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

On Wed, Jul 20, 2022 at 11:09 PM Amit Langote <amitlangote09(at)gmail(dot)com> wrote:
> On Wed, Jul 20, 2022 at 12:37 AM Andres Freund <andres(at)anarazel(dot)de> wrote:
> > On 2022-07-19 20:40:11 +0900, Amit Langote wrote:
> > > 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?
> >
> > The easiest way is to just call an external function for the implementation of
> > the step. But yes, otherwise you need to handcraft it.
>
> Ok, thanks.
>
> So I started updating llvm_compile_expr() for handling the new
> ExprEvalSteps that the patch adds to ExecExprInterp(), but quickly
> realized that code could have been consolidated into less code, or
> IOW, into fewer new ExprEvalSteps. So, I refactored things that way
> and am now retrying adding the code to llvm_compile_expr() based on
> new, better consolidated, code.
>
> Here's the updated version, without the llvm pieces, in case you'd
> like to look at it even in this state. I'll post a version with llvm
> pieces filled in tomorrow. (I have merged the different patches into
> one for convenience.)

And here's a version with llvm pieces filled in.

Because I wrote all of it while not really understanding how the LLVM
constructs like blocks and branches work, the only reason I think
those llvm_compile_expr() additions may be correct is that all the
tests in jsonb_sqljson.sql pass even if I add the following line at
the top:

set jit_above_cost to 0;

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