Enables increment operators in Python with a bytecode hack

By default, Python supports neither pre-increments (like ++x) nor post-increments (like x++). However, the first ones are syntactically correct since Python parses them as two subsequent +x operations, where + is the unary plus operator (same with --x and the unary minus). They both have no effect, since in practice -(-x) == +(+x) == x.

This module turns the ++x-like expressions into x += 1 at the bytecode level. Increments and decrements of collection items and object attributes are supported as well, for example:

some_dict = {'key': 42}
++some_dict['key']
assert some_dict['key'] == 43

Unlike the x += 1 statement, ++x remains to be an expression, so it works fine inside other expressions, if/while conditions, lambda functions, and list/dict comprehensions:

array[++index] = new_value

if --connection.num_users == 0:
    connection.close()

button.add_click_callback(lambda: ++counter)
# No need for the `global counter` statement inside lambda

index = 0
indexed_cells = {++index: cell for row in table for cell in row}

See tests for more sophisticated examples.

[How it works] [Installation]

I don't claim that allowing increments is good for real projects: such code may become less readable, confuse new developers, and behave differently if copied to environments without this module. I've made this module for fun, as a demonstration of Python flexibility and bytecode manipulation techniques.

However, some situations where increments simplify code do exist (see examples from the Python's standard library). Also, having the increment expressions seems consistent with PEP 572 "Assignment Expressions" that introduced the x := value expressions in Python 3.8+.

Python compiles all source code to a low-level bytecode executed on the Python's stack-based virtual machine. Each bytecode instruction consumes a few items from the stack, does something with them, and pushes the results back to the stack.

The ++x expressions are compiled into two consecutive UNARY_POSITIVE instructions that do not save the intermediate result in between (same with --x and two UNARY_NEGATIVE instructions). No other expressions produce a similar bytecode pattern.

plusplus replaces these patterns with the bytecode for x += 1, then adds the bytecode for storing the resulting value to the place where the initial value was taken.

This is what happens for the y = ++x line:

A similar but more complex transformation happens for the code with subscription expressions like value = ++dictionary['key']. We need the instructions from the yellow boxes to save the initial location and recall it when the increment is done (see the explanation below):

This bytecode is similar to what the string dictionary['key'] += 1 compiles to. The only difference is that it keeps an extra copy of the incremented value, so we can return it from the expression and assign it to the value variable.

Arguably, the least clear part here is the second yellow box. Actually, it is only needed to reorder the top 4 items of the stack. If we need to reorder the top 2 or 3 items of the stack, we can just use the ROT_TWO and ROT_THREE instructions (they do a circular shift of the specified number of items of the stack). If we had a ROT_FOUR instruction, we would be able to just replace the second yellow box with two ROT_FOURs to achieve the desired order.

However, ROT_FOUR was removed in Python 3.2 (since it was rarely used by the compiler) and recovered back only in Python 3.8. If we want to support Python 3.3 - 3.7, we need to use a workaround, e.g. the BUILD_TUPLE and UNPACK_SEQUENCE instructions. The first one replaces the top N items of the stack with a tuple made of these N items. The second unpacks the tuple putting the values on the stack right-to-left, i.e. in reverse order. We use them to reverse the top 4 items, then swap the top two to achieve the desired order.

[Source code]

The first way to enable the increments is to use a decorator that would patch the bytecode of a given function.

The decorator disassembles the bytecode, patches the patterns described above, and recursively calls itself for any nested bytecode objects (this way, the nested function and class definitions are also patched).

The bytecode is disassembled and assembled back using the MatthieuDartiailh/bytecode library.

[Source code]

The Python import system allows loading modules not only from files but from any reasonable place (e.g. there was a module that enables importing code from Stack Overflow answers). The only thing you need is to provide module contents, including its bytecode.

We can leverage this to implement a wrapping loader that imports the module as usual but patching its bytecode as described above. To do this, we can create a new MetaPathFinder and install it to sys.meta_path.

[Source code]

• This way, it would be impossible to distinguish applying two unary operators consequently (like ++x) from applying them in separate places of a program (like in the snippet below). It is important to not change behavior in the latter case.

  x = -value
  y = -x

• Overriding operators via magic methods (such as __pos__() and __neg__()) do not work for built-in Python types like int, float, etc. unless you use other hacks like in forbiddenfruit or dontasq. Using more hacks complicates porting this module to other Python versions and interpreters.

• You would need to override these methods for each built-in/numpy/user-defined number type. In contrast, plusplus works for all types automatically.

• pytest does its own bytecode modifications in tests, adding the code to save intermediate expression results to the assert statements. This is necessary to show these results if the test fails (see pytest docs).

  By default, this breaks the plusplus patcher because the two UNARY_POSITIVE instructions become separated by the code saving the result of the first UNARY_POSITIVE.

  We fix that by removing the code saving some of the intermediate results, which does not break the pytest introspection.

  [Source code]

You can install this module with pip:

pip install plusplus

Add a decorator:

from plusplus import enable_increments

@enable_increments
def increment_and_return(x):
    return ++x

This enables increments for all code inside the function, including nested function and class definitions.

In package/__init__.py, make this call before you import submodules:

from plusplus import enable_increments

enable_increments(__name__)

# Import submodules here
...

This enables increments in the submodules, but not in the package/__init__.py code itself.

The same approach could be used to implement the assignment expressions for the Python versions that don't support them. For example, we could replace the x <-- value expressions (two unary minuses + one comparison) with actual assignments (setting x to value).

• cpmoptimize — a module that optimizes a Python code calculating linear recurrences, reducing the time complexity from O(n) to O(log n).
• dontasq — a module that adds functional-style methods (such as .where(), .group_by(), .order_by()) to built-in Python collections.

Copyright © 2021 Alexander Borzunov