Re: using array nature for microoptimization

[Reuben Thomas <rrt@...>]

> In Lua 5, tables are implemented as having two parts: an array part,
> consisting of integer-keyed values, and a general hash table part.  A
> key is a candidate for the array part if the array part is mostly
> consecutive.  (This way assigning to t[20000] will not allocate a 20000
> element array part, unless t[1]...t[19999] is already mostly used.)

What happens is that the array part is used for numeric indices 1..n such
that at least half the entries in 1..n are non-nil.

> The array part seems much faster.

It should be, there's no hash function to call nor chains to follow.

> which I think was entirely stored as a hash table.

> Also, if you plan on doing a lot of unpack() calls on prepared tables
> you construct, add an "n=3" to your {"start", "b", {}, n=3} tables.
> That was good for another 5% speedup....

Interesting.

Have you put this on the Wiki?

-- 
http://www.mupsych.org/~rrt/ | perfect, a.  unsatirizable