Обсуждение: [HACKERS] Re: [PORTS] GEQO broken on 6-6-97?!?

Darren King wrote:
>
> > Ok, it seems to be coming from
> >
> >            if(!path->p_ordering.ord.sortop) {
> >                         break;
> >                     }
> > in geqo_paths.c:geqo_rel_paths() & prune.c:prune_rel_paths()
> >
> > So as you can reproduce it, could you check with GEQO ON that
> > if (!path->p_ordering.ord.sortop) then compute_joinrel_size()
> > returns <= 0. If it's true then we can add check to geqo_eval.c
> > and allpaths.c:
> >
>
> I added a debug statement at the end of geqo_rel_paths() to print
> the rel->size and they were all positive, ranging in value from
> 73 to 2147483647 (MAX_INT).
        ^^^^^^^^^^
Thanks for help, Darren! I hope that cause found now:
compute_joinrel_size() shouldn't be called for non-JoinPath path!
In old opt-r we have:

        if (IsA_JoinPath(cheapest))
            {
                rel->size = compute_joinrel_size(cheapest);
            }
        else
            elog(WARN, "non JoinPath called");

- - so I changed geqo_rel_paths():

    cheapest = (JoinPath*)set_paths(rel, path);
    if ( IsA_JoinPath (cheapest) )
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        rel->size = compute_joinrel_size(cheapest);

    else - does nothing!

In this case compute_rel_size() in gimme_tree() will be called
(I added if ( new_rel->size <= 0 ) there yesterday).

CVS changed.

Vadim

------------------------------

> Thanks for help, Darren! I hope that cause found now:
> compute_joinrel_size() shouldn't be called for non-JoinPath path!
> In old opt-r we have:
>
>         if (IsA_JoinPath(cheapest))
>             {
>                 rel->size = compute_joinrel_size(cheapest);
>             }
>         else
>             elog(WARN, "non JoinPath called");
>
> - so I changed geqo_rel_paths():
>
>     cheapest = (JoinPath*)set_paths(rel, path);
>     if ( IsA_JoinPath (cheapest) )
>     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
>         rel->size = compute_joinrel_size(cheapest);
>
>     else - does nothing!
>
> In this case compute_rel_size() in gimme_tree() will be called
> (I added if ( new_rel->size <= 0 ) there yesterday).

Fixes geqo for my test data.  Will test some more cases when I get
time, but this appears to have nailed the optimizer problem.


Darren   darrenk@insightdist.com

------------------------------