HAX

HAX lets you write compiled bytecode inline with pure Python. It was originally built for exploring new improvements to CPython's compiler and peephole optimizer.

Installation

HAX supports CPython 3.7-3.10 on all platforms.

To install, just run:

$ pip install hax

Example

Consider the following function; it accepts a sequence of items, and returns a list with each item repeated twice:

def doubled(items):
    out = []
    for item in items:
        out += item, item
    return out

For example, doubled((0, 1, 2)) returns [0, 0, 1, 1, 2, 2].

We can make this function faster by keeping out on the stack (instead of in a local variable) and using the LIST_APPEND op to build it. HAX makes it simple to inline these instructions:

from hax import *

@hax
def doubled(items):

    BUILD_LIST(0)

    for item in items:

        LOAD_FAST("item")
        DUP_TOP()
        LIST_APPEND(3)
        LIST_APPEND(2)

    RETURN_VALUE()

With the help of labeled jump targets (HAX_LABEL), the function can be further sped up by rewriting the for-loop in bytecode, removing all temporary variables, and operating entirely on the stack:

from hax import *

@hax
def doubled(items):

    BUILD_LIST(0)

    LOAD_FAST("items")
    GET_ITER()
    HAX_LABEL("loop")
    FOR_ITER("return")

    DUP_TOP()
    LIST_APPEND(3)
    LIST_APPEND(2)
    JUMP_ABSOLUTE("loop")

    HAX_LABEL("return")
    RETURN_VALUE()

It's important to realize that the functions HAX provides (BUILD_LIST, LOAD_FAST, ...) aren't just "emulating" their respective bytecode instructions; the @hax decorator detects them, and completely recompiles doubled's code to use the actual ops that we've specified here!

These performance improvements are impossible to get from CPython's compiler and optimizer alone.

brandtbucher/hax

HAX

Installation

Example