, which pops a value and a list
off the stack and appends the value to the list. This is used in
generating code for list comprehensions:

[x for x in somelist]

generates the following code for the body of the loop:

...
LOAD_FAST 1#a local is generated to hold the new list
LOAD_FAST 2 #the iteration variable, x
LIST_APPEND
...

Whereas if you were to try to write the comprehension directly, you
would get:

LOAD_FAST 1
LOAD_ATTR 0 #an index into the constant table: “append”
LOAD_FAST 2
CALL_FUNCTION 1 #remove x and the append function from the top of the
stack and push the result of the call
POP_TOP

This is much slower. In part, it’s the cost of doing the attribute
lookup each time, which is why it’s common to see code like a = [] ap =
a.append for x in …..: ap(x + …) return a But the function call is
naturally slower than the simpler LIST_APPEND operation. The attached
patch tries to determine statically if a local is all circumstances
holds a list, and if so, generates LIST_APPEND whenever user code would
call local.append.  It’s not perfect — in particular, I could track
local types on a more fine-grained level than per-function. But it’s a
start.