Next article: Friday Q&A 2009-03-27: Objective-C Message Forwarding
Previous article: Friday Q&A 2009-03-13: Intro to the Objective-C Runtime
Tags: fridayqna objectivec
Welcome back to another Friday Q&A. This week I'd like to take Joshua Pennington's idea and elaborate on a particular facet last week's topic of the Objective-C runtime, namely messaging. How does messaging work, and what exactly does it do? Read on!
Definitions
Before we get started on the mechanisms, we need to define our terms. A lot of people are kind of unclear on exactly what a "method" is versus a "message", for example, but this is critically important for understanding how the messaging system works at the low level.
- Method: an actual piece of code associated with a class, and which is given a particular name. Example:
- (int)meaning { return 42; }
- Message: a name and a set of parameters sent to an object. Example: sending "meaning" and no parameters to object
0x12345678
. - Selector: a particular way of representing the name of a message or method, represented as the type
SEL
. Selectors are essentially just opaque strings that are managed so that simple pointer equality can be used to compare them, to allow for extra speed. (The implementation may be different, but that's essentially how they look on the outside.) Example:@selector(meaning)
. - Message send: the process of taking a message and finding and executing the appropriate method.
Methods
The next thing that we need to discuss is what exactly a method is at the machine level. From the definition, it's a piece of code given a name and associated with a particular class, but what does it actually end up creating in your application binary?
Methods end up being generated as straight C functions, with a couple of extra parameters. You probably know that self
is passed as an implicit parameter, which ends up being an explicit parameter. The lesser-known implicit parameter _cmd
(which holds the selector of the message being sent) is a second such implicit parameter. Writing a method like this:
- (int)foo:(NSString *)str { ...
int SomeClass_method_foo_(SomeClass *self, SEL _cmd, NSString *str) { ...
What, then, happens when we write some code like this?
int result = [obj foo:@"hello"];
int result = ((int (*)(id, SEL, NSString *))objc_msgSend)(obj, @selector(foo:), @"hello");
What that ridiculous piece of code after the equals sign does is take the objc_msgSend
function, defined as part of the Objective-C runtime, and cast it to a different type. Specifically, it casts it from a function that returns id
and takes id
, SEL
, and variable arguments after that to a function that matches the prototype of the method being invoked.
To put it another way, the compiler generates code that calls objc_msgSend
but with parameter and return value conventions matched to the method in question.
Messaging
A message send in code turns into a call to objc_msgSend
, so what does that do? The high-level answer should be fairly apparent. Since that's the only function call present, it must look up the appropriate method implementation and then call it. Calling is easy: it just needs to jump to the appropriate address. But how does it look it up?
The Objective-C header runtime.h
includes this as part of the (now opaque, legacy) objc_class
structure members:
struct objc_method_list **methodLists OBJC2_UNAVAILABLE;
struct objc_method_list {
struct objc_method_list *obsolete OBJC2_UNAVAILABLE;
int method_count OBJC2_UNAVAILABLE;
#ifdef __LP64__
int space OBJC2_UNAVAILABLE;
#endif
/* variable length structure */
struct objc_method method_list[1] OBJC2_UNAVAILABLE;
} OBJC2_UNAVAILABLE;
objc_method
structs. That one is in turn defined as:
struct objc_method {
SEL method_name OBJC2_UNAVAILABLE;
char *method_types OBJC2_UNAVAILABLE;
IMP method_imp OBJC2_UNAVAILABLE;
} OBJC2_UNAVAILABLE;
So even though we're not supposed to touch these structs (don't worry, all the functionality for manipulating them is provided through functions in elsewhere in the header), we can still see what the runtime considers a method to be. It's a name (in the form of a selector), a string containing argument/return types (look up the @encode
directive for more information about this one), and an IMP
, which is just a function pointer:
typedef id (*IMP)(id, SEL, ...);
objc_msgSend
has to do is look up the class of the object you give it (available by just dereferencing it and obtaining the isa
member that all objects contain), get the class's method list, and search through the method list until a method with the right selector is found. If nothing is there, search the superclass's list, and so on up the hierarchy. Once the right method is found, jump to the IMP of method in question.
One more detail needs to be considered here. The above procedure would work but it would be extremely slow. objc_msgSend
only takes about a dozen CPU cycles to execute on the x86 architecture, which makes it clear that it's not going through this lengthy procedure every single time you call it. The clue to this is another objc_class
member:
struct objc_cache *cache OBJC2_UNAVAILABLE;
struct objc_cache {
unsigned int mask /* total = mask + 1 */ OBJC2_UNAVAILABLE;
unsigned int occupied OBJC2_UNAVAILABLE;
Method buckets[1] OBJC2_UNAVAILABLE;
}
Method
structs, using the selector as the key. The way objc_msgSend
really works is by first hashing the selector and looking it up in the class's method cache. If it's found, which it nearly always will be, it can jump straight to the method implementation with no further fuss. Only if it's not found does it have to do the more laborious lookup, at the end of which it inserts an entry into the cache so that future lookups can be fast.
(There is actually one more detail beyond this which ends up being extremely important: what happens when no method can be found for a given selector. But that one is so important that it deserves its own post, so look for it next week.)
Conclusion
That wraps up this week's edition. Come back next week for more. Have a question? Think Objective-C's messaging system should be done differently? Post below.
Remember, Friday Q&A is powered by your ideas. If you have an idea for a topic, tell me! Post your idea in the comments, or e-mail them directly to me (I'll use your name unless you ask me not to).
Comments:
As he mentions in his articles on DTrace/Instruments, the implication of this is that you can't attach a DTrace probe to the return of objc_msgSend.
http://www.friday.com/bbum/2009/12/18/objc_msgsend-part-1-the-road-map/
Comments RSS feed for this page
Add your thoughts, post a comment:
Spam and off-topic posts will be deleted without notice. Culprits may be publicly humiliated at my sole discretion.
The very beginning of objc_msgSend() has two special cases.
The first checks to see if the receiver is nil and short circuits the call. But it doesn't just return. There is actually a hook for handling messages to nil buried in there. It isn't terribly useful, though, as there are a couple of 10s of thousands of message to nil as an average app does its thing (http://www.friday.com/bbum/2008/01/02/objective-c-logging-messages-to-nil/).
The second only occurs under GC. In this case, the selectors for -retain, -release, and -autorelease are overwritten and cause a short-circuit that very quickly returns self. Trivia: -retainCount is also short-circuited and, thus, -retainCount will return an absurdly large number under GC.
Note, also, that objc_msgSend() tail calls the real IMP. This is what enables method implementations to just be C functions. objc_msgSend() goes to great lengths to not muck with the arguments on the stack. Instead, it just figures out what address should be jumped to just like a normal C function.