[Freeciv-Dev] freeciv profile data

[Jason Short <jshort@xxxxxxxxxxxxxx>]

I made a few profiles of current freeciv.

I compiled freeciv with default debugging.  I also added in the -pg flag 
to compilation (I added it to CFLAGS in configure.in).  Then I ran 
server/civserver with the autogame.rc, and generated profile data with 
gprof via "gprof server/civserver > output".
All data is in http://freeciv.org/~jdorje/profile/.

The first profile is output-orig, derived from current CVS.  I haven't 
examined it too closely, but I imagine there's much to learn.  Anyone 
interested in optimizing things should look over it to see where the 
bottlenecks are.


The other profiles aren't that useful, but I made them while I was at 
it.  Note that any numerical comparisons are inaccurate since only one 
sample was taken.
Next I generated output-gentopo, with current CVS + current 
gen-topologies patch.  This is worthless to compare to the other since 
execution is different.  I don't know why, but I will investigate. 
There are some trends, however: every map operation (that calls 
map_pos_to_index) is somewhat more expensive.
Obviously really_generate_warmap is a candidate for optimization.  So I 
looked at this function, and found an easy optimization.  WARMAP_COST 
and WARMAP_SEACOST are called more than necessary.  Currently these are 
just a few operations each, but under gen-topologies they have a few 
extra operations.  There are two calls: WARMAP_[SEA]COST(x, y) and 
WARMAP_[SEA]COST(x1, y1).  I made a patch to cache both values, then 
profiled the result.  This gave a significant decrease in the runtime of 
that function and a very slight overall decrease.  I'm not sure if 
cacheing either or both are worth it.  I'll send more info to RT.
As a second optimization, I turned map_get_tile() into an inline 
function.  (I first tried turning it into a macro but freeciv wouldn't 
compile because the global 'map' variable conflicts with some local 
variables.)  This resulted in a 6% decrease in runtime.
jason