arm64 objectForKey性能异常？

Question

To make a short story long, I was looking at a performance problem in my 32-bit iPhone app. In the process I removed the 3rd-party library that was limiting me to 32-bit and when I built for arm64 I saw a 2-fold speed improvement in a section of code unrelated to said library.

I was under the impression that simply adding arm64 to the Valid Architectures was not going to bring a significant increase in performance so I wonder if my case is just an anomaly.

In the process of whittling my app down to a few dozen lines that show the performance difference I lost some of the gain, but it's still significant. Seemingly minor changes like the number of characters in dictionary keys and the mix of the number of objects in the keys make a big difference.

The following code is the entirety of the sample app - placed in viewDidLoad. Built with arm64 using Xcode 5.1, running on my iPhone 5s (iOS 7.1) average time to retrieve 2075 dictionary objects is about 0.6 secs, built armv7s, about 1.0 secs.

Is there a simple explanation to the performance improvement that can be exploited in general?

#define NUM_DICT_ENTRIES 2075
NSMutableDictionary *aDict = [[NSMutableDictionary alloc] init];
NSDictionary *keyDictionary;
for (int i = 0; i < NUM_DICT_ENTRIES; i++) {
    if (arc4random_uniform(2)) {
        keyDictionary = [NSDictionary dictionaryWithObjectsAndKeys:
                         [[NSProcessInfo processInfo] globallyUniqueString], @"entry",
                         [[NSProcessInfo processInfo] globallyUniqueString], @"category", nil];
    } else {
        keyDictionary = [NSDictionary dictionaryWithObjectsAndKeys:
                         [[NSProcessInfo processInfo] globallyUniqueString], @"entry",
                         [[NSProcessInfo processInfo] globallyUniqueString], @"category",
                         [[NSProcessInfo processInfo] globallyUniqueString], @"article", nil];
    }
    NSDictionary *d = [NSDictionary dictionaryWithObjectsAndKeys:[[NSProcessInfo processInfo] globallyUniqueString],@"xyzzy", nil];
    [aDict setObject:d forKey:keyDictionary];
}

#define NUM_ITERATIONS 10

NSTimeInterval runTime = 0;

for (int i = 0; i < NUM_ITERATIONS; i++) {
    NSDate *start = [NSDate date];

    for (NSDictionary *keyDictionary in aDict) {
        [[aDict objectForKey:keyDictionary] objectForKey:@"xyzzy"];
    }

    runTime += [[NSDate date] timeIntervalSinceDate:start];
}

NSLog(@"average of %d iterations = %f", NUM_ITERATIONS, runTime/NUM_ITERATIONS);

Answer 1

It's hard to say what improved for this test specifically. There are many changes in arm64 vs armv7 which can affect performance. Some examples:

• arm64 uses Objective-C tagged pointer objects for some types. Code that uses those types may be faster or use less memory.
• arm64 stores Objective-C retain counts for most objects in the object itself instead of in a separate table. Code that performs lots of retain/release traffic may be faster.
• The arm64 instruction set is more efficient for some operations. Code that performs the right sort of math or memory manipulations may be faster.
• arm64's 64-bit pointers occupy more memory. Pointer-heavy memory-constrained code may be slower.

Instruments traces of the two versions might uncover the performance differences.