Re: sub-optimal code for packed boolean arrays -- bug or inherent limitation

["Jeffrey R. Carter" <spam.jrcarter.not@acm.nospam.org>]

tmoran@acm.org wrote:
> Using -O2 with the old gnat 3.15p on a 3GHz dual core Pentium (but only
> programming to use one core) under W2k, I find that the versions with
>         B(I) := not B(I);
> and
>         B(I) := B(I) xor True;
> take about 14 nanoseconds, while the version with
>         One(i) := True;
>         B   := B xor One;
> takes about 6.  With 1e9 calls, that's an 8 second difference.

Yes. But at 6 mins/move, is 8 secs enough to search another ply? That's 
the important time. Since the OP said it might be the difference between 
searching 10 and 11 plys, that implies 36 secs/ply, so the answer would 
seem to be no.

Actually, Flip is only 1 of the 5 basic operations, and it's all the 
basic operations that are called 1e8 to 1e9 times per move, so the 
savings are more like 1.6 secs.

A real improvement would be to parallelize the algorithm and run it on a 
multiprocessor.

-- 
Jeff Carter
"Hello! Smelly English K...niggets."
Monty Python & the Holy Grail
08