[Bug-gnubg] This may amuse some people

[Jim Segrave]

Last night I decided to actually use gnubg as opposed to working on
it, I thought I'd play an early evening 7 pointer to see how much I'd
forgotten about backgammon in the last month. I was horrified - it was
unbelievably slow. I ran the calibration test and got 3300 evals/sec,
it used to be 7700 on the same machine.

I immediately fired up a copy I've been using as a reference while
playing with rollouts,  built from the June 28 source - that ran
normally - 7500 evals/sec.

I wondered if the python code had affected things - FreeBSD 4.x thread
handling is less than optimal, so I rebuilt the current source without
python. Still slow.

I started doing a binary search in the source, downloading and
building the July 15 version - slow. Tried the July 6 version -
slow. July 1 - slow June 29 - slow. 

I began to think I was going crazy, so I downloaded the June 28
version (the exact same CVS timestamps as my reference). It was slow.

I went to bed, totally puzzled.

I was thinking about it and suddenly remembered that just after 28
June, there was a sort of clean-up of compiler warnings. I looked at
the script I use to fetch and build gnubg:

...
export CFLAGS="-g -Wall"
...
./configure --without-python 2>&1 | tee configure.out
gmake 2>&1 | tee make.out
...

The CFLAGS setting was added when I was looking for compiler
warnings. Light dawns, change to:
export CFLAGS="-g -Wall -O2"

and things are back to normal.
Sigh. I never would have expected a factor of 2 degradation if you
turn off optimisation.

-O3 helps a bit, that gives about 7800 evals/sec

-- 
Jim Segrave           address@hidden