shayan-taheri / nibbler

Runtime Python bytecode optimizer. ⚡️

---

nibbler is a runtime bytecode optimizer.

It explores the concept of using existing Python syntax features such as type annotations and decorators to speed up code execution by running additional bytecode optimization passes that make use of runtime context provided through these means.

Optimization passes

• inline
  Inlines parameter-less calls to functions that are decorated with @nibbler.inline.
• constantize_globals
  Copies the value of globals that were marked constant (with a Constant type annotation or with the @nibbler.constant decorator) into the co_consts tuple of functions that would normally have to access the global namespace, which speeds up variable access. This also applies to builtins (any, all, print, ...).
• precompute_conditionals
  Strips out conditionals that test constants which the peephole optimizer doesn't pick up on.
• global_to_fast
  Transforms global variable loads to local variable loads if a local variable with the same name exists (mostly a cleanup pass for inline)
• peephole
  Invokes the Python peephole optimizer with additional context.

Usage

from typing import Iterable

from nibbler import Constant, Nibbler

DEBUG: Constant[bool] = False

nibbler = Nibbler(globals())


@nibbler.inline
def square(number: int, base: int) -> int:
    result = number ** base
    return result


@nibbler.nibble
def sequential_square(numbers: Iterable[int]) -> int:
    product = 0
    base = 2
    for number in numbers:
        square()

        if DEBUG:
            print(result)

        product += result

    print(f"Result: {product}")

    return product


sequential_square(range(4))

↓

Result: 14

Examining the function bytecode reveals which optimizations nibbler has performed:

  2           0 LOAD_CONST               1 (0)
              2 STORE_FAST               1 (product)

  3           4 LOAD_CONST               2 (2)
              6 STORE_FAST               2 (base)

  4           8 SETUP_LOOP              28 (to 38)
             10 LOAD_FAST                0 (numbers)
             12 GET_ITER
        >>   14 FOR_ITER                20 (to 36)
             16 STORE_FAST               3 (number)

  5          18 LOAD_FAST                3 (number)
             20 LOAD_FAST                2 (base)
             22 BINARY_POWER
             24 STORE_FAST               4 (result)

  6          26 LOAD_FAST                1 (product)
             28 LOAD_FAST                4 (result)
             30 INPLACE_ADD
             32 STORE_FAST               1 (product)
             34 JUMP_ABSOLUTE           14
        >>   36 POP_BLOCK

  8     >>   38 LOAD_CONST               5 (<built-in function print>)
             40 LOAD_CONST               3 ('Result: ')
             42 LOAD_FAST                1 (product)
             44 FORMAT_VALUE             0
             46 BUILD_STRING             2
             48 CALL_FUNCTION            1
             50 POP_TOP

  9          52 LOAD_FAST                1 (product)
             54 RETURN_VALUE

• The square function was inlined (inline) (18-24)
• Conditional (if DEBUG) was stripped out, because DEBUG was declared a constant (precompute_conditionals) (26)
• The print function was promoted to a function-level constant (constantize_globals) (38)

Installation

pip3 install nibbler

FAQ

• Is this production ready?
  Hell no.
• Why is it called nibbler?
  ¯\_(ツ)_/¯

Prior Art

• fatoptimizer
• PEP 510
• PEP 511
• falcon