Re: PG 15 (and to a smaller degree 14) regression due to ExprEvalStep size

Hi,

On 2022-06-23 16:38:12 +0900, Michael Paquier wrote:
> On Tue, Jun 21, 2022 at 05:41:07PM -0400, Andrew Dunstan wrote:
> > On 2022-06-21 Tu 17:25, Andres Freund wrote:
> >> On 2022-06-21 17:11:33 -0400, Andrew Dunstan wrote:
> >>> I and a couple of colleagues have looked it over. As far as it goes the
> >>> json fix looks kosher to me. I'll play with it some more.
> >>
> >> Cool.
> >>
> >> Any chance you could look at fixing the "structure" of the generated
> >> expression "program". The recursive ExecEvalExpr() calls are really not ok...
>
> By how much does the size of ExprEvalStep go down once you don't
> inline the JSON structures as of 0004 in [1]? And what of 0003?

0004 gets us back to 64 bytes, if 0003 is applied first. 0003 alone doesn't
yield a size reduction, because obviously 0004 is the bigger problem. Applying
just 0004 you end up with 88 bytes.

> The JSON portions seem like the largest portion of the cake, though both are
> must-fixes.

Yep.

> > Yes, but I don't guarantee to have a fix in time for Beta2.
>
> IMHO, it would be nice to get something done for beta2. Now the
> thread is rather fresh and I guess that more performance study is
> required even for 0004, so..

I don't think there's a whole lot of performance study needed for 0004 - the
current state is obviously wrong.

I think Andrew's beta 2 comment was more about my other architectural
complains around the json expression eval stuff.

> Waiting for beta3 would a better move at this stage. Is somebody confident
> enough in the patches proposed?

0001 is the one that needs to most careful analysis, I think. 0002 I'd be fine
with pushing after reviewing it again. For 0003 David's approach might be
better or worse, it doesn't matter much I think. 0004 is ok I think, perhaps
with the exception of quibbling over some naming decisions?

Greetings,

Andres Freund