> Henson,
> Do you have any idea?

Let me be honest about this one.

I think the root issue is that none of us working on the patch has
enough confidence -- enough real control -- over the tuple slot and JIT
side of it. I wrote the code, it ran fine for me, and I assumed that
meant it was fine; but I have never felt that I fully understand and
control that area. So I cannot point to a specific cause with confidence
yet, and I would rather say that plainly than guess.

One practical note: there have been no tuple slot or JIT changes in the
patch since v47, and none are planned. So that part has been stable, and
reviewing the tuple slot / JIT code against v47 is perfectly viable --
there is no need to wait for a newer revision to look at that area.

What I think would genuinely harden the patch is for someone who knows
tuple slots and JIT well to review -- or rewrite -- that part. I would
very much welcome that, and I am happy to provide the design direction
behind the current code, or any other information a reviewer would need.

To be fair, the same caveat applies to the existing planner code the
patch builds on. I feel I have a reasonable grip on the rest of the
patch, but not on those two areas.

I will keep digging on my side (trying to reproduce the JIT crash with
JIT forced on against the current versions), and report back with
anything concrete I find.

Regards,
Henson

Reply via email to