Hi,
On 17.08.2022 04:45, Jonathan S. Katz wrote:
On 8/15/22 10:14 PM, Andres Freund wrote:
I pushed a few cleanups to https://github.com/anarazel/postgres/commits/json
while I was hacking on this (ignore that it's based on the meson tree, that's
just faster for me). Some of them might not be applicable anymore, but it
might still make sense for you to look at.
With RMT hat on, this appears to be making progress. A few questions / comments for the group:
1. Nikita: Did you have a chance to review Andres's changes as well?
Yes, I have reviewed Andres's changes, they all are ok.
Then I started to do on the top of it other fixes that help to avoid
subtransactions when they are not needed. And it ended in the new
refactoring of coercion code. Also I moved here from v6-0003 fix of
ExecEvalJsonNeedSubtransaction() which considers more cases.
On 16.08.2022 05:14, Andres Freund wrote:
>> But for JIT I still had to construct additional ExprState with a
>> function compiled from subexpression steps.
> Why did you have to do this?
I simply did not dare to implement compilation of recursively-callable
function with additional parameter stepno. In the v8 patch I did it
by adding a switch with all possible jump addresses of EEOP_SUBTRANS
steps in the beginning of the function. And it really seems to work
faster, but needs more exploration. See patch 0003, where both
variants preserved using #ifdef.
The desciprion of the v7 patches:
0001 Simplify JsonExpr execution Andres's changes + mine: - Added JsonCoercionType enum, fields like via_io replaced with it - Emit only context item steps in JSON_TABLE_OP case - Skip coercion of NULLs to non-domain types (is it correct?)
0002 Fix returning of json[b] domains in JSON_VALUE: simply rebase of v6 onto 0001
0003 Add EEOP_SUBTRANS executor step v6 + new recursive JIT
0004 Split JsonExpr execution into steps simply rebase of v6 + used LLMBuildSwitch() in EEOP_JSONEXPR
--
Nikita Glukhov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
