Basic ObjC /C++ performance ala Mike Ash

Rich @siegel asked on twitter:

How expensive is it to call a block, versus a method, versus a function? Asking for a friend.

— Rich Siegel (@siegel) March 14, 2013

and @danielpunkass pointed to a @mikeash blog post with some tests:

@rentzsch @siegel I was reminded by @mikeash of this old gem: mikeash.com/pyblog/perform… Maybe just needs an update for blocks…

— Daniel Jalkut (@danielpunkass) March 14, 2013

I was curious about this so I added that test and a test of the new @autoreleasepool {} construct. The modified code is here (@mikeash feel free to grab it and use it on your site) and the results of running this on my older MacBook Air 1.8GHz Core i7 with 4GB 1333MHz DDR3 RAM and SSD (other apps running in the background) are as follows:

Name	Iterations	Total time (sec)	Time per (ns)
IMP-cached message send	1000000000	0.2	0.2
C++ virtual method call	1000000000	0.4	0.4
Block invocation	1000000000	0.5	0.5
Integer division	1000000000	2.3	2.3
Objective-C message send	1000000000	2.7	2.7
16 byte memcpy	100000000	0.4	4.4
Floating-point division	100000000	0.6	6.0
Float division with int conversion	100000000	0.6	6.3
New Autorelease construct	10000000	0.2	17.1
NSAutoreleasePool alloc/init/release	10000000	0.6	63.5
16 byte malloc/free	100000000	6.5	64.7
NSInvocation message send	10000000	0.9	91.3
NSObject alloc/init/release	10000000	1.6	165.0
16MB malloc/free	100000	0.1	880.0
NSButtonCell creation	1000000	3.2	3201.7
Read 16-byte file	100000	0.8	8250.8
pthread create/join	10000	0.2	23009.4
Zero-second delayed perform	100000	2.8	28038.3
NSButtonCell draw	100000	4.7	47448.6
1MB memcpy	10000	0.8	78902.2
Write 16-byte file	10000	2.1	209280.7
Write 16-byte file (atomic)	10000	3.9	392459.6
Read 16MB file	100	0.8	8050127.5
NSTask process spawn	1000	67.3	67305116.9
Write 16MB file (atomic)	30	2.1	70342873.5
Write 16MB file	30	2.2	72666382.1

Built it with:

cc -v -lstdc++ -framework Cocoa perf.mm

and the dev tools are:
Apple LLVM version 4.2 (clang-425.0.24) (based on LLVM 3.2svn)
Target: x86_64-apple-darwin12.2.0
Thread model: posix

Xcode 4.6 basically.

If Mike reads this: feel free to grab any of this you want and copy it to your site. Just put it here because that was easiest.

Not sure my tests are exactly right so feel free to make suggestions etc.  One thing I didn’t test but might add another test for is how handling local variables used in the block impacts things; likewise local __block variables in the block, but the whole point of this test suite seemed to be keeping things as simple as possible so I stuck with that.

UPDATE:

@schwa asked about iOS and since I had the day off I went ahead and did those as well, though I had to remove some tests I didn’t feel like fixing for iOS (zero delay perform selector messes with the RunLoop which caused crashes, no NSButton class, NSTask (?), and the file I/O tests weren’t going to work as is).  Started with the single view iOS Application Xcode template and added a text view to jam the results into.  This was run on an iPhone 5 running iOS 6.0.2:

Name	Iterations	Total time (sec)	Time per (ns)
C++ virtual method call	1000000000	0.8	0.8
Block invocation	1000000000	2.8	2.8
IMP-cached message send	1000000000	4.6	4.6
Objective-C message send	1000000000	14.7	14.7
Float division with int conversion	100000000	1.8	17.8
Floating-point division	100000000	1.8	17.8
Integer division	1000000000	21.6	21.6
16 byte memcpy	100000000	6.1	61.0
New Autorelease construct	10000000	0.8	83.4
16 byte malloc/free	100000000	46.8	467.5
NSAutoreleasePool alloc/init/release	10000000	5.2	519.2
NSObject alloc/init/release	10000000	12.6	1259.0
NSInvocation message send	10000000	13.5	1346.3
16MB malloc/free	100000	1.5	15374.9
pthread create/join	10000	1.2	116415.9
1MB memcpy	10000	3.9	394341.5

Advertisement