Testing "if you use a default allocator, and default callbacks, CF will generate an NS collection instead". (https://twitter.com/Catfish_Man/status/455592839334723584). We currently only have that on Mavericks, but since it the same codebase I'm 100% sure this will be in iOS 8 as well.

CoreFoundationSurprises.m

// Code using the default allocator and callbacks now actually returns an NSArray (__NSArrayM), and no longer a toll-free-bridged object.
NSArray *test  =  CFBridgingRelease(CFArrayCreateMutable(kCFAllocatorDefault, 0, &kCFTypeArrayCallBacks));
NSArray *test3 =  CFBridgingRelease(CFArrayCreate(kCFAllocatorDefault, NULL, 0, &kCFTypeArrayCallBacks));

// This is custom and will always return __NSCFArray
NSArray *test2 =  CFBridgingRelease(CFArrayCreateMutable(kCFAllocatorDefault, 0, NULL));

// This will return __NSArrayI.
NSArray *test4 = [NSArray array];

/*
lldb (Mavericks, 10.9.2)
(lldb) po test
<__NSArrayM 0x10010ea80>(
)


(lldb) po test2
<__NSCFArray 0x10010ed50>(
)


(lldb) po test3
<__NSArrayI 0x100103fd0>(
)


(lldb) po test4
<__NSArrayI 0x100103fd0>(
)


lldb (iOS 7.1, Simulator):

(lldb) po test
<__NSCFArray 0x113f04720>(
)


(lldb) po test2
<__NSCFArray 0x113f04750>(
)


(lldb) po test3
<__NSCFArray 0x113f04780>(
)


(lldb) po test4
<__NSArrayI 0x10e232820>(
)
*/

TIL: The old “use CF because it’s faster” thing was arguably not true before, but very much so isn’t now. Hah. Looks like we've some things to do in http://pspdfkit.com. We're using CF extensively for performance reasons (especially for the glyph text parsing), which was very much important for iOS 4-6, but less so for 7 and might even negate in iOS 8.

This might be enabled for a future version of iOS since the objc_msgSend overhead is negligible in the arm64 runtime