Jean-Marc Lasgouttes <[EMAIL PROTECTED]> writes:
>>>>>> "Andre" == Andre Poenitz <[EMAIL PROTECTED]> writes:
>
| Andre> On Mon, Mar 11, 2002 at 06:21:03PM +0100, Lars Gullik Bjønnes
| Andre> wrote:
>>> I felt that the layout lookup became a bit slow with the new layout
>>> as string stuff, so I did this patch. Is the added complexity worth
>>> it?
>
| Did you see any evidence of that in gprof?
>
| Andre> Hm... I doubt that there is a lot of difference when there are
| Andre> just 30 items in the vector/map... Do you have any hard
| Andre> numbers?
>
| It looks suspicious, indeed. As Lars points out, everything here
| should be in constants and local optimizations.
>
| Andre> Apart from that I'd rather try to come up with some "really
| Andre> fast" 'conststring' class which could improve most of the
| Andre> handling of this 'strings used for convenience reasons' stuff,
| Andre> both for memory and runtime.
>
| What would such a class do to be fast?
In the lyxtextclass senario there is little a conststring class could
do... comparisons are O(n) on the string anyway.
One further optimization, over that patch that I sent, is to use a
hash_map if it exists. That would give O(1) lookup.
There is also the possiblity of using a "tuning" lists, and list where
often used items are moved forward, so that lookup is faster next time
the same iterm is wanted, but is still O(n) in the general case.
--
Lgb
