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

Thanks for looking at this.

On Wed, 6 Jul 2022 at 12:32, Andres Freund <andres@anarazel.de> wrote:
>
> On 2022-06-29 11:40:45 +1200, David Rowley wrote:
> > Another small thing which I considered doing was to put the
> > hash_fcinfo_data field as the final field in
> > ScalarArrayOpExprHashTable so that we could allocate the memory for
> > the hash_fcinfo_data in the same allocation as the
> > ScalarArrayOpExprHashTable.  This would reduce the pointer
> > dereferencing done in saop_element_hash() a bit further.  I just
> > didn't notice anywhere else where we do that for FunctionCallInfo, so
> > I resisted doing this.
>
> I think that'd make sense - it does add a bit of size calculation magic, but
> it shouldn't be a problem. I'm fairly sure we do this in other parts of the
> code.

I've now adjusted that.  I also changed the hash_finfo field to make
it so the FmgrInfo is inline rather than a pointer. This saves an
additional dereference in saop_element_hash() and also saves a
palloc().

I had to adjust the palloc for the ScalarArrayOpExprHashTable struct
into a palloc0 due to the FmgrInfo being inlined.  I considered just
zeroing out the hash_finfo portion but thought it wasn't worth the
extra code.

> Are you good pushing this? I'm fine with you doing so wether you adapt it
> further or not.

Pushed.

David