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

Hi,

On Tue, Jul 19, 2022 at 4:09 AM Andrew Dunstan <andrew(at)dunslane(dot)net> wrote:
> On 2022-07-15 Fr 17:07, Andres Freund wrote:
> > Perhaps you could post your current state? I might be able to help resolving
> > some of the problems.
>
> Ok. Here is the state of things. This has proved to be rather more
> intractable than I expected. Almost all the legwork here has been done
> by Amit Langote, for which he deserves both my thanks and considerable
> credit, but I take responsibility for it.
>
> I just discovered today that this scheme is failing under
> "force_parallel_mode = regress". I have as yet no idea if that can be
> fixed simply or not.

The errors Andrew mentions here had to do with a bug of the new
coercion evaluation logic. The old code in ExecEvalJsonExpr() would
skip coercion evaluation and thus also the sub-transaction associated
with it for some JsonExprs that the new code would not and that didn't
sit well with the invariant that a parallel worker shouldn't try to
start a sub-transaction.

That bug has been fixed in the attached updated version.

> Apart from that I think the main outstanding issue
> is to fill in the gaps in llvm_compile_expr().

About that, I was wondering if the blocks in llvm_compile_expr() need
to be hand-coded to match what's added in ExecInterpExpr() or if I've
missed some tool that can be used instead?

--
Thanks, Amit Langote
EDB: http://www.enterprisedb.com