Friday saw the second edition of the UTAS Computing Society Lightning Talks, if you haven’t seen them already, I highly recommend that you check them out — this semester’s were at a very high standard indeed, and I wish I’d printed out more certificates for good talks :). My talk was a demonstration of using metaprogramming in Python, though that’s not what it seemed to be about.
I went to the Apple University Consortium’s Cocoa Workshop at the University of New South Wales in February of this year, it was a heap of fun, and we learnt heaps whilst there. One of the key distinguising features of Cocoa is its use of verbose English method and attribute names, the idea being that each line of code should make a reasonable amount of sense when read aloud, hence:
NSString *str = [[NSString alloc] initWithString @"Hello World!"]
does indeed allocate memory to hold a string object, and initialises the newly-allocated memory with a string containing “Hello World!” (this code is highly redundant!). Supposedly such a naming scheme allows coders to write code that is easily maintainable by the original coder, and easily learnable by people who pick up the code for the first time.
On the other hand, my friends, collectively known as Maclab (named after the room at UTAS we inhabit) have developed a rather unique vocabularly, which in particular involves replacing as many words as possible with either ‘thrust’ or ‘fork’, so “Thrustingly thrust the forking forker” is not an uncommon utterance amongst my friends. If this is indeed their usual mode of conversation, then Cocoa’s way of identifying methods and attributes is not necessarily going to be a particularly intiuitive one. So, clearly, we need a version of cocoa that meets their needs.
So, conveniently, Apple provide a comprehensive version of the Cocoa API, thanks to the PyObjC project. We can therefore use the Python bindings for Cocoa facilitate our new version of Cocoa. Since Cocoa has a very consistent naming scheme, we can simply perform string replacement to translate from our maclab language to the standard cocoa language, using a routine somewhat like this:
def translate(inp): ''' Translates an input string from key language to value language ''' for i in LANGUAGE: if i.islower(): inp = inp.replace(i, LANGUAGE[i]) inp = inp.replace(rtitle(i), rtitle(LANGUAGE[i])) else: inp = inp.replace(i, LANGUAGE[i]) return inp def rtitle(i): return i.upper() + i[1:]
Here, LANGUAGE is a dictionary, with keys in the language code will be written in and values being the target language (in this case, Cocoa). There’s not all that much of a sophisticated nature going on in here. Now that we have a method by which we can translate our attribute accesses, we can get to the meat of the the code.
To achieve the new API, we need to use a technique that I will call proxying. This involves the use of objects whose sole purpose is to intercept attribute accesses and calls to an underlying object. In this case, the point of intercepting the calls and accesses is to perform translation from our new objects to standard Cocoa objects. In Python we can do this by overriding the standard attribute access and call methods.
First up is __getattr__, the attribute accessor method — for this, we are passed a string; the name of the attribute that we’re looking for, which we translate, and then attempt to access upon the method on the underlying object (in this case, self.__u__). There is one slight hitch: in certain cases, we may not want to translate the attribute name. This is true, in particular, of the attribute that represents the underlying object. Hence we provide a REAL_ATTRS list, for which we use the default __getattr__ method for. This results in code that looks something like this:
def __getattribute__(self,name): if name in REAL_ATTRS: return object.__getattribute__(self,name) else: new_objectname = "self.__u__.%s" % translate(name) new_object = eval(new_objectname) return CocomoProxy(new_object)
Notice that we use eval to perform the lookup? It turns out that __getattr__ doesn’t work universally, whereas . notation does — so we use that for less failover.
Being able to call methods on the objects is important, but slightly more difficult — we want behaviour to be maintained, so we need to make sure that proper Cocoa objects are passed as arguments, rather than the Proxy objects that you may have originally dealt with. We can do this with Python’s argument unpacking — we build up a list of arguments, and unproxy them as necessary:
def __call__(self,*a, **k): new_a = [i.__u__ if type(i) == CocomoProxy else i for i in a] new_k = dict( (translate(i), k[i].__u__ if type(k[i]) == CocomoProxy else k[i]) for i in k) return CocomoProxy(self.__u__(*new_a,**new_k))
We may also need to deal with iterators. This can be done using a standard generator function, thusly:
def __iter__(self): for i in self.__u__: yield CocomoProxy(i)
Finally, there may be legitimate reasons for extracting Cocoa objects, these include printing strings, so we provide an accessor method called no_really:
def no_really(self): return self.__u__
And that’s the entire implementation! The final thing we need to do is provide a pre-proxied version of the base module for Cocoa. Let’s call it GypsyMagic.
So now that we have a working bridge from Maclab English to Cocoa English, we can take this sample code that puts some stuff into an array, and then prints it:
import AppKit hworld = AppKit.NSString.alloc().initWithString_("Hello, World!") arr = AppKit.NSMutableArray.alloc().init() arr.addObject_(hworld) arr.addObject_("Boop!") for i in arr: print i
And write it in the far more palatable:
from cocomo import GypsyMagic hworld = GypsyMagic.OGMouthWords.subsume().makeGogoWithMouthWords_("Hello, World!") arr = GypsyMagic.OGForkableTrinketHolder.subsume().makeGogo() arr.thrustinglyThrustForker_(hworld) arr.thrustinglyThrustForker_("Boop!") for i in arr: print i.no_really()
If you’re interested in seeing how it all fits together, see Cocomo’s website.