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

Hi,

On 2022-08-11 13:08:27 -0400, Jonathan S. Katz wrote:
> On 8/10/22 9:27 AM, Amit Langote wrote:
> > On Wed, Aug 10, 2022 at 3:57 AM Andres Freund <andres(at)anarazel(dot)de> wrote:
> > > One way this code could be drastically simplified is to force all
> > > type-coercions to go through the "io coercion" path, which could be
> > > implemented as a single execution step (which thus could trivially
> > > start/finish a subtransaction) and would remove a lot of the complicated code
> > > around coercions.
> >
> > Could you please clarify how you think we might do the io coercion
> > wrapped with a subtransaction all as a single execution step? I
> > would've thought that we couldn't do the sub-transaction without
> > leaving ExecInterpExpr() anyway, so maybe you meant the io coercion
> > itself was done using some code outside ExecInterpExpr()?
> >
> > The current JsonExpr code does it by recursively calling
> > ExecInterpExpr() using the nested ExprState expressly for the
> > coercion.

The basic idea is to rip out all the type-dependent stuff out and replace it
with a single JSON_IOCERCE step, which has a parameter about whether to wrap
things in a subtransaction or not. That step would always perform the coercion
by calling the text output function of the input and the text input function
of the output.

> With RMT hat on, Andres do you have any thoughts on this?

I think I need to prototype how it'd look like to give a more detailed
answer. I have a bunch of meetings over the next few hours, but after that I
can give it a shot.

Greetings,

Andres Freund