I guess that creating 500e06 sub-arrays is the most memory consuming part.

I can't tell how efficient the usage of @d is, because it depends on how sparse the entries are used.

32 bits means 4.2e9 potential entries but with 500e6/256 = 1953125 in worst case you'll end up with 2200 empty entry ratio in @d. I bet a hash is more memory efficient then.

But it really depends what you are trying to achieve with all that data.

update

> which don't occur more than 256 times each.

your "test" doesn't reflect that.