Which has faster performance indexesOfObjectsPassingTest or filteredArrayUsingPredicate?

When needing to filter an NSArray to get a subset of the items in the array returned, which method is quicker more frequently and in edge cases?


The following tests (compiled in Release mode, executed on a Mac Pro) indicate that filteredArrayUsingPredicate is slower than indexesOfObjectsPassingTest if you use a "textual" predicate, but faster if you use block-based predicate. The fasted method in my test was a simple (fast-enumeration) loop that adds all matching objects to a mutable array.

Results for filtering an array of 10,000,000 dictionaries, where about 50% match the predicate:

8.514334 (predicateWithFormat)
4.422550 (predicateWithBlock)
5.170086 (indexesOfObjectsPassingTest)
3.154015 (fast-enumeration + mutable array)

Of course the results may be different for other predicates.

#import <Foundation/Foundation.h>

NSUInteger filter1(NSArray *a)
{
    NSPredicate *pred = [NSPredicate predicateWithFormat:@"num > 1000 AND foo == 'bar'"];
    NSArray *filtered = [a filteredArrayUsingPredicate:pred];
    return [filtered count];
}

NSUInteger filter2(NSArray *a)
{
    NSPredicate *pred = [NSPredicate predicateWithBlock:^BOOL(NSDictionary *obj, NSDictionary *bindings) {
        return ([obj[@"num"] intValue] > 1000 && [obj[@"foo"] isEqualToString:@"bar"]);
    }];
    NSArray *filtered = [a filteredArrayUsingPredicate:pred];
    return [filtered count];
}

NSUInteger filter3(NSArray *a)
{
    NSIndexSet *matching = [a indexesOfObjectsPassingTest:^BOOL(NSDictionary *obj, NSUInteger idx, BOOL *stop) {
        return ([obj[@"num"] intValue] > 1000 && [obj[@"foo"] isEqualToString:@"bar"]);
    }];
    NSArray *filtered = [a objectsAtIndexes:matching];
    return [filtered count];
}

NSUInteger filter4(NSArray *a)
{
    NSMutableArray *filtered = [NSMutableArray array];
    for (NSDictionary *obj in a) {
        if ([obj[@"num"] intValue] > 1000 && [obj[@"foo"] isEqualToString:@"bar"]) {
            [filtered addObject:obj];
        }
    }
    return [filtered count];
}

void testmethod(NSArray *a, NSUInteger(*method)(NSArray *a))
{
    @autoreleasepool {
        NSDate *t1 = [NSDate date];
        NSUInteger count = method(a);
        NSDate *t2 = [NSDate date];
        NSLog(@"%f", [t2 timeIntervalSinceDate:t1]);
    }
}

int main(int argc, const char * argv[])
{
    @autoreleasepool {
        NSMutableArray *a = [NSMutableArray array];
        for (int i = 0; i < 10000000; i++) {
            [a addObject:@{@"num": @(arc4random_uniform(2000)), @"foo":@"bar"}];
        }
        testmethod(a, filter1);
        testmethod(a, filter2);
        testmethod(a, filter3);
        testmethod(a, filter4);
    }
    return 0;
}

I tested this problem with the brand new Xcode 6 performance tests (Objective-C) with the test cases below. I got the following results indicating that the enumerationBlock with the flag NSEnumerationConcurrent is the fastest filtering method for large arrays:

testPerformancePredicateWithFormat - measured [Time, seconds] average: 0.189
testPerformancePredicateWithBlock - measured [Time, seconds] average: 0.093
testPerformanceEnumerationBlock - measured [Time, seconds] average: 0.092
testPerformanceIndexesOfObjectsPassingTest - measured [Time, seconds] average: 0.082
testPerformanceFastEnumeration - measured [Time, seconds] average: 0.068
testPerformanceEnumerationConcurrent - measured [Time, seconds] average: 0.036

Here the tests:

#import <XCTest/XCTest.h>

@interface TestPMTests : XCTestCase
@property(nonatomic, copy)NSArray *largeListOfDictionaries;
@end

@implementation TestPMTests

- (void)setUp {
    [super setUp];

    self.largeListOfDictionaries = [NSMutableArray array];

    // Initialize a large array with ~ 300.000 entries as Dictionaries of at least one key value pair {"id":"<any id>"}
}

- (void)testPerformancePredicateWithFormat {
    NSString *ID = @"204440e5-4069-48e8-a405-88882a5ba27e";
    NSPredicate *pred = [NSPredicate predicateWithFormat:@"SELF.id == %@", ID];

    [self measureBlock:^{
        NSArray *filtered = [self.largeListOfDictionaries filteredArrayUsingPredicate:pred];
        NSLog(@"Count: %d", filtered.count);
    }];
}

- (void)testPerformancePredicateWithBlock {
    NSString *ID = @"204440e5-4069-48e8-a405-88882a5ba27e";
    NSString *kID = @"id";
    NSPredicate *pred = [NSPredicate predicateWithBlock:^BOOL(NSDictionary *d, NSDictionary *bindings) {
        return [d[kID] isEqualToString:ID];
    }];

    [self measureBlock:^{
        NSArray *filtered = [self.largeListOfDictionaries filteredArrayUsingPredicate:pred];
        NSLog(@"Count: %d", filtered.count);
    }];
}

- (void)testPerformanceIndexesOfObjectsPassingTest {
    NSString *ID = @"204440e5-4069-48e8-a405-88882a5ba27e";
    NSString *kID = @"id";

    [self measureBlock:^{
        NSIndexSet *matchingIndexes = [self.largeListOfDictionaries indexesOfObjectsPassingTest:^BOOL(NSDictionary *d, NSUInteger idx, BOOL *stop) {
            return [d[kID] isEqualToString:ID];
        }];
        NSArray *filtered = [self.largeListOfDictionaries objectsAtIndexes:matchingIndexes];
        NSLog(@"Count: %d", filtered.count);
    }];
}

- (void)testPerformanceFastEnumeration {
    NSString *ID = @"204440e5-4069-48e8-a405-88882a5ba27e";
    NSString *kID = @"id";

    [self measureBlock:^{
        NSMutableArray *filtered = [NSMutableArray array];
        for (NSDictionary *d in self.largeListOfDictionaries) {
            if ([d[kID] isEqualToString:ID]) {
                [filtered addObject:d];
            }
        }
        NSLog(@"Count: %d", filtered.count);
    }];
}

- (void)testPerformanceEnumerationBlock {
    NSString *ID = @"204440e5-4069-48e8-a405-88882a5ba27e";
    NSString *kID = @"id";

    [self measureBlock:^{
        NSMutableArray *filtered = [NSMutableArray array];
        [self.largeListOfDictionaries enumerateObjectsUsingBlock:^(NSDictionary *d, NSUInteger idx, BOOL *stop) {
            if ([d[kID] isEqualToString:ID]) {
                [filtered addObject:d];
            }
        }];
        NSLog(@"Count: %d", filtered.count);
    }];
}

- (void)testPerformanceEnumerationConcurrent {
    NSString *ID = @"204440e5-4069-48e8-a405-88882a5ba27e";
    NSString *kID = @"id";

    [self measureBlock:^{
        NSMutableArray *filtered = [NSMutableArray array];
        [self.largeListOfDictionaries enumerateObjectsWithOptions:NSEnumerationConcurrent usingBlock:^(NSDictionary *d, NSUInteger idx, BOOL *stop) {
            if ([d[kID] isEqualToString:ID]) {
                [filtered addObject:d];
            }
        }];
        NSLog(@"Count: %d", filtered.count);
    }];
}  

UPDATE

I changed the following in -testPerformanceEnumerationConcurrent:

dispatch_sync(queue, ^{
    [filtered addObject:d];
});

And the results are still better for the concurrent version that in all other tests.

-[TestPMTests testPerformancePredicateWithFormat average: 0.134
-[TestPMTests testPerformancePredicateWithBlock] average: 0.079
-[TestPMTests testPerformanceEnumerationBlock] average: 0.079
-[TestPMTests testPerformanceIndexesOfObjectsPassingTest] average: 0.068
-[TestPMTests testPerformanceFastEnumeration] average: 0.054
-[TestPMTests testPerformanceEnumerationConcurrent] average: 0.029