AA performance again

[Robert Fraser]

bearophile wrote:
> Sometimes the simplest code is enough to show a problem. Other gentle people in the #D channel say those timings aren't random:
> http://leonardo-m.livejournal.com/64284.html
> 
> The D versions eat 425 MB with 10 million longs and about 757 MB RAM with dmd with 20 million longs.
> 
> Bye,
> bearophile

If the bottleneck is GC, then it might be interesting to see how a 
dictionary type that uses less memory (i.e. Judy or another tree-based 
implementation) compares instead of using a hash (obviously with a lot 
of pre-allocation).

If the bottleneck is in aaA.d, then maybe that just needs to be 
optimized, and I'm sure at least the Tango folks have the knowledge to 
do that. I'm guessing they've done all they can already, though.