Re: [gnugo-devel] paul_3_16.1

[Paul Pogonyshev]

Dan wrote:
> I ran the regressions with paul_3_16.1, the corner pattern
> matcher. I got two unexpected failures:
>
> ./regress.sh . nngs.tst
> 500 unexpected FAIL: Correct 'J17', got 'C12'
> ./regress.sh . strategy5.tst
> 229 unexpected FAIL: Correct 'D7', got 'N10'

oh, i feel really stupid :( i couldn't reproduce the bug (it is
compiler/system/everything else dependent - memory corruption).
exactly the same type like the one i fixed before submitting the
patch. please, try this:

change num_stones[] allocation in matchpat.c to

static int real_num_stones[BOARDMAX + (DELTA(MAX_BOARD, 0) - POS(MIN_BOARD, 
0))];
static int *const num_stones = real_num_stones
                             + (DELTA(MAX_BOARD, 0) - POS(MIN_BOARD, 0));

btw, if you could check these lines, it would be very useful in
preventing a same bug in the future :) in fact, this would surely
work:

static int real_num_stones[2 * BOARDMAX];
static int *const num_stones = real_num_stones + BOARDMAX;

and maybe saving some bytes of memory doesn't cost those long
lines.

> Of itself two regression failures is not a reason to postpone
> the patch. However the compiler warnings are fairly obnoxious.

hoshi cleaning patch eliminates all warnings but two. about the
remaining two (`value' used in komoku.sgf) i'd like to hear your
opinion on what to do. those `value' directives has no impact on
move valuation, so they can be removed. on the other hand, they
were put in for some reason, so probably something in the patterns
should be fixed (set pattern class to `j'?). in the latter case,
i'd have to add support for `value' into the new matcher.

Paul