Maps are slow in hashlink

[apxapob]

Here is my test code:

function mapTest() {
    var start = Timer.stamp();
    var N = 1000000;

    var map:Map<String, Int> = [];
    for (i in 0...N){//set
      map.set(i+'', i);
    }
    trace("1 step time:", Timer.stamp() - start);
    for (key in map.keys()){//get
      var a = map.get(key);
      a++;
    }
    trace("2 step time:", Timer.stamp() - start);
    for (i in 0...N){//remove
      map.remove(i+'');
    }
    trace("total time:", Timer.stamp() - start);
  }

And here are my results for N = 1 000 000:
1 step time: 0.400216579437256
2 step time: 0.55933403968811
total time: 20.15642786026

Hashlink:
1000 => total time: 0.000833272933959961
10000 => total time: 0.0115585327148438
100000 => total time: 0.325639247894287
1000000 => total time: 20.6692686080933
10000000 => total time: ???

the same code compiled for Javascript:
1000 => total time: 0.0005000000000000004
10000 => total time: 0.001600000001490104
100000 => total time: 0.015400000002235181
1000000 => total time: 0.16799999999999998
10000000 => total time: 2.606199999999255

Also I tried IntMaps they're faster at setting and getting but slower at removing.

Anyway set and get are also slow. Javascript completes the whole test 2 times faster than Hashlink only step 1.

[ncannasse]

Thanks for the benchmark results, that's interesting indeed to see how both implementations are scaling wrt large number of elements. I think it would be better first to benchmark IntMaps to factor our String allocation and hashing, but I agree that the benchmark result might be related to the current implementation of maps, here:
https://github.com/HaxeFoundation/hashlink/blob/master/src/std/maps.h

I wonder what kind of structure is used on JS VM and how it compares also in terms of memory usage.