Home > Articles

  • Print
  • + Share This
This chapter is from the book

Messaging

Whereas objects in Objective-C are little more than slightly specialized C structures, the efficient and highly flexible message dispatch system is at the heart of Objective-C. It combines true object encapsulation and the dynamicism of languages such as Ruby or Smalltalk. Not only are Objective-C messages powerful, they are also relatively cheap, only around twice the cost of a C function call and within an order of magnitude of basic machine operations. Even an unoptimized message send is around 10 times faster than keyed access via NSString, and 50 times faster than object-creation, despite the fact that in the current Objective-C runtime, an Objective-C selector, is really just a C string.

The reason that messaging via string selectors is so quick is that the compiler, linker, and runtime conspire to guarantee that every C string representing an Objective-C selector has a unique address, and therefore the Objective-C messenger function objc_msgSend() does not have to concern itself with the string that the selectors point at, but just uses the pointer itself as an uninterpreted unique integer value. In fact, as Brad Cox writes in Object-Oriented Programming: An Evolutionary Approach, this selector-uniquing process was the main driver for converting Objective-C from a set of C macros to an actual preprocessor, which then made it possible to create a distinct syntax.

Example 3.15 Generate and test dictionar y-backed accessor method (statically or dynamically)

#import <Foundation/Foundation.h>
#import <objc/runtime.h>
#define dictAccessor( objectType, var, setVar, someDict )     -(objectType*)var { return someDict[@""#var]; }     -(void)setVar:(objectType*)newValue {         someDict[@""#var]=newValue;     }@interface MyObject : NSObject
@property (retain) NSMutableDictionary *dict;
@end
@interface MyObject(notimplemented)
@property (retain) NSString *a;
@property (retain) NSString *b;
@end
@implementation MyObject
-(instancetype)init {
  self=[super init];
  self.dict=[NSMutableDictionary new];
  return self;
}
-(void)addDictAccessorForKey:(NSString*)key
{
  SEL selector=NSSelectorFromString( key );
  id (^block)()=^{
    return self.dict[key];
  };
  imp=imp_implementationWithBlock( block );
   class_addMethod([self class], selector, imp , "@:");
}
dictAccessor( NSString, b, setB , self.dict )
@end
int main()
{
        MyObject *m=[MyObject new];
        
        [m addDictAccessorForKey:@"a"];
        m.dict[@"a"]=@"Hello";
        m.b=@"World!";
        NSLog(@"m.a: %@ m.b: %@",m.a,m.b);
        return(0);
}

On Mac OS X 10.11 with Xcode 7.3.1, the code in Example 3.16 prints selector: "hasPrefix:", but the compiler already warns that cast of type "SEL" to "char *" is deprecated; use sel_getName instead. In the GNU runtime, selectors are structure that reference both the message name and its type encoding.

Example 3.16 Printing a selector as a C string using Apple’s runtime

#import <Foundation/Foundation.h>

int main()
{
  SEL a=@selector(hasPrefix:);
  printf("selector: %s\n",(char*)a);
  return 0;
}

IMP Caching

Although developers new to Objective-C tend to worry most about message sending, for example, compared to C++ virtual function invocation, the Objective-C messenger function objc_msgSend() (or objc_msg_lookup() in GNU-objc) has been highly optimized and is usually not a bottleneck.

In the rare cases that it does become a factor, it is possible to retrieve the function pointer from the runtime and call that instead. The technique is known as IMP caching because the type definition of an Objective-C method pointer is called an IMP (implementation method pointer, or just IMPlementation). IMP caching can be useful in a tight loop with a fixed receiver when the method itself is trivial and therefore message dispatch is a major contributor. Example 3.17 shows a greater than 2.5-times improvement in runtime from 2.8 ns to 1.08 ns after subtracting loop overhead.

Example 3.17 Replacing a plain message send with an IMP-cached message send

#import <MPWFoundation/MPWFoundation.h>
@interface MyInteger : NSObject
@property (assign) int intValue;
@end

@implementation MyInteger
@end

int main()
{
  MyInteger *myObject=[MyInteger new];
  int a=0;
  myObject.intValue=42;
  for ( int i=0; i<1000; i++) {
    a+=[myObject intValue];
  }
  
  IMP intValueFun=[myObject methodForSelector:@selector(intValue)];
  for ( int i=0; i<1000; i++) {
    a+=(int)intValueFun( myObject, @selector(intValue) );
  }
}

Due to the dynamic nature of Objective-C, there is no automatic way of determining at compile time whether this optimization is safe, which is one reason the Objective-C compiler doesn’t do it for you. Fortunately, it is usually very easy for a developer to make that determination. While there are numerous ways for the IMP to change during execution (for example, loading a bundle that includes a category, and using runtime functions to add, remove, or change method implementations or even change the class of the object in question), all of these are rare events that happen fairly predictably.

It is the developer’s job to ensure that either none of these events happen, or alternately, that they do not have an impact on the computation.

A special case that needs to be considered when doing IMP caching is the nil receiver. The Objective-C messenger quietly ignores messages to nil, simply returning zero instead of dispatching the message. This short-circuiting protects receivers from having to worry about a nil self pointer, and sender from having to special case nil-receivers. IMP caching breaks this protection on several counts: If the receiver is nil when requesting the IMP, a NULL function pointer will be returned, and invoking such a NULL function pointer will crash the program. On the other hand, if a correct function pointer was obtained from an earlier, non-nil object pointer, calling that function pointer will call a method with a nil self pointer. Any instance variable access from within that method will also crash the program.

So you will need to ensure both that you are not getting a NULL IMP and that you don’t call an IMP with a nil receiver.

IMP caching can be particularly useful when sending messages to “known” objects such as delegates or even self. Example 3.18 shows part of the actual header of the object cache discussed in the “Mutability and Caching” section of this chapter. In addition to the cache itself ( objs, cacheSize ) and the current pointer into the cache objIndex, it also maintains IMP pointers for all the message sent in the -getObject method from Example 3.10, allowing the actual -getObject to run without once invoking the messenger. In addition, it makes the IMP for the -getObject method itself available in a @public instance variable, along with a GETOBJECT() C-preprocessor macro to invoke it. The GETOBJECT macro is actually slightly less code to write than a normal alloc-init-autorelease, is 8% faster even with a cache miss, is 15 times faster with a cache hit, and last but not least decouples the user of the cache from the specific class used.

Example 3.18 Definition and use of an object cache for integer objects

@interface MPWObjectCache : MPWObject
{
    id          *objs;
    int         cacheSize,objIndex;
    Class       objClass;
    SEL         allocSel,initSel,reInitSelector;
    IMP         allocImp,initImp,reInitImp,releaseImp;
    IMP         retainImp,autoreleaseImp;
    IMP         retainCountImp,removeFromCacheImp;
    @public
    IMP         getObject;
}

+(instancetype)cacheWithCapacity:(int)newCap class:(Class)newClass;
-(instancetype)initWithCapacity:(int)newCap class:(Class)newClass;
-getObject;
#define GETOBJECT( cache )
    ((cache)->getObject( (cache), @selector(getObject)))
...
@end
integerCache=[[MPWObjectCache alloc] initWithCapacity:20
          class:[MPWInteger class]];
MPWInteger *integer=GETOBJECT( integerCache );
[integer setIntValue:2];

If IMP caching is insufficient and you have the source code of the method you need to call available, you can always turn it into a C function, an inline function, or even a preprocessor Macro.

Considering how little of a problem dynamic dispatch is in practice, and how easy it is to remove the problem in the rare cases it does come up, it is a little surprising how much emphasis the Swift team has placed on de-emphasizing and removing dynamic dispatch from Swift for performance reasons.

Forwarding

While close to C function call speeds on one end, Objective-C messages are flexible enough to take the place of reified messaging and control structures on the other end. For example, Cocoa does not have to use the Command pattern because messages carry enough runtime information to be reified, stored, and introspected about so something like the NSUndoManager can be built using the fast built-in messaging system.

For your own projects, I would always recommend mapping any requirements for dynamic runtime behavior onto the messaging infrastructure if at all possible, and with a full reflective capabilities what is possible is very broad. The code in Example 3.19 will execute the message to the object in question as a Unix shell command, so [object ls] will execute the ls command, and [object date] the date command. A more elaborate example would translate message arguments to script arguments.

Example 3.19 Mapping sent messages to shell commands

#import <Foundation/Foundation.h>
@interface Shell:NSObject
@end
@interface Shell(notimplemented)
-(void)ls;
@end
@implementation Shell

-(void)forwardInvocation:(NSInvocation*)invocation {
  system( [NSStringFromSelector( [invocation selector])
           fileSystemRepresentation] );
}
-(void)dummy {}
-methodSignatureForSelector:(SEL)sel
{
   NSMethodSignature *sig=[super methodSignatureForSelector:sel];
   if (!sig) {
      sig=[super methodSignatureForSelector:@selector(dummy)];
   }
   return sig;
}
@end

int main()
{
    Shell *sh=[Shell new];
    [sh ls];
    return 0;
}

Example 3.20 reads the file that is named by the sent message instead of executing it, and perhaps somewhat more realistically, Example 3.21 looks up the selector in a local dictionary.

Example 3.20 Mapping sent messages to file contents

#import <Foundation/Foundation.h>
@interface Filer:NSObject
@end
@interface Filer(notimplemented)
-(NSString*)hello;
@end
@implementation Filer

-(void)forwardInvocation:(NSInvocation*)invocation {
  NSString *filename=NSStringFromSelector( [invocation selector]);
  NSString *contents=[[NSString alloc]
                                initWithContentsOfFile:filename
                                encoding:NSISOLatin1StringEncoding
                                error:nil];
  [invocation setReturnValue:&contents];
}
-(NSString*)dummy { return @""; }
-methodSignatureForSelector:(SEL)sel
{
   NSMethodSignature *sig=[super methodSignatureForSelector:sel];
   if (!sig) {
      sig=[super methodSignatureForSelector:@selector(dummy)];
   }
   return sig;
}
@end

int main()
{
    Filer *filer=[Filer new];
    NSLog(@"filer: %@",[filer hello]);
    return 0;
}

Example 3.21 Mapping sent messages to dictionary keys

-(void)forwardInvocation:(NSIvocation*)invocation {
     id result=[[self dictionary] objectForKey:
               NSStringFromSelectr([invocation selector])];
     [invocation setReturnValue:&result];
}

Uniformity and Optimization

Although there is no actual performance benefit for the implementations of Examples 3.19 to 3.21, the benefit comes from using the fastest plausible interface, an interface that can be kept the same all the way from reading files (3.20) via using runtime introspection to look up keys (3.21), generating accessors to a keyed store at runtime or compile time (3.15) or switching to an accessor for an actual instance variable, and finally IMP caching that message send. You don’t have to start out fast, but you have to use interfaces that allow you to become fast should the need arise.

The more I have followed Alan’s advice to focus on the messages, the better my programs have become, and the easier it has been to make them go fast.

  • + Share This
  • 🔖 Save To Your Account