Diet Python and some ranting

Posted on by Chris Warburton

Long time no post (in fact, I haven’t posted at all since starting my fourth year at Sheffield Uni!). I keep writing big ranty things, meaning to post them, but haven’t got around to it yet. Ah well, maybe soon (exams at the moment though :( )

As part of my biannual exam revision procrastination I’ve started to rejig my Diet Python language. The idea behind Diet Python is pretty simple: take all of the syntactic sugar out of Python. Thus Diet Python is just a subset of Python, which is just as powerful as the complete language. Diet Python isn’t meant to be programmed in, however; it exists to simplify Python programs so that they can be handled more easily by automated tools. For example, in Python the following two lines are equivalent:

a + b
a.__add__(b)

So the Diet Python equivalent is this:

a.__add__(b)

It is thus completely compatible with Python, but there is no need to bother with “+” (or the associated nodes in the Abstract Syntax Tree). The same applies to other Python syntax such as:

my_list[index]
my_list.__getitem__(index)

In Diet Python we can get rid of the [] subscription notation, without losing the ability to grab elements from lists (or any other type of object which implements the [] syntax, which is done via the __getitem__ method). Thus if we have a Python implementation (CPython, Jython, PyPy, IronPython, etc.) then it is also a Diet Python implementation (plus some extra stuff that Diet Python won’t use), but more interestingly if we implement Diet Python then we’ve actually implemented the whole of Python in terms of features, just not the syntax. This can be overcome easily by using a translator to turn Python’s nice, sugary syntax into Diet Python’s awkward, canonical syntax, which is exactly what I’ve done.

Diet Python originally started as a simple test case for my Python pretty printer (or “decompiler”), which turns a Python Abstract Syntax Tree, produced by Python 2.x’s built-in “compiler” module, into valid Python code which implements the AST’s functionality (ie. compile some Python into an AST, stick that into the decompiler to get some Python code, then compile that to an AST and the two ASTs should be the same (as long as every transformation is reversible, is reversed, doesn’t lose information and is done naively, that is ;) ).

The decompiler itself was an experiment to get used to the PyMeta pattern matching framework (which I’ve since used in a University project to test my code), and since PyMeta, as an implementation of Alessandro Warth’s OMeta, should be nicely extensible via subclassing, I wanted both an experiment in PyMeta and an experiment in extending my experiment in PyMeta to really get to grips with it.

Unfortunately subclassing in PyMeta has proven difficult, which might be a bug in the implementation (I’ll have to check up on that). Making a pattern matcher, for example to decompile Python ASTs, in PyMeta allows anyone to make a similar pattern matcher based on it quite easily through Python’s object system. For example if you want to get rid of every “print” statement in some code, you take the decompiler (which is a Python class object), then you write down the grammar rules which differ from the original (in this example every Print and Printnl node should be translated into ’’ (ie. an empty string, and thus no code)), then you tell the decompiler to make a grammar out of your rules, and it will give you a new Python class which implements a pattern matcher using the decompiler’s rules + your new ones (where the new ones override the decompiler’s ones in case of conflicts).

This is all well and good, however the REALLY cool thing about OMeta and thus PyMeta is that their operation, ie. turning written rules into parsers for those rules, is written in (O/Py)Meta (which is why they are Meta). Thus it is possible to take OMeta and, by writing some OMeta rules, change the way that OMeta works, we could call it OMeta’. Now OMeta’ can be changed by writing rules in either OMeta or OMeta’, to produce another pattern matcher which we can call OMeta’’, and so on. This, however, doesn’t seem to work in PyMeta, despite trying multiple ways and looking through the source code (which is written in OMeta) over and over again. Sad face :(

Ah well, this limitation has resulted in a bit of hackiness when it comes to the Python decompiler and Diet Python translators. Firstly, PyMeta has no syntax for comments, which is annoying. It should be simple to subclass PyMeta to make a PyMeta’ which supports comments, but since I can’t subclass PyMeta without losing its bootstrapping, I’m stuck with using Python to remove comments before passing the rules to PyMeta. The second hack is that doing tree operations requires recursion. Whilst PyMeta has recursion built in, it’s not available in the most suitable way for my experiments. Once again, subclassing PyMeta should solve this, but I can’t, so I’ve had to monkey-patch the AST nodes (ie. pollute their namespaces with functions and attributes) then call these from inside the grammar. What this results in is every node instantiating their own pattern matcher on themselves, which happens recursively down the trees.

Unfortunately the “type” system of Python 2.x rears its ugly head here, where historical implementation decisions have left Python with 2 object hierarchies (which, I believe, was one of the main motivations for making Python 3). The object system which is the most familiar, since it’s used in Python code, has the class “object” as the eventual ancestor of everything, such that every class is a subclass of object, or a subclass of a subclass of object, etc. This would be a “pure” object system, except that “everything” isn’t quite everything. Many core pieces of Python, for example text strings, are not descendents of “object” at all, and are not subclasses of anything, or indeed classes. Instead they are “types”, where each “type” seems to be isolated from everything else, written by hand in C, utterly inextensible, cannot be subclassed, and basically brings to mind those nightmarish things that Java programmers call “basic types” (shudders, washes mouth out with soap). Since they are in their own little statically-compiled-C world there is no way to monkey patch them with the required functions and attributes, so that every string, number, None and probably more require custom code in the pattern matchers. Shit. This also brings with it that great friend of everybody who loves to waste time known as combinatorial explosion. In other words, instead of doing a substitution like:

apply_recursively ::= <anything>:a => a.recurse()

(apply_recursively is defined as taking anything and calling it a, then outputting the value of running a.recurse())

We have to do something like:

apply_recursively ::= <anything>:a ?(not issubclass(a.__class__, Node)) => a
                    | <anything>:a => a.recurse()

(apply_recursively is defined as taking anything and calling it a, as long as it is not descended from Node, and outputting it’s value, or else taking anything and calling it a and outputting the value of a.recurse())

And of course, since this is our friend combinatorial explosion, we cannot write:

apply_recursively ::=<anything>:a1 <anything>:a2 &lt;anything&gt;:a3 &lt;anything&gt;:a4 =&gt; a1.recurse() + a2.recurse() + a3.recurse() + a4.recurse()

Oh no, if there’s the chance that any of those might be “types” (winces) then we are forced to write instead:

apply_recursively ::= <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a1__class__, Node) or issubclass(a2__class__, Node) or issubclass(a3__class__, Node) or issubclass(a4__class__, Node))) => a1 + a2 + a3 + a4
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a1.__class__, Node) or issubclass(a2.__class__, Node) or issubclass(a3.__class__, Node)) and issubclass(a4.__class__, Node)) => a1 + a2 + a3 + a4.recurse()
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a1.__class__, Node) or issubclass(a2.__class__, Node) or issubclass(a4.__class__, Node)) and issubclass(a3.__class__, Node)) => a1 + a2 + a3.recurse() + a4
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a1.__class__, Node) or issubclass(a2.__class__, Node)) and issubclass(a3.__class__, Node) and issubclass(a4.__class__, Node)) => a1 + a2 + a3.recurse() + a4.recurse()
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a1.__class__, Node) or issubclass(a4.__class__, Node) or issubclass(a3.__class__, Node)) and issubclass(a2.__class__, Node)) => a1 + a2.recurse() + a3 + a4
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a1.__class__, Node) or issubclass(a3.__class__, Node)) and issubclass(a2.__class__, Node) and issubclass(a4.__class__, Node)) => a1 + a2.recurse() + a3 + a4.recurse()
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a1.__class__, Node) or issubclass(a4.__class__, Node)) and issubclass(a3.__class__, Node) and issubclass(a2.__class__, Node)) => a1 + a2.recurse() + a3.recurse() + a4
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a1.__class__, Node)) and issubclass(a2.__class__, Node) and issubclass(a3.__class__, Node) and issubclass(a4.__class__, Node)) => a1 + a2.recurse() + a3.recurse() + a4.recurse()
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a4.__class__, Node) or issubclass(a2.__class__, Node) or issubclass(a3.__class__, Node)) and issubclass(a1.__class__, Node)) => a1.recurse() + a2 + a3 + a4
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a3.__class__, Node) or issubclass(a2.__class__, Node)) and issubclass(a1.__class__, Node) and issubclass(a4.__class__, Node)) => a1.recurse() + a2 + a3 + a4.recurse()
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a4.__class__, Node) or issubclass(a2.__class__, Node)) and issubclass(a3.__class__, Node) and issubclass(a1.__class__, Node)) => a1.recurse() + a2 + a3.recurse() + a4
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a2.__class__, Node)) and issubclass(a1.__class__, Node) and issubclass(a3.__class__, Node) and issubclass(a4.__class__, Node)) => a1.recurse() + a2 + a3.recurse() + a4.recurse()
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(issubclass(a1.__class__, Node) and issubclass(a2.__class__, Node) and not (issubclass(a3.__class__, Node) or issubclass(a4.__class__, Node))) => a1.recurse() + a2.recurse() + a3 + a4
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(not (issubclass(a3.__class__, Node)) and issubclass(a2.__class__, Node) and issubclass(a1.__class__, Node) and issubclass(a4.__class__, Node)) => a1.recurse() + a2.recurse() + a3 + a4.recurse()
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(issubclass(a1.__class__, Node) and issubclass(a2.__class__, Node) and issubclass(a3.__class__, Node) and not (issubclass(a4.__class__, Node))) => a1.recurse() + a2.recurse() + a3.recurse() + a4
                    | <anything>:a1 <anything>:a2 <anything>:a3 <anything>:a4 ?(issubclass(a1.__class__, Node) or issubclass(a2.__class__, Node) or issubclass(a3.__class__, Node) or issubclass(a4.__class__, Node)) => a1.recurse() + a2.recurse() + a3.recurse() + a4.recurse()

Which, even if you’ve never programmed before, should look like a bloody stupid hoop to have to jump through.

So, there’s a little insight into how even high-level, meta, abstract things can be hindered by ancient, low-level implementation artifacts, and possibly an insight into the angry posts I was making to Indenti.ca whilst writing this stuff six months ago ;)

My code, as always, is in git, and now that I’ve turned the Diet Python translator into a tree transform it should be much easier to strip away more and more layers of unnecessary Python (and thus pave the way for some interesting programming experiments!)