What is the cost of using autorelease in Cocoa?
There are two costs:
(Assuming you have an option to avoid autoreleased objects.) You effectively unnecessarily extend the lifetime of your objects. This can mean that your memory footprint grows -- unnecessarily. On a constrained platform, this can mean that your application is terminated if it exceeds a limit. Even if you don't exceed a limit, it may cause your system to start swapping, which is very inefficient.
The additional overhead of finding the current autorelease pool, adding the autoreleased object to it, and then releasing the object at the end (an extra method call). This may not be a large overhead, but it can add up.
Best practice on any platform is to try to avoid autorelease if you can.
To answer the questions:
Ultimately should I use a strategy where everything is autoreleased and using retain/release should be the exception to the rule for specific cases?
Quite the opposite.
Or should I generally be using retain/release with autorelease being the exception for returned objects from convenience methods like [NSString stringWithEtc...] ?
You should always use retain/release if you can -- in the case of NSString there is typically no need to use stringWithEtc methods as there are initWithEtc equivalents.
See also this question.
I have to disagree with Jim Puls - I think that not using Autorelease makes debugging more difficult, because you are more likely to find yourself accidentally leaking memory. Of course Clang static analyzer can pick up some of these instances, but for me, the slight overhead costs in habitually using autorelease are far overshadowed by my code being less likely to be buggy.
And then, only if I have a tight loop I need to optimize will I start looking at performance. Otherwise this is all just premature optimization, which is generally considered to be a bad thing.
I'm surprised nobody has mentioned this yet. The biggest reason to avoid autoreleased objects when you can has nothing to do with performance. Yes, all of the performance concerns mentioned here are absolutely valid, but the biggest downside to autorelease is that it makes debugging significantly more difficult.
If you have an overreleased object that's never autoreleased, it's trivially easy to track down. If you have a user-reported crash that happens intermittently with a backtrace somewhere south of NSPopAutoreleasePool, good luck...
I generally use autoreleased objects these days because they tend to result in simpler, easier to read code. You declare and initialize them, then let the drop out of scope. Mechanically they exist quite a bit longer, but from the perspective of the person writing the code it is equivalent to a stack declared object in C++ automatically being destructed when the function returns and its frame is destroyed.
While there is an efficiency loss, in most cases it is not significant. The bigger issue is the more extant objects and later memory recovery can lead to a more fragmented address space. If that is an issue it usually is fairly simple to go in and switch to manual retain/release in a few hot methods and improve it.
As others have said, readability trumps performance in nonperformance sensitive code. There are a number of cases where using autoreleased objects leads to more memory fragmentation, but in any case where the object will outlive the pool it will not. In those cases the only price you are paying is finding the cost of finding the correct autorelease pool.